text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
author:
- ZEUS Collaboration
date: July 2003
title: ' Measurement of the open-charm contribution to the diffractive proton structure function'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A novel interferometric method – SLIVER (Super Localization by Image inVERsion interferometry) – is proposed for estimating the separation of two incoherent point sources with a mean squared error that does not deteriorate as the sources are brought closer. The essential component of the interferometer is an image inversion device that inverts the field in the transverse plane about the optical axis, assumed to pass through the centroid of the sources. The performance of the device is analyzed using the Cramér-Rao bound applied to the statistics of spatially-unresolved photon counting using photon number-resolving and on-off detectors. The analysis is supported by Monte-Carlo simulations of the maximum likelihood estimator for the source separation, demonstrating the superlocalization effect for separations well below that set by the Rayleigh criterion. Simulations indicating the robustness of SLIVER to mismatch between the optical axis and the centroid are also presented. The results are valid for any imaging system with a circularly symmetric point-spread function.'
address: |
$^1$Department of Electrical and Computer Engineering,\
National University of Singapore, 4 Engineering Drive 3,\
Singapore 117583\
$^2$Department of Physics,\
National University of Singapore, 2 Science Drive 3, Singapore 117551
author:
- 'Ranjith Nair$^{1,*}$ and Mankei Tsang$^{1, 2}$'
title: Interferometric superlocalization of two incoherent optical point sources
---
[10]{}
, “[XXXI. I]{}nvestigations in optics, with special reference to the spectroscope,” The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science **8**, 261–274 (1879).
M. Born and E. Wolf, *Principles of Optics: Electromagnetic Theory of Propagation, Interference and Diffraction of Light* (Cambridge University, 1999).
E. Bettens, D. Van Dyck, A. Den Dekker, J. Sijbers, and A. Van den Bos, “Model-based two-object resolution from observations having counting statistics,” Ultramicroscopy **77**, 37–48 (1999).
S. Ram, E. S. Ward, and R. J. Ober, “Beyond [R]{}ayleigh’s criterion: a resolution measure with application to single-molecule microscopy,” Proceedings of the National Academy of Sciences of the United States of America **103**, 4457–4462 (2006).
S. V. Aert, A. den Dekker, D. V. Dyck, and A. van den Bos, “High-resolution electron microscopy and electron tomography: resolution versus precision,” Journal of Structural Biology **138**, 21 – 33 (2002).
H. L. [Van Trees]{}, *Detection, Estimation, and Modulation Theory : Part I* (Wiley-Interscience 1st Ed, 2001), 1st ed.
M. Tsang, R. Nair, and X.-M. Lu, “Quantum theory of superresolution for two incoherent optical point sources,” (2015). `http://arxiv.org/abs/1511.00552`.
S. Weisenburger and V. Sandoghdar, “Light microscopy: an ongoing contemporary revolution,” Contemporary Physics **56**, 123–143 (2015).
J. Chao, S. Ram, E. S. Ward, and R. J. Ober, “Ultrahigh accuracy imaging modality for super-localization microscopy,” Nat Meth **10**, 335–338 (2013).
M. Tsang, “Quantum limits to optical point-source localization,” Optica **2**, 646–653 (2015).
C. W. Helstrom, *Quantum Detection and Estimation Theory* (Academic Press, 1976).
A. S. Holevo, *Probabilistic and Statistical Aspects of Quantum Theory* (Edizioni Della Normale, Pisa, Italy, 2011).
M. G. Paris, “Quantum estimation for quantum technology,” International Journal of Quantum Information **7**, 125–137 (2009).
L. Mandel and E. Wolf, *Optical Coherence and Quantum Optics* (Cambridge University, 1995).
J. H. Shapiro, “The quantum theory of optical communications,” IEEE Journal of Selected Topics in Quantum Electronics **15**, 1547 –1569 (2009).
J. W. Goodman, *Statistical Optics* (John Wiley & Sons, 1985).
N. Sandeau and H. Giovannini, “Increasing the lateral resolution of 4pi fluorescence microscopes,” J. Opt. Soc. Am. A **23**, 1089–1095 (2006).
K. Wicker and R. Heintzmann, “Interferometric resolution improvement for confocal microscopes,” Optics Express **15**, 12206–12216 (2007).
K. Wicker, S. Sindbert, and R. Heintzmann, “Characterisation of a resolution enhancing image inversion interferometer,” Optics Express **17**, 15491–15501 (2009).
D. Weigel, R. Foerster, H. Babovsky, A. Kiessling, and R. Kowarschik, “Enhanced resolution of microscopic objects by image inversion interferometry,” Optics Express **19**, 26451–26462 (2011).
D. Weigel, H. Babovsky, A. Kiessling, and R. Kowarschik, “[Investigation of the resolution ability of an image inversion interferometer]{},” [Optics Communications]{} **[284]{}**, [2273–2277]{} ([2011]{}).
D. Weigel, T. Elsmann, H. Babovsky, A. Kiessling, and R. Kowarschik, “Combination of the resolution enhancing image inversion microscopy with digital holography,” Optics Communications **291**, 110 – 115 (2013).
D. Weigel, H. Babovsky, A. Kiessling, and R. Kowarschik, “[Imaging properties of different types of microscopes in combination with an image inversion interferometer]{},” [Optics Communications]{} **[332]{}**, [301–310]{} ([2014]{}).
D. Weigel, H. Babovsky, A. Kiessling, and R. Kowarschik, “Widefield microscopy with infinite depth of field and enhanced lateral resolution based on an image inverting interferometer,” Optics Communications **342**, 102 – 108 (2015).
J. W. Goodman, *Introduction to Fourier Optics* (Roberts and Company Publishers, 2005).
S. Kay and Y. C. Eldar, “[Rethinking biased estimation]{},” [IEEE Signal Processing Magazine]{} **[25]{}**, [133–136]{} ([2008]{}).
R. S. Kennedy, “A near-optimum receiver for the binary coherent state channel,” Tech. Rep. 108, MIT RLE Quarterly Progress Report (1973).
S. J. Dolinar, “An optimum receiver for the binary coherent state channel,” Tech. Rep. 111, MIT RLE Quarterly Progress Report (1973).
R. Nair, S. Guha, and S.-H. Tan, “[Realizable receivers for discriminating coherent and multicopy quantum states near the quantum limit]{},” [Physical Review A]{} **[89]{}**, [032318]{} ([2014]{}).
U. Kubitscheck (Ed.), *Fluorescence Microscopy: from Principles to Biological Applications* (John Wiley & Sons, 2013).
Introduction
============
Rayleigh’s criterion for resolution of two incoherent point sources, which asserts that a minimum separation between the sources equal to the diffraction-limited spot size is necessary for them to be resolvable, has perhaps been the most influential resolution criterion in the history of optics despite its heuristic character [@Ray1879xxxi; @BW99Principles]. Based as it is on the eye as a detection instrument, Rayleigh’s criterion neglects the possibility of better resolution using light of greater intensity or using a longer observation period. Pioneering studies on resolution limits as a function both of the source separation and the mean number of observed photons were made by Bettens *et al.* in [@BDD+99] and more recently by Ram *et al.* in [@RWO06] (See also [@VDD+02] in the context of electron microscopy). In [@BDD+99], it was shown that for a given mean photon number, ideal continuum image-plane photon counting can locate the centroid, i.e., the midpoint, of the two radiating sources with a finite precision (depending on the size of the diffraction-limited spot in the image plane and on the number of photons detected) no matter how small their separation is. Based on the Cramér-Rao (CR) lower bound of estimation theory [@VanTreesI], it was also shown in [@BDD+99; @RWO06] that any (unbiased) estimate of the separation between the sources based on image-plane photon counting must suffer a divergent mean squared error for a given mean photon number as the separation tends to zero. This phenomenon was dubbed *Rayleigh’s curse* in [@TNL15] as it suggests a fundamental limitation in resolving incoherent point sources even when the role of the average detected photon number is taken into account. In the past few decades, several far-field super-resolution techniques that circumvent the Rayleigh limit have revolutionized microscopy (see [@WS15] for a review), but these rely on placing restrictions on the emissions from the sources and do not fundamentally challenge the criterion (or Rayleigh’s curse) for independently and incoherently radiating sources. Other developments, e.g.,[@CRW+13], use clever techniques to approach the performance of ideal continuum image-plane photon counting using imperfect detectors.
In [@TNL15], following preliminary work in [@Tsa15], the problem of resolving two incoherent point sources was approached anew from the perspective of quantum estimation theory using the quantum Cramér-Rao (QCR) bound [@Hel76; @Hol11; @Par09]. This bound provides a fundamental limit to the accuracy of estimating the source separation optimized over all possible measurement techniques allowed by quantum mechanics. Under a weak-source assumption, it was found in [@TNL15] that the QCR lower bound on the minimum mean squared error (MSE) of estimating the separation of two point sources is *independent* of that separation. Further, a linear optics-based measurement – spatial-mode demultiplexing (SPADE) – was proposed and shown in principle to approach the quantum bound for all values of the separation. These results are in stark contrast to the performance of image-plane photon counting mentioned above, as the divergent behavior of the minimum MSE with decreasing separation – Rayleigh’s curse – is conspicuously absent.
In this paper, we revisit the resolution problem from the point of view of the semiclassical theory of photodetection [@MW95; @Sha09; @Goo85Statistical]. For sources emitting thermal radiation, for linear propagation of light through the imaging system, and for linear-optics processing followed by photon counting, the semiclassical theory of photodetection may be used to obtain the same results as a fully quantum analysis would give [@MW95; @Sha09]. This is because the thermal state has a positive $P$-representation, and its counting statistics may be obtained by statistical averaging over those of coherent states via Mandel’s rule [@MW95]. The SLIVER scheme of this paper as well as the schemes of [@TNL15] satisfy the above necessary conditions for the applicability of the semiclassical theory. We emphasize that, for the state of radiation and measurements considered here and in [@TNL15], “semiclassical” does not imply “approximate” – the fully quantum and semiclassical treatments are in quantitative agreement.
Besides its greater familiarity to the optics community, a semiclassical analysis of the resolution problem offers several advantages. Firstly, the analysis can be carried out for arbitrary source strengths. Indeed, the work in [@BDD+99; @RWO06] assumes the sources are weak enough so that the counting statistics are nearly Poissonian. Similarly, the quantum analysis in [@TNL15] assumed, for mathematical tractability, that the sources are weak enough so that the state of light could be assumed to be confined to the zero and one-photon subspaces. Thus, a semiclassical treatment will incorporate the effects of multi-photon events neglected in [@BDD+99; @RWO06] and in the quantum analysis of [@TNL15]. Such an analysis would also reveal whether the superresolution effect predicted in [@TNL15] for weak sources persists as the source strength is increased. A second advantage of the semiclassical approach is that intuitions from semiclassical optics can be harnessed to both understand the reasons for superresolution as well as to design new measurement schemes which can also be analyzed semiclassically provided that they only involve linear optics and photodetection, possibly along with the use of other classical sources of light.
Accordingly, our contributions in this paper are three-fold:-
1. We set up the problem of resolving two incoherently radiating point sources in the framework of semiclassical photodetection theory. The sources can be of arbitrary strength and imaging systems with two-dimensional circularly-symmetric point-spread functions are studied. The thermal source model is applicable to a wide range of physical scenarios ranging from optical astronomy to fluorescence microscopy.
2. We propose a new interferometric scheme for estimating source separation that we call SLIVER (Super Localization by Image inVERsion interferometry) and that yields finite resolution for arbitrarily small values of the source separation and for arbitrary source strengths. Devices employing image inversion have been proposed, studied and demonstrated previously in the microscopy literature for various related applications [@SG06; @WH07; @WSH09; @Wei+11; @WBK+11; @WEB+13; @WBK+14; @WBK+15], but the fundamental limits on the capability of such devices for improving lateral resolution seem not to have been explored before. Through an analysis of the photodetection statistics of SLIVER in the semiclassical framework, we show both using the Cramér-Rao bound and through explicit simulation that our method manifestly alleviates Rayleigh’s curse.
3. We offer an explanation for the superlocalization effect of SLIVER at small values of separation in the language of estimation theory applied to the photodetection statistics for thermal light.
Source and system model
=======================
Consider the depiction in Fig. 1 of the focused image of two incoherent point sources in the image plane of an imaging system coordinatized by ${{\boldsymbol}{\rho}}= (x,y)$. We assume that the imaging system is spatially invariant and coordinates have been rescaled so that the image is of unit magnification [@Goo05Fourier]. We also assume that the (possibly complex-valued) normalized amplitude point-spread function (PSF) $\psi({{\boldsymbol}{\rho}})$ satisfying $$\begin{aligned}
\int \operatorname{\,d}{{\boldsymbol}{\rho}}\, {\left | \psi({{\boldsymbol}{\rho}}) \right |}^2 =1\end{aligned}$$ is circularly symmetric, so that $\psi({{\boldsymbol}{\rho}})$ depends only on the magnitude ${\left | {{\boldsymbol}{\rho}}\right |}$ for all ${{\boldsymbol}{\rho}}$. It is straightforward to make $\psi({{\boldsymbol}{\rho}})$ spatially invariant and circularly symmetric in an imaging system using two lenses [@Goo05Fourier]. The prototypical example of such a PSF is of course the Airy disk pattern resulting from a circular pupil. The images of the sources 1 and 2 in the image plane are assumed to be centered at $\mp \frac{{{\mathbf}{d}}}{2}$ respectively, giving rise in the image plane to a combined field with complex amplitude $$\begin{aligned}
\label{ipf}
E({{\boldsymbol}{\rho}}) = A_1\psi{\left(}{{\boldsymbol}{\rho}}+\frac{{{\mathbf}{d}}}{2}{\right)}+ A_2\psi{\left(}{{\boldsymbol}{\rho}}-\frac{{{\mathbf}{d}}}{2}{\right)}.\end{aligned}$$ in normalized units of $\sqrt{\mbox{photons}\cdot\mbox{m}^{-2}}$. We will assume throughout this paper, except in Section 5.3, that the centroid of the sources, i.e., their midpoint, has already been located, perhaps by image-plane photon counting [@RWO06]. It is taken to lie at the origin of the image plane. In Sec. 5.3, we consider the effect of a small error in centroid localization. Our initial results suggest that our measurement scheme is robust to such perturbations.
The pair of dimensionless complex numbers $A = (A_1,A_2) \in \mathbb{C}^2$ are the source amplitudes. The thermal and mutually incoherent nature of the sources dictates that the source amplitudes are circular-complex Gaussian random variables having the first and second moments [@Goo85Statistical]:- $$\begin{aligned}
&\mathbb{E}[A_{\mu}] = 0 \label{S1}\\
&\mathbb{E}[A_{\mu}A_{\nu}] = 0 \label{S2}\\
&\mathbb{E}[A_{\mu}^*A_{\mu}] = \epsilon_\mu \label{S3} \\
&\mathbb{E}[A_{1}^*A_2] = 0 \label{S4}\end{aligned}$$ for $\mu,\nu \in \{1,2\}$ ranging over the two sources. According to the above relations, the real and imaginary parts of the $\{A_{\mu}\}$ comprise four statistically independent zero-mean Gaussians of variance $\epsilon_1/2$ (for the components of $A_1$) or $\epsilon_2/2$ (for the components of $A_2$). Eqs. (\[S1\])-(\[S4\]) are as appropriate for completely incoherent sources with respective strengths (i.e., mean photon number) $\epsilon_1$ and $\epsilon_2$, which can assume any values.
Given that the centroid has been located, the parameter of interest is the separation $d ={\left | {{\mathbf}{d}}\right |}$ between the two sources. The angle between the line joining the sources and the $x$-axis is another unknown parameter, but the measurement method proposed here works regardless of the value of that angle, which need not be known beforehand in practice. We will thus focus just on the separation $d$ and apply the tools of single-parameter estimation theory in our analysis.
The SLIVER measurement and its statistics
=========================================
The SLIVER measurement, illustrated schematically in Fig. 2 and described in detail below, essentially involves the separation of the image-plane field $E({{\boldsymbol}{\rho}})$ of Eq. into its *symmetric* and *antisymmetric* parts with respect to inversion in the image plane about the optical axis. By definition, the symmetric and antisymmetric parts of $E({{\boldsymbol}{\rho}})$ are given respectively by $$\begin{aligned}
\label{Esa}
\begin{split}
E_s({{\boldsymbol}{\rho}}) &:= \frac{E({{\boldsymbol}{\rho}})+E(-{{\boldsymbol}{\rho}})}{2} = \frac{S}{2}{\left[}\psi{\left(}{{\boldsymbol}{\rho}}+ \frac{{{\mathbf}{d}}}{2}{\right)}+ \psi{\left(}{{\boldsymbol}{\rho}}- \frac{{{\mathbf}{d}}}{2}{\right)}{\right]}\\
E_a({{\boldsymbol}{\rho}}) &:= \frac{E({{\boldsymbol}{\rho}})-E(-{{\boldsymbol}{\rho}})}{2} = \frac{D}{2}{\left[}\psi{\left(}{{\boldsymbol}{\rho}}+ \frac{{{\mathbf}{d}}}{2}{\right)}- \psi{\left(}{{\boldsymbol}{\rho}}- \frac{{{\mathbf}{d}}}{2}{\right)}{\right]}.
\end{split}\end{aligned}$$ In writing the right-hand sides of Eqs. , we have used the fact that $\psi({{\boldsymbol}{\rho}})$ is circularly symmetric and have defined the sum and difference $$\begin{aligned}
S &= A_1 + A_2 \\
D &= A_1 - A_2\end{aligned}$$ of the amplitudes of the two sources. The relations $$\begin{aligned}
\label{SD}
\begin{split}
&\mathbb{E}[S] = \mathbb{E}[D] = 0 \\
&\mathbb{E}[S^2] =\mathbb{E}[D^2] = \mathbb{E}[SD] = 0 \\
&\mathbb{E}[S^{*} S] = \mathbb{E}[D^{*} D] = \epsilon_1 + \epsilon_2 \\
&\mathbb{E}[S^*D] = \epsilon_1 - \epsilon_2
\end{split}\end{aligned}$$ hold for these new random variables, which are thus also identically distributed circular-complex Gaussians and statistically independent of each other when $\epsilon_1 = \epsilon_2$.
The field components of Eq. may be obtained in spatially separated planes using an *image inversion interferometer* (III) that works by splitting the input field using a 50-50 beamsplitter, spatially inverting the field from one output about the optical axis and recombining the two beams at a second 50-50 beamsplitter. The two beams output from the second beamsplitter then have the field patterns of Eq. . The entire setup is thus essentially a balanced Mach-Zehnder interferometer with an inversion device in one arm (Fig. 2). The critical element in such setups is the inversion device – a device that implements the transformation $E_{\mathrm}{out}({{\boldsymbol}{\rho}}) = E_{\mathrm}{in}(-{{\boldsymbol}{\rho}})$. Inversion devices have been proposed and demonstrated for various applications in the fluorescence microscopy community and a considerable literature on image inversion microscopy exists [@SG06; @WH07; @WSH09; @Wei+11; @WBK+11; @WEB+13; @WBK+14; @WBK+15]. However, to the best of our knowledge, a fundamental statistical analysis of the kind we make here of the capabilities of an III for lateral resolution has not been done (See the discussion in Section 6). The spatial inversion itself can be accomplished in a variety of ways, e.g., using a $4f$-arrangement with lenses [@WH07], with a 3-D setup using plane mirrors [@WSH09], or using two concave mirrors [@Wei+11]. We will be concerned here only with the outputs of such a device rather than its detailed implementation, which may vary according to the application.
Suppose now that the output beams from the III are directed to different photodetectors. Conditioned on the values of $S$ and $D$, the intensity patterns $I_s({{\boldsymbol}{\rho}}_s)$ and $I_a({{\boldsymbol}{\rho}}_a)$ of the fields on their respective detector planes are $$\begin{aligned}
I_s({{\boldsymbol}{\rho}}_s) &= \frac{{\left | S \right |}^2}{4}{\left[}{\left | \psi{\left(}{{\boldsymbol}{\rho}}_s + \frac{{{\mathbf}{d}}}{2}{\right)}\right |}^2 + {\left | \psi{\left(}{{\boldsymbol}{\rho}}_s - \frac{{{\mathbf}{d}}}{2}{\right)}\right |}^2 +2I_{{\mathrm}{int}}({{\boldsymbol}{\rho}}_s, {{\mathbf}{d}}) {\right]}, \\
I_a({{\boldsymbol}{\rho}}_a) &= \frac{{\left | D \right |}^2}{4}{\left[}{\left | \psi{\left(}{{\boldsymbol}{\rho}}_a + \frac{{{\mathbf}{d}}}{2}{\right)}\right |}^2 + {\left | \psi{\left(}{{\boldsymbol}{\rho}}_a - \frac{{{\mathbf}{d}}}{2}{\right)}\right |}^2 -2I_{{\mathrm}{int}}({{\boldsymbol}{\rho}}_a, {{\mathbf}{d}}) {\right]}, \end{aligned}$$ where $$\begin{aligned}
I_{{\mathrm}{int}}({{\boldsymbol}{\rho}}, {{\mathbf}{d}}) =\operatorname{Re\,}\psi^*{\left(}{{\boldsymbol}{\rho}}+ \frac{{{\mathbf}{d}}}{2}{\right)}\psi{\left(}{{\boldsymbol}{\rho}}- \frac{{{\mathbf}{d}}}{2}{\right)}\end{aligned}$$ is an interference term. Conditioned on the values of $S$ and $D$, semiclassical detection theory dictates that the photocounts on the two detector planes are independent inhomogeneous Poisson processes with the rate functions $I_s({{\boldsymbol}{\rho}}_s)$ and $I_a({{\boldsymbol}{\rho}}_a)$ respectively [@MW95; @Goo85Statistical; @Sha09]. It follows that the spatially-unresolved integrated photocounts $N_s$ and $N_a$ at each detector are Poisson random variables with the means
$$\begin{aligned}
\label{sameans}
\mathbb{E}[N_s | S] &= \frac{{\left | S \right |}^2}{2} \int \operatorname{\,d}{{\boldsymbol}{\rho}}_s I_s({{\boldsymbol}{\rho}}_s) = \frac{{\left | S \right |}^2}{2}{\left[}1+ \delta(d) {\right]}, \\
\mathbb{E}[N_a | D] &= \frac{{\left | D \right |}^2}{2} \int \operatorname{\,d}{{\boldsymbol}{\rho}}_a I_a({{\boldsymbol}{\rho}}_a) = \frac{{\left | D \right |}^2}{2}{\left[}1- \delta(d) {\right]},\end{aligned}$$
where $$\begin{aligned}
\int \operatorname{\,d}{{\boldsymbol}{\rho}}I_{{\mathrm}{int}}({{\boldsymbol}{\rho}}, {{\mathbf}{d}}) &= \operatorname{Re\,}\int \operatorname{\,d}{{\boldsymbol}{\rho}}\,\psi^*({{\boldsymbol}{\rho}})\psi({{\boldsymbol}{\rho}}-{{\mathbf}{d}}) \label{delta}\\
&=\operatorname{Re\,}\iint\limits_{-\infty}^{\infty} \operatorname{\,d}x \operatorname{\,d}y \,\psi^*(x,y)\psi(x-d,y) \label{E2}\\
&\equiv \delta(d).\end{aligned}$$ The circular symmetry of $\psi({{\boldsymbol}{\rho}})$ has been used to get Eq. , and we have retained only the separation $d$ as argument. According to Eqs. , both ${\left | S \right |}^2$ and ${\left | D \right |}^2$ are exponentially distributed with mean $(\epsilon_1+ \epsilon_2)$, so the photocounts $N_s$ and $N_a$ integrated over the photodetector surfaces are distributed according to Bose-Einstein distributions [@MW95; @Goo85Statistical; @Sha09] with means $$\begin{aligned}
\bar{N}_s &= \frac{\epsilon_1+\epsilon_2}{2} {\left[}1 +\delta(d){\right]}= \epsilon_{{\mathrm}{ave}} {\left[}1 +\delta(d){\right]}, \\
\bar{N}_a &= \frac{\epsilon_1 + \epsilon_2}{2} {\left[}1 -\delta(d){\right]}= \epsilon_{{\mathrm}{ave}} {\left[}1 -\delta(d){\right]}, \label{Na}\end{aligned}$$ for $\epsilon_{{\mathrm}{ave}} = (\epsilon_1 + \epsilon_2)/2$. If $\epsilon_1 = \epsilon_2$ , the photocounts are also statistically independent because $S$ and $D$ are independent in that case.
The SLIVER method consists of optimal statistical processing of the sequence of observed photocount measurements of $N_s$ and (or) $N_a$ over $M$ independent shots using the same sample to obtain a good estimate $\hat{d}$ of the separation $d$ – see Fig. 2 and Secs. 4 and 5. As we show below, the photocount from the antisymmetric component of the input field is much more informative in this regard.
{width="0.9\columnwidth"}
Cramér-Rao bounds for the separation $d$
========================================
Number-resolved photon counting
-------------------------------
Given the Bose-Einstein distribution $$\begin{aligned}
\label{Ps}
P_s(n) &:=\mbox{Pr}[N_s = n] = \frac{1}{\bar{N}_s +1}{\left(}\frac{\bar{N}_s }{\bar{N}_s+1} {\right)}^n,\end{aligned}$$ of number-resolved photon counting of the inversion-symmetric field, we can calculate the fundamental Cramér-Rao (CR) lower bound of estimation theory [@VanTreesI] on the mean squared error (MSE) $\mathbb{E}[\hat{d}(N_s) -d]^2$ of any unbiased estimator $\hat{d}(N_s)$ of $d$ that is a function of the photocount $N_s$. An unbiased estimator is one that satisfies $\mathbb{E}[\hat{d}(N_s)] = d$ for all $d$ in some interval of interest. The statistical expectations here are taken over the distribution of $N_s$. The CR bound is itself the reciprocal of the Fisher information ${\mathcal}{J}_d^{{\mathrm}{(sym-pc)}}$ on the parameter $d$ defined as $$\begin{aligned}
{\mathcal}{J}_d^{{\mathrm}{(sym-pc)} }= \mathbb{E}_{N_s}{\left[}\frac{\partial \log P_s(n)}{\partial d} {\right]}^2.\end{aligned}$$ The superscript indicates that the Fisher information pertains to photon counting of the inversion-symmetric field. This quantity turns out to equal $$\begin{aligned}
{\mathcal}{J}_d^{{\mathrm}{(sym-pc)} }& = \sum_{n=0}^{\infty} P_s(n) {\left[}\frac{\partial \log P_s(n)}{\partial d} {\right]}^2\\
& = {\left[}\frac{\partial \bar{N}_s}{\partial d} {\right]}^2 \sum_{n=0}^{\infty} P_s(n) {\left[}\frac{\partial \log P_s(n)}{\partial \bar{N}_s} {\right]}^2 \\
&= {\left[}\frac{\partial \bar{N}_s}{\partial d} {\right]}^2 \, \frac{1}{\bar{N}_s (\bar{N}_s +1)} \label{septerms}\\
&= \frac{\epsilon_{\mathrm}{ave}\gamma^2(d)}{1 +\delta(d)} \left\{1 + \epsilon_{\mathrm}{ave}{\left[}1 + \delta(d){\right]}\right\}^{-1}. \label{Jspc}\end{aligned}$$ Here we have defined $$\begin{aligned}
\label{gamma}
\gamma(d) = \delta'(d) =- \operatorname{Re\,}\iint\limits_{-\infty}^\infty \operatorname{\,d}x \operatorname{\,d}y\, \psi^*(x,y) \frac{\partial \psi(x-d,y)}{\partial x}.\end{aligned}$$ In a similar fashion, we find the Fisher information of photon counting of the inversion-antisymmetric field to be $$\begin{aligned}
\label{Japc}
{\mathcal}{J}_d^{({\mathrm}{asym-pc})} &= \frac{\epsilon_{\mathrm}{ave}\gamma^2(d)}{1 -\delta(d)} \left\{1 + \epsilon_{\mathrm}{ave}{\left[}1 - \delta(d){\right]}\right\}^{-1}.\end{aligned}$$ The CR lower bounds for these individual measurements then respectively read $$\begin{aligned}
\mathbb{E}[\hat{d}(N_s) -d]^2 &\geq \frac{1}{{\mathcal}{J}_d^{{\mathrm}{(sym-pc)} }} = \frac{1+\delta(d)}{\epsilon_{\mathrm}{ave}\gamma^2(d)} + {\left[}\frac{1+\delta(d)}{\gamma(d)}{\right]}^2 , \\
\mathbb{E}[\hat{d}(N_a) -d]^2 &\geq \frac{1}{{\mathcal}{J}_d^{{\mathrm}{(asym-pc)} }} = \frac{1-\delta(d)}{\epsilon_{\mathrm}{ave}\gamma^2(d)} + {\left[}\frac{1-\delta(d)}{\gamma(d)}{\right]}^2. \end{aligned}$$ For a fixed $d$, both bounds contain a term that displays the familiar shot-noise scaling with respect to the combined source strength $\epsilon_{\mathrm}{ave}$. Of greater interest here, however, is the behavior of the bounds as the separation reduces to zero for fixed source strengths. Since $\gamma(d) \rightarrow 0$ and $\delta(d) \rightarrow 1$ as $d \rightarrow 0$, we have $$\begin{aligned}
\lim_{d \rightarrow 0} \frac{1+\delta(d)}{\gamma^2(d)} = \infty,\end{aligned}$$ so that any unbiased estimator based on bulk direct detection of the inversion-symmetric field suffers a divergent mean squared error for separations $d \rightarrow 0$. Thus, Rayleigh’s curse that plagues spatially-resolved image-plane photon counting remains in effect here [@BDD+99; @RWO06]. However, note that $$\begin{aligned}
\lim_{d \rightarrow 0} \frac{ 1 -\delta(d)}{ \gamma^2(d)} = -\frac{\delta'(0)}{2\gamma(0) \gamma'(0)} = - \frac{1}{2\gamma'(0)} \equiv \frac{1}{2(\Delta k)^2}.\end{aligned}$$ Here we have used $$\begin{aligned}
\gamma'(d)& =- \operatorname{Re\,}\iint\limits_{-\infty}^\infty \operatorname{\,d}x \operatorname{\,d}y\,\frac{\partial \psi^*(x+d,y)}{\partial x} \frac{\partial \psi(x,y)}{\partial x}, \\
\gamma'(0) & = - \iint\limits_{-\infty}^\infty \operatorname{\,d}x \operatorname{\,d}y\,{\left | \frac{\partial \psi(x,y)}{\partial x} \right |}^2 \\
& \equiv - (\Delta k)^2, \label{Deltak2}\end{aligned}$$ where $ (\Delta k)^2$ is the squared spectral width of the PSF. Remarkably therefore, direct detection of the *inversion-antisymmetric field does not suffer Rayleigh’s curse* and a limiting MSE $$\begin{aligned}
\label{MSEmin}
\frac{1}{M{\mathcal}{J}_0^{{\mathrm}{(asym-pc)} }} = \frac{1}{2M\epsilon_{\mathrm}{ave}(\Delta k)^2} \equiv \frac{1}{N(\Delta k)^2}\end{aligned}$$ is potentially achievable for small $d$, where $M$ is the number of independent measurements, and $$\begin{aligned}
\label{N}
N= 2M\epsilon_{\mathrm}{ave} \end{aligned}$$ is the the mean number of photons emitted over the $M$ shots. The limiting Fisher information value $2M \epsilon_{\mathrm}{ave}(\Delta k)^2$ is similar to the quantum limit obtained in [@TNL15] and achieved by the SPADE and binary SPADE schemes therein. An advantage of SLIVER over those schemes is that there is no need to couple the image-plane field into a waveguide(s) with mode profile(s) tailored to the PSF.
If $\epsilon_1=\epsilon_2$, $N_s$ and $N_a$ are statistically independent so that the two Fisher information terms of eqs. and may be added to give the total Fisher information. The reciprocal of this quantity is then the CR bound for any unbiased estimate $\hat{d}(N_s,N_a)$ of $d$ based on both photocounts. In the remainder of the paper, however, we will only consider measurements in the *antisymmetric* output port.
On-off detection of the inversion-antisymmetric field
-----------------------------------------------------
Consider bucket direct detection at the antisymmetric port with an *on-off detector*, i.e., a detector that only distinguishes between no photon and one or more photons. The no-click probability and click probability are respectively $$\begin{aligned}
P_a(0) &= \frac{1}{1+ \epsilon_{{\mathrm}{ave}}{\left[}1- \delta(d){\right]}} \\
P_a(>0) &= \frac{\epsilon_{{\mathrm}{ave}}{\left[}1- \delta(d){\right]}}{1+ \epsilon_{{\mathrm}{ave}}{\left[}1- \delta(d){\right]}}.\end{aligned}$$ The Fisher information for this binary measurement can be shown to be $$\begin{aligned}
\label{Joo}
{\mathcal}{J}^{{\mathrm}{(asym-oo)}}_d &= \frac{\epsilon_{{\mathrm}{ave}}\gamma^2(d)}{ 1 -\delta(d)} \left\{1 + \epsilon_{{\mathrm}{ave}}{\left[}1 - \delta(d){\right]}\right\}^{-2}
< {\mathcal}{J}_d^{({\mathrm}{asym-pc})},\end{aligned}$$ which is expected as on-off detection is a coarser measurement than photon counting. However, it is interesting that on-off detection also evades Rayleigh’s curse for small $d$ and the Fisher information attains the optimal value at $d=0$. Note also that the Fisher information of Eq. is negligibly different from that of Eq. for $\epsilon_{\mathrm}{ave} \ll 1$. These behaviors are expected as multi-photon arrivals in the image-plane are rare in the regime $\epsilon_{\mathrm}{ave} \ll 1$ while multi-photon detections at the antisymmetric output port are rare in the $\epsilon_{{\mathrm}{ave}}{\left[}1- \delta(d){\right]}\ll 1$ regime, i.e., the regime of $d \approx 0$. Equations and indicate that the two antisymmetric detection methods should exhibit superresolution for *arbitrary* values of source strength beyond the $\epsilon_{\mathrm}{ave} \ll 1$ regime.
![Fisher information obtainable using the SLIVER method: - The Fisher information for photon counting $J_d^{{\mathrm}{sym-pc}}$ at the symmetric output port of the image inversion interferometer (blue curve), photon counting at the antisymmetric output port of the image inversion interferometer ($J_d^{{\mathrm}{asym-pc}}$ – dashed green curve) and for on-off detection at the antisymmetric output port ($J_d^{{\mathrm}{asym-oo}}$ – dotted red curve) are plotted against the source separation $d$. The plots are normalized with respect to the maximum value $2 \epsilon_{\mathrm}{ave} (\Delta k)^2$ of $J_d^{{\mathrm}{asym-pc}}$ and $J_d^{{\mathrm}{asym-oo}}$, attained at $d=0$. The superlocalization effect, consistent with the quantum bound of [@TNL15], is evident for $d \sim 0$. The circular Gaussian PSF of Eq. has been assumed and the plots are independent of the value of the half-width $\sigma$. The average source strength $\epsilon_{\mathrm}{ave}= 10^{-3}$ photons.](Figure3.pdf){width="70.00000%"}
We illustrate the above results for the circular Gaussian PSF $$\begin{aligned}
\label{Gaussianpsf}
\psi_G({{\boldsymbol}{\rho}}) = \frac{1}{(2\pi\sigma^2)^{1/2}}\exp{\left(}-\frac{{\left | {{\boldsymbol}{\rho}}\right |}^2}{4\sigma^2}{\right)}.\end{aligned}$$ The PSF-dependent quantities appearing in the CR bound then become:- $$\begin{aligned}
\delta_G(d) &= \exp{\left(}-\frac{d^2}{8\sigma^2}{\right)}, \\
\gamma_G(d) &= -\frac{d}{4\sigma^2}\,\exp{\left(}-\frac{d^2}{8\sigma^2}{\right)}, \\
(\Delta k)_G^2 &= \frac{1}{4\sigma^2}.\end{aligned}$$ The Fisher information quantities of Eqs. , , and pertaining to photon counting at the symmetric and antisymmetric ports, and on-off detection at the antisymmetric port of the interferometer are plotted as a function of the separation $d$ in Figures 3-4. The source strength is $\epsilon_{\mathrm}{ave}=10^{-3}$ photons and $0.5$ photons in Figures 3 and 4 respectively. The plots are normalized to the maximum value of $J_d^{{\mathrm}{asym-pc}}$ attained at $d=0$. We see that the information obtainable from the antisymmetric detection methods is maximum at $d=0$ and decreases thereafter. The information from detection at the symmetric port is zero for $d=0$ and reaches a peak around $d=2-2.5 \sigma$ and decreases thereafter.
![Fisher information obtainable using the SLIVER method: - The Fisher information for photon counting $J_d^{{\mathrm}{sym-pc}}$ at the symmetric output port of the image inversion interferometer (blue curve), photon counting at the antisymmetric output port of the image inversion interferometer ($J_d^{{\mathrm}{asym-pc}}$ – dashed green curve) and for on-off detection at the antisymmetric output port ($J_d^{{\mathrm}{asym-oo}}$ – dotted red curve) are plotted against the source separation $d$. The plots are normalized with respect to the maximum value $2 \epsilon_{\mathrm}{ave} (\Delta k)^2$ of $J_d^{{\mathrm}{asym-pc}}$ and $J_d^{{\mathrm}{asym-oo}}$, attained at $d=0$. The superlocalization effect, consistent with the quantum bound of [@TNL15], is evident for $d \sim 0$. The circular Gaussian PSF of Eq. has been assumed and the plots are independent of the value of the half-width $\sigma$. The average source strength $\epsilon_{\mathrm}{ave}= 0.5$ photons, so that superresolution persists outside the regime $\epsilon_{\mathrm}{ave} \ll 1$ and is more marked for number-resolved measurements.](Figure4.pdf){width="70.00000%"}
Maximum-Likelihood (ML) estimates of $d$: Monte-Carlo analysis
==============================================================
To further support the predictions from the CR bounds, we present Monte-Carlo simulations of the MSE for on-off detection and photon counting at the antisymmetric output port of the interferometer. The circular Gaussian PSF of Eq. is assumed and the simulation results are independent of the half-width $\sigma$.
On-off detection in the antisymmetric output port
-------------------------------------------------
Consider direct detection of the output $E_a({{\boldsymbol}{\rho}})$ in $M$ detection windows using an on-off detector. The ensuing measurement record is a bit string ${\mathbf}{k} = (k_1, \ldots, k_M)$, with $k_m=0$ if the detector did not fire in the $m$-th detection window and $k_m=1$ if it did. The maximum likelihood (ML) estimator [@VanTreesI] for $d$ is then:- $$\begin{aligned}
\label{MLestimatoroo}
\hat{d}_{\mathrm}{ML} = \left\{
\begin{array}{ll}
2 \sigma \sqrt{-2 \ln {\left(}1 - \frac{K}{(M-K) \epsilon_{\mathrm}{ave}}{\right)}} & {\mathrm}{if}\;\; \frac{K}{M-K} < \epsilon_{\mathrm}{ave}.\\
2 \sigma & \mbox{otherwise.}
\end{array}
\right.\end{aligned}$$ Here, $K = \sum_{m=1}^M k_m$ is the total number of clicks observed in the detector during the measurement, which is a sufficient statistic for generating the estimate. The second case above is necessary because the equation for the maximum likelihood estimate of $d$ has no solution if $K/(M-K) \geq \epsilon_{\mathrm}{ave}$, in which case the estimate is set to an arbitrary value. As $M$ increases, the probability of such “large deviations” goes to zero.
![Simulated mean squared errors for the maximum likelihood estimator $\hat{d}_{\mathrm}{ML}$ of Eq. as a function of the source separation for $M = 100, 200, 400$ measurements. Each measurement consists of a binary value indicating whether or not an on-off detector in the antisymmetric output port of the interferometer registered a click. The MSE curves are each scaled by $[2M\epsilon_{\mathrm}{ave} (\Delta k)^2]^{-1}$, the value of the Cramér-Rao bound for that $M$ at $d=0$. Each data point was obtained as an average of $10^5$ Monte-Carlo runs. The Cramér-Rao bound for on-off detection in the antisymmetric output port (solid blue curve) normalized to unity at its minimum value of $[2M\epsilon_{\mathrm}{ave} (\Delta k)^2]^{-1}$ is also shown. A circular Gaussian PSF (Eq. ) was assumed and the source strength $\epsilon_{\mathrm}{ave} = 0.2$ photons. The simulated results are independent of the half-width $\sigma$.](Figure5.pdf){width="\textwidth"}
Figure 5 shows the simulated MSE of the ML estimate for a source strength of $\epsilon_{\mathrm}{ave} = 0.2$ photons. The MSE is shown scaled relative to the minimum value $[2M\epsilon_{\mathrm}{ave} (\Delta k)^2]^{-1}$ of the CR bound for $M$-values of $100$, 200 and $400$. As the separation increases beyond the half-width, the curves for different $M$ approach the Cramér-Rao bound for on-off detection in the antisymmetric output port. We see that the ML estimate actually beats the CR bound for small $d \lesssim 0.3 \sigma$. This is because the ML estimate is biased for finite $M$, as may be expected from the highly nonlinear nature of the estimator of Eq. . We recall here that bias and MSE are independent characteristics of an estimator. Many minimum MSE estimators are biased [@VanTreesI], and techniques for reducing MSE by introducing bias are well-known in the signal processing literature [@KE08]. Since MSE is the relevant metric for most applications, the behavior of the MSE observed here is a welcome – though intriguing – feature. For all values of the separation, the performance – when not superior to that predicted by the CR bound – is within a factor of 2 of it. The behavior of the simulated MSE of SLIVER with respect to the CR bound is very similar to that reported for the SPADE and binary SPADE schemes in [@TNL15].
For PSFs other than the circular Gaussian, the ML estimator takes a different form than that in Eq. , but depends on the measurement result only through the quantity ${\left(}\frac{K}{M-K} {\right)}$.
Photon counting in the antisymmetric output port
------------------------------------------------
For number-resolved photon counting in the antisymmetric output port, the measurement record ${\mathbf}{p} = (p_1,\ldots, p_M)$ is now a vector of non-negative integers. The total number of photons counted – $P = \sum_{m=1}^M p_m$ – is a sufficient statistic and the ML estimator for $d$ is:- $$\begin{aligned}
\label{MLestimatorpc}
\hat{d}_{\mathrm}{ML} = \left\{
\begin{array}{ll}
2 \sigma \sqrt{-2 \ln {\left(}1 - \frac{P}{M \epsilon_{\mathrm}{ave}}{\right)}} & {\mathrm}{if}\;\; \frac{P}{M} < \epsilon_{\mathrm}{ave}.\\
2 \sigma & \mbox{otherwise,}
\end{array}
\right.\end{aligned}$$ where the last value is again an arbitrary assignment needed in the event that a solution to the equation for the ML estimate for $d$ does not exist.
![Simulated mean squared errors for the maximum likelihood estimator $\hat{d}_{\mathrm}{ML}$ of Eq. as a function of the source separation for $M = 50, 100, 150$ measurements. Each measurement counts the number of photons in the antisymmetric port, i.e., is a measurement of $N_a$ of Sec. 2. The MSE curves are each scaled by $[2M\epsilon_{\mathrm}{ave} (\Delta k)^2]^{-1}$, the value of the Cramér-Rao bound for that $M$ at $d=0$. Each data point was obtained as an average of $10^5$ Monte-Carlo runs. The Cramér-Rao bound for photon counting in the antisymmetric output port (solid blue curve) normalized to unity at its minimum value of $[2M\epsilon_{\mathrm}{ave} (\Delta k)^2]^{-1}$ is also shown. A circular Gaussian PSF (Eq. ) was assumed, the source strength $\epsilon_{\mathrm}{ave} = 5$ photons, and the simulated results are independent of the half-width $\sigma$. ](Figure6.pdf)
Figure 6 shows the results for the MSE of the ML estimate for a source strength $\epsilon_{\mathrm}{ave} = 5$ photons, well outside the $\epsilon_{\mathrm}{ave} \ll 1$ regime. For such a source, an on-off detector would need a very large $M$ to collect useful statistics because of detector saturation. The simulated MSE is shown scaled relative to the minimum value $[2M\epsilon_{\mathrm}{ave} (\Delta k)^2]^{-1}$ of the CR bound for $M$-values of 50, 100 and 150. The ML estimate again beats the CR bound for small $d \lesssim 0.1 \sigma$ owing to the biasedness of the estimator of Eq. . However, the curves for each $M$ closely approach the CR bound for photon counting in the antisymmetric output port for separations $d \gtrsim 0.5\sigma$. For all values of the separation, the observed performance – when not superior to that predicted by the CR bound – is within a factor of 2 of it.
For PSFs other than the circular Gaussian, the ML estimator takes a different form than that in Eq. , but depends on the measurement result only through the quantity ${\left(}\frac{P}{M} {\right)}$.
Effect of centroid mismatch: On-off detection in the antisymmetric output port
------------------------------------------------------------------------------
We have so far assumed that the centroid has been located at the optical axis, so that the vector $\bf{c}$ from the origin of the image plane to the midpoint of $\bf{d}$ is zero (Fig. 1). A detailed study of the operation and performance of SLIVER when this assumption is removed will be given elsewhere. Here, we briefly evaluate the performance of on-off detection in the antisymmetric port for nonzero $\bf{c}$. It can be shown that the no-click and click probability are given by $$\begin{aligned}
\label{onoffmismatch}
P_a(0) &= {\left[}1 + \alpha \epsilon_1 + \beta \epsilon_2 + {\left(}\alpha \beta - \gamma^2 {\right)}\epsilon_1 \epsilon_2{\right]}^{-1}, \\
P_a(>0) &= 1 - P_a(0), \nonumber\end{aligned}$$ where $$\begin{aligned}
\alpha &= \frac{1-\delta{\left(}{\left | 2\bf{c} - \bf{d} \right |}{\right)}}{2}, \nonumber\\
\beta &= \frac{1-\delta{\left(}{\left | 2\bf{c} + \bf{d} \right |}{\right)}}{2}, \label{albetgam}\\
\gamma & = \frac{\delta{\left(}{\left | \bf{d} \right |} {\right)}- \delta{\left(}{\left | 2\bf{c} \right |}{\right)}}{2}. \nonumber\end{aligned}$$
Consider the case of $\bf{c}$ parallel to the separation vector $\bf{d}$ (Fig. 7). For the Gaussian PSF of Eq. , let ${\left | \bf{c} \right |} = \xi \sigma,$ so that $\xi$ is a normalized misalignment factor. For such a configuration, and for fixed $\bf{c}$ with $\xi = 0.1$, simulation results for the MSE of the estimator of Eq. are shown in Fig. 8. Note that the estimator used cannot depend on the centroid position, as this is unknown in practice. We have assumed $\epsilon_1 = \epsilon_2 = \epsilon_{\mathrm}{ave} = 0.2$ photons.
![Simulated mean squared errors for the estimator Eq. for a misaligned SLIVER system using on-off detection in the antisymmetric port shown as a function of the source separation for $M = 100, 200, 400$ measurements. We have assumed that the centroid is a distance ${\left | \bf{c} \right |}\equiv \xi \sigma$ away from the inversion axis, that $\bf{c}$ and $\bf{d}$ are parallel, and that $\xi=0.1$. The MSE curves are each scaled by $[2M\epsilon_{\mathrm}{ave} (\Delta k)^2]^{-1}$, the value of the Cramér-Rao bound for the *ideal* aligned measurement ($\xi =0$) and that $M$ at $d=0$. Each data point was obtained as an average of $10^5$ Monte-Carlo runs. The Cramér-Rao bound for on-off detection in the antisymmetric output port of an aligned SLIVER system (solid blue curve) normalized to unity at its minimum value of $[2M\epsilon_{\mathrm}{ave} (\Delta k)^2]^{-1}$ is also shown. A circular Gaussian PSF (Eq. ) was assumed and the source strengths $\epsilon_1= \epsilon_2 = \epsilon_{\mathrm}{ave} = 0.2$ photons. The simulated results are independent of the half-width $\sigma$.](Figure8.pdf)
Comparing to Fig. 5, we see that the misalignment has little effect on the MSE for separations $d \gtrsim 0.5 \sigma$. For smaller separations, the MSEs in Fig. 8 are greater than those in Fig. 5 and the MSE does not go to zero for zero separation. However, the MSEs still beat the CR bound for the *aligned* SLIVER system, suggesting they also beat the CR bound – yet to be determined – for the misaligned system.
Figure 9 displays the simulated MSEs for the same source and system parameters, with $M=100$ and for the misalignment factor values $\xi = 0.1, 0.2, 0.3$ and 0.4. We see that the performance degrades gracefully with increasing misalignment, remains within an order of magnitude of the CR bound of the aligned system, and does not display the divergent behavior of image-plane photon counting [@BDD+99; @RWO06; @TNL15].
A detailed study of SLIVER under misalignment – including derivations of the formulas in this section, Cramér-Rao bounds, and performance analyses – will be given elsewhere. Here, we simply note the important point that the MSE of a misaligned SLIVER system is still finite for arbitrarily small separations, and Rayleigh’s curse has not returned.
![Simulated mean squared errors for the estimator Eq. for misaligned SLIVER as a function of the source separation for $M = 100$ measurements and for misalignment factors $\xi = 0.1, 0.2, 0.3$ and 0.4. The MSE curves are each scaled by $[2M\epsilon_{\mathrm}{ave} (\Delta k)^2]^{-1}$, the value of the Cramér-Rao bound for the *ideal* aligned measurement ($\xi =0$) at $d=0$. Each data point was obtained as an average of $10^5$ Monte-Carlo runs. The Cramér-Rao bound for on-off detection in the antisymmetric output port of an aligned SLIVER system normalized to unity at its minimum value is also shown. A circular Gaussian PSF (Eq. ) was assumed and $\epsilon_1= \epsilon_2 = \epsilon_{\mathrm}{ave} = 0.2$ photons. The simulated results are independent of the half-width $\sigma$.](Figure9.pdf)
Discussion and conclusions
==========================
Motivated by the greater familiarity and facility of analysis afforded by the semiclassical theory of photodetection over the full quantum theory, we have set up the problem of resolving two incoherently radiating point sources in this framework. This theory is quantitatively exact for such sources and for measurements involving only linear-optics processing prior to photodetection. The sources are assumed to be of arbitrary strength and circularly-symmetric PSF’s directly related to real imaging systems are considered. Since the analysis in [@TNL15] was largely confined to a one-dimensional model of the point-spread function with a weak-source assumption, the model used here is of interest both theoretically and in practice (See also the remarks in Sec. V therein). We have not made full semiclassical analyses of the binary SPADE and SPADE protocols of [@TNL15] here, as these are rather involved and will be given elsewhere. Preliminary calculations indicate that the results from [@TNL15] for $\epsilon_{\mathrm}{ave} \ll 1$ agree with those from the semiclassical analysis to $O(\epsilon_{\mathrm}{ave}^2)$.
We have also proposed and analyzed a novel interferometric scheme – the SLIVER method of Sec. 3 – for estimating the separation between the sources once their centroid has been located. The scheme alleviates Rayleigh’s curse and provides superresolution at all values of source strengths. The essential ingredient of the scheme is an image inversion interferometer (III) – a device that separates the symmetric and antisymmetric parts of the image-plane field relative to inversion about the centroid of the sources. The SLIVER method, unlike the binary SPADE and SPADE schemes, does not require coupling the image-plane field into one or more waveguides with mode profiles tailored to the PSF of the imaging system. As such, it appears to be more readily implementable in practice.
An explanation of the superlocalization effect of SLIVER for estimating $d \approx 0$ may be given in terms of photon statistics as follows. If both sources were superimposed at the centroid (so that $d=0$), the field $E_a({{\boldsymbol}{\rho}})$ in the antisymmetric output port of the III is identically zero despite their incoherence. If $d$ is small but nonzero, the mean photon number $\bar{N}_a$ is finite but still small as per Eq. . The Fisher information on the mean of a Bose-Einstein distribution equals $[\bar{N}_a(\bar{N}_a +1)]^{-1}$ and hence is very large in this region of $d$. This extreme sensitivity to changes in $d$ is partially offset by the weak sensitivity of $\bar{N_a}$ on $d$ via $\gamma(d)$ which is close to zero for $d \approx 0$ (cf. the term in square brackets in Eq. ). These two effects compensate each other to give a finite value of the Fisher information for separations close to zero. In the symmetric port, on the other hand, the mean output is large at $d \approx 0$ and the Fisher information is correspondingly small, so that the second effect dominates and superlocalization is not obtained. Similar remarks apply to on-off detection in the two output ports. It is rather remarkable that appropriate linear-optics processing followed by spatially-unresolved on-off detection can alleviate the divergent MSE behavior plaguing ideal spatially-resolved image-plane photon counting for closely separated sources. Somewhat intriguingly, receivers using the same toolkit of operations (linear-optics processing with bulk on-off detection) also approach the quantum limits of binary [@Ken73; @Dol73] and $M$-ary [@NGT14] communication with laser (coherent-state) light.
As mentioned in Sec. 3, various imaging modalities employing image inversion have been proposed and demonstrated previously in the microscopy literature [@SG06; @WH07; @WSH09; @Wei+11; @WBK+11; @WEB+13; @WBK+14; @WBK+15]. The importance of the antisymmetric output of the III for improving lateral resolution has also been noticed, e.g., in [@WH07]. However, the claimed resolution enhancements in these works are in terms of the reduction of effective widths of the PSF by fixed factors, which in turn are related to the intensity patterns at the two outputs of the III (or their difference). Thus, from a statistical viewpoint, it appears that these works focus on the spatial structure of the first moment of the photocount random process (i.e, the intensity) at the two output channels of the interferometer and on the information about the separation $d$ extractable from it. In contrast, we have focused on spatially unresolved detection at the antisymmetric output port of the III and on the Cramér-Rao bound as the fundamental limit to the mean squared error of any (unbiased) estimate of the separation. By considering statistically optimal estimators, we can exploit the information available in the full probability distribution of the measurement rather than just its first moment. Such subtler estimators are necessitated by the limited number of photons that can be collected in most imaging applications. The maximum likelihood estimator, which approaches the Cramér-Rao bound in the limit of large $M$ [@VanTreesI], was shown to yield performance in close agreement to the bound and sometimes exceeding it.
In order to focus on the spatial aspects of the resolution problem, we have, as in [@TNL15], suppressed the explicit temporal dependence of the field in our analysis. In effect, we have assumed that the light emitted by the sources is in a single quasimonochromatic temporal mode in each detection window. Such a situation can be easily realized using a pseudothermal light source [@Goo85Statistical], but is in general unrealistic for many astronomical and biological imaging situations in which the coherence time of the emitted radiation is typically much shorter than the duration of the detection window. An approximate treatment of such cases can be given using the concept of the number of coherence cells [@Goo85Statistical]. This approach effectively redefines the number of shots $M$ according to the coherence properties of the light, with exact results obtainable in certain cases. These issues will be explored in detail elsewhere.
The effects of loss and non-unity quantum efficiency of the detectors can be incorporated in our analysis by appropriate scaling of the source strengths. Since spatially homogeneous losses simply scale the semiclassical field amplitudes, thermal states remain thermal states under such attenuation.Thus, a system resolving sources of strength $\epsilon_{\mathrm}{ave}$ using a detector with quantum efficiency $\eta$ can be modeled as a system with source strength $\eta\epsilon_{\mathrm}{ave}$ and an ideal detector.
We have assumed through most of this paper that the centroid of the sources has been located at the optical axis. In optical astronomy applications, the centroid position may be available from previous observational records, or else a long observation time should be available to determine the centroid by, say, image-plane photon counting. In microscopy applications, the time available for measurements may be more limited, e.g., because of photobleaching in fluorescent samples[@Kub13], and lack of knowledge of the centroid position can be a significant impediment to implementing SLIVER. To address this problem, one can first perform image-plane photon counting using a portion of the available light to determine the centroid position to within a small fraction of the PSF width before implementing SLIVER. The simulations in Sec. 5.3 indicate that the MSE of separation estimation should still be well below that achievable using image-plane photon counting. A more general study of the effects of centroid mismatch is best carried out in the framework of multi-parameter estimation theory and will be given elsewhere.
Apart from exploring these issues, various extensions and refinements of the schemes of this paper suggest themselves. For example, spatially-resolved detection in one or both output ports of the interferometer can only improve the performance of SLIVER. In any case, the results here should hopefully spur investigations into these issues in order to reap the superlocalization gain promised by the SLIVER method.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by the Singapore National Research Foundation under NRF Grant No. NRF-NRFF2011-07.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Applying the method of the paper [@CT], we perform a quantum version of the Drinfeld-Sokolov reduction in Reflection Equation algebras and braided Yangians, associated with involutive and Hecke symmetries of general forms. This reduction is based on the Cayley-Hamilton identity valid for the generating matrices of these algebras.'
author:
- |
Dimitri Gurevich[^1]\
[*LAMAV, Université de Valenciennes, 59313 Valenciennes, France*]{}\
Pavel Saponov[^2]\
[*National Research University Higher School of Economics,*]{}\
[*20 Myasnitskaya Ulitsa, Moscow 101000, Russian Federation*]{}\
[*and*]{}\
[*Institute for High Energy Physics, NRC “Kurchatov Institute”*]{}\
[*Protvino 142281, Russian Federation*]{}\
Dmitry Talalaev[^3]\
[*Moscow State University, Faculty of Mechanics and Mathematics*]{}\
[*119991 Moscow, Russian Federation*]{}
title: 'Drinfeld-Sokolov reduction in quantum algebras'
---
addtoreset[equation]{}[section]{}
[**Keywords:**]{} Reflection Equation algebra, braided Yangian, second canonical form, quantum elementary symmetric elements, quantum power sums, Cayley-Hamilton identity
Introduction
============
In the seminal paper [@DS] it was shown that any connection operator $\pa_u-M(u)$, where $\pa_u=\frac{d}{du}$ and $M(u)$ is a $N\times N$ matrix smoothly depending of the parameter $u$, can be reduced by means of the gauge transformations $$\pa_u-M(u)\,\mapsto\, g(u) (\pa_u-M(u)) g(u)^{-1}=\pa_u-g(u) M(u) g(u)^{-1}+g(u) \pa_u(g(u)^{-1})$$ to the form $\pa_u-M_{can}(u)$, where $M_{can}(u)$ looks like (\[can\]). This form of operators and matrices will be called [*canonical*]{} or, more precisely, [*second canonical*]{}. Also, the authors of [@DS] found that the reduced Poisson structure can be identified with the second Gelfand-Dickey one. In [@FF] this Poisson reduction was quantized in terms of the BRST cohomology and the quantum object was identified with $W$-algebras.
A natural question arises: what are suitable operators and their canonical forms related to the quantum groups or other quantum algebras. Thus, in [@FR; @FRS] a version of the Drinfeld-Sokolov reduction is considered, where the operator $\pa_u$ is replaced by that $D_q f(x)=f(qx)$, $q\not=1$ and the gauge transformations are replaced by $$D_q-M(u)\,\mapsto \,g(qu)(D_q-M(u))g(u)^{-1}.$$ However, this structure is not a deformation of the classical one since for $q=1$ the corresponding operators differ from the classical opertors.
In this note we apply another approach, suggested in [@CT]. There, the second canonical form was found for a matrix $M(u)$ subject to the relation =, \[Gaud\] where $P$ is the $N^2\times N^2$ matrix of the usual flip, $M(u)=\sum_{k\geq 0} M[k] u^{-k}$ is an $N\times N$ matrix, and, as usual, $M_1=M\ot I$, $M_2=I\ot M$. Note that the Hamiltonians of the rational Gaudin model are expressed via such a matrix.
Besides, the matrix $M(u)$ satisfies the Cayley-Hamilton (CH) identity with modified matrix powers of $M(u)$. Namely, the terms $M(u)^k$ have been replaced by some combinations of this matrix and its derivatives in $u$. By using these modified powers one can show that the operators $\pa_u-M(u)$ and $\pa_u-M_{can}(u)$ are in a sense similar (see [@CT] for detail). Namely, in this sense the term “Drinfeld-Sokolov reduction” was used in [@CT].
In [@CF] a version of the CH identity for the matrices $L(u)$ which generates the Drinfeld’s Yangian $\Y(gl(N))$ was found (below, $L(u)$ and other similar matrices will be called [*generating matrices*]{}). The powers of this matrix are also understood in a modified sense: namely, in the products of the matrix $L(u)$ the arguments of the factors are shifted. The use of the CH identity allows one to define the second canonical form of the generating matrix $L(u)$ and to establish the similarity of the initial and reduced forms of this matrix.
The aim of this note is to define an analogous Drinfeld-Sokolov (DS) reduction for the generating matrices of some quantum algebras. Namely, we perform this reduction for the generating matrices of the Reflection Equation algebras, corresponding to constant (i.e. independent of the parameters) involutive or Hecke symmetries $R$ and for the [*braided Yangians*]{}, also arising from these symmetries. Such type Yangian-like algebras have been recently introduced [@GS2; @GS3].
We mention here one of possible applications of this result of particular interest. The quantum algebras under consideration have direct relevance to the theory of quantum integrable systems due to the presence of typical commutative families. We hope that the construction of the second canonical form will allow us to establish its relations with new versions of the KZ equation.
[**Acknowledgements:**]{} The work of D.T. was partially supported by the grant of the Simons foundation and the RFBR grant 17-01-00366 A. The work of P.S. has been funded by the Russian Academic Excellence Project ’5-100’ and was also partially supported by the RFBR grant 16-01-00562.
Quantum Matrix algebras
=======================
Let us briefly describe the quantum algebras under consideration. First, recall that by a current $R$-matrix one usually means an operator $R(u,v)$ depending on parameters $u$ and $v$ and subject to the so-called Quantum Yang-Baxter equation $$R_{12}(u,v) R_{23}(u,w) R_{12}(v,w) =R_{23}(v,w) R_{12}(u,w) R_{23}(u,v),$$ where $R_{12}(u,v)=R(u,v)\ot I$ and $R_{23}(u,v)=I\ot R(u,v)$. If $R$ is independent of the parameters it is also called [*a braiding*]{}. In this case we shall assume the operator $R:\vv\to\vv$ to be either involutive $R^2=I$ or to satisfy the Hecke condition $$(q\, I-R)( \qq\, I+R)=0, \quad q\in \C, \quad q^2\not=1.$$ Here, $V$ is a vector space of the dimension $N$ (over the field $\C$). These braidings are respectively called [*involutive and Hecke symmetries*]{}.
The best known are the Hecke symmetries coming from the quantum groups $U_q(sl(N))$. However, there are known numerous examples of involutive and Hecke symmetries which are deformations neither of usual nor of super-flips (see [@GPS2] and the references therein).
All symmetries, we are dealing with, are assumed to be [*skew-invertible*]{} and to have the bi-rank $(m|0)$. For the definitions of the notions “skew-invertible” and “bi-rank” the reader is referred to [@GS2; @GS3]. We want only to mention that for a skew-invertible braiding $R$, an $R$-trace $${\rm Tr}_R:\End(V)\to \C$$ in a sense coordinated with $R$ can be defined. Moreover, for any matrix $A$ the expression ${\rm Tr}_R \, A$ is also well-defined. The properties of such an $R$-traces can be found in [@O; @GPS2].
Let us respectively define two quantum matrix algebras by the following systems of relations on their generators R T\_1T\_2-T\_1 T\_2 R=0,T=t\_i\^j\_[1i,j N]{} \[RTT\] R L\_1 RL\_1-L\_1 R L\_1 R=0,L=l\_i\^j\_[1i,j N]{} \[RE\] where $R$ is assumed to be a skew-invertible involutive or Hecke symmetry. The former algebra is called an RTT one, the latter one is called the Reflection Equation (RE) algebra. Their detailed consideration can be found in [@GPS2]. Here, we want only to observe that if $R$ is a deformation of the usual flip $P$, the dimensions of the homogenous components of both algebras are classical, i.e. equal to those in ${\rm Sym}(gl(N))$ (if $R$ is a Hecke symmetry, the value of $q$ is assumed to be generic).
Now, let us exhibit Yangian-like algebras associated with current $R$-matrices. First, observe that the current $R$-matrices, we are dealing with, are constructed from involutive or Hecke symmetries via the so-called Baxterization procedure described in [@GS2; @GS3]. This procedure results in the following $R$-matrix R(u,v) = R - g(u,v) I,g(u,v)=g(u,v)=. \[curR\] If $R$ is an involutive symmetry, $g(u,v)$ is defined by the former formula. If $R$ is a Hecke symmetry, $g(u,v)$ is defined by the latter one.
The corresponding Yangian-like algebras, introduced in [@GS2; @GS3] are also of two types. They are respectively defined by the following systems R(u,v) T\_1(u) T\_2(v)=T\_1(v) T\_2(u) R(u,v), \[ya\] R(u,v) L\_1(u) R L\_1(v)=L\_1(v) R L\_1(u) R(u,v), \[brya\] where the generating matrices $T(u)$ and $L(u)$ are assumed to possess a series expansion as the matrix $M(u)$ above. We call the former (resp., latter) algebras [*the Yangians of RTT type*]{} (resp., [*the braided Yangians*]{}). Note that the braided Yangians are defined similarly to the RE algebras, but with the current $R$-matrices at the outside positions.
Let us introduce the following notation L\_[1]{}(u)=L\_1(u),L\_[k]{}(u)=R\_[k-1]{} L\_(u) R\^[-1]{}\_[k-1]{},k2, \[note\] where we write $R_i$ instead of $R_{i\, i+1}$ (recall that $R_{i\, i+1}$ stands for the operator $R$ acting at the $i$-th and $i+1$-th positions in the tensor product $V^{\ot k}$, $i+1\leq k$). In the RE algebras we use the same notation for the generating matrix $L$ which is independent of the parameters.
By using this notation we can present the defining relations of an RE algebra in the form $$R_1 \, L_{\overline 1} \, L_{\overline 2}- L_{\overline 1} \, L_{\overline 2}\, R_1=0.$$ Moreover, in this algebra the following holds: $$R_k \, L_{\overline k} \, L_{\overline{k+1}}- L_{\overline k} \, L_{\overline{k+1}}\, R_k=0, \quad \forall\,k\ge 1.$$
This notation enables us to define analogs of symmetric polynomials in the RTT and RE algebras in a uniform way. Thus, the [*power sums*]{} are respectively defined as follows $$p_k(T)={\rm Tr}_{R(12\dots k)} R_{k-1}R_{k-2}\dots R_2R_1T_1\,T_2\dots T_k,$$ p\_k(L)=[Tr]{}\_[R(12…k)]{} R\_[k-1]{}R\_[k-2]{}…R\_2R\_1 L\_L\_[2]{}…L\_. \[ps\] Here the notation ${\rm Tr}_{R(12\dots k)}$ means that the $R$-traces are applied at the positions $1,2, \dots ,k$. Note that in the RE algebra the formula (\[ps\]) can be simplified to $p_k(L)={\rm Tr}_R L^k$ whereas for the power sums $p_k(T)$ in the RTT algebra such a transformation is not possible.
In a similar manner the “quantum powers" of the matrices $T$ and $L$ can be defined: $$T^{[k]}:={\rm Tr}_{R(2\dots k)} R_{k-1}R_{k-2}R_2\dots R_1T_1T_2\dots T_k,$$ $$L^{[k]}:={\rm Tr}_{R(2\dots k)} R_{k-1}R_{k-2}\dots R_2R_1 L_{\overline{1}}L_{\overline{2}}\dots L_{\overline{k}}.$$ However, if the former formula cannot be simplified, the latter one can be reduced to the usual one: $L^{[k]}=L^k$.
Also, exhibit analogs of elementary symmetric polynomials in both algebras $$e_0(T)=1,\qquad e_k(T):={\rm Tr}_{R(1\dots k)} (\A^{(k)} T_{1} \,T_{ 2}\dots T_{ k}),\quad k\geq 1,$$ e\_0(L)=1,e\_k(L):=[Tr]{}\_[R(1…k)]{} (\^[(k)]{} L\_[1]{} L\_[2]{}…L\_[k]{}),k1. \[ele\] Here $\A^{(k)}: V^{\ot k} \to V^{\ot k}$, $k\ge 1$ are the skew-symmetrizers (i.e. the projectors of skew-symmetrization) which are recursively defined by \^[(1)]{} = I,\^[(k)]{} = \^[(k-1)]{}(q\^[k-1]{} I-(k-1)\_q R\_[k-1]{})\^[(k-1)]{},k2, \[q-asym\] Hereafter, we use the standard notation: $k_q=\frac{q^k-q^{-k}}{q-q^{-1}}$.
Note that if the bi-rank of $R$ is $(m|0)$, $m\geq 2$, the skew-symmetrizers $\A^{(k)}$ are trivial for $k> m$ and the rank of the skew-symmetrizer $\A^{(m)}$ is equal to 1.
Analogs of matrix powers, power sums and elementary symmetric polynomials can be also defined for generalized Yangians of both classes (see [@GS2; @GS3]). We exhibit them below for the braided Yangians. The point is that only for the braided Yangians and RE algebras we can use the CH identity for performing the DS reduction in a way similar to that of [@CT]. This is due to the fact that in these algebras the powers of the generating matrices are classical (up to shifts of the arguments in the generating matrices) and the structure of the CH identity is close to the classical one.
Observe that analogs of the CH identity for the generating matrices of quantum matrix algebras associated with couples of braidings were presented in [@IOP]. Note also, that quantum analogs of the elementary symmetric polynomials and power sums can be defined in the frameworks of the so-called half-quantum algebras as defined in [@IO]. However, for proving their commutativity it is necessary to impose more strong relations giving rise to one of the quantum matrix algebras.
DS Reduction in RE algebras
===========================
Let $R$ be again a skew-invertible involutive or Hecke symmetry. Denote $\LL(R)$ the RE (\[RE\]). As was shown in [@GPS1], the generating matrix $L$ of the algebra $\LL(R)$ meets the [*quantum CH identity*]{} $Q(L)=0$, where the [*characteristic polynomial*]{} $Q(t)$ reads Q(t)=t\^m-q t\^[m-1]{} e\_1(L)+q\^2 t\^[m-2]{}e\_2(L)+…+(-q)\^[m-1]{} t e\_[m-1]{}(L)+(-q)\^[m]{} e\_m(L)=0. \[char\] Here, the factors $e_k(L)$ are the quantum elementary symmetric polynomials defined by (\[ele\]).
We stress a very important property of the polynomial (\[char\]): its coefficients belong to the center $Z(\LL(R))$ of the algebra $\LL(R)$. Let us introduce“eigenvalues“ $\{\mu_i\}_{1\le i\le m}$ of the matrix $L$ in a natural way $$e_1(L) = \mu_1+...+\mu_m,\quad\dots\quad e_m(L) = \mu_1\cdot\dots \cdot \mu_m.$$ These ”eigenvalues" are elements of an algebraic extension of the center $Z(\LL(R))$. Consider the quotient algebra $$\LL(R)/\langle e_1(L)-\al_1,\dots,e_m(L)-\al_m \rangle,\qquad \al_i\in \C,$$ where $<I>$ stands for the ideal generated by a set $I\subset \LL(R)$. This quotient is a quantum analog of an orbit (or a union of orbits) in the coadjoint representation of the group $GL(N)$. In [@GS1] there was considered the problem: for which values of $\al_i$ this quotient is an analog of a regular orbit. If it is so, we introduce the diagonal matrix $diag(\mu_1,\dots,\mu_m)$, where the elements $\mu_i$ solve the system $e_1(L)=\al_1,\dots,e_m(L)=\al_m$, and treat this matrix as the [*first canonical form*]{} of the generating matrix $L$.
Let us define the [*second canonical form*]{} of the matrix $L$: L\_[can]{}=(
[ccccc]{} 0&1&0&...&0\
0&0&1&...&0\
...&...&...&...&...\
0&0&0&...&1\
a\_m& a\_[m-1]{} &a\_[m-2]{}&...&a\_1
) , \[can\] where $$a_k=-(-q)^k e_k(L).$$
Following [@CT] we show that the matrices $L$ and $\Lc$ are in a sense similar. Let $v\in V$ be an arbitrary non-trivial vector written as a one-row matrix. Then we introduce the following $N\times N$ matrix C=(
[c]{} v\
vL\
...\
vL\^[m-1]{}
). \[CC\]
The following relation holds true CL= C. \[simi\]
In this sense we say that the matrices $L$ and $\Lc$ are [*similar*]{}.
In order to justify the term “similar” it would be desirable to show that at least for some vectors $v$ the matrix $C$ is invertible in the skew-field of the algebra $\LL(R)$. A similar problem is also open for other quantum algebras considered below.
Along with the RE algebra $\LL(R)$ define its quadratic-linear deformation $\hLL(R)$ by the following system R \_1 R\_1-\_1 R \_1 R=R\_1-\_1 R,=\_i\^j\_[1i,j N]{}. \[mRE\]
If the algebra $\LL(R)$ is a braided analog of the algebra ${\rm Sym}(gl(N))$, then $\hLL(R)$ is a braided analog of the enveloping algebra $U(gl(N))$. Namely, if $R$ is a Hecke symmetry which is a deformation of the usual flip (for instance, that coming from the quantum group $U_q(sl(N))$), the algebras $\LL(R)$ and $\hLL(R)$ turn into ${\rm Sym}(gl(N))$ and $U(gl(N))$ respectively as $q\to 1$. By using the fact that the algebras $\LL(R)$ and $\hLL(R)$ are isomorphic to each other if $q\not= \pm 1$, it is possible to get a characteristic polynomial $\hQ(t)$ for the generating matrix $\hL$ of the algebra $\hLL(R)$. Namely, we have (see [@GS2; @GS3]) $${\hat Q}(t)= {\rm Tr}_{R(1\dots m)} \left(\A^{(m)}(tI-\hL_{\overline 1})((q^2t-q) I-\hL_{\overline 2})\dots ((q^{2(m-1)} t-q^{m-1}(m-1)_q) I-\hL_{\overline m})\right).$$
In the algebra $\hLL$ the following matrix identity takes place ${\hat Q}(L)=0$.
Passing to the limit $q\to 1$ we get the characteristic polynomial for the generating matrix[^4] $M$ of the algebra $U(gl(N))$ (here $N=m$) $${\cal Q}(t)= {\rm Tr}_{(1\dots N)} \left(\A^{(N)}(tI-M_1)((t-1) I-M_2)\dots ((t-N+1) I-M_N)\right),$$ where $\A^{(N)}$ is the usual skew-symmetrizer in $V^{\otimes N}$ and ${\rm Tr}$ is the usual trace.
In the algebra $U(gl(N))$ the following holds ${\cal Q}(M)=0$.
Note that the famous Capelli determinant is defined by a similar formula. The same claim is valid for any algebra $\hLL$ provided $R$ be an involutive symmetry, which can be approximated by a Hecke symmetry.
Besides, it is possible to introduce the second canonical forms $\hL_{can}$ and $M_{can}$ of the matrices $\hL$ and $M$ respectively generating the algebras $\hLL(R)$ and $U(gl(N))$, and to perform a DS reduction of the matrices $\hL$ and $M$. Upon replacing the matrix $L$ in (\[CC\]) by $\hL$ and $M$ respectively, we get formulae similar to (\[simi\]). However, in the case of the algebra $\hLL$ we have first to normalize the polynomial ${\hat Q}(t)$ in order to get a monic one. Thus, if this monic polynomial is $L^m+b_{1} L^{m-1}+\dots + b_m I$, we have to put $a_k=-b_k$ in the matrix (\[can\]).
There are other matrices with entries from the algebras under consideration for which the CH identities exist. First, consider the enveloping algebra $U(gl(N))$. Its generating matrix $M$ belongs to the so-called Kirillov’s quantum family algebra (see [@K]) (U(gl(N))(V))\^[gl(N)]{}. \[fam\] Upon replacing $V$ by other irreducible $U(gl(N))$-modules, we get other quantum family algebras. For their generating matrices a characteristic polynomial can be found. Note that in the algebra $\hLL(R)$ a q-analog of (\[fam\]) is $$(\hLL\ot \End(V))^{U_q(sl(N))}.$$ Changing $\hLL$ for $\LL$, we get a “$q$-family algebra" for the algebra $\LL(R)$.
DS reduction in braided Yangians
================================
Now, consider a braided Yangians $\YR$. Using the notations (\[note\]) we can represent the defining system of this algebra as follows $$R(u,v)\, L_{\overline 1}(u)\, L_{\overline 2}(v)=L_{\overline 1}(v)\, L_{\overline 2}(u)\, R(u,v).$$
First, we assume $R$ to be an involutive symmetry and the corresponding $R$-matrix $R(u,v)$ to be given by the first formula (\[curR\]).
In this case we get the recurrence formula defining the skew-symmetrizers ${\cal A}^{(k)}$ by putting $q=1$ in (\[q-asym\]). Thus, we have $${\cal A}^{(k)} =\frac{1}{k} \, {\cal A}^{(k-1)}\left( I-(k-1) R_{k-1}\right) {\cal A}^{(k-1)}.$$
Let us respectively introduce the [*quantum elementary symmetric*]{} elements and the [*quantum matrix powers*]{} of $L(u)$ as follows: $$e_0(u)=1,\quad e_k(u) = {\rm Tr}_{ R(1\dots k)}\left( {\cal A}^{(k)} L_{\overline 1}(u)L_{\overline 2}(u-1)\dots L_{\overline k}(u-(k-1))\right),\quad k\geq 1,$$ $$L^{[0]}(u)=I,\quad L^{[k]}(u) = L(u-(k-1))\dots L(u-1)L(u),\quad k\geq 1.$$ Then the following CH identity is valid for the generating matrix $L(u)$ (see [@GS2; @GS3]) \_[k=0]{}\^m (-1)\^k L\^[\[m-k\]]{}(u-k) e\_k(u)=0. \[CH1\]
Consider the operator $L(u)e^{\pa_u}$, where $\pa_u=\frac{d}{du}$. Our current aim is to give an explicit canonical form $L_{can}(u)e^{\pa_u}$ of this operator and to show that the operators $L(u)e^{\pa_u}$ and $L_{can}(u)e^{\pa_u}$ are similar in the sense of formula (\[simi\]).
Note, that as $L_{can}(u)$ we use the matrix transposed to (\[can\]). This is motivated by the fact that the coefficients $e_p(u)$ in the CH identity (\[CH1\]) are on the right hand side from the factors $L^{[k]}(u)$. (Note that in the RE algebras these canonical forms are equivalent due to the centrality of the coefficients $\sigma_k(L)$.)
So, the matrix $L_{can}(u)$ reads L\_[can]{}(u)=(
[cccccc]{} 0&0&...&0&0&a\_m(u)\
1&0&...&0&0&a\_[m-1]{}(u)\
0&1&...&0&0&a\_[m-2]{}(u)\
...&...&...&...&...&...\
0&0&...&1&0&a\_2(u)\
0&0&...&0&1&a\_1(u)
) . \[can1\]
Let again $v\in V$ be an arbitrary non-trivial vector. Constitute an $N\times N$ matrix $$C(u)=\left(v, L(u)v, L^{[2]}(u+1)v\dots L^{[k]}(u+k-1))v\dots L^{[m-1]}(u+m-2)v \right)$$ where $v$ stands for the corresponding column.
The following holds L(u)e\^[\_u]{} C(u)= C(u) L\_[can]{}(u)e\^[\_u]{}, \[simi1\] provided the entries $a_k(u)$ are $$a_k(u)=(-1)^{k+1} \, e_k(u+m-1).$$
[**Proof.**]{} By using the evident relation $$e^{\pa_u}f(u)=f( u+1)e^{\pa_u},$$ and canceling the operator $e^{\pa_u}$, we can present the relation (\[simi1\]) as L(u)C( u+1)= C(u) L\_[can]{}(u). \[simi2\]
The equality of the corresponding columns from (\[simi2\]), except the last ones, immediately follows from the relation L(u)L\^[\[k\]]{}(u+k)=L\^[\[k+1\]]{}(u+k). \[add\] This of the last columns follows from the CH identity (\[CH1\]).
------------------------------------------------------------------------
Now, assume $R(u,v)$ to be defined by the second formula (\[curR\]). Let us respectively define the [*quantum elementary symmetric*]{} elements and [*quantum matrix powers*]{} by $$e_0(u)=1,\quad e_k(u) = {\rm Tr}_{R(1\dots k)}\left(\A^{(k)} L_{\overline 1}(u)L_{\overline 2}(q^{-2}u)\dots L_{\overline k}(q^{-2(k-1)}u)\right),\quad k\geq 1,$$ $$L^{[0]}(u)=I,\quad L^{[k]}(u)= L(q^{-2(k-1)} u) L(q^{-2(k-2)} u)\dots L(u), \quad k\geq 1.$$
Then the following form of the CH identity is valid for the generating matrix $L(u)$ (see [@GS2; @GS3]) \_[k=0]{}\^m (-q)\^k L\^[\[m-k\]]{}(q\^[-2k]{}u) e\_k(u)=0. \[CH2\]
Consider the operator $L(u)q^{2u \pa_u}$ and constitute the following matrix C(u)=(v, L(u)v, L\^[\[2\]]{}(q\^2u)v…L\^[\[k\]]{}(q\^[2(k-1)]{}u)v …L\^[\[m-1\]]{}(q\^[2(m-2)]{}u)v ), where $v$ stands for the corresponding column.
The following identity holds L(u)q\^[2u\_u]{} C(u)= C(u) L\_[can]{}(u)q\^[2u\_u]{}, \[simi3\] where the matrix $L_{can}(u)$ is given by (\[can1\]) provided the entries $a_k(u)$ are defined by $$a_k(u)=-(-q)^{k} \, e_k(q^{2(m-1)}u).$$
[**Proof.**]{} Now, we use the following relation $$q^{2u\pa_u}f(u)=f(q^2 u)q^{2u\pa_u}.$$ Then, upon canceling the factor $q^{2u\pa_u}$, we rewrite the relation (\[simi3\]) as L(u)C(q\^[2]{}u)= C(u) L\_[can]{}(u). \[simi4\]
Also, instead of (\[add\]) we use the relation $$L(u)L^{[k]}(q^{2k} u)=L^{[k+1]}(q^{2k} u).$$ This relation entails the equality of the corresponding columns from (\[simi4\]), except the last ones. The equality of the last columns follows from (\[CH2\]).
------------------------------------------------------------------------
In conclusion, we want to remark that the above version of the DS reduction is valid in a more general setting than the conventional one. Indeed, it is not possible to reduce the generating matrix $M$ of the enveloping algebra $U(gl(N))$ to a simpler form by usual transformations $M\to g\, M\, g^{-1}$, whereas a version of the reduction based on the CH identity is valid.
[MRPP]{}
A.Chervov, G.Falqui [*Manin matrices and Talalaev’s formula*]{}, J.Phys. A41 (2008) N9, 28 pp.
A.Chervov, D.Talalaev [*The KZ equation, G-opers, the quantum Drinfe’d-Sokolov reduction and the quantum Cayley-Hamilton identity*]{} J. Math. Sci. (N.Y.) 158 (2008), 904-911.
V.Drinfeld, V.Sokolov [*Lie algebras and equations of Korteweg-de Vries type*]{} (Russian) Current problems in math. Vol 24, 81-180, Moscow, 1984.
B.Feigin, E.Frenkel [*Quantization of the Drinfeld-Sokolov reduction*]{} Phys. Lett. B 246 (1990), 75-81.
E.Frenkel, N.Reshetikhin, M.Semenov-Tian-Shansky [*Drinfeld-Sokolov reduction for difference operators and deformations of W-algebras. I. The case of Virasoro algebra*]{}, CMP 192 (1998), 605-629.
E.Frenkel, N.Reshetikhin [*Quantum affine algebras and deformations of the Virasoro and W-algebras*]{}, CMP 178, 237-264.
D. Gurevich, P. Pyatov, P. Saponov [*Hecke symmetries and characteristic relations on reflection equation algebras*]{}, Lett. Math. Phys. 41 (1997), no. 3, 255–264.
D. Gurevich, P. Pyatov, P. Saponov [*Representation theory of (modified) Reflection Equation Algebra of the $GL(m|n)$ type*]{}, Algebra and Analysis, [**20**]{} (2008), 70–133.
D. Gurevich, P. Saponov [*Generic super-orbits in $gl(m|n)^*$ and their braided counterparts*]{}, J. of Geometry and Physics 60 (2010), 1411–1423
D. Gurevich, P. Saponov [*From Reflection Equation Algebra to Braided Yangians*]{}, Proceedings of International Conference on Mathematical Physics (Grozny, Russia), to be published, arXiv:1612.05929.
D. Gurevich, P. Saponov [*Generalized Yangians and their Poisson counterparts*]{}, Theoret. and Math. Pgysics 192 (2017), 1243–1257.
A.Isaev, O.Ogievetsky [*Half-quantum linear algebra*]{} in: Symmetries and groups in contemporary physics, pp 479–486, Nankai Ser. Pure Appl. Math. Theoret. Phys., 11, World Sci. Publ., Hackensack, NJ, 2013.
A.Isaev, O.Ogievetsky, P.Pyatov, [*On quantum matrix algebras satisfying the Cayley-Hamilton-Newton identities*]{}, J. Phys. A 32 (1999), no. 9, L115–L121.
A.Kirillov [*Introduction to family algebras*]{}, Mosc.Math. J (2001), 49-63.
O.Ogievetsky, [*Uses of Quantum Spaces*]{}, 3rd cycle. Bariloche (Argentine), 2000, pp.72, cel-00374419.
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
[^4]: Note, that the defining relations on the generators $m_i^j$ of the algebra $U(gl(N))$ $$m_i^j\, m_k^l- m_k^l\, m_i^j =m_i^l\delta_k^j- m_k^j\delta_i^l,$$ can be written in a matrix form with the use of the generating $N\times N$ matrix $M=\|m_i^j\|$: $$P\, M_1\, P\,M_1-M_1\, P\, M_1\, P=P\,M_1-M_1\, P.$$
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Cyril ODASSO\
\
Ecole Normale Supérieure de Cachan, antenne de Bretagne,\
Avenue Robert Schuman, Campus de Ker Lann, 35170 Bruz (FRANCE).\
and\
IRMAR, UMR 6625 du CNRS, Campus de Beaulieu, 35042 Rennes cedex (FRANCE)\
---
**Abstract**: We study the Navier-Stokes equations in dimension $3$ (NS3D) driven by a noise which is white in time. We establish that if the noise is at same time sufficiently smooth and non degenerate in space, then the weak solutions converge exponentially fast to equilibrium.
We use a coupling method. The arguments used in dimension two do not apply since, as is well known, uniqueness is an open problem for NS3D. New ideas are introduced. Note however that many simplifications appears since we work with non degenerate noises.
[**Key words**]{}: Stochastic three-dimensional Navier-Stokes equations, Markov transition semi-group, invariant measure, ergodicity, coupling method, exponential mixing, galerkin approximation.
Introduction {#introduction .unnumbered}
============
We are concerned with the stochastic Navier–Stokes equations on a three dimensional bounded domain (NS3D) with Dirichlet boundary conditions. These equations describe the time evolution of an incompressible fluid subjected to a determinist and a random exterior force and are given by $$\label{EqIntroNS}
\Espace
\left\{
\begin{array}{l}
dX+\nu(-\Delta)X \, dt+(X,\nabla)X\,dt+\nabla p \, dt = \phi(X)dW+ f \, dt,\\
\begin{array}{rclll}
\left(\textrm{div } X\right)(t,\xi) &=& 0, &\textrm{ for } \xi\in D,&t>0,\\
X(t,\xi) &=& 0,&\textrm{ for } \,\xi\in\partial D,&t>0,\\
X(0,\xi)&=& x_0(\xi), &\textrm{ for } \xi\in D.&
\end{array}
\end{array}
\right.$$ Here $D$ is an open bounded domain of $\R^3$ with smooth boundary $\partial D$ or $D=(0,1)^3$. We have denoted by $X$ the velocity, by $p$ the pressure and by $\nu$ the viscosity. The external force field acting on the fluid is the sum of a random force field of white noise type $\phi(X)dW$ and a determinist one $f \, dt$.
In the deterministic case ($\phi=0$), there exists a global weak solution (in the PDE sense) of when $x_0$ is square integrable, but uniqueness of such solution is not known. On another hand, there exists a unique local strong solution when $x_0$ is smooth, but global existence is an open problem (see [@temam-survey] for a survey on these questions). In the stochastic case, there exists a global weak solution of the martingale problem, but pathwise uniqueness or uniqueness in law remain open problems (see [@flandoli-cetraro] for a survey on the stochastic case).
The main result of the present article is to establish that, if $\phi$ is at the same time sufficiently smooth and non degenerate, then the solutions converge exponentially fast to equilibrium. More precisely, given a solution, there exists a stationary solution (which might depends on the given solution), such that the total variation distance between the laws of the given solution and of the stationary solution converges to zero exponentially fast.
Due to the lack of uniqueness, it is not straightforward to define a Markov evolution associated to . Some recent progress have been obtained in this direction. In [@DebusscheNS3D], [@ODASSOMarkov], under conditions on $\phi$ and $f$ very similar to ours, it is shown that every solution of limit of Galerkin approximations verify the weak Markov property. Uniqueness in law is not known but we think that this result is a step in this direction. Our result combined with this result implies that the transition semi-group constructed in [@DebusscheNS3D] is exponentially mixing.
Note also that recently, a Markov selection argument has allowed the construction of a Markov evolution in [@FlandoliRomito2]. Our result does not directly apply since we only consider solutions which are limit of Galerkin approximations. However, suitable modifications of our proof might imply that under suitable assumptions on the noise, the Markov semi-group constructed in [@FlandoliRomito2] is also exponentially mixing.
Our proof relies on coupling arguments. These have been introduced recently in the context of stochastic partial differential equations by several authors (see [@H], [@K], [@KS], [@KS2], [@KS3], [@Matt], [@ODASSO1], [@ODASSO3] and [@S]). The aim was to prove exponential mixing for degenerate noise. It was previously observed that the degeneracy of the noise on some subspace could be compensated by dissipativity arguments [@BKL], [@EMS], [@KS00]. More recently, highly degenerate noise noises have been considered in [@HM04], [@MattPar].
In all these articles, global well posedness of the stochastic equation is strongly used in many places of the proof. As already mentioned, this is not the case for the three dimensional Navier-Stokes equations considered here. Thus substantial changes in the proof have to be introduced. However, we require that the noise is sufficiently non degenerate and many difficulties of the above mentioned articles disappear.
The main idea is that coupling of solutions can be achieved for initial data which are small in a sufficiently smooth norm. A coupling satisfying good properties is constructed thanks to the Bismut-Elworthy-Li formula. Another important ingredient in our proof is that any weak solution enters a small ball in the smooth norm and that the time of entering in this ball admits an exponential moment. We overcome the lack of uniqueness of solutions by working with Galerkin approximations. We prove exponential mixing for these with constants which are controlled uniformly. Taking the limit, we obtain our result for solutions which are limit of Galerkin approximations.
Preliminaries and main result
=============================
Weak solutions
--------------
Here $\mathcal L(K_1;K_2)$ (resp $\mathcal L_2(K_1;K_2)$) denotes the space of bounded (resp Hilbert-Schmidt) linear operators from the Hilbert space $K_1$ to $K_2$.
We denote by ${\left\vert\cdot\right\vert}$ and $(\cdot,\cdot)$ the norm and the inner product of $L^2(D;\R^3)$ and by ${\left\vert\cdot\right\vert}_p$ the norm of $L^p(D;\R^3)$. Recall now the definition of the Sobolev spaces $H^p(D;\R^3)$ for $p\in \N$ $$\Espace
\left\{
\begin{array}{l}
H^p(D;\R^3)=\left\{X\in L^2(D;\R^3)\,\left|\,\partial_\alpha X \in L^2(D;\R^3)\,\textrm{ for } {\left\vert\alpha\right\vert}\leq p\right.\right\},\\
{\left\vertX\right\vert}^2_{H^p}=\sum_{{\left\vert\alpha\right\vert}\leq p}{\left\vert\partial_\alpha X\right\vert}^2.
\end{array}
\right.$$ It is well known that $(H^p(D;\R^3),{\left\vert\cdot\right\vert}_{H^p})$ is a Hilbert space. The Sobolev space $H^1_0(D;\R^3)$ is the closure of the space of smooth functions on $D$ with compact support by ${\left\vert\cdot\right\vert}_{H^1}$. Setting $
{\left\VertX\right\Vert}={\left\vert\nabla X\right\vert},
$ we obtain that ${\left\Vert\cdot\right\Vert}$ and ${\left\vert\cdot\right\vert}_{H^1}$ are two equivalent norms on $H^1_0(D;\R^3)$ and that $(H^1_0(D;\R^3),{\left\Vert\cdot\right\Vert})$ is a Hilbert space.
Let $H$ and $ V $ be the closure of the space of smooth functions on $D$ with compact support and free divergence for the norm ${\left\vert\cdot\right\vert}$ and ${\left\Vert\cdot\right\Vert}$, respectively.
Let $\pi$ be the orthogonal projection in $L^2(D;\R^3)$ onto the space $H$. We set $$A = \pi\left(-\Delta\right), \, D(A) = V \cap H^2(D;\R^3), \,
B(u,v)=\pi \left((u,\nabla)v\right)\,\textrm{ and
}\,B(u)=B(u,u).$$ Let us recall the following useful identities $$\Espace\left\{ \begin{array}{rclll}
(B(u,v),v)&=&0, &u,\,v \in V,\\
(B(u,v),w)&=&-(B(u,w),v),&u,\,v,\,w\in V.
\end{array}
\right.$$
As is classical, we get rid of the pressure and rewrite problem in the form $$\label{EqNS}
\Espace
\left\{
\begin{array}{rcl}
dX+ \nu A X dt+B(X) dt &=& \phi(X)dW+f \, dt,\\
X(0)&=& x_0,
\end{array}
\right.$$ where $W$ is a cylindrical Wiener process on $H$ and with a slight abuse of notations, we have denoted by the same symbols the projections of $\phi$ and $f$.
It is well-known that $\left(A,\Dr(A)\right)$ is a self-adjoint operator with discrete spectrum. See [@Cons], [@Temam]. We consider $(e_n)_n$ an eigenbasis of $H$ associated to the increasing sequence $(\mu_n)_n$ of eigenvalues of $\left(A,\Dr(A)\right)$. It will be convenient to use the fractionnal power $\left(A^s,\Dr(A^s)\right)$ of the operator $(A,\Dr(A))$ for $s\in \R$ $$\Espace
\left\{
\begin{array}{rcl}
\Dr(A^s)&=&\left\{X=\sum_{n=1}^\infty x_n e_n\,\left|\,\sum_{n=1}^\infty\mu_n^{2s}{\left\vertx_n\right\vert}^2<\infty\right.\right\},\\
A^s X&=&\sum_{n=1}^\infty \mu_n^s x_n e_n\;\textrm{ where }\;X=\sum_{n=1}^\infty x_n e_n.
\end{array}
\right.$$ We set for any $s\in \R$ $${\left\VertX\right\Vert}_s={\left\vertA^\frac{s}{2}X\right\vert},\quad
\H_s=\Dr(A^\frac{s}{2}).$$ It is obvious that $(\H_s,{\left\Vert\cdot\right\Vert}_s)$ is a Hilbert space, that $(\H_0,{\left\Vert\cdot\right\Vert}_0)=(H,{\left\vert\cdot\right\vert})$ and that $(\H_1,{\left\Vert\cdot\right\Vert}_1)=(V,{\left\Vert\cdot\right\Vert})$. Moreover, recall that, thanks to the regularity theory of the Stokes operator, $\H_s$ is a closed subspace of $H^s(D,\R^3)$ and ${\left\Vert\cdot\right\Vert}_s$ is equivalent to the usual norm of $H^s(D;\R^3)$ when $D$ is an open bounded domain of $\R^3$ with smooth boundary $\partial D$. When $D=(0,1)^3$, it remains true for $s\leq 2$.
Let us define $$\Espace
\left\{
\begin{array}{rcl}
\mathcal X&=&L^\infty_{\textrm{loc}}(\R^+;H)\cap L^2_{\textrm{loc}}(\R^+;V)\cap C(\R^+;\H_s),\\
\mathcal W&=&C(\R^+;\H_{-2}),\\
{\Omega_*}&=& \mathcal X\times \mathcal W,
\end{array}
\right.$$ where $s$ is any fixed negative number. Remark that the definition of $\mathcal X$ is not depending on $s<0$. Let ${X_*}$ (resp ${W_*}$) be the projector ${\Omega_*}\to\mathcal X$ (resp ${\Omega_*}\to\mathcal W$). The space ${\Omega_*}$ is endowed with its borelian ${\sigma}$-algebra ${\F^*}$ and with $\left({\F_t^*}\right)_{t\geq 0}$ the filtration generated by $({X_*},{W_*})$.
Recall that $W$ is said to be a $\left(\mathcal F_t\right)_{t}$–cylindrical Wiener process on $H$ if $W$ is $\left(\mathcal F_t\right)_{t}$–adapted, if $W(t+\cdot)-W(t)$ is independant of $\mathcal F_t$ for any $t\geq 0$ and if $W$ is a cylindrical Wiener process on $H$. Let $E$ be a Polish space. We denote by $P(E)$ the set of probability measure on $E$ endowed with the borelian ${\sigma}$–algebra.
A probability measure $\P_{\lambda}$ on $\left({\Omega_*},{\F^*}\right)$ is said to be a weak solution of with initial law $\lambda\in P(H)$ if the three following properties hold.
- The law of ${X_*}(0)$ under $\P_{\lambda}$ is $\lambda$.
- The process ${W_*}$ is a $({\F_t^*})_{t}$–cylindrical Wiener process on $H$ under $\P_{\lambda}$.
- We have $\P_{\lambda}$-almost surely $$\label{Def_weak}
\Espace
\begin{array}{r}
({X_*}(t),\psi)+\nu\int_0^t({X_*}(s),A\psi)ds+\int_0^t(B(X_*(s)),\psi)ds\quad\quad\quad\quad\quad\\
=({X_*}(0),\psi)+t\left(f,\psi\right)
+\int_0^t(\psi,\phi({X_*}(s))d{W_*}(s)),
\end{array}$$ for any $t\in\R^+$ and any $\psi$ smooth mapping on $D$ with compact support and divergence zero.
When the initial value $\lambda$ is not specified, $x_0$ is the initial value of the weak solution $\P_{x_0}$ (i.e. $\lambda$ is equal to $\delta_{x_0}$ the Dirac mass at point $x_0$).
These solutions are weak in both probability and PDE sense. On the one hand, these are solutions in law. Existence of solutions in law does not imply that, given a Wiener process $W$ and an initial condition $x_0$, there exist a solution $X$ associated to $W$ and $x_0$. On the other hand, these solutions live in $H$ and it is not known if they live in $\H_1$. This latter fact causes many problems when trying to apply Ito Formula on $F({X_*}(t))$ when $F$ is a smooth mapping. Actually, we do not know if we are allowed to apply it.
That is the reason why we do not consider any weak solution but only those which are limit in distribution of solutions of Galerkin approximations of . More precisely, for any $N\in\N$, we denote by $P_N$ the eigenprojector of $A$ associated to the first $N$ eigenvalues. Let $(\Omega,\F,\P)$ be a probability space and $W$ be a cylindrical Wiener process on $H$ for $\P$. We consider the following approximation of $$\label{galerkin}
\Espace
\left\{
\begin{array}{rcl}
dX_N+ \nu A X_N dt+P_NB(X_N) dt &=& P_N\phi(X_N)dW+P_Nfdt,\\
X_N(0)&=& P_N x_0.
\end{array}
\right.$$ In order to have existence of a weak solution, we use the following assumption.
\[HGal\] The mapping $\phi$ is bounded Lipschitz $H\to \mathcal L_2\left(H;\H_1 \right)$ and $f\in H$.
We set $$B_1=\sup_{x\in H}{\left\vert\phi(x)\right\vert}_{\mathcal L_2\left(H;\H_1
\right)}^2+\frac{{\left\vertf\right\vert}^2}{\nu \mu_1}.$$
It is easily shown that, given $x_0\in H$, has a unique solution $X_N=X_N(\cdot,x_0)$. Proceeding as in [@FlandoliGatarek], we can see that the laws $(\P_{x_0}^N)_N$ of $(X_N(\cdot,x_0),W)$ are tight in a well chosen functional space. Then, for a subsequence $(N_k)_k$, $(X_{N_k},W)$ converges in law to $\P_{x_0}$ a weak solution of . Hence we have existence of the weak solutions of , but uniqueness remains an open problem.
We only consider weak solutions constructed in that way. This allows to make some computations and to obtain many estimates. For instance, when trying to estimate the $L^2$-norm of ${X_*}(t)$ under a weak solution $\P_{x_0}$, we would like to apply the Ito Formula on ${\left\vert{X_*}\right\vert}^2$. This would give $$d{\left\vert{X_*}\right\vert}^2+2\nu{\left\Vert{X_*}\right\Vert}^2dt=2\left({X_*},\phi({X_*})d{W_*}\right)+2(f,{X_*})dt+{\left\vert\phi({X_*}(t))\right\vert}^2_{\mathcal
L_2(H;H)}dt.$$ Integrating and taking the expectation, we would deduce that, if $f=0$ and $\phi$ constant, $$\E_{x_0}\left({\left\vert{X_*}(t)\right\vert}^2+2\nu\int_0^t{\left\Vert{X_*}(s)\right\Vert}^2dt\right)={\left\vertx_0\right\vert}^2+t{\left\vert\phi\right\vert}^2_{\mathcal
L_2(H;H)}.$$ Unfortunately, those computations are not allowed. However, analogous computations are valid if we replace $\P_{x_0}$ by $\P_{x_0}^N$, which yields $$\E\left({\left\vertX_N(t)\right\vert}^2+2\nu\int_0^t{\left\VertX_N(s)\right\Vert}^2dt\right)={\left\vertP_N x_0\right\vert}^2+t{\left\vertP_N\phi\right\vert}^2_{\mathcal L_2(H;H)}.$$ Then, we take the limit and we infer from Fatou Lemma and from the semi-continuity of ${\left\vert\cdot\right\vert}$, ${\left\Vert\cdot\right\Vert}$ in $\H_s$ that $$\E_{x_0}\left({\left\vert{X_*}(t)\right\vert}^2+2\nu\int_0^t{\left\Vert{X_*}(s)\right\Vert}^2dt\right)\leq{\left\vertx_0\right\vert}^2+t{\left\vert\phi\right\vert}^2_{\mathcal
L_2(H;H)},$$ provided $f=0$ and $\phi$ constant and provided $\P_{x_0}$ is limit in distribution of solutions of .
Let $\P'$ and $Y$ be a probability measure and a random variable on $({\Omega_*},{\F^*})$, respectively. The distribution $\Dr_{\P'}(Y)$ denotes the law of $Y$ under $\P'$.
A weak solution $\P_{\mu}$ with initial law $\mu$ is said to be stationary if, for any $t\geq 0$, $\mu$ is equal to $\Dr_{\P_{\mu}}({X_*}(t))$.
We define $$\left(\Pcal_t^N\psi\right)(x_0)=\E\left(\psi(X_N(t,x_0))\right)=\E_{x_0}^N\left(\psi({X_*}(t)\right),$$ where $\E^N_{x_0}$ is the expectation associated to $\P_{x_0}^N$.
It is easily shown that $X_N(\cdot,x_0)$ verifies the strong Markov property, which obviously implies that $(\Pcal_t^N)_{t\in \R^+}$ is a Markov transition semi-group on $P_NH$. Ito Formula on ${\left\vertX_N(\cdot,x_0)\right\vert}^2$ gives $$d{\left\vertX_N\right\vert}^2+2\nu{\left\VertX_N\right\Vert}^2dt=2\left(X_N,\phi(X_N)dW\right)+2(X_N,f)dt+{\left\vertP_N\phi(X_N)\right\vert}^2dt,$$which yields, by applying arithmetico-geometric inequality and Hypothesis \[HGal\], $$\label{Eq1.10}
d{\left\vertX_N\right\vert}^2+\nu{\left\VertX_N\right\Vert}^2dt\leq 2\left(X_N,\phi(X_N)dW\right)+cB_1dt.$$ Integrating and taking the expectation, we obtain $$\label{Eq1.10bis}
\E\left({\left\vertX_N(t)\right\vert}^2\right)\leq e^{-\nu\mu_1
t}{\left\vertx_0\right\vert}^2+\frac{c}{\nu\mu_1}B_1.$$ Hence, applying the Krylov-Bogoliubov Criterion (see [@DPZ1]), we obtain that $(\Pcal_t^N)_t$ admits an invariant measure $\mu_N$ and that every invariant measure has a moment of order two in $H$. Let $X_0^N$ be a random variable whose law is $\mu_N$ and which is independent of $W$, then $X_N=X_N(\cdot,X_0^N)$ is a stationary solution of . Integrating , we obtain $$\E{\left\vertX_N(t)\right\vert}^2+\nu\E\int_0^t{\left\VertX_N(s)\right\Vert}^2ds\leq\E{\left\vertX_N(0)\right\vert}^2+cB_1t.$$ Since the law of $X_N(s)$ is $\mu_N$ for any $s\geq 0$ and since $\mu_N$ admits a moment of order $2$, it follows $$\label{Eq1.11}
\int_{P_N H}{\left\Vertx\right\Vert}^2\,\mu_N(dx)\leq \frac{c}{\nu}B_1.$$ Moreover the laws $(\P_{\mu_N}^N)_N$ of $(X_N(\cdot,X_0^N),W)$ are tight in a well chosen functional space. Then, for a subsequence $(N_k')_k$, $\P_{\mu_{N_k}}^{N_k}$ converges in law to $\P_{\mu}$ a weak stationary solution of with initial law $\mu$ (See [@FlandoliGatarek] for details). We deduce from that $$\int_{H}{\left\Vertx\right\Vert}^2\,\mu(dx)\leq \frac{c}{\nu}B_1,$$ which yields (see [@FlandoliRomito]) $$\label{Eq1.12}
\P_{\mu}\left({X_*}(t)\in\H_{1}\right)=1\;\textrm{ for any
}\;t\geq 0.$$ We do not know if $X_*(t)\in\H_1$ for all $t$ holds $\P_\mu$–almost surely. This would probably imply strong uniqueness $\mu$–almost surely. Remark that it is not known in general if $\mu$ is an invariant measure because, due to the lack of uniqueness, it is not known if defines a Markov evolution. We will see below that this is the case under suitable assumptions.
Exponential convergence to equilibrium
--------------------------------------
In the present article, the covariance operator $\phi$ of the noise is assumed to be at the same time sufficiently smooth and non degenerate with bounded derivatives. More precisely, we use the following assumption.
\[H0\] There exist ${\varepsilon}>0$ and a family $(\phi_n)_n$ of continuous mappings $H\to\R$ with continuous derivatives such that $$\Espace
\left\{
\begin{array}{l}
\phi(x)dW=\sum_{n=1}^\infty\phi_n(x)e_n dW_n\quad\textrm{ where }\quad W=\sum_{n=0}^\infty W_n e_n,\\
\kappa_0=\sum_{n=1}^\infty\sup_{x\in H}{\left\vert\phi_n(x)\right\vert}^2\mu_n^{1+{\varepsilon}}<\infty.
\end{array}
\right.$$ Moreover there exists $\kappa_1$ such that for any $x$, $\eta\in\H_2$ $$\sum_{n=1}^\infty{\left\vert\phi_n'(x)\cdot \eta\right\vert}^2\mu_n^{2}<\kappa_1{\left\Vert\eta\right\Vert}_ 2^2.$$ For any $x\in H$ and $N\in\N$, we have $\phi_n(x)>0$ and $$\label{EqH1.4}
\kappa_2=\sup_{x\in H}{\left\vert\phi^{-1}(x)\right\vert}_{\mathcal
L\left(\H_3;H\right)}^2<\infty,$$ where $$\phi(x)^{-1}\cdot h=\sum_{n=1}^\infty \phi_n(x)^{-1} h_n
e_n\quad\textrm{ for }\quad h=\sum_{n=0}^\infty h_n e_n.$$
For instance, $\phi=A^{-\frac{s}{2}}$ fulfills Hypothesis \[H0\] provided $s\in\left(\frac{5}{2},3\right]$.
We set $$B_0=\kappa_0+\kappa_1+\kappa_2+{\left\vertf\right\vert}^2.$$
If the noise is additive, Hypothesis \[H0\] simplifies. Indeed in this case, we do not need to assume that $\phi$ and $A$ commute. This requires a different but simpler proof of Lemma \[PropSmallH3\] below.
Another situation where we can get rid of the assumption that the noise is diagonal is when the viscosity $\nu$ is sufficiently large. The proof is simpler in that case.
It is easily shown that Hypothesis \[H0\] and $f\in H$ imply Hypothesis \[HGal\]. Therefore, solutions of are well-defined and, for a subsequence, they converge to weak solution of .
The aim of the present article is to establish that, under Hypothesis \[H0\] and under a condition of smallness of ${\left\Vertf\right\Vert}_{\varepsilon}$, the law of ${X_*}(t)$ under a weak solution $\P_{x_0}$ converges exponentially fast to equilibrium provided $\P_{x_0}$ is limit in distribution of solutions of .
Before stating our main result, let us recall some definitions. Let $E$ be Polish space. The set of all probability measures on $E$ is denoted by $\mathcal P(E)$. The set of all bounded measurable (resp uniformly continuous) maps from $E$ to $\R$ is denoted by $B_b(E;\R)$ (resp $UC_b(E;\R)$). The total variation ${\left\Vert\mu\right\Vert}_{var}$ of a finite real measure $\lambda$ on $E$ is given by $${\left\Vert\lambda\right\Vert}_{var}= \sup \left\{{\left\vert\lambda(\Gamma)\right\vert}\;|\; \Gamma \in \Br(E) \right\},$$ where we denote by $\Br(E)$ the set of the Borelian subsets of $E$.
The main result of the present article is the following. Its proof is given in section $4$ after several preliminary results.
\[Th\] Assume that Hypothesis \[H0\] holds. There exists ${{\delta^0}}$, $C$ and $\gamma>0$ only depending on $\phi$, $D$, ${\varepsilon}$ and $\nu$ such that, for any weak solution $\P_{\lambda}$ with initial law $\lambda \in \Pcal(H)$ which is limit of solutions of , there exists a weak stationary solution $\P_{\mu}$ with initial law $\mu$ such that $$\label{EqTh}
{\left\Vert\Dr_{\P_{\lambda}}({X_*}(t))-\mu\right\Vert}_{var}\leq C e^{-\gamma
t}\left(1+\int_{H}{\left\vertx\right\vert}^2\,\lambda(dx)\right),$$ provided ${\left\Vertf \right\Vert}_{\varepsilon}^2\leq {{\delta^0}}$ and where ${\left\Vert\cdot\right\Vert}_{var}$ is the total variation norm associated to the space $\H_s$ for $s<0$.
Moreover, for a given $\P_{\lambda}$, $\mu$ is unique and $\P_{\mu}$ is limit of solutions of .
It is well known that ${\left\Vert\cdot\right\Vert}_{var}$ is the dual norm of ${\left\vert\cdot\right\vert}_\infty$ which means that for any finite measure $\lambda'$ on $\H_s$ for $s<0$ $${\left\Vert\lambda'\right\Vert}_{var}=\sup_{{\left\vertg\right\vert}_\infty\leq 1}
{\left\vert\int_{\H_s} g(x)\,\lambda'(dx)\right\vert},$$ where the supremum is taken over $g\in UC_b(\H_s)$ which verifies ${\left\vertg\right\vert}_\infty\leq 1$. Hence is equivalent to $$\label{EqTh2}
{\left\vert\E_{\lambda}\left(g({X_*}(t))\right)-\int_{H}g(x)\,\mu(dx)\right\vert}\leq
C{\left\vertg\right\vert}_\infty\left(1+\int_{H}{\left\vertx\right\vert}^2\,\lambda(dx)\right),$$ for any $g\in UC_b(\H_s)$.
Remark that if $\lambda'$ is a finite measure of $\H_{s_0}$, then the value of the total variation norm of $\lambda'$ associated to the space $\H_s$ is not depending of the value of $s\leq s_0$.
Hence, since $\Dr_{\P_{\lambda}}({X_*}(t))$ is a probability measure on $H$ then (resp ) remains true when ${\left\Vert\cdot\right\Vert}_{var}$ is the total variation norm associated to the space $H$ (resp for any $g\in B_b(H; \R)$).
Moreover, we see below that, under suitable assumptions, if $\lambda$ is a probability measure on $\H_2$, then $\Dr_{\P_{\lambda}}({X_*}(t))$ is still a probability measure on $\H_2$. It follows that (resp ) remains true when ${\left\Vert\cdot\right\Vert}_{var}$ is associated to $\H_2$ (resp for any $g\in B_b(\H_2; \R)$).
[ ]{}
Our method is not influenced by the size of the viscosity $\nu$. Then, for simplicity in the redaction, we now assume that $\nu=1$.
Markov evolution
----------------
Here, we take into account the results of [@DebusscheNS3D], [@ODASSOMarkov] and we rewrite Theorem \[Th\]. This section is not necessary in the understanding of the proof of Theorem \[Th\].
Let $(N_k')_k$ be an increasing sequence of integer. In [@DebusscheNS3D], [@ODASSOMarkov], it is established that it is possible to extract a subsequence $(N_k)_k$ of $(N_k')_k$ such that, for any $x_0\in\H_2$, $\P_{x_0}^{N_k}$ converges in distribution to a weak solution $\P_{x_0}$ of provided the following assumption holds.
\[HCst\] There exist ${\varepsilon},\delta>0$ such that the mapping $\phi$ is bounded in $\mathcal L_2\left(H;\H_{1+{\varepsilon}} \right)$. Moreover, for any $x$, $\textrm{ker }\phi(x)=\{0\}$ and there exits a bounded map $\phi^{-1}: H \to\mathcal L
\left(\H_{3-\delta}; H\right)$ such that for any $x\in H$, $$\phi(x)\cdot \phi^{-1}(x)\cdot h=h\quad\textrm{ for any } h\in
\H_{3-\delta}.$$ Moreover $f\in V$.
The method to extract $(N_k)_k$ is based on the investigation of the properties of the Kolmogorov equation associated to perturbed by a very irregular potential.
It follows that $(\P_{x_0})_{x_0\in \H_2}$ is a weak Markov family, which means that for any $x_0\in \H_2$ $$\label{Eq1.2A}
\P_{x_0}\left({X_*}(t)\in\H_2\right)=1\;\textrm{ for any } t\geq
0.$$ and that, for any $t_1<\dots<t_n$, $t>0$ and any $\psi\in
B_b(\H_2;\R)$ $$\label{Eq1.2B}
\E_{x_0}\left(\left.\psi(X_*(t+t_n))\right|X_*(t_1),\dots,X_*(t_n)\right)=\mathcal
P_t\psi(X_*(t_n)),$$ where $$\left(\Pcal_t\psi\right)(x_0)=\E_{x_0}\left(\psi({X_*}(t))\right).$$ Note that was known only for a stationary solution (see [@FlandoliRomito]).
Assume that Hypothesis \[H0\] holds. If we strengthen into $$\kappa_2=\sup_{x\in H}{\left\vert\phi^{-1}(x)\right\vert}_{\mathcal
L\left(\H_{3-\delta};H\right)}^2<\infty,$$ for some $\delta>0$, then Hypothesis holds.
Hence, we immediately deduce the following corollary from Theorem \[Th\].
\[ThDebussche1\] Assume that Hypothesis \[H0\] and \[HCst\] hold. Then there exits a unique invariant measure $\mu$ for $\left(\Pcal_t\right)_{t\in \R^+}$ and $C,\gamma>0$ such that for any $\lambda \in \Pcal(\H_2)$ $$\label{EqThDebussche}
{\left\Vert\Pcal_t^*\lambda-\mu\right\Vert}_{var}\leq C e^{-\gamma
t}\left(1+\int_{\H_2}{\left\vertx\right\vert}^2\,\lambda(dx)\right),$$ provided ${\left\Vertf \right\Vert}_{\varepsilon}^2\leq {{\delta^0}}$ and where ${\left\Vert\cdot\right\Vert}_{var}$ is the total variation norm associated to the space $\H_2$.
Assume that Hypothesis \[HCst\] holds. Let $\P_{x_0}$ and $\P'_{x_0}$ be two weak solutions of which are limit in distribution of solutions of . Then we build $(\Pcal_t)_t$ and $(\Pcal_t')_t$ as above associated to $\P_{x_0}$ and $\P'_{x_0}$, respectively. It follows that there exists $\mu$ and $\mu'$ such that and hold for $((\Pcal_t)_t,\P_{x_0},\mu)$ and $((\Pcal_t')_t,\P'_{x_0},\mu')$. Although we have uniqueness of the invariant measures $\mu$ and $\mu'$ associated to $(\Pcal_t)_t$ and $(\Pcal_t')_t$, we do not know if $\mu$ and $\mu'$ are equal.
Coupling methods
----------------
The proof of Theorem \[Th\] is based on coupling arguments. We now recall some basic results about coupling. Moreover, in order to explain the coupling method in the case of non degenerate noise, we briefly give the proof of exponential mixing for equation .
Let $(\lambda_1,\lambda_2)$ be two distributions on a polish space $(E,\mathcal{E})$ and let $(\Omega,\mathcal{F},\P)$ be a probability space and let $(Z_1,Z_2)$ be two random variables $(\Omega,\mathcal{F}) \to (E,\mathcal{E})$. We say that $(Z_1,Z_2)$ is a coupling of $(\lambda_1,\lambda_2)$ if $\lambda_i=\Dr(Z_i)$ for $i=1,2$. We have denoted by $\Dr(Z_i)$ the law of the random variable $Z_i$.
Next result is fundamental in the coupling methods, the proof is given for instance in the Appendix of [@ODASSO1].
\[lem\_coupling\] Let $(\lambda_1,\lambda_2)$ be two probability measures on $(E,\mathcal{E})$. Then $${\left\Vert\lambda_1-\lambda_2\right\Vert}_{var}= \min \P(Z_1\not = Z_2).$$ The minimum is taken over all couplings $(Z_1,Z_2)$ of $(\lambda_1,\lambda_2)$. There exists a coupling which reaches the minimum value. It is called a maximal coupling.
[ ]{}
Let us first consider the case of the solutions of . Assume that Hypothesis \[H0\] holds. Let $N\in\N$ and $(x_0^1,x_0^2)\in\R^2$. Combining arguments from [@KS1], [@Matt], it can be shown that there exists a decreasing function $p_N(\cdot)>0$ such that $$\label{toy3}
{\left\Vert\left(\Pcal_1^N\right)^* \delta_{x_0^2}-\left(\Pcal_1^N\right)^* \delta_{x_0^1}\right\Vert}_{var}
\leq 1-p_N\left({\left\vertx_0^1\right\vert}+{\left\vertx_0^2\right\vert}\right).$$ Applying Lemma \[lem\_coupling\], we build a maximal coupling $(Z_1,Z_2)=(Z_1(x_0^1,x_0^2),Z_2(x_0^1,x_0^2))$ of $(\left(\Pcal_1^N\right)^* \delta_{x_0^1},\left(\Pcal_1^N\right)^* \delta_{x_0^2})$. It follows $$\label{toy4}
\P\left(Z_1=Z_2\right)\geq p_N\left({\left\vertx_0^1\right\vert}+{\left\vertx_0^2\right\vert}\right)>0.$$ Let $(W,\widetilde W)$ be a a couple of independent cylindrical Wiener processes and $\delta>0$. We denote by $X_N(\cdot,x_0)$ and $\widetilde X_N(\cdot,x_0)$ the solutions of associated to $W$ and $\widetilde W$, respectively. Now we build a couple of random variables $(V_1,V_2)=(V_1(x_0^1,x_0^2),V_2(x_0^1,x_0^2))$ on $P_NH$ as follows $$\label{toy5}
\Espace
(V_1,V_2)=
\left\{
\begin{array}{ll}
(X_N(\cdot,x_0),X_N(\cdot,x_0))& \textrm{ if } x_0^1=x_0^2=x_0,\\
(Z_1(x_0^1,x_0^2),Z_2(x_0^1,x_0^2))& \textrm{ if } (x_0^1,x_0^2)\in B_H(0,\delta)\backslash\{x_0^1=x_0^2\},\\
(X_N(\cdot,x_0^1),\widetilde X_N(\cdot,x_0^2))& \textrm{ else},
\end{array}
\right.$$ where $B_H(0,\delta)$ is the ball of $H\times H$ with radius $\delta$.
Then $(V_1(x_0^1,x_0^2),V_2(x_0^1,x_0^2))$ is a coupling of $(\left(\Pcal_1^N\right)^* \delta_{x_0^1},\left(\Pcal_1^N\right)^*
\delta_{x_0^2})$. It can be shown that it depends measurably on $(x_0^1,x_0^2)$. We then build a coupling $(X^1,X^2)$ of $(\Dr(X_N(\cdot,x_0^1)),\Dr(X_N(\cdot,x_0^2)))$ by induction on $\N$. We first set $X^i(0)=x_0^i$ for $i=1,2$. Then, assuming that we have built $(X^1,X^2)$ on $\{0,1,\dots,k\}$, we take $(V_1,V_2)$ as above independent of $(X^1,X^2)$ and set $$X^i(k+1)=V_i(X^1(k),X^2(k))\quad\textrm{ for } i=1,2.$$ Taking into account , it is easily shown that the time of return of $(X^1,X^2)$ in $B(0,4(c/\mu_1)B_1)$ admits an exponential moment. We choose $\delta=4(c/\mu_1)B_1$. It follows from , that, $(X^1(n),X^2(n))\in B(0,\delta)$ implies that the probability of having $(X^1,X^2)$ coupled (i.e. equal) at time $n+1$ is bounded below by $p_N(2\delta)>0$. Finally, remark that if $(X^1,X^2)$ are coupled at time $n+1$, then they remain coupled for any time after. Combining these three properties and using the fact that $(X^1(n),X^2(n))_{n\in\N}$ is a discrete strong Markov process, it is easily shown that $$\label{toy6}
\P\left(X^1(n)\not=X^2(n)\right)\leq C_Ne^{-\gamma_N
n}\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right),$$ with $\gamma_N>0$.
Recall that $(X^1,X^2)$ is a coupling of $(\Dr(X_N(\cdot,x_0^1)),\Dr(X_N(\cdot,x_0^2)))$ on $\N$. It follows that $(X^1(n),X^2(n))$ is a coupling of $((\Pcal_n^N)^* \delta_{x_0^1},(\Pcal_n^N)^* \delta_{x_0^2})$. Combining Lemma \[lem\_coupling\] and , we obtain, for $n\in\N$, $${\left\Vert\left(\Pcal_n^N\right)^* \delta_{x_0^2}-\left(\Pcal_n^N\right)^* \delta_{x_0^1}\right\Vert}_{var}\leq C_Ne^{-\gamma_N n}\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right).$$Setting $n=\lfloor t\rfloor$ and integrating $(x_0^2,x_0^1)$ over $((\Pcal_{t-n}^N)^*\lambda)\otimes\mu_N$ where $\mu_N$ is an invariant measure, it follows that, for any $\lambda\in P(P_NH)$, $$\label{toy7}
{\left\Vert\left(\Pcal_t^N\right)^*\lambda-\mu_N\right\Vert}_{var}\leq C_N e^{-\gamma_N t}\left(1+\int_{P_N H}{\left\vertx\right\vert}^2\,\lambda(dx)\right).$$ This result is useless when considering equation since the constants $C_N$, $\gamma_N$ strongly depend on $N$. If one tries to apply directly the above arguments to the infinite dimensional equation , one faces several difficulties. First it is not known whether $\P_{x_0}$ is Markov. We only know that, as explained in section $1.3$, a Markov transition semi-group can be constructed. This is a major difficulty since this property is implicitely used in numerous places above. Another strong problem is that Girsanov transform is used in order to obtain . Contrary to the two dimensional case, no Foias-Prodi estimate is available for the three dimensional Navier-Stokes equations and the Girsanov transform should be done in the infinite dimensional equation. This seems impossible. We will show that we are able to prove an analogous result to by a completely different argument. However, this will hold only for small initial data in $\H_2$. Another problem will occur since it is not known whether solutions starting in $\H_2$ remain in $\H_2$.
We remedy the lack of Markov property by working only on Galerkin approximations and prove that holds with constants uniform in $N$. As already mentioned, we prove that is true for $x_0^1$, $x_0^2$ in a small ball of $\H_2$ and uniformly in $N$. Then, following the above argument, it remains to prove that the time of return in this small ball admits an exponential moment. Note that the smallness assumption on $f$ is used at this step. In the following sections, we prove
\[ThN\] Assume that Hypothesis \[H0\] holds. Then there exist ${{\delta^0}}={{\delta^0}}(B_0,D,{\varepsilon},\nu)$, $C=C(\phi,D,{\varepsilon},\nu)>0$ and $\gamma=\gamma(\phi,D,{\varepsilon},\nu)>0$ such that if ${\left\Vertf \right\Vert}_{\varepsilon}^2\leq {{\delta^0}}\,$ holds, then, for any $N\in\N$, there exists a unique invariant measure $\mu_N$ for $\left(\Pcal_t^N\right)_{t\in \R^+}$. Moreover, for any $\lambda
\in \Pcal(P_N H)$ $$\label{EqThN}
{\left\Vert\left(\Pcal_t^N\right)^*\lambda-\mu_N\right\Vert}_{var}\leq C e^{-\gamma t}\left(1+\int_{P_N H}{\left\vertx\right\vert}^2\,\lambda(dx)\right).$$
We now explain why this result implies Theorem \[Th\].
Let $\lambda\in P(H)$ and $X_{\lambda}$ be a random variable on $H$ whose law is $\lambda$ and which is independant of $W$. Since ${\left\Vert\cdot\right\Vert}_{var}$ is the dual norm of ${\left\vert\cdot\right\vert}_\infty$, then implies that $$\label{EqTh3}
{\left\vert\E\left(g(X_N(t,X_{\lambda}))\right)-\int_{P_NH}g(x)\,\mu_N(dx)\right\vert}\leq
C{\left\vertg\right\vert}_\infty\left(1+\int_{H}{\left\vertx\right\vert}^2\,\lambda(dx)\right),$$ for any $g\in UC_b(\H_s)$ for $s<0$.
Assume that, for a subsequence $(N_k')_k$, $X_N(t,X_{\lambda})$ converges in distribution in $\H_s$ to the law ${X_*}(t)$ under the weak solution $\P_\lambda$ of . Recall that the family $(\P_{\mu_N}^N)_N$ is tight. Hence, for a subsequence $(N_k)_k$ of $(N_k')_k$, $\P_{\mu_{N_k}}$ converges to $\P_{\mu}$ a weak stationary solution of with initial law $\mu$. Taking the limit, follows from , which yields Theorem \[Th\].
Coupling of solutions starting from small initial data
======================================================
The aim of this section is to establish the following result. A result analogous to but uniform in $N$.
\[Prop\_I\_0\] Assume that Hypothesis \[H0\] holds and that $f\in H$. Then there exist $(T,\delta)\in(0,1)^2$ such that, for any $N\in\N$, there exists a coupling $(Z_1(x_0^1,x_0^2),Z_2(x_0^1,x_0^2))$ of $((\Pcal_T^N)^*\delta_{x_0^1},(\Pcal_T^N)^*\delta_{x_0^2})$ which measurably depends on $(x_0^1,x_0^2)\in\H_2$ and which verifies $$\label{I_0bis}
\P\left(Z_1(x_0^1,x_0^2)=Z_2(x_0^1,x_0^2)\right)\geq\frac{3}{4}$$ provided $$\label{2.5}
{\left\Vertx_0^1\right\Vert}_2^2\vee{\left\Vertx_0^2\right\Vert}_2^2\leq \delta.$$
Assume that Hypothesis \[H0\] holds and that $f\in H$. Let $T\in(0,1)$. Applying Lemma \[lem\_coupling\], we build $(Z_1(x_0^1,x_0^2),Z_2(x_0^1,x_0^2))$ as the maximal coupling of $(\Pcal_T^*\delta_{x_0^1},\Pcal_T^*\delta_{x_0^2})$. Measurable dependance follows from a slight extension of Lemma $1.17$ (see [@ODASSO1], remark $A.1$).
In order to establish Proposition \[Prop\_I\_0\], it is sufficient to prove that there exists $c(B_0,D)$ not depending on $T\in(0,1)$ and on $N\in\N$ such that $$\label{I_0ter}
{\left\Vert\left(\Pcal_T^N\right)^*\delta_{x_0^2}-\left(\Pcal_T^N\right)^*\delta_{x_0^1}\right\Vert}_{var}\leq c(B_0,D)\sqrt
T,$$ provided $$\label{2.5ter}
{\left\Vertx_0^1\right\Vert}_2^2\vee{\left\Vertx_0^2\right\Vert}_2^2\leq B_0 T^3.$$ Then it suffices to choose $T\leq 1/(4c(B_0,D))^2$ and $\delta=B_0T^3$.
Since ${\left\Vert\cdot\right\Vert}_{var}$ is the dual norm of ${\left\vert\cdot\right\vert}_\infty$, is equivalent to $$\label{I_0}
{\left\vert\E\left( g(X_N(T,x_0^2))-g(X_N(T,x_0^1))\right)\right\vert}\leq
8{\left\vertg\right\vert}_\infty c(B_0,D)\sqrt T.$$ for any $g\in UC_b(P_N H)$.
It follows from the density of $C^1_b(P_N H)\subset UC_b(P_N H)$ that, in order to establish Proposition \[Prop\_I\_0\], it is sufficient to prove that holds for any $N\in\N$, $T\in(0,1)$ and $g\in C_b^1(P_N H)$ provided holds.
The proof of under this condition is splitted into the next three subsections.
A priori estimate
-----------------
For any process $X$, we define the $\H_1$–energy of $X$ at time $t$ by $$ E_X^{\H_1}(t)={\left\VertX(t)\right\Vert}^2+ \int_{0}^t{\left\VertX(s)\right\Vert}_2^2ds.$$ Now we establish the following result which will be useful in the proof of \[I\_0\].
\[Prop\_nrj\] Assume that Hypothesis \[H0\] holds and that $f\in H$. There exist $K_0=K_0(D)$ and $c=c(D)$ such that for any $T\leq 1$ and any $N\in\N$, we have $$\P\left(\sup_{(0,T)} E_{X_N(\cdot,x_0)}^{\H_1} >K_0\right)\leq c\left(1+\frac{B_0}{K_0}\right)\sqrt T,$$ provided ${\left\Vertx_0\right\Vert}^2\leq B_0 T$.
Let $X_N=X_N(\cdot,x_0)$. Ito Formula on ${\left\Vert{X_N}\right\Vert}^2$ gives $$\label{8.1}
d{\left\Vert{X_N}\right\Vert}^2+2 {\left\Vert{X_N}\right\Vert}_2^2dt=dM_{\H_1}+I_{\H_1}dt+{\left\Vert{P_N \phi}({X_N})\right\Vert}^2_{\mathcal L_2(H;\H_1)}dt
+I_fdt,$$ where $$\Espace
\left\{
\begin{array}{l}
I_{\H_1}=-2\left(A{X_N},B({X_N})\right),\;I_f=2\left(A{X_N},f\right),\\
\;M_{\H_1}(t)=2\int_0^t\left(A{X_N}(s),\phi({X_N}(s))dW(s)\right).
\end{array}
\right.$$ Combining a Hölder inequality, a Agmon inequality and a arithmetico-geometric inequality gives $$\label{8.2}
I_{\H_1}\leq 2{\left\Vert{X_N}\right\Vert}_2{\left\vert{X_N}\right\vert}_\infty{\left\Vert{X_N}\right\Vert}\leq
c{\left\Vert{X_N}\right\Vert}_2^\frac{3}{2}{\left\Vert{X_N}\right\Vert}^\frac{3}{2} \leq
\frac{1}{4}{\left\Vert{X_N}\right\Vert}_2^2+c{\left\Vert{X_N}\right\Vert}^6.$$ Similarly, using Poincaré inequality and Hypothesis \[H0\], $$\label{8.2bis}
I_{f}\leq \frac{1}{4}{\left\Vert{X_N}\right\Vert}_2^2+c{\left\vertf\right\vert}^2\leq
\frac{1}{4}{\left\Vert{X_N}\right\Vert}_2^2+cB_0.$$ We deduce from , , , Hypothesis \[H0\] and Poincaré inequality that $$\label{8.3}
d{\left\Vert{X_N}\right\Vert}^2+ {\left\Vert{X_N}\right\Vert}_2^2dt\leq dM_{\H_1}+cB_0 dt
+c{\left\Vert{X_N}\right\Vert}^2 \left({\left\Vert{X_N}\right\Vert}^4-4K_0^2\right)\,dt,$$ where $$\label{2.9bis}
K_0=\sqrt{\frac{\mu_1}{8c}}.$$ Setting $${\sigma}_{\H_1}=\inf\left\{t\in(0,T)\,\left|\, {\left\Vert{X_N}(t)\right\Vert}^2> 2 K_0 \right.\right\},$$ we infer from ${\left\Vertx_0\right\Vert}^2\leq B_0 T$ that for any $t\in(0,{\sigma}_{\H_1})$ $$\label{8.4}
E_{X_N}^{\H_1}(t)\leq cB_0T +M_{\H_1}(t).$$ We deduce from Hypothesis \[H0\] and from Poincaré inequality that $\phi(x)^*A$ is bounded in $\mathcal L(\H_1;\H_1)$ by $cB_0$. It follows that for any $t\in (0,{\sigma}_{\H_1})$ $$\left<M_{\H_1}\right>(t)=4\int_0^t{\left\Vert{P_N
\phi}({X_N}(s))^*A{X_N}(s)\right\Vert}^2dt\leq
cB_0 \int_0^t{\left\Vert{X_N}(s)\right\Vert}^2ds\leq 2cK_0B_0 T.$$ Hence a Burkholder-Davis-Gundy inequality gives $$\E\left(\sup_{(0,{\sigma}_{\H_1})}M_{\H_1}\right)\leq
c\E\sqrt{\left<M_{\H_1}\right>({\sigma}_{\H_1})}\leq
c\sqrt{K_0B_0T}\leq c(K_0+B_0)\sqrt T.$$ It follows from and $T\leq 1$ that $$\E\left(\sup_{(0,{\sigma}_{\H_1})}E_{X_N}^{\H_1}\right)\leq
c(B_0+K_0)\sqrt T,$$ which yields, by a Chebyshev inequality, $$\P\left(\sup_{(0,{\sigma}_{\H_1})}E_{X_N}^{\H_1}> K_0\right)\leq c\left(1+\frac{B_0}{K_0}\right)\sqrt T.$$Now, since $\sup_{(0,{\sigma}_{\H_1})}E_{X_N}^{\H_1}\leq K_0$ implies ${\sigma}_{\H_1}=T$, we deduce Lemma \[Prop\_nrj\].
Estimate of the derivative of $X_N$ {#Preuve_eta}
-----------------------------------
Let $N\in\N$ and $(x_0,h)\in (\H_2)^2$. We are concerned with the following equation $$\label{Eqeta}
\Espace
\left\{
\begin{array}{rcl}
d{\eta_N}+ A {\eta_N} \;dt+{P_N \widetilde B}({X_N},{\eta_N})\, dt &=& P_N({ \phi}'({X_N})\cdot{\eta_N})\, dW,\\
{\eta_N}(s,s,x_0)\cdot h &=& P_N h,
\end{array}
\right.$$ where $\widetilde
B({X_N},{\eta_N})=B({X_N},{\eta_N})+B({\eta_N},{X_N})$, $X_N=X_N(\cdot,x_0)$ and ${\eta_N}(t)={\eta_N}(t,s,x_0)\cdot h$ for $t\geq s$.
Existence and uniqueness of the solutions of are easily shown. Moreover if $g\in C_b^1(P_N H)$, then, for any $t\geq 0$, we have $$\label{Eqetabis}
\left(\nabla\left(\Pcal_t^Ng\right)(x_0),h\right)=\E\left(\nabla g(X_N(t,x_0)),\eta_N(t,0,x_0)\cdot h\right).$$ For any process $X$, we set $$\label{2.12bis}
{\sigma}(X)=\inf\left\{t\in
(0,T)\,\left|\,\int_0^t{\left\VertX(s)\right\Vert}_2^2ds\geq K_0+1\right.\right\},$$ where $K_0$ is defined in Lemma \[Prop\_nrj\]. We establish the following result.
\[Prop\_eta\] Assume that Hypothesis \[H0\] holds and that $f\in H$. Then there exists $c=c(B_0,D)$ such that for any $N\in \N$, $T\leq 1$ and $(x_0,h)\in (\H_2)^2$ $$\E\int_0^{{\sigma}(X_N(\cdot,x_0))} {\left\Vert{\eta_N}(t,0,x_0)\cdot h\right\Vert}_3^2 \,dt\leq
c{\left\Verth\right\Vert}_2^2.$$
For a better readability, we set ${\eta_N}(t)={\eta_N}(t,0,x_0)\cdot h$ and ${\sigma}={\sigma}(X_N(\cdot,x_0))$. Ito Formula on ${\left\Vert{\eta_N}(t)\right\Vert}_2^2$ gives $$\label{5.1}
d{\left\Vert{\eta_N}\right\Vert}_2^2+2 {\left\Vert{\eta_N}\right\Vert}_3^2dt=dM_{{\eta_N}}+I_{\eta_N} \,dt+
{\left\VertP_N \left(\phi'({X_N})\cdot {\eta_N}\right)\right\Vert}^2_{\mathcal L_2(U;\H_2)}dt,$$ where $$\Espace
\left\{
\begin{array}{rcl}
M_{{\eta_N}}(t)&=&2\int_0^t \left(A^2{\eta_N},({P_N \phi}'({X_N})\cdot {\eta_N})\, dW\right)\,ds,\\
I_{\eta_N}&=&-2\left(A^\frac{3}{2}{\eta_N},A^\frac{1}{2}\widetilde
B({X_N},{\eta_N})\right).
\end{array}
\right.$$ It follows from Hölder inequalities, Sobolev Embedding and a arithmetico-geometric inequality $$I_{\eta_N}\leq c {\left\Vert{\eta_N}\right\Vert}_3{\left\Vert{\eta_N}\right\Vert}_2{\left\Vert{X_N}\right\Vert}_2\leq {\left\Vert{\eta_N}\right\Vert}_3^2+c{\left\Vert{\eta_N}\right\Vert}_2^2{\left\Vert{X_N}\right\Vert}_2^2.$$ Hence, we deduce from and Hypothesis \[H0\] $$d{\left\Vert{\eta_N}\right\Vert}_2^2+ {\left\Vert{\eta_N}\right\Vert}_3^2dt\leq dM_{{\eta_N}}+c{\left\Vert{\eta_N}\right\Vert}_2^2{\left\Vert{X_N}\right\Vert}_2^2+B_0{\left\Vert{\eta_N}\right\Vert}^2_2dt.$$ Integrating and taking the expectation, we obtain $$\label{5.2}
\E\left( \mathcal E({\sigma},0){\left\Vert{\eta_N}({\sigma})\right\Vert}_2^2
+ \int_0^{\sigma}\mathcal E({\sigma},t){\left\Vert{\eta_N}(t)\right\Vert}_3^2\,dt\right)
\leq{\left\Verth\right\Vert}^2_2,$$ where $$\mathcal E(t,s)=e^{-B_0t-c\int_s^t{\left\Vert{X_N}(r)\right\Vert}_2^2\,dr}.$$
Applying the definition of ${\sigma}$, we deduce $$\label{5.3}
\E\int_0^{\sigma}{\left\Vert{\eta_N}(t)\right\Vert}_3^2 \,dt\leq
{\left\Verth\right\Vert}_2^2\exp\left(c(K_0+1)+B_0 T\right),$$ which yields Lemma \[Prop\_eta\].
Proof of
---------
Let $\psi\in C^\infty\left(\R; [0,1]\right)$ such that $$\psi=0 \;\textrm{ on } \;(K_0 +1,\infty)\quad \textrm{ and }\quad \psi=1 \;\textrm{ on }\; (-\infty,K_0 ).$$ For any process $X$, we set $$\psi_{X}=\psi\left(\int_0^{T}{\left\Vert{X}(s)\right\Vert}_2^2ds\right).$$ Remark that $$\label{2.10}
{\left\vert\E\left( g({X_N}(T,x_0^2)) -g({X_N}(T,x_0^1))\right)\right\vert}\leq I_0+{\left\vertg\right\vert}_\infty\left(I_1+I_2\right),$$ where $$\Espace
\left\{
\begin{array}{lcl}
I_0 &=& {\left\vert\E\left(g({X_N}(T,x_0^2))\psi_{{X_N(\cdot,x_0^2)}}-g({X_N}(T,x_0^1))\psi_{{X_N(\cdot,x_0^1)}}\right)\right\vert},\\
I_i &=&\P\left(\int_0^T{\left\Vert{X_N}(s,x_0^i)\right\Vert}_2^2ds>K_0 \right).
\end{array}
\right.$$ For any $\theta\in[1,2]$, we set $$\Espace
\left\{
\begin{array}{ll}
x_0^\theta=(2-\theta)x_0^1+(\theta-1)x_0^2,& X_\theta={X_N}(\cdot,x_0^\theta),\\
\eta_\theta(t)=\eta_N(t,0,x_0^\theta),&{\sigma}_\theta={\sigma}(X_\theta).
\end{array}
\right.$$ Recall that ${\sigma}$ was defined in . For a better readability, the dependance on $N$ has been omitted. Setting $$h=x_0^2-x_0^1,$$ we have $$\label{2.11}
I_0\leq \int_1^2{\left\vertJ_\theta\right\vert}d\theta\quad
J_\theta=\left(\nabla \E\left(g(X_{\theta}(T))\psi_{X_\theta}\right),h\right).$$ To bound $J_\theta$, we apply a truncated Bismut-Elworthy-Li formula (See appendix \[Preuve\_2.15\]) $$\label{2.15}
J_\theta=\frac{1}{T}J_{\theta,1}'+2J_{\theta,2}',$$ where $$\Espace
\left\{
\begin{array}{lcl}
J_{\theta,1}'&=&\E \left(g(X_\theta(T))\psi_{X_\theta}\int_0^{{\sigma}_\theta} (\phi^{-1}(X_\theta(t))\cdot \eta_\theta(t)\cdot h,dW(t))\right),\\
J_{\theta,2}'&=&\E\left(g(X_\theta(T))\psi'_{X_\theta} \int_0^{{\sigma}_\theta} \left(1-\frac{t}{T}\right)\left(AX_\theta(t),A (\eta_\theta(t)\cdot h)\right)\,dt\right),\\
\psi'_X&=&\psi'\left(\int_0^{T}{\left\VertX_\theta(s)\right\Vert}_2^2ds\right).
\end{array}
\right.$$ It follows from Hölder inequality that $${\left\vertJ_{\theta,2}'\right\vert}\leq {\left\vertg\right\vert}_\infty{\left\vert\psi'\right\vert}_\infty \sqrt{\E\int_0^{{\sigma}_\theta} {\left\VertX_\theta(t)\right\Vert}_2^2 \,dt }
\sqrt{\E\int_0^{{\sigma}_\theta} {\left\Vert\eta_\theta(t)\cdot h\right\Vert}_2^2 \,dt }.$$ and from Hypothesis \[H0\] that $${\left\vertJ_{\theta,1}'\right\vert}\leq {\left\vertg\right\vert}_\infty B_0\sqrt{\E\int_0^{{\sigma}_\theta} {\left\Vert\eta_\theta(t)\cdot h\right\Vert}_3^2 \,dt }.$$ Hence for any $T\leq 1$ $$\label{2.16}
{\left\vertJ_\theta\right\vert}\leq c(B_0,D){\left\vertg\right\vert}_\infty \frac{1}{T}\sqrt{\E\int_0^{{\sigma}_\theta} {\left\Vert\eta_\theta(t)\cdot h\right\Vert}_3^2 \,dt }.$$ Combining and Lemma \[Prop\_eta\], we obtain $${\left\vertJ_\theta\right\vert}\leq c(B_0,D){\left\vertg\right\vert}_\infty\frac{{\left\Verth\right\Vert}_2}{T},$$ which yields, by and , $$I_0\leq c(B_0,D){\left\vertg\right\vert}_\infty\sqrt{T}.$$ Since $B_0T^3\leq B_0T$, we can apply Lemma \[Prop\_nrj\] to control $I_1+I_2$ in if holds. Hence follows provided holds, which yields Proposition \[Prop\_I\_0\].
Time of return in a small ball of $\H_2$ {#Preuve_Lyap}
========================================
Assume that Hypothesis \[H0\] holds. Let $N\in\N$ and $T,\delta,Z_1,Z_2$ be as in Proposition \[Prop\_I\_0\]. Let $(W,\widetilde W)$ be a couple of independant cylindrical Wiener processes on $H$. We denote by $X_N(\cdot,x_0)$ and $\widetilde X_N(\cdot,x_0)$ the solutions of associated to $W$ and $\widetilde W$, respectively. We build a couple of random variables $(V_1,V_2)=(V_1(x_0^1,x_0^2),V_2(x_0^1,x_0^2))$ on $P_NH$ as follows $$\label{toy5bisbis}
\Espace
(V_1,V_2)=
\left\{
\begin{array}{ll}
(X_N(\cdot,x_0),X_N(\cdot,x_0))& \textrm{ if } x_0^1=x_0^2=x_0,\\
(Z_1(x_0^1,x_0^2),Z_2(x_0^1,x_0^2))& \textrm{ if } (x_0^1,x_0^2)\in B_{\H_2}(0,\delta)\backslash\{x_0^1=x_0^2\},\\
(X_N(\cdot,x_0^1),\widetilde X_N(\cdot,x_0^2))& \textrm{ else},
\end{array}
\right.$$ We then build $(X^1,X^2)$ by induction on $T\N$. Indeed, we first set $X^i(0)=x_0^i$ for $i=1,2$. Then, assuming that we have built $(X^1,X^2)$ on $\{0,T,2T,\dots,nT\}$, we take $(V_1,V_2)$ as above independent of $(X^1,X^2)$ and we set $$X^i((n+1)T)=V_i(X^1(nT),X^2(nT))\quad\textrm{ for } i=1,2.$$ It follows that $(X^1,X^2)$ is a discrete strong Markov process and a coupling of $(\Dr(X_N(\cdot,x_0^1)),\Dr(X_N(\cdot,x_0^2)))$ on $T\N$. Moreover, if $(X^1,X^2)$ are coupled at time $nT$, then they remain coupled for any time after.
We set $$\label{3.1bis}
\tau=\inf\left\{t\in
T\N\backslash\{0\}\,\left|\,
{\left\VertX^1(t)\right\Vert}_2^2\vee{\left\VertX^2(t)\right\Vert}_2^2\leq \delta \right.\right\}.$$ The aim of this section is to establish the following result.
\[Prop\_Lyap\] Assume that Hypothesis \[H0\] holds. There exist $ \delta^3= \delta^3(B_0,D,{\varepsilon},\delta)$, $\alpha=\alpha(\phi,D,{\varepsilon},\delta)>0$ and $K"=K"(\phi,D,{\varepsilon},\delta)$ such that for any $(x_0^1,x_0^2)\in H\times H$ and any $N\in\N$ $$\E\left(e^{\alpha \tau}\right)\leq K"\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right),$$ provided ${\left\Vertf\right\Vert}^2_{\varepsilon}\leq \delta^3$.
The result is based on the fact that, in the absence of noise and forcing term, all solutions go to zero exponentially fast in $H$. A similar idea is used for the two-dimensional Navier-Stokes equations in [@KS1]. The proof is based on the following four Lemmas. The first one allows to control the probability that the contribution of the noise is small. Its proof strongly uses the assumption that the noise is diagonal in the eigenbasis of $A$. As already mentioned, in the additive case, the proof is easy and does not need this assumption.
\[PropSmallH3\] Assume that Hypothesis \[H0\] holds. For any $t, {M} >0$, there exists $p_0(t, {M} )=p_0(t, {M} ,{\varepsilon},({\left\vert\phi_n\right\vert}_\infty)_n,D)>0$ such that for any adapted process $X$ $$\P\left(\sup_{(0,t)}{\left\VertZ\right\Vert}_2^2\leq {M} \right)\geq p_0(t, {M}
),$$ where $$Z(t)=\int_0^te^{-A(t-s)}\phi(X(s))dW(s).$$
It is proved in section $3.1$.
Then, using this estimate and the smallness assumption on the forcing term, we estimate the moment of the first return time in a small ball in $H$.
Let $\delta_3>0$. We set $$\tau_{L^2}=\tau\wedge\inf\left\{t\in T\N^*\,\left|\, {\left\vertX^1(t)\right\vert}^2\vee{\left\vertX^2(t)\right\vert}\geq \delta_3 \right.\right\}.$$
\[lem7.2\] Assume that Hypothesis \[H0\] holds. Then, for any $\delta_3>0$, there exist $C_3(\delta_3)$, $C_3'(\delta_3)$ and $\gamma_3(\delta_3)$ such that for any $(x_0^1,x_0^2)\in (\H_2)^2$ $$ \E\left(e^{\gamma_3
\tau_{L^2}}\right)\leq
C_3\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right),$$provided $${\left\vertf\right\vert}\leq C_3'.$$
The proof is postponed to section \[Preuve7.2\].
Then, we need to get a finer estimate in order to control the time necessary to enter a ball in stronger topologies. To prove the two next lemmas, we use an argument similar to one used in the determinist theory (see [@Temam2], chapter $7$).
\[lem7.4\] Assume that Hypothesis \[H0\] holds. Then, for any $\delta_4$, there exist $p_4(\delta_4)>0$, $C_4'(\delta_4)>0$ and $R_4(\delta_4)>0$ such that for any $x_0$ verifying ${\left\vertx_0\right\vert}^2\leq R_4$, we have for any $T\leq 1$ $$\P\left({\left\VertX_N(T,x_0)\right\Vert}^2\leq \delta_4\right)\geq p_4,$$provided $${\left\vertf\right\vert}\leq C_4'.$$
The proof is postponed to section \[Preuve7.4second\].
\[lem7.5\] Assume that Hypothesis \[H0\] holds. Then, for any $\delta_5$, there exist $p_5(\delta_5)>0$, $C_5'(\delta_5)>0$ and $R_5(\delta_5)>0$ such that for any $x_0$ verifying ${\left\Vertx_0\right\Vert}^2\leq R_5$ and for any $T\leq 1$ $$ \P\left({\left\VertX_N(T,x_0)\right\Vert}_2^2\leq
\delta_5\right)\geq p_5.$$provided $${\left\Vertf\right\Vert}_{\varepsilon}\leq C_5'.$$
The proof is postponed to section \[Preuve7.5\].
: We set $$\delta_5=\delta,\quad \delta_4=R_5(\delta_5),\quad
\delta_3=R_4(\delta_4),\quad p_4=p_4(\delta_4), \quad
p_5=p_5(\delta_5), \quad p_1=\left(p_4p_5\right)^ 2,$$ and $$\delta^3=C_3'(\delta_3)\wedge C_4'(\delta_4)\wedge C_5'(\delta_5).$$ By the definition of $\tau_{L^2}$, we have $${\left\vertX^1(\tau_{L^2})\right\vert}^2\vee{\left\vertX^2(\tau_{L^2})\right\vert}^2\leq
R_4(\delta_4).$$ We distinguish three cases.
The first case is ${\left\VertX^1(\tau_{L^2})\right\Vert}_2^2\vee{\left\VertX^2(\tau_{L^2})\right\Vert}_2^2\leq\delta$, which obviously yields $$\label{7.5bis1}
\P\left(\min_{k=0,\dots,2}\max_{i=1,2}{\left\VertX^i(\tau_{L^2}+kT)\right\Vert}_2^2\leq\delta\,\left|\,
\left(X^2(\tau_{L^2}),X^2(\tau_{L^2})\right)\right.\right)\geq
p_1.$$
We now treat the case $x_0=X^1(\tau_{L^2})=X^2(\tau_{L_2})$ with ${\left\Vertx_0\right\Vert}_ 2^2>\delta$. Combining Lemma \[lem7.4\] and Lemma \[lem7.5\], we deduce from the weak Markov property of $X_N$ that $$\P\left({\left\VertX_N(2T,x_0)\right\Vert}_2^2\leq \delta\right)\geq p_5p_4,$$ provided ${\left\vertx_0\right\vert}^2\leq R_4$. Recall that, in that case, $X^1(\tau_{L^2}+2T)=X^2(\tau_{L^2}+2T)$. Hence, since the law of $X^1(\tau_{L^2}+2T)$ conditioned by $(X^1(\tau_{L^2}),X^2(\tau_{L_2}))$ is $\Dr(X_N(2T,x_0))$, it follows $$\P\left(\max_{i=1,2}{\left\VertX^i(\tau_{L^2}+2T)\right\Vert}_2^2\leq\delta\,\left|\,
\left(X^1(\tau_{L^2}),X^2(\tau_{L_2})\right)\right.\right)\geq
p_4p_5\geq p_1,$$and then
The last case is $X^1(\tau_{L^2})\not=X^2(\tau_{L_2})$ and ${\left\VertX^1(\tau_{L^2})\right\Vert}_2^2\vee{\left\VertX^2(\tau_{L^2})\right\Vert}_2^2>\delta$. In that case, $(X^1(\tau_{L^2}+T),X^2(\tau_{L^2}+T))$ conditioned by $(X^1(\tau_{L^2}),X^2(\tau_{L_2}))$ are independent. Hence, since the law of $X^i(\tau_{L^2}+T)$ conditioned by $(X^1(\tau_{L^2}),X^2(\tau_{L_2}))=(x_0^1,x_0^2)$ is $\Dr(X_N(T,x_0^i))$, it follows from Lemma \[lem7.4\] that $$\P\left(\max_{i=1,2}{\left\VertX^i(\tau_{L^2}+T)\right\Vert}_1^2\leq\delta_4\,\left|\,
\left(X^1(\tau_{L^2}),X^2(\tau_{L_2})\right)\right.\right)\geq
p_4^2.$$ Then, we distinguish the three cases $({\left\VertX^i(\tau_{L^2}+T)\right\Vert}_1^2)_{i=1,2}$ in the small ball of $\H_2$, equal or different and we deduce from Lemma \[lem7.5\] by the same method $$\P\left(\min_{k=1,2}\max_{i=1,2}{\left\VertX^i(\tau_{L^2}+kT)\right\Vert}_2^2\leq\delta\,\left|\,
\left(X^1(\tau_{L^2}+T),X^2(\tau_{L_2}+T)\right)\right.\right)\geq
p_5^2,$$ provided $$\max_{i=1,2}{\left\VertX^i(\tau_{L^2}+T)\right\Vert}_1^2\leq\delta_4.$$ Combining the two previous inequalities, we deduce for the latter case. We have thus proved that is true almost surely.
Integrating , we obtain $$\label{7.5bis}
\P\left(\min_{k=0,\dots,
2}\max_{i=1,2}{\left\VertX^i(\tau_{L^2}+kT)\right\Vert}_2^2\leq\delta\right)\geq
p_1.$$
Combining Lemma \[lem7.2\] and , we conclude.
Probability of having a small noise
-----------------------------------
We now establish Lemma \[PropSmallH3\].
We deduce from Hölder inequality and from $\sum_n\mu_n^{-2}<\infty$ that Hypothesis \[H0\] implies the following fact: for any ${\varepsilon}_0\in (0,{\varepsilon})$, there exists $\alpha
\in (0,1)$, a family $(\bar \phi_n)_n$ of measurable maps $H\to
\R$ and a family $(b_i)_i$ of positive numbers such that $$\label{3.6bis}
\Espace \left\{
\begin{array}{l}
\phi(x)\cdot e_n =b_n\bar \phi_n(x)e_n,\\
\sup_{x\in H}{\left\vert\bar \phi_n(x)\right\vert}\leq 1,
\quad B^*=\sum_n\,\mu_n^{1+{\varepsilon}_0}\,\left( b_n\right)^{2(1-\alpha)}<\infty.
\end{array}
\right.$$ For simplicity we restrict our attention to the case $t=1$. The generalization is easy.
Remark that $$Z(t)=\sum_n b_n Z_n(t)e_n,$$ where $$Z_n(t)=\int_0^te^{- \mu_n(t-s)}\bar \phi_n(X(s))dW_n(s),\;\textrm{ where }\;W=\sum_nW_ne_n.$$ It follows from ${\left\VertZ\right\Vert}^2_2=\sum_n b_n^2 \mu_n {\left\vert\sqrt{
\mu_n}Z_n\right\vert}^2$ and from that $$\label{Small3.1}
\P\left(\sup_{(0,1)}{\left\VertZ\right\Vert}^2_2 \leq B^* {M} \right)\geq
\P\left(\sup_{(0,1)}{\left\vert\sqrt{ \mu_n}Z_n\right\vert}^2 \leq {M} \,\mu_n^{{\varepsilon}_0}\left( b_n\right)^{-2\alpha} ,
\;\forall\; n\right).$$ Setting $$W_n'(t)=\sqrt{ \mu_n}W_n\left(\frac{t}{ \mu_n}\right),$$ we obtain $(W_n')_n$ a family of independent brownian motions. Moreover we have $$\sqrt{ \mu_n}Z_n(t)=Z_n'( \mu_n t),$$where $$Z_n'(t)=\int_0^te^{-(t-s)}\psi_n(s)dW_n'(s), \quad \psi_n(s)=\bar \phi_n\left(X\left(\frac{s}{ \mu_n}\right)\right).$$ Hence, it follows from that $$\label{Small3.3}
\P\left( \sup_{(0,1)}{\left\VertZ\right\Vert}^2_2 \leq B^* {M} \right)\geq
\P\left(\sup_{(0, \mu_n)}{\left\vertZ_n'\right\vert}^2 \leq {M} \,\mu_n^{{\varepsilon}_0}\left( b_n\right)^{-2\alpha} ,
\;\forall\; n\right).$$ Let $W_{n,i}'=W_n'(i+\cdot)-W_n'(i)$ on $(0,1)$. We set $$\Espace
M_{n,i}(t)=
\left\{
\begin{array}{lcl}
0 &\textrm{ if }& t\leq 0,\\
\int_0^{1\wedge t}e^{s}\,\psi_n(i+s)dW_{n,i}'(s)&\textrm{ if }& t\geq 0.
\end{array}
\right.$$ Remark that $$Z_n'(t)=\sum_{i=1}^\infty e^{-(t-i)}M_{n,i}(t-i),$$ which yields for any $q\in\N$ $$\label{Small3.4}
\sup_{(0,q)}{\left\VertZ_n'\right\Vert}_2\leq \left(\frac{e}{e-1}\right)\max_{i=0,...,q-1}\sup_{(0,1)}{\left\vertM_{n,i}\right\vert}.$$ Remark that $(W_{n,k}')_{n,k}$ is a family of independant brownian motions on $(0,1)$. It follows that $(M_{n,k})_{n,k}$ are martingales verifying $\left<M_{n,k},M_{n',k'}\right>=0$ if $(n,k)\not =(n',k')$.
Hence, combining a Theorem by Dambis, Dubins and Schwartz (Theorem 4.6 page 174 of [@KARATZAS]) and a Theorem by Knight (Theorem 4.13 page 179 of [@KARATZAS]), we obtain a family $(B_{n,k})_{n,k}$ of independent brownian motions verifying $$\label{Small3.5}
M_{n,k}(t)=B_{n,k}(\left<M_{n,k}\right>(t)).$$
In the two previous Theorem (Theorem 4.6 page 174 and Theorem 4.13 page 179 of [@KARATZAS]), it is assumed that $\P\left(<M>(\infty)<\infty\right)=0.$ However, as explained in Problem 4.7 of [@KARATZAS], the proof is easily adapted to the case $\P\left(<M>(\infty)<\infty\right)>0.$
Remarking that for any $t\in (0,1)$ $$\left<M_{n,k}\right>(t)=\int_0^{t}{\left\vert\psi_n(k+s)\right\vert}^2ds\leq 1,$$ we deduce from and that for any $q\in\N^*$ $$\sup_{(0,q)}{\left\VertZ_n'\right\Vert}_2\leq \left(\frac{e}{e-1}\right)\max_{i=0,...,q-1}\sup_{(0,1)}{\left\vertB_{n,i}\right\vert}.$$ Hence it follows from that $$\P\left(\sup_{(0,1)}{\left\VertZ\right\Vert}^2_2 \leq {M} \right)\geq
\P\left(\sup_{(0,1)}{\left\vertB_{n,i}\right\vert}^2\leq c {M} \,\mu_n^{{\varepsilon}_0}\left( b_n\right)^{-2\alpha},\;
\forall \,n,\;\forall \,i\leq \mu_n+1\right),$$ where $c=\left(\frac{e-1}{ e}\right)^2\frac{1}{B^*}$.
We deduce from the independence of $(B_{n,k})_{n,k}$ that $$\label{Small3.6}
\P\left(\sup_{(0,1)}{\left\VertZ\right\Vert}^2_2 \leq {M} \right)\geq
\prod_{n\in\N^*}\left( P\left( c {M} \,\mu_n^{{\varepsilon}_0}\left( b_n\right)^{-2\alpha}\right)^{ \mu_n+1}\right),$$ where $$P(d_0)=\P\left(\sup_{(0,1)}{\left\vertB_{1,1}\right\vert}^2 \leq d_0 \right).$$
Recall there exists a family $(c_p)_p$ such that $$\E\left(\sup_{(0,1)}{\left\vertB_{1,1}\right\vert}^{2p}\right)\leq c_p.$$ It follows from Chebyshev inequality and from $1-x\geq e^{-ex}$ for any $x\leq e^{-1}$ that for any $d_0\leq d_p=\left(e^{-1}c_p\right)^\frac{1}{p}$ $$P(d_0)\geq 1-c_pd_0^{-p}\geq e^{-ec_pd_0^{-p}}.$$ Applying , we obtain for any $p>0$ $$\label{Small3.7}
\P\left(\sup_{(0,1)}{\left\VertZ\right\Vert}^2_2 \leq {M} \right)\geq
C_p( {M} )\exp\left(-\frac{c_p'}{ {M} ^p}
\sum_{n> \N(p, {M} )}\left(\frac{ \mu_n+1}{\mu_n^{{\varepsilon}_0p}}\right)b_n^{2\alpha p}
\right),$$ where $$\Espace
\left\{
\begin{array}{lcl}
N(p, {M} )&=&\sup\left\{n\in\N\backslash\{0\}\,\left|\, {M} \,\mu_n^{{\varepsilon}_0}\left( b_n\right)^{-2\alpha}
\leq d_p \right.\right\},\\
C_p( {M} )
&=&\Pi_{n\leq\N(p, {M} )}\left( P\left(c {M} \,\mu_n^{{\varepsilon}_0}\left( b_n\right)^{-2\alpha}\right)^{ \mu_n+1}\right).
\end{array}
\right.$$ Choosing $p$ sufficiently high, we deduce from [**H0**]{} that $$\sum_{n}\left(\frac{ \mu_n+1}{\mu_n^{{\varepsilon}_0p}}\right)b_n^{2\alpha p}\leq C_p'<\infty,$$ which yields, by , that for any $ {M} >0$ and for $p$ sufficiently high $$\label{Small3.7bis}
\P\left(\sup_{(0,1)}{\left\VertZ\right\Vert}^2_2 \leq {M} \right)\geq
C_p( {M} )\exp\left(-\frac{c_p"}{ {M} ^p} \right),$$ Remark that for any $p$, ${\varepsilon}_0$ we have $N(p, {M} )<\infty$. Moreover, it is well-known that for any $d_0>0$, $P(d_0)>0$, which yields $C_p( {M} )>0$ and then Lemma \[PropSmallH3\].
Proof of Lemma \[lem7.2\] {#Preuve7.2}
-------------------------
For simplicity in the redaction, we restrict our attention to the case $f=0$. The generalisation is easy.
Recall $$\E{\left\vertX_N(t)\right\vert}^2\leq e^{- \mu_1 t}{\left\vertx_0\right\vert}^2+\frac{c}{\mu_1}B_0.$$Since $(X^1,X^2)$ is a coupling of $(\Dr(X_N(\cdot,x_0^1)),\Dr(X_N(\cdot,x_0^1)))$ on $T\N$, we obtain $$\label{7.1}
\E\left({\left\vertX^1(nT)\right\vert}^2+{\left\vertX^2(nT)\right\vert}^2\right)\leq e^{- \mu_1
nT}\left({\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right)+2\frac{c}{\mu_1}B_0.$$ Since $(X^1,X^2)$ is a strong Markov process, it can be deduced that there exist $C_6$ and $\gamma_6$ such that for any $x_0\in H$ $$\label{7.2bis}
\E\left(e^{\gamma_6 \tau_{L^2}'}\right)\leq C_6\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right),$$ where $$\tau_{L^2}'=\inf\left\{t\in T\N\backslash\{0\}\,\left|\, {\left\vertX^1(t)\right\vert}^2+{\left\vertX^2(t)\right\vert}^2\geq 4cB_0 \right.\right\}.$$ Taking into account , a standard argument gives that, in order to establish Lemma \[lem7.2\], it is sufficient to prove that there exist $(p_7,T_7)$ such that $$\label{7.2.1}
\P\left({\left\vertX_N(t,x_0)\right\vert}^2\leq \delta_3\right)\geq p_7(\delta_3,t)>0,$$ provided $N\in\N$, $t\geq T_7(\delta_3)$ and ${\left\vertx_0\right\vert}^2\leq 4cB_0$.
We set $$Z(t)=\int_0^te^{-A(t-s)}\phi(X_N(s))dW(s),\quad Y_N=X_N-P_NZ,\quad
M=\sup_{(0,t)}{\left\VertZ\right\Vert}_2^2.$$
Assume that there exist $M_7(\delta_3)>0$ and $T_7(\delta_3)$ such that $$\label{7.2.2}
M\leq M_7(\delta_3)\quad\textrm{ implies }\quad
{\left\vert{Y_N}(t)\right\vert}^2\leq \frac{\delta_3}{4},$$ provided $t\geq T_7(\delta_3)$ and ${\left\vertx_0\right\vert}^2\leq 4cB_0$. Then results from Lemma \[PropSmallH3\] with $$M=\min\left\{M_7(\delta_3),\frac{\delta_3}{4}\right\}.$$ We now prove . Remark that $$\label{7.2.3}
\frac{d}{dt}{Y_N}+ A {Y_N}+P_N B({Y_N}+{P_N Z})=0.$$ Taking the scalar product of with ${Y_N}$, it follows that $$\label{7.2.4bis}
\frac{d}{dt}{\left\vert{Y_N}\right\vert}^2+2 {\left\Vert{Y_N}\right\Vert}^2=-2({Y_N},B({Y_N}+{P_N Z})).$$ Recalling that $(B(y,x),x)=0$, we obtain $$-2({Y_N},B({Y_N}+{P_N Z}))=-2({Y_N},({Y_N},\nabla){P_N Z})-2({Y_N},B({P_N Z})).$$
We deduce from Hölder inequalities and Sobolev embedding that $$-(z,(x,\nabla)y)\leq c{\left\Vertz\right\Vert}{\left\Vertx\right\Vert}{\left\Verty\right\Vert}.$$ Hence it follows from that $$\frac{d}{dt}{\left\vert{Y_N}\right\vert}^2+2 {\left\Vert{Y_N}\right\Vert}^2\leq c{\left\VertZ\right\Vert}^2{\left\Vert{Y_N}\right\Vert}+c{\left\VertZ\right\Vert}{\left\Vert{Y_N}\right\Vert}^2,$$ which yields, by an arithmetico-geometric inequality, $$\frac{d}{dt}{\left\vert{Y_N}\right\vert}^2+2 {\left\Vert{Y_N}\right\Vert}^2\leq cM^\frac{1}{2}{\left\Vert{Y_N}\right\Vert}^2+cM^\frac{3}{2}.$$ It follows that $ M\leq \frac{1}{c^2}$ implies $$\label{7.2.4}
\frac{d}{dt}{\left\vert{Y_N}\right\vert}^2+ {\left\Vert{Y_N}\right\Vert}^2\leq cM^\frac{3}{2}\quad \textrm{ on }\; (0,t).$$ Integrating, we deduce from ${\left\vertx_0\right\vert}^2\leq 4cB_0$ that $${\left\vert{Y_N}(t)\right\vert}^2\leq 4c e^{- \mu_1t}B_0+c\left(\frac{M^\frac{3}{2}}{ \mu_1}\right).$$ Choosing $t$ sufficiently large and $M$ sufficiently small we obtain which yields and then Lemma \[lem7.2\].
In order to avoid a lengthy proof, we have not splitted the arguments in several cases as in the proof of Proposition \[Prop\_Lyap\]. The reader can complete the details.
Proof of Lemma \[lem7.4\] {#Preuve7.4second}
-------------------------
We use the decomposition $X_N=Y_N+P_NZ$ defined in section \[Preuve7.2\] and set $$M=\sup_{(0,T)}{\left\VertZ\right\Vert}_2^2.$$ Integrating , we obtain for $M$ satisfying the same assumption $M\leq\frac{1}{c^2}$ $$\frac{1}{T}\int_0^T{\left\VertY_N(t)\right\Vert}^2dt\leq \frac{1}{T}{\left\vertx_0\right\vert}^2+cM^{\frac{3}{2}},$$which yields, by a Chebyshev inequality, $$\label{7.7.2}
\lambda\left(t\in (0,T)\,\left|\,{\left\VertY_N(t)\right\Vert}^2\leq\frac{2}{T}{\left\vertx_0\right\vert}^2+2cM^{\frac{3}{2}}\right.\right)\geq\frac{T}{2},$$ where $\lambda$ denotes the Lebesgue measure on $(0,T)$.
Setting $$\tau_{\H_1}=\inf\left\{t\in (0,T)\,\left|\, {\left\VertY_N(t)\right\Vert}^2\leq
\frac{2}{T}{\left\vertx_0\right\vert}^2+2cM^{\frac{3}{2}} \right.\right\},$$ we deduce from and the continuity of $Y_N$ that $$\label{7.7.3}
{\left\VertY_N(\tau_{\H_1})\right\Vert}^2\leq
\frac{2}{T}{\left\vertx_0\right\vert}^2+2cM^{\frac{3}{2}}.$$
Taking the scalar product of $2AY$ and , we obtain $$\label{7.7.7}
\frac{d}{dt}{\left\VertY_N\right\Vert}^2+2 {\left\VertY_N\right\Vert}_2^2=-2(AY_N,B(Y_N+P_NZ)).$$ It follows from Hölder inequalities, Sobolev Embeddings and Agmon inequality that $$-2(Ay,\widetilde B(x,z))\leq
c{\left\Verty\right\Vert}_2{\left\Vertz\right\Vert}^\frac{1}{2}{\left\Vertz\right\Vert}_2^\frac{1}{2}{\left\Vertx\right\Vert},$$ where $\widetilde B(x,y)=(x,\nabla)y+(y,\nabla)x$. Hence, we obtain by applying arithmetico-geometric inequalities $$\Espace
\left\{
\begin{array}{lcccl}
-2(A{Y_N},B({Y_N}))&\leq& c{\left\Vert{Y_N}\right\Vert}_2^\frac{3}{2}{\left\Vert{Y_N}\right\Vert}^\frac{3}{2}&\leq &\frac{1}{4}{\left\Vert{Y_N}\right\Vert}^2_2+c{\left\Vert{Y_N}\right\Vert}^6,\\
-2(A{Y_N},B({P_N Z}))&\leq& c{\left\VertY_N\right\Vert}_2{\left\VertZ\right\Vert}^\frac{3}{2}{\left\VertZ\right\Vert}_2^\frac{1}{2}&\leq&
\frac{1}{4}{\left\Vert{Y_N}\right\Vert}^2_2+c{\left\VertZ\right\Vert}_2^4 ,\\
-2(A{Y_N},\widetilde B({Y_N},{P_N Z}))&\leq& c{\left\VertY_N\right\Vert}_2^\frac{3}{2}{\left\VertY_N\right\Vert}^\frac{1}{2}{\left\VertZ\right\Vert}&\leq &
c{\left\VertZ\right\Vert}{\left\VertY_N\right\Vert}_2^2
\end{array}
\right.$$ Remarking that $B({Y_N}+{P_N Z})=B({Y_N})+\widetilde B({Y_N},{P_N
Z})+B({P_N Z})$, it follows from that $M\leq\frac{1}{4c}$ implies $$\label{7.7.8}
\frac{d}{dt}{\left\Vert{Y_N}\right\Vert}^2+ {\left\Vert{Y_N}\right\Vert}_2^2\leq
c{\left\Vert{Y_N}\right\Vert}^2\left({\left\Vert{Y_N}\right\Vert}^4-4K_0^2\right) +cM^2,$$ where $K_0$ is defined in . Let us set $${\sigma}_{\H_1}=\inf\left\{t\in(\tau_{\H_1},T)\,\left|\, {\left\Vert{Y_N}(t)\right\Vert}^2> 2 K_0 \right.\right\},$$ and remark that on $(\tau_{\H_1},{\sigma}_{\H_1})$, we have $$\label{7.7.9}
\frac{d}{dt}{\left\Vert{Y_N}\right\Vert}^2+ {\left\Vert{Y_N}\right\Vert}_2^2\leq
cM^2.$$ Integrating, we obtain that $$\label{7.7.10}
{\left\Vert{Y_N}({\sigma}_{\H_1})\right\Vert}^2+ \int_{\tau_{\H_1}}^{{\sigma}_{\H_1}}{\left\Vert{Y_N}(t)\right\Vert}_2^2dt\leq {\left\Vert{Y_N}(\tau_{\H_1})\right\Vert}^2
+cM^2.$$ Combining and , we obtain that, for $M$ and ${\left\vertx_0\right\vert}^2$ sufficiently small, $${\left\Vert{Y_N}({\sigma}_{\H_1})\right\Vert}^2\leq \frac{\delta_4}{4}\wedge K_0,$$ which yields ${\sigma}_{\H_1}= T$. It follows that $$\label{7.7.6}
{\left\Vert X_N( T)\right\Vert}^2\leq \delta_4,$$ provided $M$ and ${\left\vertx_0\right\vert}^2$ sufficiently small. It remains to use Lemma \[PropSmallH3\] to get Lemma \[lem7.4\].
Proof of Lemma \[lem7.5\] {#Preuve7.5}
-------------------------
It follows from that $$ \int_0^T{\left\VertY_N(t)\right\Vert}_2^2dt\leq {\left\Vertx_0\right\Vert}^2+
cM^2,$$provided $M\leq \frac{1}{4c}$ and $
{\left\Vertx_0\right\Vert}^2+
cM^2\leq K_0.
$
Applying the same argument as in the previous subsection, it is easy to deduce that there exists a stopping times $\tau_{\H_2}\in (0,T)$ such that $$\label{7.7.3bis}
{\left\VertY_N(\tau_{\H_2})\right\Vert}_2^2\leq \frac{2}{T}\left({\left\Vertx_0\right\Vert}^2+
cM^2\right),$$ provided $M$ and ${\left\Vertx_0\right\Vert}$ are sufficiently small.
Taking the scalar product of and $2A^2{Y_N}$, we obtain $$\label{7.15}
\frac{d}{dt}{\left\Vert{Y_N}\right\Vert}_2^2+2 {\left\Vert{Y_N}\right\Vert}_3^2= -2\left(A^\frac{3}{2}{Y_N},A^\frac{1}{2}B({Y_N}+{P_N Z})\right).
$$ Applying Hölder inequality, Sobolev Embeddings $\H_2\subset L^\infty$ and $\H_1\subset L^4$ and arithmetico-geometric inequality, we obtain $$-2\left(A^\frac{3}{2}y,A^\frac{1}{2}B(x,y)\right)\leq
c{\left\Verty\right\Vert}_3{\left\Vertx\right\Vert}_2{\left\Verty\right\Vert}_2\leq
\frac{1}{4} {\left\Verty\right\Vert}_3^2+c\left({\left\Vertx\right\Vert}_2^4+{\left\Verty\right\Vert}_2^4\right).$$ Hence we deduce from and from $B({Y_N}+{P_N Z})=B({Y_N})+B({Y_N},{P_N Z})+B({P_N Z})$ $$\label{7.16}
\frac{d}{dt}{\left\Vert{Y_N}\right\Vert}_2^2+ {\left\Vert{Y_N}\right\Vert}_3^2\leq
c{\left\Vert{Y_N}\right\Vert}_2^2({\left\Vert{Y_N}\right\Vert}_2^2-2K_1)+c{\left\VertZ\right\Vert}_2^4 ,$$ where $K_1$ is defined as $K_0$ in but with a different $c$. We set $${\sigma}_{\H_2}=\inf\left\{t\in(\tau_{\H_2},T)\,\left|\, {\left\Vert{Y_N}(t)\right\Vert}_2^2> 2 K_1 \right.\right\},$$ Integrating , we obtain $${\left\Vert{Y_N}({\sigma}_{\H_2})\right\Vert}_2^2+ \int_{\tau_{\H_2}}^{{\sigma}_{\H_2}}{\left\Vert{Y_N}(t)\right\Vert}_3^2dt\leq
{\left\Vert{Y_N}(\tau_{\H_2})\right\Vert}_2^2+cM^2.$$Taking into account and choosing ${\left\Vertx_0\right\Vert}^2$ and $M^2$ sufficiently small, we obtain $${\left\Vert{Y_N}({\sigma}_{\H_2})\right\Vert}_2^2\leq \frac{\delta}{4}\wedge K_1.$$ It follows that ${\sigma}_{\H_2}=T$ and that $$\label{7.7.6bis}
{\left\Vert X_N( T)\right\Vert}^2\leq \delta,$$ provided $M$ and ${\left\Vertx_0\right\Vert}$ sufficiently small, which yields .
Proof of Theorem \[Th\]
=======================
As already explained, Theorem \[Th\] follows from Proposition \[ThN\]. We now prove Proposition \[ThN\]. Let $(x_0^1,x_0^2)\in(\H_2)^2$. Let us recall that the process $(X^1,X^2)$ is defined at the beginning of section $3$.
Let $\delta>0$, $T\in (0,1)$ be as in Proposition \[Prop\_I\_0\] and $\tau$ defined in , setting $$\tau_1=\tau,\quad \tau_{k+1}=\inf\left\{t>\tau_{k}\,\left|\, {\left\VertX^1(t)\right\Vert}_2^2\vee{\left\VertX^2(t)\right\Vert}_2^2\leq \delta \right.\right\}.$$ it can be deduced from the strong Markov property of $(X^1,X^2)$ and from Proposition \[Prop\_Lyap\] that $$\E\left(e^{\alpha \tau_{k+1}}\right)\leq K"\E\left(e^{\alpha \tau_k}\left(1+{\left\vertX^1(\tau_k)\right\vert}^2+{\left\vertX^2(\tau_k)\right\vert}^2\right)\right),$$ which yields, by the Poincaré inequality, $$\Espace
\left\{
\begin{array}{rcl}
\E\left(e^{\alpha \tau_{k+1}}\right)&\leq& cK"(1+2\delta)\E\left(e^{\alpha \tau_k}\right),\\
\E\left(e^{\alpha \tau_{1}}\right)&\leq& K"\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right).
\end{array}
\right.$$ It follows that there exists $K>0$ such that $$\E\left(e^{\alpha \tau_{k}}\right)\leq K^k\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right).$$Hence, applying Jensen inequality, we obtain that, for any $\theta\in(0,1)$ $$\label{tau_k}
\E\left(e^{\theta\alpha \tau_{k}}\right)\leq K^{\theta k}\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right).$$ We deduce from Proposition \[Prop\_I\_0\] and from that $$\P\left(X^1(T)\not=X^2(T)\right)\leq\frac{1}{4},$$ provided $(x_0^1,x_0^2)$ are in the ball of $(\H^2)^2$ with radius $\delta$.
Setting $$k_0=\inf\left\{k\in\N\,\left|\,X^1(\tau_k+T)=X^2(\tau_k+T)\right.\right\},$$ it follows that $k_0<\infty$ almost surely and that $$\label{k_0}
\P\left(k_0>n\right)\leq \left(\frac14\right)^n.$$ Let $\theta\in (0,1)$. We deduce from Schwartz inequality that $$\E\left(e^{\frac{\theta}{2}\alpha \tau_{k_0}}\right)
=\sum_{n=1}^\infty\E\left(e^{\frac{\theta}{2}
\alpha\tau_{n}}1_{k_0=n}\right) \leq\sum_{n=1}^\infty
\sqrt{\P\left(k_0\geq n\right)\E\left(e^{\theta\alpha
\tau_{n}}\right)}.$$ Combining and , we deduce $$\E\left(e^{\frac{\theta}{2}\alpha \tau_{k_0}}\right)\leq
\left(\sum_{n=0}^\infty\left(\frac{K^\theta}{2}\right)^n\right)
\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right).$$ Hence, choosing $\theta\in(0,1)$ sufficiently small, we obtain that there exists $\gamma>0$ non depending on $N\in\N$ such that $$\label{exp}
\E\left(e^{\gamma\tau_{k_0}}\right)\leq 4\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right).$$ Recall that if $(X^1,X^2)$ are coupled at time $t\in T\N$, then they remain coupled for any time after. Hence $X^1(t)=X^2(t)$ for $t>\tau_{k_0}$. It follows $$\P\left(X^1(nT)\not=X^2(nT)\right)\leq 4e^{-\gamma nT}\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right).$$ Since $(X^1(nT),X^2(nT))$ is a coupling of $((\Pcal_{nT}^N)^*\delta_{x_0^1},(\Pcal_{nT}^N)^*\delta_{x_0^2})$, we deduce from Lemma \[lem\_coupling\] $$\label{2.1}
{\left\Vert\left(\Pcal_{nT}^N\right)^*\delta_{x_0^1}-\left(\Pcal_{nT}^N\right)^*\delta_{x_0^2}\right\Vert}_{var}
\leq 4 e^{-\gamma n T}\left(1+{\left\vertx_0^1\right\vert}^2+{\left\vertx_0^2\right\vert}^2\right),$$ for any $n\in\N$ and any $(x_0^1,x_0^2)\in(\H_2)^2$.
Recall that the existence of an invariant measure $\mu_N\in P(P_N H)$ is justified in section 1.3. Let $\lambda\in P(H)$ and $t\in\R^+$. We set $n=\lfloor\frac{t}{T}\rfloor$ and $C=4e^{\gamma T}$. Integrating $(x_0^1,x_0^2)$ over $((\Pcal_{t-nT}^N)^*\lambda)\otimes\mu_N$ in , we obtain $${\left\Vert\left(\Pcal_{t}^N\right)^*\lambda-\mu_N\right\Vert}_{var}
\leq C e^{-\gamma t}\left(1+\int_{H}{\left\vertx_0\right\vert}^2\,\lambda(dx)\right),$$ which establishes .
Proof of {#Preuve_2.15}
==========
For simplicity in the redaction, we omit $\theta$ and $N$ in our notations.
Remark that $$\label{2.12}
J=\left(\nabla \E\left(g(X(T))\psi_X\right),h\right)=J_1+2J_2,$$ where $$\Espace
\left\{
\begin{array}{lcl}
J_1&=&\E\left(\left(\nabla g(X(T)),\eta(T,0)\cdot h\right)\psi_X\right),\\
J_2&=&\E\left( g(X(T))\psi'_X\int_0^T\left(AX(t),A(\eta(t,0)\cdot h)\right)ds\right).
\end{array}
\right.$$ According to [@Nualart], let us denote by $D_s F$ the Malliavin derivative of $F$ at time $s$. We have the following formula of the Malliavin derivative of the solution of a stochastic differential equation $$ D_sX(t)=1_{t\geq
s}\eta(t,s)\cdot\phi(X(s)),$$which yields $$\label{A2}
\int_0^t D_sX(t)\cdot m(s)\,ds=G(t)\cdot m,$$ where $$G(t) \cdot m= \int_0^t\eta(t,s)\cdot\phi(X(s))\cdot m(s)\,ds.$$ The uniqueness of the solutions gives $$\eta(t,0)\cdot h=\eta(t,s)\cdot (\eta(s,0)\cdot h) \textrm{ for any } 0\leq s \leq t,$$which yields $$\eta(T,0)\cdot h=\frac{1}{T}\int_0^T\eta(t,s)\cdot (\eta(s,0)\cdot
h)\,ds.$$ Setting $$w(s)=\phi^{-1}(X(s))\cdot \eta(s,0)\cdot h,$$ we infer from $$\label{A3}
\eta(T,0)\cdot h=\frac{1}{T}\,G(T)\cdot
w=\frac{1}{T}\int_0^TD_sX(T)\cdot w\,ds,$$ which yields $$\left(\nabla g(X(T)),\eta(T,0)\cdot
h\right)=\frac{1}{T}\int_0^T\left(\nabla g(X(T)),D_sX(T)\cdot
w\right)ds.$$ Remark that $$\left(D_sg(X(T)),w\right)=\left(\nabla g(X(T)),D_sX(T)\cdot
w\right).$$ It follows $$\left(\nabla g(X(T)),\eta(T,0)\cdot
h\right)=\frac{1}{T}\int_0^T\left(D_sg(X(T)),w\right)ds,$$ which yields $$\label{2.13ter}
J_1=\frac{1}{T}\E\int_0^T \psi_X\left(D_{s} g(X(T)), w\right)ds.$$ Recall that the Skohorod integral is the dual operator of the Malliavin derivative (See [@Nualart]). It follows $$\label{2.14}
J_1=\frac{1}{T}\E \left(g(X(T))\int_0^T \psi_X(w(t),dW(t))\right).$$ Recall the formula of integration of a product $$\label{6.1}
\int_0^T \psi_X(w(t),dW(t))=\psi_X \int_0^T (w(t),dW(t))- \int_0^T
\left(D_{s} \psi_X,w(s)\right)ds.$$ Remark that $$D_s\psi_X=2\psi_X'\int_0^T AD_sX(t) \cdot (AX(t))\,dt,$$ which yields, by $\left(AX(t),A D_{s}X(t)\cdot
w(s)\right)=\left(w(s),A D_{s}X(t)\cdot (AX(t))\right)$, $$ \int_0^T\left(D_{s} \psi_X,w(s)\right)ds=2\psi'_X \int_0^T\int_0^T \left(AX(t),A D_{s}X(t)\cdot w(s)\right)\,dtds.$$ We deduce from that $$\label{6.2}
\int_0^T\left(D_{s} \psi_X,w(s)\right)ds=2\psi'_X \int_0^T t\left(AX(t),A \eta(t,0)\cdot h\right)\,dt.$$ Remark that $$\psi_X=\psi'_X=0\quad\textrm{ if }\quad{\sigma}<T.$$ Hence combining and , we obtain $$\int_0^T \psi_X(w(t),dW(t))=\psi_X \int_0^{\sigma}(w(t),dW(t))-
2\psi'_X \int_0^{\sigma}t\left(AX(t),A \eta(t,0)\cdot h\right)\,dt.$$ Thus, follows from and .
[9]{}
[ ]{}
J. Bricmont, A. Kupiainen and R. Lefevere, [*Exponential mixing for the 2D stochastic Navier-Stokes dynamics*]{}, Commun. Math. Phys. [**230**]{}, No.1, 87-132, 2002.
P. Constantin and C. Foias, [*Navier-Stokes Equations*]{}, University of Chicago Press, Chicago, IL, 1988.
G. Da Prato, A. Debussche, [*Ergodicity for the 3D stochastic Navier-Stokes equations*]{} J. Math. Pures Appl. (9) 82 (2003), no. 8, 877–947.
G. Da Prato and J. Zabczyk, [*Stochastic equations in infinite dimensions,*]{} Encyclopedia of Mathematics and its Applications, Cambridge University Press, 1992.
G. Da Prato and J. Zabczyk, [*Ergodicity for Infinite Dimensional Systems,*]{} London Mathematical Society Lecture Notes, n.229, Cambridge University Press, 1996.
[ ]{}
A. Debussche, C. Odasso, [*Ergodicity for the weakly damped stochastic Non-linear Schrödinger equations*]{}, Journal of Evolution Equations , Birkhäuser Basel, vol. 5, no. 3, pp. 317-356.
A. Debussche, C. Odasso, [*Markov solutions for the 3D stochastic Navier Stokes equations with state dependent noise*]{}, Journal of Evolution Equations, Birkhäuser Basel, vol. 6, no. 2, pp. 305-324.
W. E, J.C. Mattingly, Y. G. Sinai, [*Gibbsian dynamics and ergodicity for the stochastically forced Navier-Stokes equation*]{}, Commun. Math. Phys. [**224**]{}, 83–106, 2001.
[ ]{}
F. Flandoli [*Irreducibilty of the 3-D stochastic Navier–Stokes equation,*]{} Journal of Functional Ananlysis, [**149**]{}, 160-177, 1997.
F. Flandoli, [*An introduction to 3D stochastic Fluid Dynamics*]{}, CIME lecture notes 2005.
F. Flandoli and D. Gatarek, [*Martingale and stationary solutions for stochastic Navier–Stokes equations*]{}, PTRF, 102, 367-391, 1995.
[ ]{}
F. Flandoli and M. Romito, [*Partial regularity for the stochastic Navier–Stokes equations*]{}, Trans. Am. Math. Soc. 354, No 6, 2207-2241, 2002.
F. Flandoli and M. Romito, [*Markov selections for the 3D stochastic Navier- Stokes equations*]{}, in preparation.
Foias, C. and Temam, R., [*Gevrey class regularity for the solutions of the Navier-Stokes equations*]{}, J. Funct. Anal. 87 (1989), no. 2, 359–369.
[ ]{} M. Hairer, [*Exponential Mixing Properties of Stochastic PDEs Through Asymptotic Coupling*]{}, Proba. Theory Related Fields,[**124**]{}, 3 :345-380, 2002.
M. Hairer, J. C. Mattingly [*Ergodic properties of highly degenerate 2D stochastic Navier-Stokes equations*]{} C. R. Math. Acad. Sci. Paris 339 (2004), no. 12, 879–882. M. Hairer, J. Mattingly, [*Ergodicity of the 2D Navier-Stokes equations with degenerate forcing*]{}, preprint.
[ ]{}
W.D. Henshaw, H.O. Kreiss, L.G. Reyna, [*Smallest scale estimate for the Navier-Stokes equations for incompressible fluids*]{}, Arch. Rationnal Mech. Anal. [**112**]{} (1990), no. 1, 21-44.
G. Huber, P. Alstrom, [*Universal Decay of vortex density in two dimensions*]{}, Physica A [**195**]{}, 448-456, 1993.
Karatzas I. and Shreve S. E., [*Brownian motion and stochastic calculus, Second edition*]{}, Graduate Texts in Mathematics, 113. Springer-Verlag, New York, 1991.
S. Kuksin, [*On exponential convergence to a stationary mesure for nonlinear PDEs*]{}, The M. I. Viishik Moscow PDE seminar, Amer. Math. Soc. Trans. (2), vol 206, Amer. Math. Soc., 2002.
S. Kuksin, A. Shirikyan, [*Stochastic dissipative PDE’s and Gibbs measures*]{}, Commun. Math. Phys. [**213**]{}, 291–330, 2000.
S. Kuksin and A. Shirikyan,[*Ergodicity for the randomly forced 2D Navier-Stokes equations*]{}, Math. Phys. Anal. Geom. [**4**]{}, 2001.
S. Kuksin, A. Shirikyan, [*A coupling approach to randomly forced randomly forced PDE’s I*]{}, Commun. Math. Phys. [**221**]{}, 351–366, 2001.
S. Kuksin, A. Piatnitski, A. Shirikyan, [*A coupling approach to randomly forced randomly forced PDE’s II*]{}, Commun. Math. Phys. [**230**]{}, No.1, 81-85, 2002.
S. Kuksin, A. Shirikyan, [*Coupling approach to white-forced nonlinear PDEs*]{}, J. Math. Pures Appl. 1 (2002) pp. 567-602.
[ ]{}
J. Mattingly, [*Exponential convergence for the stochastically forced Navier-Stokes equations and other partially dissipative dynamics*]{}, Commun. Math. Phys. [**230**]{}, 421-462, 2002.
J. Mattingly, E. Pardoux, [*Ergodicity of the 2D Navier-Stokes Equations with Degenerate Stochastic Forcing* ]{}, preprint 2004.
[ ]{}
Nualart, [*Malliavin Calculus and related topic*]{}, Probability and its Applications, 1995, Springer.
C. Odasso, [*Ergodicity for the stochastic Complex Ginzburg–Landau equations*]{}, to appear in Annales de l’institut Henri-Poincaré, Probabilités et Statistiques.
C. Odasso, [*Exponential Mixing for Stochastic PDEs: The Non-Additive Case,*]{} preprint available on http://www.bretagne.ens-cachan.fr/math/people/cyril.odasso/.
A. Shirikyan, [*Exponential mixing for 2D Navier-Stokes equation pertubed by an unbounded noise*]{}, J. Math. Fluid Mech. [**6**]{}, no. 2, 169–193, 2004.
A. Shirikyan, [*Analyticity of solutions of randomly perturbed two-dimensional Navier-Stokes equations*]{}, Uspekhi Mat. Nauk [**57**]{} (2002), no. 4(346), 151–166; translation in Russian Math. Surveys [**57**]{} (2002), no. 4, 785–799.
R. Temam, [*Navier–Stokes Equations. Theory and Numerical Analysis.*]{}, North-Holland, Amsterdam-New York-Oxford,1977.
R. Temam [*Navier-Stokes equations and nonlinear functional analysis*]{}, Philadelphia (PA US) : SIAM , 1995 , CBMS-NSF regional conference series in applied mathematics ; 66 ISBN 0-89871-340-4
R. Temam, [*Some developments on Navier-Stokes Equations in the second half of the 20th Century*]{}, “in Développement des Mathématiques au cours de la seconde moitié du XXème siécle”, J.P. Pier ed., Birkhauser Verlag, Basel, 2000.
[ ]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We review the use of techniques of positive currents for the study of parameter spaces of one-dimensional holomorphic dynamical systems (rational mappings on $\mathbb{P}^1$ or subgroups of the Möbius group $\mathrm{PSL}(2, \mathbb{C})$). The topics covered include: the construction of bifurcation currents and the characterization of their supports, the equidistribution properties of dynamically defined subvarieties on parameter space.'
address: |
CMLS\
École Polytechnique\
91128 Palaiseau\
France
author:
- Romain Dujardin
title: Bifurcation currents and equidistribution in parameter space
---
Introduction {#introduction .unnumbered}
============
Let $(f_{\lambda})_{{\lambda}\in{\Lambda}}$ be a holomorphic family of dynamical systems acting on the Riemann sphere ${{\mathbb{P}^1}}$, parameterized by a complex manifold ${\Lambda}$. The “dynamical systems" in consideration here can be polynomial or rational mappings on ${{\mathbb{P}^1}}$, as well as groups of Möbius transformations. It is a very basic idea that the product dynamics $\widehat f$ acting on ${\Lambda}\times {{\mathbb{P}^1}}$ by $\widehat{f}({\lambda},z) =({\lambda}, f_{\lambda}(z))$ is an important source of information on the bifurcation theory of the family. The input of techniques from higher dimensional holomorphic dynamics into this problem recently led to a number of interesting new results in this area, especially when the parameter space ${\Lambda}$ is multidimensional. Our purpose in this paper is to review these recent developments.
The main new idea that has arisen from this interaction is the use of positive closed currents. We will see that the consideration of the dynamics of $\widehat f$ gives rise to a number of interesting currents on ${\Lambda}\times {{\mathbb{P}^1}}$ and ${\Lambda}$. Positive currents have an underlying measurable structure, so it would be fair to say that we are studying these parameter spaces at a measurable level, somehow in the spirit of the ergodic theoretic approach to dynamics[^1].
A basic way in general to construct and study dynamical currents is to view them as limits of sequences of dynamically defined subvarieties. This will be another major theme in this paper.
We will try as much as possible to emphasize the similarities between methods of higher dimensional dynamics, of the study of families of rational maps and that of Möbius subgroups. We will also state a number of open questions, to foster further developments of this theory.
An interesting outcome of these methods is the possibility of studying “higher codimensional" phenomena –like the property for a rational map of having several periodic critical points. These phenomena are difficult to grasp using elementary complex analysis techniques because of the failure of Montel’s theorem in higher dimension. In the same vein, we will see that when $\dim({\Lambda})>1$, the bifurcation locus of a family of rational maps possesses a hierarchical structure, which may conveniently be formalized using bifurcation currents. When $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ is the family of polynomials of degree $d\geq 3$, the smaller of these successive bifurcation loci is the right analogue of the Mandelbrot set in higher degree, with whom it shares many important properties.
$\diamond$
[**Contents.**]{} Let us now outline the contents of this article. Section \[sec:prologue\] is of general nature. We explain how the non-normality of a sequence of holomorphic mappings $f_n:{\Lambda}{\rightarrow}X$ between complex manifolds is related to certain closed positive currents of bidegree (1,1) on ${\Lambda}$. We also show that the preimages under $f_n$ of hypersurfaces of $X$ tend to be equidistributed. This will provide –at least at the conceptual level– a uniform framework for many of the subsequent results.
Since these facts are not so easy to extract from the literature, we explain them in full detail, therefore the presentation is a bit technical. The reader who wants to dive directly into holomorphic dynamics is advised to skip this section on a first reading.
Sections \[sec:bifcur\] and \[sec:higher\] are devoted to the study of bifurcation currents for polynomials and rational maps on ${{\mathbb{P}^1}}$, which is the most developed part of the theory. In Section \[sec:bifcur\] we present two (related) constructions of bifurcation currents of bidegree (1,1): the “absolute" bifurcation current ${T_{\mathrm{bif}}}$ and the bifurcation current associated to a marked critical point. In both cases, the support of the bifurcation current is equal to the corresponding bifurcation locus. We also show that these currents describe the asymptotic distribution of families of dynamically defined codimension 1 subsets of parameter space. More precisely we will be interested in the families of hypersurfaces ${\mathrm{PerCrit}}(n,k)$ (resp. ${\mathrm{Per}}(n,w)$) defined by the condition that a critical point satisfies $f^n(c)=f^k(c)$ (resp. $f$ possesses a periodic $n$-cycle of multiplier $w$).
In Section \[sec:higher\], we study “higher" bifurcation currents, which are obtained by taking exterior products of the previous ones. We will develop the idea that the supports of these currents define a dynamically meaningful filtration of the bifurcation locus, and try to characterize them precisely. We also explain why bifurcation currents should display some laminar structure in parts of parameter space, and give some results in this direction.
Many of the proofs will be sketched, the reader being referred to the original papers for complete arguments. Let us also mention a recent set of lectures notes by Berteloot [@berteloot; @survey] which covers most of this material with greater detail (and complete proofs).
In Section \[sec:kleinbif\] we introduce currents associated to bifurcations of holomorphic families of subgroups of ${\mathrm{PSL}(2,\mathbb C)}$, which is in a sense the counterpart of Section \[sec:bifcur\] in the Kleinian groups setting. The existence of such a counterpart is in accordance with the so-called [*Sullivan dictionary*]{} between rational and Kleinian group dynamics, nevertheless its practical implementation requires a number of new ideas. To be specific, let $(\rho_{\lambda})$ be a holomorphic family of representations of a given finitely generated group into ${\mathrm{PSL}(2,\mathbb C)}$ (satisfying certain natural assumptions). We construct a bifurcation current on ${\Lambda}$ associated to a random walk on $G$. As before, this current is supported precisely on the “bifurcation locus” of the family, and it describes the asymptotic distribution of natural codimension 1 subsets of parameter space. We will see that the key technical ingredient here is the ergodic theory of random products of matrices.
In Section \[sec:further\] we outline some possible extensions of the theory. An obvious generalization would be to consider rational mappings in higher dimension. A basic difficulty is that in that setting the understanding of bifurcation phenomena is still rather poor.
We do not include a general discussion about plurisubharmonic (psh for short) functions and positive currents. Good reference sources for this are the books by Demailly [@demailly Chap. I and III] and Hörmander [@hormander Chap. 4]. See also [@cantat; @survey] in this volume for a short presentation. We do not require much knowledge in holomorphic dynamics, except for the basic properties of the maximal entropy measure [@lyubich; @flm].
$\diamond$
[**Bibliographical overview.**]{} Let us briefly review the main references that we will be considering in the paper. Bifurcation currents were introduced by DeMarco in [@demarco1]. In this paper she constructs a current ${T_{\mathrm{bif}}}$ on any holomorphic family of rational maps, whose support is the bifurcation locus. This current is defined in terms of the critical points. In [@demarco2], she proves a formula for the Lyapunov exponent of a rational map on ${{\mathbb{P}^1}}$ (relative to its maximal entropy measure), and deduces from this that ${T_{\mathrm{bif}}}$ is the $dd^c$ of the Lyapunov exponent function.
In [@bas-ber1], Bassanelli and Berteloot generalize DeMarco’s formula to higher dimensional rational maps, and initiate the study of the higher exterior powers of the bifurcation current (associated to rational mappings on ${{\mathbb{P}^1}}$), by showing that $\operatorname{Supp}({T_{\mathrm{bif}}}^k)$ is accumulated by parameters possessing $k$ indifferent cycles. In [@preper], Favre and the author study the asymptotic distribution of the family of hypersurfaces ${\mathrm{PerCrit}}(n,k)$. The structure of the space of polynomials of degree $d$ is also investigated, with emphasis on the higher dimensional analogue of the boundary of the Mandelbrot set. A finer description is given in the particular case of cubic polynomials in [@cubic].
Several equidistribution theorems for the family of hypersurfaces ${\mathrm{Per}}(n,w)$ are obtained by Bassanelli and Berteloot in [@bas-ber2; @bas-ber3]. [@bas-ber2] also discusses the laminarity properties of bifurcation currents in the space of quadratic rational maps.
In [@buff; @epstein], Buff and Epstein develop a method based on transversality ideas to characterize the supports of certain “higher" bifurcation currents. This was recently generalized by Gauthier [@gauthier], leading in particular to Hausdorff dimension estimates for the supports of these currents, which generalize Shishikura’s famous result that the boundary of the Mandelbrot set has dimension 2.
Bifurcation currents for families of subgroups of ${\mathrm{PSL}(2,\mathbb C)}$, satisfying properties similar to the above, were designed by Deroin and the author in [@kleinbif].
Related developments in higher dimensional holomorphic dynamics can be found in [@ds-pl; @pham].
$\diamond$
We close this introduction with a few words on the connexions between these ideas and the work of Milnor. Alone or with coauthors, he wrote a number of papers, most quite influential, on parameter spaces of polynomials and rational functions [@milnor; @cubic; @milnor; @quadratic; @milnor; @components; @milnor; @bicritical; @milnor; @smooth; @bkm]. Common features among these articles include the emphasis on multidimensional issues, and the role played by subvarieties of parameter space, like the ${\mathrm{PerCrit}}(n,k)$ and the ${\mathrm{Per}}(n,w)$. I hope he will appreciate the way in which these ideas reappear here.
Many thanks to Serge Cantat, Charles Favre, and Thomas Gauthier for their useful comments.
Prologue: normal families, currents and equidistribution {#sec:prologue}
========================================================
Let ${\Lambda}$ be a complex manifold of dimension $d$, and $X$ be a compact Kähler manifold of dimension $k$, endowed with Kähler form $\omega$. Let $(f_n)$ be a sequence of holomorphic mappings from ${\Lambda}$ to $X$. In this section we explain a basic construction relating the non normality of the sequence $(f_n)$ and certain positive (1,1) currents on ${\Lambda}$. When applied to particular situations it will give rise to various bifurcation currents. This construction is also related to higher dimensional holomorphic dynamics since we may take ${\Lambda}= X$ and $f_n$ be the family of iterates of a given self-map on $X$.
The problems we consider are local on ${\Lambda}$ so without loss of generality we assume that ${\Lambda}$ is an open ball in ${\mathbb{C}}^{d}$. We say that the family $(f_n)$ is [*quasi-normal*]{} if for every subsequence of $(f_n)$ (still denoted by $(f_n)$) there exists a further subsequence $(f_{n_j})$ and an analytic subvariety $E\subset {\Lambda}$ such that $(f_{n_j})$ is a normal family on ${\Lambda}\setminus E$ (see [@ivashkovich] for a discussion of this and other related notions).
A normality criterion
---------------------
The following result is a variation on well-known ideas, but it is apparently new.
\[thm:volume\] Let ${\Lambda}$ and $X$ be as above, and $(f_n)$ be a sequence of holomorphic mappings from ${\Lambda}$ to $X$. If the sequence of bidegree (1,1) currents $f_n^*\omega$ has locally uniformly bounded mass on ${\Lambda}$, then the family $(f_n)$ is quasi-normal on ${\Lambda}$.
Recall that the mass of a positive current of bidegree (1,1) in an open set ${\Omega}\subset {\mathbb{C}}^d$ is defined by ${{\bf M}}_{\Omega}(T) =
\sup {\left\vert{\left\langle T,\varphi\right\rangle}\right\vert}$, where $\varphi$ ranges among test $(d-1, d-1)$ forms $\sum\varphi_{I,J}dz_I\wedge d\overline{z}_J$ with ${\left\Vert\varphi_{I,J}\right\Vert}_{L^\infty} \leq 1$.
A few comments are in order here. First; if $d=1$, the result is a well-known consequence of Bishop’s criterion for the normality of a sequence of analytic sets [@chirka §15.5] (see Lemma \[lem:dim1\] below for the proof). The point here is that if $d>1$ our assumption does *not* imply that the volumes of the graphs of the $f_n$ are locally bounded (see [@ivashkovich Example 5.1]). Secondly, it is clear that the converse of Theorem \[thm:volume\] is false, i.e. there exist quasi-normal families such that $f_n^*\omega$ has unbounded mass. For this, take any sequence of holomorphic mappings ${\mathbb{D}}{\rightarrow}{{\mathbb{P}^1}}$, converging on compact subsets of ${\mathbb{D}}^*$ to $z\mapsto \exp(1/z)$.
\[lem:dim1\] Theorem \[thm:volume\] holds when $\dim({\Lambda}) =1$.
Let $\Gamma(f_n)\subset {\Lambda}\times X$ be the graph of $f_n$. Let $\pi_1$, $\pi_2$ be the first and second projections on ${\Lambda}\times X$. Let $\beta$ be the standard Kähler form on ${\Lambda}\subset {\mathbb{C}}$. Then if $U\Subset {\Lambda}$, the volume of $\Gamma(f_n)\cap {\pi_1{^{-1}}(U)}$ relative to the product Hermitian structure equals $$\int_{\Gamma(f_n)\cap {\pi_1{^{-1}}(U)} }\pi_1^*\beta + \pi_2^*\omega =\operatorname{vol}_{\Lambda}(U) + \int_U {\left(\pi_2\circ \left(\pi_1{ \arrowvert_{\Gamma(f_n)}}{^{-1}}\right)\right)}^*\omega = \operatorname{vol}_{\Lambda}(U) + \int_U f_n^*\omega.$$ Therefore our assumption implies that the volumes of the analytic sets $\Gamma(f_n)\cap {\pi_1{^{-1}}(U)}$ are uniformly bounded[^2]. By Bishop’s theorem one can extract a subsequence $n_j$ such that the $\Gamma(f_{n_j})\cap {\pi_1{^{-1}}(U)}$ converge in the Hausdorff topology to a one-dimensional analytic set $\Gamma$ of ${\pi_1{^{-1}}(U)}$.
We claim that $\Gamma$ is the union of a graph and finitely many vertical curves. Here vertical means that it projects to a point on ${\Lambda}$. Indeed, first note that by Lelong’s lower bound on the volume of an analytic set [@chirka §15.3], the volume of any analytic subset of $X$ is uniformly bounded below. Since $\operatorname{vol}_{{\Lambda}\times X}(\Gamma)\leq \liminf \operatorname{vol}_{{\Lambda}\times X}(\Gamma(f_{n_j}))$, this implies that $\Gamma$ contains only finitely many vertical components. Let $E\subset{\Lambda}$ be the projection of these components. We claim that the $f_{n_j}$ converge locally uniformly outside $E$. Indeed let $V\subset {\Lambda}$ be a connected open subset disjoint from $E$. Since the $\Gamma(f_{n_j})$ converge in the Hausdorff topology, we see that $\Gamma\cap \pi_1{^{-1}}(V)$ is non empty. Now since $\pi_1$ is proper we infer that $\pi_1{ \arrowvert_{\Gamma\cap \pi_1{^{-1}}(V)}}$ is a branched covering, which must be of degree 1 (for if not, generic fibers of $\pi_1$ would intersect $\Gamma(f_{n_j})$ in several points for large $j$). We conclude that $\Gamma$ is a graph over $V$, of some $f:V{\rightarrow}X$, and that the $f_{n_j}$ converge uniformly to $f$ there.
The possibility of vertical components of $\Gamma$ over a locally finite set is known as the [*bubbling phenomenon*]{}. An important consequence of the proof is that there exists a constant $\delta_0$ (any number smaller than the infimum of the volumes of 1-dimensional subvarieties of $X$ will do) such that if $\limsup \int_V f_n^*\omega \leq \delta_0$, then no bubbling occurs in $V$. An easy compactness argument yields the following:
\[cor:derivative\] Let $V$ be a one dimensional disk and let $\delta_0>0$ be as above. For every $K\Subset V$, there exists a constant $C(V,K,\delta_0)$ so that if $f:V{\rightarrow}X$ is a holomorphic mapping satisfying $\int_V f^*\omega \leq\delta_0$, then ${\left\Vertdf\right\Vert}_{L^\infty(K)}\leq C$.
We now prove the theorem. The idea, based on an argument from [@ds-pl] (see also [@fatou Prop. 5.7]), is to use a slicing argument together with a theorem due to Sibony and Wong [@sibony-wong]. For convenience let us state this result first.
\[thm:sibonywong\] let $g$ be a holomorphic function defined in the neighborhood of the origin in ${\mathbb{C}}^d$, which admits a holomorphic continuation to a neighborhood of $\bigcup_{L\in E} L\cap B(0,R)$, where $E\subset {\mathbb{P}}^{d-1}$ is a set of lines through the origin, of measure $\geq 1/2$ (relative to the Fubini-Study volume on ${\mathbb{P}}^{d-1}$).
Then there exists a constant $C_{SW}>0$ such that $g$ extends to a holomorphic function of $B(0, C_{SW}R)$, and furthermore $$\label{eq:sibonywong}
\sup_{B(0, C_{SW}R)}\nolimits {\left\vertg\right\vert} \leq
\sup_{\bigcup_{L\in E} L\cap B(0,R)}\nolimits {\left\vertg\right\vert}.$$
Recall that ${\Lambda}$ was supposed to be an open subset in $\mathbb{C}^d$. Denote by $\beta$ the standard Kähler form on ${\mathbb{C}}^d$. Let $T_n = f_n^*\omega$ and consider a subsequence (still denoted by $n$) such that $T_{n}$ converges to a current $T$ on ${\Lambda}$. Let $\sigma_T = T\wedge \beta^{d-1}$ be the trace measure of $T$. For every $p\in {\Lambda}$,$ {\frac{1}{c_{2d-2}r^{2d-2}}} \sigma_T(B(p, r))$ converges as $r{\rightarrow}0$ to the Lelong number of $T$ at $p$, denoted by $\nu(T, p)$ (here $c_{2d-2}$ is the volume of the unit ball in ${\mathbb{C}}^{2d-2}$). By Siu’s semi-continuity theorem [@demailly], for each $c>0$, $E_c(T)= {\left\{p, \ \nu(T,c)\geq c\right\}}$ is a proper analytic subvariety of ${\Lambda}$. Fix $c=\delta_0/4$, where $\delta_0$ is as above. We show that $(f_n)$ is a normal family on ${\Lambda}\setminus E_c(T)$.
Indeed, let $p\notin E_c(T)$. Then for $r<r(\delta_0)$ (which will be fixed from now on), $${\frac{1}{c_{2d-2}r^{2d-2}}}\int_{B(p, r)} T\wedge \beta^{d-1} \leq \frac{\delta_0}{3},$$ hence for large $n$, $${\frac{1}{c_{2d-2}r^{2d-2}}}\int_{B(p, r)} T_{n}\wedge \beta^{d-1} \leq \frac{\delta_0}{2}.$$ Let now $\alpha_p = dd^c\log{\left\Vertz-p\right\Vert}$. By Crofton’s formula [@demailly Cor. III.7.11], $\alpha_p^{d-1} = \int_{{\mathbb{P}}^{d-1}}[L]dL$ is the average of the currents of integrations along the complex lines through $p$ (w.r.t. the unitary invariant probability measure on ${\mathbb{P}}^{d-1}$). By a well-known formula due to Lelong [@chirka §15.1], for every $r_1<r$, $${\frac{1}{c_{2d-2}r ^{2d-2}}} \sigma_{T_{n}}(B(p, r ))-{\frac{1}{c_{2d-2}r_1^{2d-2}}} \sigma_{T_{n}} (B(p, r_1)) =
\int_{r_1<{\left\Vertz-p\right\Vert}< r} T_{n}\wedge \alpha_p^{d-1}.$$ Since $T_{n}$ is a smooth form it has zero Lelong number at $p$ and we can let $r_1$ tend to zero. We conclude that for every large enough $n$, $\int_{B(p,r)}T_{n}\wedge \alpha_p^{d-1} \leq \frac{\delta_0}{2}$. Applying Crofton’s formula we see that $$\int_{{\mathbb{P}}^{d-1}} \left(\int_{L\cap B(p,r)} f_{n}^*\omega\right) \leq \frac{\delta_0}{2},$$ therefore there exists a set $E_n$ of lines of measure at least $1/2$ such that if $L\in E_n$ then $\int_{L\cap B(p,r)} f_{n}^*\omega<\delta_0$. By Corollary \[cor:derivative\] above, for each such line $L\in E_n$, the derivative of $f_{n}{ \arrowvert_{L\cap B(p,r)}}$ is locally uniformly bounded. Extract a further subsequence so that $f_n(p)$ converges to some $x\in X$. Thus, reducing $r$ if necessary, we can assume that for $L\in E_n$, $f_{n}{ \arrowvert_{L\cap B(p,r)}}$ takes its values in a fixed coordinate patch containing $x$, which may be identified to a ball in ${\mathbb{C}}^k$. By Theorem \[thm:sibonywong\], there exists a constant $C_{SW}$ such that $f_n{ \arrowvert_{B(p,C_{SW}r)}}$ takes its values in the chart, with the same bound on the derivative. This implies that $(f_n)$ is a normal family in $B(p,C_{SW}r)$, thereby concluding the proof.
\[rmk:normal\] The proof shows that if it can be shown that the Lelong numbers of the cluster values of $(f_n)^*\omega$ are smaller than $\delta_0$ (for instance if the potentials are uniformly bounded), then the family is actually normal.
Equidistribution in codimension 1
---------------------------------
We now turn to the case where the mass of $f_n^*\omega$ tends to infinity, and show that the preimages of hypersurfaces under $f_n$ of $X$ tend to equidistribute in the sense of currents. This idea goes back to the work of Russakovskii, Shiffmann and Sodin [@russakovskii-sodin; @russakovskii-shiffman]. Dinh and Sibony later gave in [@ds-tm] a wide generalization of these results. Here we present a simple instance of this phenomenon, which is inspired by (and can be deduced from) [@ds-tm].
Let ${\Lambda}$, $X$ and $(f_n)$ be as above and set $d_n = \int_{\Lambda}f_n^*\omega\wedge \beta^{d-1}$, so that $ {d_n{^{-1}}} f_n^*\omega $ is a sequence of currents of bounded mass on ${\Lambda}$. A first remark is that if $\omega'$ is another Kähler form, there exists a constant $C\geq 1$ such that $ {C}{^{-1}}\omega\leq \omega'\leq C\omega$, hence $ {d_n}{^{-1}}f_n^*\omega'$ also has bounded mass.
By definition, a holomorphic family of subvarieties $(H_a)_{a\in A}$ of dimension $l$ parameterized by a complex manifold $A$ is the data of a subvariety $H$ in $A\times X$, of dimension $\dim (A)+l$, such that for every $a\in A$, $\pi_1{^{-1}}(a)\cap H =:H_a$ has dimension $l$. Of course here we are identifying every fiber $\pi_1{^{-1}}(a)$ with $X$, using the second projection. We will only need to consider the case $l=k-1$.
We need a notion of a “sufficiently mobile" family of hypersurfaces. For this, let us assume for simplicity that $X$ is a projective manifold. We say that $(H_a)_{a\in A}$ is a [*substantial*]{} family of hypersurfaces on $X$ if the hypersurfaces $(H_a)$ are hyperplane sections relative to some embedding $\iota:X\hookrightarrow {\mathbb{P}}^N$ and there exists a positive measure $\nu$ on $A$ such that the current $\int[H_a] d\nu(a)$ has locally bounded potentials.
Notice that the family of [*all*]{} hyperplane sections relative to some projective embedding of $X$ is substantial. Indeed by Crofton’s formula there exists a natural smooth measure $dL$ on the dual projective space $\check{\mathbb{P}}^N$ (i.e. the space of hyperplanes) such that the average current of integration is the Fubini-Study form, i.e. $\int[L]dL = \omega_{FS}$. Therefore on $X$ we get that $\int [\iota{^{-1}}(L)] dL = \iota^*\omega_{FS}$, and the family is substantial. From this it follows for instance that the family of hypersurfaces of given degree in ${\mathbb{P}}^k$ is substantial.
Here is the equidistribution statement. We do not strive for maximal generality here and it is likely that some of the assumptions could be relaxed. For instance, in view of applications to random walks on groups it is of interest to obtain similar results for non compact $X$ (to deal with examples like $X=\mathrm{SL}(n,{\mathbb{C}})$, etc). One might also obtain equidistribution statements in higher codimension, by introducing appropriate dynamical degrees.
\[thm:equidist abstrait\] Let ${\Lambda}$ be a complex manifold of dimension $d$ and $X$ be a projective manifold. Let $f_n:{\Lambda}{\rightarrow}X$ be a family of holomorphic mappings such that $d_n = \int_{\Lambda}f_n^*\omega\wedge \beta^{d-1}$ tends to infinity. Let $(H_a)_{a\in A}$ be an substantial holomorphic family of hypersurfaces in $X$.
Let $\mathcal{E} $ be the set of $a\in A$ such that $${\frac{1}{d_n}}(f_n^*[H_a] - f_n^*\omega) \text{ does not converge to zero in the sense of currents.}$$ Then:
1. if the series $ \sum {d_n{^{-1}}}$ converges, then $\mathcal{E}$ is pluripolar;
2. if for every $t>0$ the series $\sum \exp(-td_n)$ converges, then $\mathcal{E}$ has zero Lebesgue measure.
As the proof easily shows, in case [*ii.*]{} Lebesgue measure can actually be replaced by any [*moderate measure*]{}, that is a measure satisfying an inequality of the form $m({\left\{u<-t\right\}})\leq C e^{-\alpha t}$ on any compact class of psh functions. This is wide class of measures which contains for instance the area measure on totally real submanifolds of maximal dimension. We refer to [@ds-tm] for details.
Under the assumptions of the theorem, if the sequence ${\frac{1}{d_n}} f_n^*\omega$ converges to some current $T$, then for $a\notin \mathcal{E}$, ${\frac{1}{d_n}}f_n^*[H_a] $ converges to $T$.
Without loss of generality, assume that $A$ is an open ball in ${\mathbb{C}}^{\dim(A)}$.
\[lem:uax\] Under the assumptions of the theorem, there exists a Kähler form $\omega$ on $X$ and a negative function $(a,x)\mapsto u(a,x)$ on $A\times X$ such that
i. for every $a\in A$, $dd^c_x u(a, \cdot) = [H_a]-\omega$;
ii. the $L^1$ norm ${\left\Vertu(a, \cdot)\right\Vert}_{L^1(X)}$ is locally uniformly bounded with respect to $a\in A$;
iii. for every $x$, $ u(\cdot,x)$ is psh on $A$.
Assuming this result for the moment, let us continue with the proof of the theorem. Suppose first that the series $ \sum {d_n{^{-1}}}$ converges. Let $m$ be a positive measure with compact support on $A$ such that psh functions are $m$-integrable. We claim that for $m$-a.e. $a$, ${\frac{1}{d_n}}(f_n^*[H_a] - f_n^*\omega)$ converges to zero. By Lemma \[lem:uax\].i. for this it is enough to prove that ${\frac{1}{d_n}}\int u_a\circ f_n$ tends to zero in $L^1_{\rm loc}({\Lambda})$ (here and in what follows we denote $u(a, \cdot)$ by $u_a$).
Let us admit the following lemma for the moment.
\[lem:bdd\] The function defined by $x\mapsto \int u_a(x) dm(a)$ is $\omega$-psh (i.e. $dd^cu_a\geq -\omega$) and bounded on $X$.
Fix ${\Lambda}'\Subset {\Lambda}$. By Fubini’s theorem and Lemma \[lem:bdd\], setting $ \widetilde{u}(x) = \int u_a(x) dm(a)$ we have that $$\int{\left({\frac{1}{d_n}}\int_{{\Lambda}'} (-{u_a\circ f_n({\lambda})}) d{\lambda}\right)}dm(a) = {\frac{1}{d_n}}\int_{{\Lambda}'} (-{\widetilde{u}\circ f_n({\lambda})})d{\lambda}\leq \frac{C }{d_n},$$ so $$m{\left({\left\{ a, \ {\frac{1}{d_n}}\int_{{\Lambda}'} {\left\vertu_a\circ f_n({\lambda})\right\vert} d{\lambda}\geq {\varepsilon}\right\}}\right)}\leq \frac{C }{{\varepsilon}d_n},$$ and the result follows from the Borel-Cantelli Lemma.
To conclude the proof of case [*i.*]{} in the theorem, we argue that if the exceptional $\mathcal{E}$ set was not pluripolar, then it would contain a non-pluripolar compact set $K$. By the work of Bedford and Taylor [@bt] there exists a Monge-Ampère measure $m=(dd^cv)^N$ supported on $K$, with $v$ bounded. It is well known that for such a measure psh functions are integrable (see [@bt] or [@demailly Prop. III.3.11]), so we are in contradiction with the previous claim.
Assume now that $\sum \exp(-td_n)$ converges for all $t$. By Lemma \[lem:bdd\] applied to $m$ the Lebesgue measure (cut-off to any compact subset of $A$) the family of negative psh functions $${\left\{a\mapsto \int_{{\Lambda}'} {u_a\circ f_n({\lambda})} d{\lambda}\right\}}_{n\geq 1}$$ is bounded in $L^1_{\rm loc}(A)$. Let $A'\Subset A''\Subset A$. It follows from an inequality due to Hörmander [@hormander Prop 4.2.9] that there exists constants $(C_0,\alpha_0)$ such that if $\varphi$ is any negative psh function on $A$ such that ${\left\Vert\varphi\right\Vert}_{L^1(A'')}\leq 1$, then $$\operatorname{vol}{\left({\left\{a\in A', \ \varphi(a)<-t\right\}}\right)} \leq C_0\exp(-\alpha_0 t).$$ It follows that there exist constants $(C,\alpha)$ independent of $n$ such that $$\operatorname{vol}{\left({\left\{ a\in A', \ {\frac{1}{d_n}}\int_{{\Lambda}'} {\left\vertu_a\circ f_n({\lambda})\right\vert} d{\lambda}>{\varepsilon}\right\}}\right)}\leq C \exp( -{\varepsilon}\alpha d_n),$$ and again the Borel-Cantelli implies that for (Lebesgue) a.e. $a$, ${\frac{1}{d_n}} u_a\circ f_n$ converges to zero in $L^1_{\rm loc}(A)$.
By definition there is an embedding $\iota:X\hookrightarrow {\mathbb{P}}^N$ and a holomorphic family $(L_a)_{a\in A}$ of hyperplanes such that $H_a=\iota{^{-1}}(L_a)$. There exists a holomorphic family of linear forms $(\ell_a)_{a\in A}$ on ${\mathbb{C}}^{N+1}$ such that ${\left\{\ell_a=0\right\}}$ is an equation of $L_a$. We normalize so that ${\left\vert\ell_a\right\vert}\leq 1$ on the unit ball. Now define $\varphi(a,\cdot)$ on ${\mathbb{P}}^N$ by $$\varphi(a,z) = \frac{\log{\left\vert\ell_a(Z)\right\vert}}{\log{\left\VertZ\right\Vert}},$$ where $Z$ is any lift of $z$ and ${\left\Vert\cdot\right\Vert}$ is the Hermitian norm. Then $\varphi$ satisfies [*i.*]{}-[*iii.*]{} relative to the family $L_a$ on ${\mathbb{P}}^N$, i.e. $dd^c_z \varphi(a, \cdot) = [L_a]-\omega_{FS}$, etc.
We now put $u(a,x) = \varphi(a, \iota(x))$ and claim that it satisfies the desired requirements (with $\omega = \iota^*\omega_{FS}$). Properties [*i.*]{} and [*iii.*]{} are immediate. If by contradiction [*ii.*]{} did not hold, then by the Hartogs Lemma [@hormander pp. 149-151] we would get a sequence $a_n{\rightarrow}a_0\in A$ such that $u_{a_n}$ diverges uniformly to $-\infty$. But if $x_0\notin H_{a_0}$ it is clear that $u$ is locally uniformly bounded near $(a_0, x_0)$, hence the contradiction.
Let $A'\Subset A$ be an open set containing $\operatorname{Supp}(m)$ According to [@ds-pl Prop. 3.9.2] there exists a constant $C$ such that for every psh function $\varphi$ on $A$, ${\left\Vert\varphi\right\Vert}_{L^1(m)}\leq C{\left\Vert\varphi\right\Vert}_{L^1(A')}$. From this we infer that for every $x$ in $X$, $\int {\left\vertu(a,x)\right\vert}dm(a)\leq C\int_{A'} {\left\vertu(a,x)\right\vert}da$, where $da$ denotes the Lebesgue measure.
Now we claim that there exists a constant $C'$ such that for any negative psh function on $A$, ${\left\Vert\varphi\right\Vert}_{L^1(A')}\leq C' {\left\Vert\varphi\right\Vert}_{L^1(\nu)}$, where $\nu$ is the measure from the definition of substantial families. Indeed, by the Hartogs lemma (see [@hormander pp.149-151]) the set $${\left\{\varphi \text{ negative psh s.t. }\int{\left\vert\varphi\right\vert}d\nu\leq1\right\}}$$ is relatively compact in $L^1_{\rm loc}$, hence bounded.
From these two facts we infer that for every $x$, $$\int {\left\vertu(a,x)\right\vert}dm(a)\leq CC'\int {\left\vertu(a,x)\right\vert} d\nu(a).$$ We now show that $x\mapsto \int u_a(x) d\nu(a)$ is uniformly bounded. For this, note first that this function is integrable, because by Lemma \[lem:uax\], ${\left\Vertu(a, \cdot)\right\Vert}_{L^1(X)}$ is locally uniformly bounded. So we can take the $dd^c$ in $x$ and we obtain that $$dd^c_x{\left(\int u_a(\cdot) d\nu(a)\right)} = \int [H_a]d\nu(a) - \omega
,$$ and we conclude by using our assumption that the local potentials of $\int [H_a]d\nu(a)$ are bounded. Finally, the same argument shows that $x\mapsto \int u_a(x) dm(a)$ is $\omega$-psh, since by uniform boundedness of ${\left\Vertu(a, \cdot)\right\Vert}_{L^1(X)}$, we can permute the $dd^c$ in $x$ and integration with respect to $m$.
We close this section by highlighting the following direct consequence of Theorem \[thm:volume\], which appears as a key step in the characterization of the supports of certain bifurcation currents.
It can also be used to obtain a coordinate-free proof of the fact that the support of the Green current of a holomorphic self-map of ${{\mathbb{P}^k}}$ coincides with the Julia set (a result originally due to Fornæss-Sibony [@fs2] and Ueda [@ueda]). More precisely, it implies that if ${\Omega}\subset {{\mathbb{P}^k}}$ is an open set disjoint from the support of the Green current $T$, then the sequence of iterates $f^n$ is normal in ${\Omega}$. Notice that the converse inclusion $\operatorname{Supp}(T)\subset {J}$ is an easy consequence of the definitions.
\[prop:support\] Let ${\Lambda}$ be a complex manifold of dimension $d$ and $X$ be a compact Kähler manifold of dimension $k$, endowed with a Kähler form $\omega$. Let $(f_n)$ be a sequence of holomorphic mappings ${\Lambda}{\rightarrow}X$, and assume that the sequence ${\frac{1}{d_n}} f_n^*\omega$ converges to a current $T$.
Assume furthermore that every test function $\varphi$ one has the following estimate $$\label{eq:mass}
{\left\langle {\frac{1}{d_n}} f_n^*\omega -T, \varphi\right\rangle} = O{\left({\frac{1}{d_n}}\right)}.$$
Then the sequence $(f_n)$ is quasi-normal outside $\operatorname{Supp}(T)$.
If $U$ is an open set disjoint where $T\equiv 0$, then implies that the sequence $(f^*_n\omega)$ has locally uniformly bounded mass on $U$. It then follows from Theorem \[thm:volume\] that $(f_n)$ is quasi-normal on $U$.
Notice that conversely, if $d_n{\rightarrow}\infty$, and $U$ is an open set such that $U\cap \operatorname{Supp}(T)\neq \emptyset$, then the sequence $(f_n)$ is [*not*]{} normal on $U$. Indeed it follows from the explicit expression of $f^*_n\omega$ in local coordinates that the $L^2$ norm of the derivative of $f_n$ tend to infinity in $U$.
Bifurcation currents for families of rational mappings on ${{\mathbb{P}^1}}$ {#sec:bifcur}
============================================================================
Generalities on bifurcations
----------------------------
Let us first review a number well-known facts on holomorphic families of rational maps. Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps $f_{\lambda}:{{\mathbb{P}^1}}{\rightarrow}{{\mathbb{P}^1}}$ of degree $d\geq 2$ parameterized by a connected complex manifold ${\Lambda}$. By definition, a [*marked critical point*]{} is a holomorphic map $c:{\Lambda}{\rightarrow}{{\mathbb{P}^1}}$ such that $f_{\lambda}'(c({\lambda}))=0$ for all ${\lambda}\in {\Lambda}$.
Given any family $(f_{\lambda})$ if $c_0$ is a given critical point at parameter ${\lambda}_0$, there exists a branched cover $\pi: \widetilde {\Lambda}{\rightarrow}{\Lambda}$ such that the family of rational mappings $(\widetilde f_{\widetilde{\lambda}})_{\widetilde {\lambda}\in \widetilde{\Lambda}}$ defined for $\widetilde {\lambda}\in
\widetilde{\Lambda}$ by $\widetilde f_{\widetilde{\lambda}} = f_{\pi(\widetilde{\lambda})}$ has a marked critical point $\widetilde c(\widetilde{\lambda})$ with $\widetilde c(\widetilde {\lambda}_0)=c_0$. Specifically, it is enough to parameterize the family by $\widetilde{{\Lambda}} =
\widehat{\mathcal C} = {\left\{({\lambda}, z)\in {\Lambda}\times {{\mathbb{P}^1}}, \ f'_{\lambda}(z)=0\right\}}$ (or its desingularization if $\widehat{\mathcal C}$ is not smooth). Then the first projection $\pi_1: \widehat{\mathcal C} {\rightarrow}{\Lambda}$ makes it a branched cover over ${\Lambda}$, and for any $\widetilde {\lambda}= ({\lambda},z)\in \widehat{\mathcal C}$, we set $\widetilde c(\widetilde{\lambda})= z$ which is a critical point for $\widetilde f_{\widetilde{\lambda}} := f_{\pi_1(\widetilde{\lambda})}$.
Therefore, taking a branched cover of ${\Lambda}$ if necessary, it is always possible to assume that all critical points are marked.
We always denote with a subscript ${\lambda}$ the dynamical objects associated to $f_{\lambda}$: Julia set, maximal entropy measure, etc.
In a celebrated paper, Mañé, Sad and Sullivan [@mss], and independently Lyubich [@lyubich-bif], showed the existence of a decomposition $${\Lambda}= \operatorname{Bif}\cup\operatorname{Stab}$$ of the parameter space ${\Lambda}$ into a (closed) bifurcation locus and a (open) stability locus, which is similar to the Fatou-Julia decomposition of dynamical space.
Stability is defined by a number of equivalent properties, according to the following theorem [@mss; @lyubich-bif].
\[thm:mss\] Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d\geq 2$ on ${{\mathbb{P}^1}}$. Let ${\Omega}\subset{\Lambda}$ be a connected open subset. The following conditions are equivalent:
i. the periodic points of $(f_{\lambda})$ do not change nature (attracting, repelling, indifferent) in $
{\Omega}$;
ii. the Julia set $J_{\lambda}$ moves under a holomorphic motion for ${\lambda}\in {\Omega}$;
iii. ${\lambda}\mapsto J_{\lambda}$ is continuous for the Hausdorff topology;
iv. for any two parameters ${\lambda},{\lambda}'$ in ${\Omega}$, $f_{\lambda}{ \arrowvert_{J_{\lambda}}}$ is conjugate to $f_{{\lambda}'}{ \arrowvert_{J_{{\lambda}'}}}$.
If in addition, the critical points ${\left\{c_1,\ldots , c_{2d-2}\right\}}$ are marked, these conditions are equivalent to
1. for any $1\leq i\leq 2d-2$ the family of meromorphic functions $(f_{\lambda}^n(c_i({\lambda})))_{n\geq 0}: {\Lambda}{\rightarrow}{{\mathbb{P}^1}}$ is normal
If these conditions are satisfied, we say that $(f_{\lambda})$ is [*$J$-stable*]{} in ${\Omega}$ (which we simply abbreviate as [*stable*]{} in this paper). The stability locus is the union of all such ${\Omega}$, and $\operatorname{Bif}$ is by definition its complement.
Another famous result is the following [@mss; @lyubich-bif].
\[thm:dense\] Let $(f_{\lambda})_{{\lambda}\in{\Lambda}}$ be as above. The stability locus $\operatorname{Stab}$ is dense in ${\Lambda}$.
If $(f_{\lambda})$ is the family of all polynomials or all rational functions of degree $d$, it is conjectured that ${\lambda}\in \operatorname{Stab}$ if and only if all critical points converge to attracting cycles (the [*hyperbolicity conjecture*]{}). More generally, the work of McMullen and Sullivan [@mcms] leads to a conjectural description of the components of the stability locus in any holomorphic family of rational maps (relying on the the so-called [*no invariant line fields conjecture*]{}).
We now explain how the bifurcation locus can be seen in a number of ways as the limit set (in the topological sense) of countable families of analytic subsets of codimension 1. This is a basic motivation for a description of the bifurcation locus in terms of positive closed currents of bidegree (1,1).
A marked critical point $c$ is said to be [*passive*]{} in ${\Omega}$ if the family $(f_{\lambda}^n(c({\lambda})))_{n\geq 0}$ is normal, and [*active*]{} at ${\lambda}_0$ if for every neighborhood $V\ni{\lambda}_0$, $c$ is not passive in $V$ (this convenient terminology is due to McMullen). The characterization [*v.*]{} of stability in Theorem \[thm:mss\] then rephrases as “a critically marked family is stable iff all critical points are passive in ${\Omega}$".
A typical example of a passive critical point is that of a critical point converging to an attracting cycle, for this property is robust under perturbations. If ${\Lambda}$ is the family of all polynomials or all rational functions of degree $d$, according to the hyperbolicity conjecture, all passive critical points should be of this type. Indeed, any component of passivity would intersect the hyperbolicity locus. On the other hand, in a family of rational mappings with a persistent parabolic point (resp. a persistent Siegel disk), a critical point attracted by this parabolic point (resp. eventually falling in this Siegel disk) is passive.
Let $(f_{\lambda})$ be the space of all polynomials or rational mappings of degree $d$ with a marked critical point $c$. It seems to be an interesting problem to study the geometry of [*hyperbolic passivity components*]{}, that is, components of the passivity locus associated to $c$, where $c$ converges to an attracting cycle. Does there exist a “center” in this component, that is a subvariety where $c$ is periodic? How does the topology of the component related to that of its center?
The following result is an easy consequence of Montel’s theorem (see e.g. [@levin1] or [@preper] for the proof).
\[thm:preperiodic dense\] Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d$ on ${{\mathbb{P}^1}}$, with a marked critical point $c$. If $c$ is active at ${\lambda}_0$ then ${\lambda}_0 = \lim_n {\lambda}_n$ where for every $n$, $c({\lambda}_n)$ is periodic (resp. falls onto a repelling cycle).
Combined with Theorem \[thm:mss\] this implies that in any (not necessarily critically marked) holomorphic family, the family of hypersurfaces, defined by the condition that a critical point is periodic (resp. preperiodic) cluster on the bifurcation locus.
Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d$ on ${{\mathbb{P}^1}}$. Then $$\operatorname{Bif}\subset\overline{{\left\{{\lambda}, \ \exists \ c({\lambda}) \text{ \rm (pre)periodic critical point} \right\}}}$$ and more precisely, $$\mathrm{Bif}=\overline{{\left\{{\lambda}, \ \exists \ c({\lambda}) \text{ \rm critical point falling non-persistently on a repelling cycle}\right\}}}.$$
At this point a natural question arises: assume that several marked critical points are active at ${\lambda}_0$. Is it possible to perturb ${\lambda}_0$ so that these critical points become simultaneously (pre)periodic? It turns out that the answer to the question is “no”, which is a manifestation of the failure of Montel’s theorem in higher dimension (see Example \[ex:douady\] below). An important idea in higher dimensional holomorphic dynamics is that the use of currents and pluripotential theory is a way to get around this difficulty. As it turns out, the theory of bifurcation currents will indeed provide a reasonable understanding of this problem.
The following simple consequence of item [*i.*]{} of Theorem \[thm:mss\], provides yet another dense codimension 1 phenomenon in the bifurcation locus.
Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d$ on ${{\mathbb{P}^1}}$. Then for every $\theta\in {\mathbb{R}}/2\pi{\mathbb{Z}}$, $$\mathrm{Bif}=\overline{{\left\{{\lambda}, \ f_{\lambda}\text{ admits a non-persistent periodic point of multiplier }e^{i\theta}\right\}}}.$$
Again one might ask: what is the set of parameters possessing *several* non-persistent indifferent periodic points?
The bifurcation current {#subs:bifcur}
-----------------------
It is a classical observation that in a holomorphic family of dynamical systems, Lyapunov exponents often depend subharmonically on parameters (this idea plays for instance a key role in [@herman]). Also, since the 1980’s, potential theoretic methods appear to play an important role in the study the quadratic family and the Mandelbrot set (see e.g. [@dh; @sibony-ucla]).
DeMarco made this idea more systematic by putting forward the following definition [@demarco2].
Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d\geq 2$. For each ${\lambda}\in {\Lambda}$, let $\mu_{\lambda}$ be the unique measure of maximal entropy of $f_{\lambda}$. Then $$\chi:{\lambda}\longmapsto \chi(f_{\lambda}) = \int \log{\left\Vert f_{\lambda}' \right\Vert}d\mu_{\lambda}$$ is a continuous psh function on ${\Lambda}$ (here the norm of the differential is relative to any Riemannian metric on ${{\mathbb{P}^1}}$).
The [*bifurcation current*]{} of the family $(f_{\lambda})$ is by definition ${T_{\mathrm{bif}}}= dd^c\chi$.
The continuity of $\chi$ was originally proven by Mañé [@mane]. Actually $\chi$ is Hölder continuous, as can easily be seen from DeMarco’s formula for $\chi$ (see below), and the joint Hölder continuity in $({\lambda},z)$ of the dynamical Green’s function (see also [@ds-survey §2.5] for another approach).
The significance of this definition is justified by the following result [@demarco1; @demarco2].
The support of ${T_{\mathrm{bif}}}$ is equal to $\operatorname{Bif}$.
The easier inclusion is the fact that $\operatorname{Supp}({T_{\mathrm{bif}}})\subset\operatorname{Bif}$, or equivalently, that $\chi$ is pluriharmonic on the stability locus. A neat way to see this is to use the following approximation formula, showing that the Lyapunov exponent of $f_{\lambda}$ can be read on periodic orbits: for every rational map $f$ of degree $d$, $$\label{eq:bdm}
\chi(f ) = \lim_{n{\rightarrow}\infty} {\frac{1}{d^n}}\sum_{p \text{ repelling of period } n} {\frac{1}{n}} \log^+{\left\vert(f^n)'(p)\right\vert}$$ (see Berteloot [@berteloot; @parma] for the proof). It thus follows from the characterization [*i.*]{} of Theorem \[thm:mss\] that if $(f_{\lambda})$ is stable on some open set ${\Omega}$, then $\chi$ is pluriharmonic there (notice that by the Hartogs Lemma the pointwise limit of a uniformly bounded sequence of pluriharmonic functions is pluriharmonic).
Let us sketch DeMarco’s argument for the converse inclusion. It is no loss of generality to assume that all critical points are marked. The main ingredient is a formula relating $\chi$ and the value of the Green’s function at critical points (see also Proposition \[prop:sum ci\] below). We do not state this formula precisily here[^3], and only note that it generalizes the well-known formula due to Przytycki [@prz] (see also Manning [@manning]) for the Lyapunov exponent of a monic polynomial $P$: $$\label{eq:lyap}
\chi(P) = \log d + \sum_{c \text{ critical}} G_P(c) .$$ Here $G_P$ denotes the dynamical Green’s function of $P$ in ${\mathbb{C}}$, defined by $$G_P(z) = \lim_{n{\rightarrow}\infty} d^{-n}\log^+{\left\vertP^n(z)\right\vert}.$$
From this one deduces that if $\chi$ is pluriharmonic in some open set ${\Omega}$, then all critical points are passive in ${\Omega}$, hence the family is stable by Theorem \[thm:mss\].
A first consequence of this result, which was a source of motivation in [@demarco1], is that if ${\Lambda}$ is a Stein manifold (e.g. an affine algebraic manifold), then the components of the stability locus are also Stein.
In the most studied quadratic family $(P_{\lambda}(z))= (z^2+{\lambda})_{{\lambda}\in {\mathbb{C}}}$, we see from that the bifurcation “current” (which is simply a measure in this case) is defined by the formula ${\mu_{\mathrm{bif}}}= dd^c G_{P_{{\lambda}}}(0)$, where $G_{P_{\lambda}}$ is the Green function. As expected, we recover the usual parameter space measure, that is the harmonic measure of the Mandelbrot set.
Marked critical points {#subs:marked}
----------------------
In this paragraph we present a construction due to Favre and the author [@preper] of a current associated to the bifurcations of a marked critical point. It would also be possible to define this current by lifting the dynamics to ${{{\mathbb{C}}^2}}\setminus {\left\{0\right\}}$ and evaluating appropriate dynamical Green’s function at the lifted critical points, in the spirit of [@demarco1]. However our construction is more instrinsic, and generalizes to other situations.
Let $(f_{\lambda},c({\lambda}))$ be a holomorphic family of rational maps of degree $d\geq 2$ with a marked critical point. Let $f_n:{\Lambda}{\rightarrow}{{\mathbb{P}^1}}$ be defined by $f_n({\lambda}) = f_{\lambda}^n(c({\lambda}))$. Let $\omega$ be a Fubini-Study form on ${{\mathbb{P}^1}}$.
In the spirit of Section \[sec:prologue\] we have the following result.
\[thm:activity support\] Let as above $(f,c)$ be a holomorphic family of rational maps of degree $d\geq 2$ with a marked critical point, and set $f_n: {\lambda}\mapsto f_{\lambda}^n(c({\lambda}))$. Then the sequence of currents $(d^{-n}f_n^*\omega)$ converges to a current $T_c$ on ${\Lambda}$. The support of $T_c$ is the activity locus of $c$.
By definition, $T_c$ is *the bifurcation current* (also referred to as the [*activity current*]{}) associated to $(f,c)$.
The convergence of the sequence of currents $(d^{-n}f_n^*\omega)$ does not follow from the general formalism of Section \[sec:prologue\]. The proof relies on equidistribution results for preimages of points under $f^n$ instead. For this, it is convenient to consider the product dynamics on ${\Lambda}\times {{\mathbb{P}^1}}$. Let ${\widehat{\Lambda}}= \Lambda \times {{\mathbb{P}^1}}$. The family $f_\lambda$ lifts to a holomorphic map ${\widehat{f}}: {\widehat{\Lambda}}\to {\widehat{\Lambda}}$ mapping $(\lambda, z) $ to $(\lambda,
f_\lambda(z))$. We denote by $\pi_1: {\widehat{\Lambda}}{\rightarrow}\Lambda$ and $\pi_2: {\widehat{\Lambda}}\to {{\mathbb{P}^1}}$ the natural projections, and let $\widehat \omega = \pi_2^*\omega$.
The following proposition follows from classical techniques in higher dimensional holomorphic dynamics.
Let $(f_{\lambda})$ be a holomorphic family of rational maps of degree $d\geq 2$, and ${\widehat{f}}$ be its lift to ${\Lambda}\times {{\mathbb{P}^1}}$ as above. Then the sequence of currents $d^{-n}{\widehat{f}}^{n*} \widehat\omega$ converges to a limit ${\widehat{T}}$ in ${\Lambda}\times {{\mathbb{P}^1}}$.
The current ${\widehat{T}}$ should be understood as “interpolating” the family of maximal measures $\mu_{\lambda}$.
Let $\widehat c = \{ (\lambda, c(\lambda)), \ {\lambda}\in {\Lambda}\}\subset {\widehat{\Lambda}}$ (resp. ${\widehat{f}}^n(\widehat c)$) be the graph of $c$ (resp. $f^n(c)$). As observed in Lemma \[lem:dim1\], $d^{-n}f_n^*\omega = (\pi_1)_* {\left(d^{-n} \widehat\omega{ \arrowvert_{{\widehat{f}}^n(\widehat c)}}\right)}$. Now ${\widehat{f}}^n$ induces a biholomorphism $\widehat c {\rightarrow}{\widehat{f}}^n(\widehat c)$, so $d^{-n} \widehat\omega{ \arrowvert_{{\widehat{f}}^n(\widehat c)}} = d^{-n}\big(({\widehat{f}}^n)^*\widehat\omega\big){ \arrowvert_{ \widehat c}}$. Thus we see that the bifurcation current $T_c$ is obtained by slicing ${\widehat{T}}$ by the hypersurface $\widehat c$ and projecting down to ${\Lambda}$: $T_c = (\pi_1)_*{\left({\widehat{T}}{ \arrowvert_{\widehat c}}\right)}$.
Making this precise actually requires a precise control on the convergence of the sequence $(d^{-n}({\widehat{f}}^{n})^* \widehat\omega)$ to ${\widehat{T}}$. This follows from the following classical computation: write $ {d}{^{-1}}{\widehat{f}}^*
{\widehat\omega}- {\widehat\omega}= dd^c g_1$. Then $ d^{-n}({\widehat{f}}^n )^* {\widehat\omega}- {\widehat\omega}= dd^c g_n$, where $g_n = \sum_{j=0}^{n-1} d^{-j} g_0\circ {\widehat{f}}^j$. Therefore $(g_n)$ converges uniformly to $g_\infty$, with ${\widehat\omega}+ dd^c g_\infty = {\widehat{T}}$.
In particular we have that ${\left\vertg_n - g_\infty\right\vert} = O(d^{-n})$, which implies that the assumption in Proposition \[prop:support\] is satisfied. Hence the family $(f^n_{\lambda}(c({\lambda}))$ is quasi-normal outside $T_c$. To see that it is actually normal, we notice that the uniform control on the potentials allows to apply Remark \[rmk:normal\].
Conversely, $\operatorname{Supp}(T_c)$ is contained in the activity locus. Indeed it follows from the explicit expression of $d^{-n}f_n^*\omega$ in local coordinates on ${\Lambda}$ that if $U\subset{\Lambda}$ is an open set intersecting $\operatorname{Supp}(T_c)$, the $L^2$ norm of the derivative of $f_n$ (relative to the spherical metric on ${{\mathbb{P}^1}}$) grows exponentially in $U$. The result follows.
Observe that the fact that $c$ is a critical point does not play any role here. We might as well associate activity/passivity loci and a bifurcation current to any holomorphically moving point $(a({\lambda}))$, and Theorem \[thm:activity support\] holds in this case (this type of considerations appear e.g. in [@baker-demarco]).
For the quadratic family $(P_{\lambda}(z)) = (z^2+{\lambda})_{{\lambda}\in {\mathbb{C}}}$, which has a marked critical point at 0, one easily checks that the current ${\widehat{T}}$ is defined in ${\mathbb{C}}\times {\mathbb{C}}$ by the formula ${\widehat{T}}= dd^c_{({\lambda},z)} G_{P_{\lambda}}(z)$, and that the bifurcation current associated to the critical point is again $dd^c_{{\lambda}} G_{\lambda}(0)={\mu_{\mathrm{bif}}}$.
More generally, DeMarco’s formula for the Lyapunov exponent of a rational map gives the relationship between these currents and the bifurcation current ${T_{\mathrm{bif}}}$ defined in §\[subs:bifcur\].
\[prop:sum ci\] Let $(f_{\lambda})$ be a family of rational maps with all critical points marked ${\left\{c_1, \ldots, c_{2d-2}\right\}}$, then ${T_{\mathrm{bif}}}=
\sum T_i$, where $T_i$ is the bifurcation current associated to $c_i$.
Equidistribution of critically preperiodic parameters
-----------------------------------------------------
We keep hypotheses as before, that is we work with a family of rational maps with a marked critical point $(f,c)$. Our goal is to show that the bifurcation current $T_c$ is the limit in the sense of currents of sequences of dynamically defined codimension 1 subvarieties.
The first result follows directly from Theorem \[thm:equidist abstrait\]. It is a quantitative version of the fact that near an activity point, $f^n_{\lambda}(c({\lambda}))$ assumes almost every value in ${{\mathbb{P}^1}}$.
\[thm:nanamarre\] Let $(f_{\lambda}, c({\lambda}))_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d\geq 2$ with a marked critical point, and $T_c$ be the associated bifurcation current. There exists a pluripolar exceptional set $\mathcal{E}\subset {{\mathbb{P}^1}}$ such that if $a\notin \mathcal{E}$, then $$\lim_{n{\rightarrow}\infty} {\frac{1}{d^n}} [H_{n,a}] {\rightarrow}T_c, \text{ where } H_{n,a}={\left\{{\lambda}, \ f^n_{\lambda}(c({\lambda})) = a\right\}}.$$
Note that $H_{n,a}$ is defined not only as a set, but as an analytic subvariety, with a possible multiplicity. It is likely that the size of the exceptional set can be estimated more precisely.
Is the exceptional set in Theorem \[thm:nanamarre\] finite, as in the case of a single mapping?
It is dynamically more significant to study the distribution of parameters for which $c$ becomes periodic (resp. preperiodic), that is, to try to make Theorem \[thm:preperiodic dense\] an equidistribution result. This is expected to be more difficult because in this case the set of targets that $f^n_{\lambda}(c({\lambda}))$ is supposed to hit (the set of periodic points, say) is both countable and moving with ${\lambda}$.
Let $e\in {\left\{0,1\right\}}$ be the cardinality of the exceptional set of $f_{\lambda}$ for generic ${\lambda}$. If $e=2$, then the family is trivial. If $e=1$, it is conjugate to a family of polynomials. Given two integers $n>m\geq 0$ denote by ${\mathrm{PerCrit}}(n,m)$ the subvariety of ${\Lambda}$ defined by the (non necessarily reduced) equation $f^n_{\lambda}(c({\lambda})) = f^m_{\lambda}(c({\lambda}))$.
It is convenient to adopt the convention that $[{\Lambda}]=0$. This means that if some subvariety $V$ like ${\mathrm{PerCrit}}(n,k)$ turns out to be equal to ${\Lambda}$, then we declare that $[V] = 0$.
The following equidistribution theorem was obtained in [@preper].
\[thm:cvg preper\] Let $(f,c)$ be a non-trivial holomorphic family of rational maps on ${{\mathbb{P}^1}}$ of degree $d\geq 2$, with a marked critical point, and denote by $e$ the generic cardinality of the exceptional set. Assume furthermore that the following technical assumption is satisfied:
- [*for every $\lambda \in \Lambda$, there exists an immersed curve $\Gamma
\subset \Lambda$ through $\lambda$ such that the complement of the set $\{ \lambda,
\, c(\lambda) \text{is attracted by a periodic cycle}\}$ is relatively compact in $\Gamma$.*]{}
Then for every sequence $0\le k(n) <
n$, we have that $$\lim_{n{\rightarrow}\infty} \frac{[{\mathrm{PerCrit}}(n,k(n))]}{d^n+d^{(1-e)k(n)}} = T_c$$
Notice that with our convention, if $c({\lambda})$ is periodic throughout the family, then both sides of the equidistribution equation vanish.
Is assumption (H) really necessary?
One might at least try to replace it by a more tractable condition like ${\Lambda}$ being an algebraic family –compare with Theorem \[thm:equidist speed\]. It is easy to see that (H) holds e.g. in the space of all polynomial or rational maps of degree $d$ (see [@preper]).
In the 1-parameter family $(z^d+{\lambda})_{{\lambda}\in{\mathbb{C}}}$ of unicritical polynomials of degree $d$, the theorem implies the equidistribution of the centers of components of the degree $d$ Mandelbrot set $$\label{eq:equidist mandelbrot}
\lim_{n{\rightarrow}\infty} {\frac{1}{d^n}} [{\mathrm{PerCrit}}(n,0)] =
\lim_{n{\rightarrow}\infty} \frac1{d^n}\sum_{f_c^n(0)=0} \delta_c = {\mu_{\mathrm{bif}}}.$$ This result had previously been proven by Levin [@levin] (see also McMullen [@mcm; @equidist]).
Another interesting approach to this type of equidistribution statements is to use arithmetic methods based on height theory (following work of Zhang, Autissier, Chambert-Loir, Thuillier, Baker-Rumely, and others). In particular a proof of along these lines was obtained, prior to [@preper], by Baker and Hsia in [@baker-hsia Theorem 8.15]).
Since the varieties ${\mathrm{PerCrit}}(n,0)$ and ${\mathrm{PerCrit}}(n,k)$ have generally many irreducible components, (e.g. ${\mathrm{PerCrit}}(n-k, 0)\subset{\mathrm{PerCrit}}(n,k)$), and since it is difficult to control multiplicities, the theorem does not directly imply that $T_c$ is approximated by parameters where $c$ is genuinely preperiodic. To ensure this, we use a little trick based on the fact that $T_c$ gives no mass to subvarieties (since it has local continuous potentials).
Denote by ${\mathrm{PreperCrit}}(n,k)\subset{\mathrm{PerCrit}}(n,k)$ be the union of irreducible components of ${\mathrm{PerCrit}}(n,k)$ (with their multiplicities) along which $c$ is strictly preperiodic at generic parameters. As sets we have that $${\mathrm{PreperCrit}}(n,k) =\overline{{\mathrm{PerCrit}}(n,k)\setminus {\mathrm{PerCrit}}(n-k,0)}.$$
\[cor:strpreper\] Under the assumptions of the theorem, if $k$ is fixed, then $${\frac{1}{d^n+d^{(1-e)(n-k)}}}[{\mathrm{PreperCrit}}({n,n-k})] {\rightarrow}T_c~.$$
$[{\mathrm{PerCrit}}({n,n-k})] - [{\mathrm{PreperCrit}}({n,n-k})]= [D_n]$ is a sequence of effective divisors supported on ${\mathrm{PerCrit}}(k, 0)$. Assume by contradiction that the conclusion of the corollary does not hold. Then $T_c$ would give positive mass to ${\mathrm{PerCrit}}(k, 0)$, which cannot happen.
Here is a heuristic geometric argument justifying the validity of Theorem \[thm:cvg preper\]. For each ${\lambda}$, (pre)periodic points equidistribute towards the maximal measure $\mu_{\lambda}$ [@lyubich]. For this one “deduces” that in ${\Lambda}\times {{\mathbb{P}^1}}$, the sequence of integration currents on the hypersurfaces $${\left\{({\lambda}, z)\in {\Lambda}\times{{\mathbb{P}^1}}, \ f^n_{\lambda}(z) = f^{k(n)}_{\lambda}(z)\right\}},$$ conveniently normalized, converge to $\widehat T$. By “restricting" this convergence to the graph $\widehat c$, one gets the desired result. The trouble here is that there is no general result showing that that the slices $\widehat{T}_n{ \arrowvert_{\widehat c}}$ converge to $\widehat{T}{ \arrowvert_{\widehat c}}$.
Assume for simplicity that $(f_{\lambda})$ is a family of polynomials of degree $d$. A psh potential of $d^{-n}{\mathrm{PerCrit}}({n,k(n)})$ is given by $d^{-n} \log{\left\vertf^n_{\lambda}(c({\lambda})) - f^{k(n)}_{\lambda}(c({\lambda}))\right\vert}$. We need to show that this sequence converges to $G_{\lambda}(c({\lambda}))$, or equivalently that $$\label{eq:cv zero}
{\frac{1}{d^n}} \log{\left\vertf^n_{\lambda}(c({\lambda})) - f^{k(n)}_{\lambda}(c({\lambda}))\right\vert} - {\frac{1}{d^n}}\log^+{\left\vertf^n_{\lambda}(c({\lambda}))\right\vert} \underset{n{\rightarrow}\infty}\longrightarrow 0 \text{ in } L^1_{\rm loc}({\Lambda})$$ converges to zero. We argue by case by case analysis depending on the behavior of $c$. For instance it is clear that holds when $c$ escapes to infinity, or is attracted to an attracting cycle. On the other hand there are parts of parameter space where the convergence is delicate to obtain directly. So instead we apply some potential-theoretic ideas (slightly reminiscent of the proof of Brolin’s theorem [@brolin]). One of these arguments is based on the maximum principle, and requires a certain compactness property leading to assumption (H).
The speed of convergence in Theorem \[thm:cvg preper\] is unknown in general. Furthermore the proof is ultimately based on compactness properties of the space of psh functions, so it is not well suited to obtain such an estimate. The only positive result in this direction is due to Favre and Rivera-Letelier [@frl], based on the method of [@baker-hsia], and concerns the unicritical family.
\[thm:frl\] Consider the unicritical family of polynomials $(z^d+{\lambda})_{{\lambda}\in {\mathbb{C}}}$. Then for any any compactly supported $C^1$ function $\varphi$, if $0\leq k(n)<n$ is any sequence, as $n{\rightarrow}\infty$ we have $${\left\vert{\left\langle {\mathrm{PerCrit}}(n,k(n)) - {T_{\mathrm{bif}}}, \varphi\right\rangle}\right\vert}\leq C {\left(\frac{n }{d^{n}}\right)}^{1/2} {\left\Vert\varphi\right\Vert}_{C^1}.$$
Equidistribution of parameters with periodic orbits of a given multiplier {#subs:multiplier}
-------------------------------------------------------------------------
Bassanelli and Berteloot studied in [@bas-ber2; @bas-ber3] the distribution of parameters for which there exists a periodic cycle of a given multiplier. For this, given any holomorphic family of rational maps $(f_{\lambda})$, we need to define the subvariety ${\mathrm{Per}}(n,w)$ of parameter space defined by the condition that $f_{\lambda}$ admits a cycle of exact period $n$ and multiplier $w$. Doing this consistently as $w$ crosses the value 1 requires a little bit of care. The following result, borrowed from [@bas-ber2], originates from the work of Milnor [@milnor; @quadratic] and Silverman [@silverman; @book].
Let $(f_{\lambda})$ be a holomorphic family of rational maps of degree $d\geq 2$. Then for every integer $n$ there exists a holomorphic function $p_n$ on ${\Lambda}\times {\mathbb{C}}$, which is polynomial on ${\mathbb{C}}$, such that
1. For any $w\in {\mathbb{C}}\setminus{\left\{1\right\}}$, $p_n({\lambda}, w)=0$ if and only if $f_{\lambda}$ admits a cycle of exact period $n$ and of multiplier $w$;
2. For $w=1$, $p_n({\lambda}, 1)=0$ if and only if $f$ admits a cycle of exact period $n$ and of multiplier 1 or a cycle of exact period $m$ whose multiplier is a primitive $r^{\rm th}$ root of unity, and $n=mr$.
We now put ${\mathrm{Per}}(n,w) = {\left\{{\lambda}, \ p_n({\lambda}, w) =0\right\}}$. The equidistribution result is the following [@bas-ber2; @bas-ber3] (recall our convention that $[{\Lambda}]=0$).
\[thm:basber\] Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d\geq 2$. Then
1. for any $w\in {\mathbb{C}}$ such that ${\left\vertw\right\vert}<1$, $\displaystyle {\frac{1}{d^n}} \left[{\mathrm{Per}}(n,w)\right]{\rightarrow}{T_{\mathrm{bif}}}$.
2. Let $d\theta$ denote the normalized Lebesgue measure on ${\mathbb{R}}/2\pi{\mathbb{Z}}$. Then for every $r>0$, $${\frac{1}{d^n}} \int_{{\mathbb{R}}/2\pi{\mathbb{Z}}} \left[{\mathrm{Per}}(n,re^{i\theta})\right] d\theta {\rightarrow}{T_{\mathrm{bif}}};$$
If moreover $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ is the family of all polynomials of degree $d$, then
i. for any $w$ such that ${\left\vertw\right\vert}\leq 1$, $\displaystyle {\frac{1}{d^n}} \left[{\mathrm{Per}}(n,w)\right]{\rightarrow}{T_{\mathrm{bif}}}$.
For fixed ${\lambda}$, the polynomial $w\mapsto p_n({\lambda},w)$ can be decomposed into $$p_n({\lambda}, w) = \prod_{i=1}^{N_d(n)} (w-w_j({\lambda})),$$ where the degree $N_d(n)$ satisfies $N_d(n)\sim \frac{d^n}{n}$ and the $w_j({\lambda})$ are the multipliers of the periodic cycles of period $n$ of $f_{\lambda}$. For simplicity, in [*i.*]{} let us only discuss the case where the multiplier $w$ equals 0. We can write $${\frac{1}{d^n}}[{\mathrm{Per}}(n,0)] = {\frac{1}{d^n}}dd^c\log{\left\vertp_n({\lambda}, 0)\right\vert} = dd^c{\left( {\frac{1}{d^n}}\sum_{i=1}^{N_d(n)} \log{\left\vert w_j({\lambda})\right\vert} \right)}.$$ We see that the potential of ${\frac{1}{d^n}}[{\mathrm{Per}}(n,0)]$ is just the average value of the logarithms of the multipliers of repelling orbits. Now for $n$ large enough, all cycles of exact period $n$ are repelling so from we see[^4] that the sequence of potentials converges pointwise to $\chi(f_{\lambda})$. Furthermore it is easy to see that this sequence is locally uniformly bounded from above so by the Hartogs lemma it converges in $L^1_{\rm loc}$. By taking $dd^c$ we see that ${\frac{1}{d^n}}[{\mathrm{Per}}(n,0)]$ converges to ${T_{\mathrm{bif}}}$.
The argument for [*ii.*]{} is similar. For simplicity assume that $r=1$. We write $$\begin{aligned}
{\frac{1}{d^n}}\int& \left[{\mathrm{Per}}(n,e^{i\theta})\right] d\theta =
{\frac{1}{d^n}} dd^c {\left( \int \log{\left\vertp_n({\lambda}, e^{i\theta})\right\vert} d\theta\right)} \\ &= dd^c{\left({\frac{1}{d^n}} \sum_{i=1}^{N_d(n)}
\int \log{\left\verte^{i\theta} -w_j({\lambda})\right\vert} d\theta \right)} = dd^c{\left( {\frac{1}{d^n}}
\sum_{i=1}^{N_d(n)} \log^+{\left\vertw_j({\lambda})\right\vert}\right)},\end{aligned}$$ where the last equality follows from the well-known formula $\log^+{\left\vertz\right\vert} = \int\log{\left\vertz- e^{i\theta}\right\vert}d\theta$. As before for each ${\lambda}$, when $n$ is large enough all points of period $n$ are repelling, so the potentials converge pointwise to $\chi$. We conclude as before.
The proof of [*iii.*]{} is more involved since for ${\left\vertw\right\vert}=1$, in the estimation of the potentials we have to deal with the possibility of cycles of large period with multipliers close to $w$. To overcome this difficulty, the authors use a global argument (somewhat in the spirit of the use of (H) in Theorem \[thm:cvg preper\]), requiring the additional assumption that $(f_{\lambda})$ is the family of polynomials.
It is a useful fact that in assertions [*i.*]{} and [*ii.*]{} of Theorem \[thm:basber\], no global assumption on ${\Lambda}$ is required (see below Theorem \[thm:tbifk per\]). On the other hand in [*iii*]{}, it is expected that the convergence of $d^{-n}[{\mathrm{Per}}(n,w)]$ to ${T_{\mathrm{bif}}}$ (even for ${\left\vertw\right\vert}>1$) holds in any family of rational mappings. Note that by using techniques similar to those of Theorem \[thm:equidist abstrait\], it can be shown that in any family of rational maps, the set of $w\in {\mathbb{C}}$ violating the convergence in [*iii*]{} is polar.
Higher bifurcation currents and the bifurcation measure {#sec:higher}
=======================================================
A crucial difference between one and higher dimensional families of rational mappings is the presence of a hierarchy of bifurcations according to the number of bifurcating critical points. These “higher bifurcation loci" are rather delicate to define precisely and the main thesis in this section is that the formalism of bifurcation currents is well suited to deal with these questions.
In this respect, let us start by suggesting a certain “dictionary” of analogies between these issues and the dynamics of holomorphic endomorphisms of projective space ${\mathbb{P}}^k$, which turns out to be a very instructive guide for the intuition. Let $f$ be a holomorphic self map of degree $d$ on ${{\mathbb{P}^k}}$. There exists a natural invariant positive closed current $T$ of bidegree (1,1) satisfying $f^*T = dT$ (the Green’s current), whose support is the Julia set of $f$ [@fs2]. In dimension 1, the dynamics of $f$ is generically expanding along the Julia set. In higher dimension, the situation is more subtle in that that “the number of directions” along which the iterates are not equicontinuous can vary from $1$ to $k$. One then introduces the following filtration of the Julia set $$J_1 =J =\operatorname{Supp}(T) \supset \cdots \supset J_q = \operatorname{Supp}(T^q) \supset \cdots \supset J_k = \operatorname{Supp}(T^k) = \operatorname{Supp}(\mu),$$ where $\mu$ is the unique measure of maximal entropy. The dynamics on $J_k$ is “repelling in all directions” according to the work of Briend and Duval [@briend-duval1]. On the other hand for $q<k$, the dynamics along $J_q\setminus J_{q+1}$ is expected to be “Fatou in codimension $q$”. It is not completely obvious how to formalise this precisely (see [@fatou] for an account). A popular way to understand this is to conjecture that $J_q\setminus J_{q+1} = \operatorname{Supp}(T^q)\setminus \operatorname{Supp}(T^{q+1})$ is filled (in a measure theoretic sense) with holomorphic disks of codimension $q$ along which the dynamics of $(f^n)$ is equicontinuous.
In this section we will try to develop a similar picture for parameter spaces of polynomial and rational maps, with deformation disks playing the role of Fatou disks, and Misiurewicz parameters replacing repelling periodic points.
For cubic polynomials with marked critical points, it is also possible to draw a rather complete dictionary with the dynamics of polynomial automorphisms of ${{{\mathbb{C}}^2}}$ (see [@cubic]).
Some general results
--------------------
In this paragraph, $(f_{\lambda})$ is a general holomorphic family of rational maps of degree $d\geq 2$. Our purpose here is to introduce the higher bifurcation currents ${T_{\mathrm{bif}}}^k$ and study some of their properties. We will try to demonstrate that their successive supports $\operatorname{Supp}({T_{\mathrm{bif}}}^k)$, $1\leq k \leq \dim({\Lambda})$ define a dynamically meaningful filtration of the bifurcation locus.
We first observe that it is harmless to assume that all critical points are marked. Indeed, let us take a branched cover $\pi:\widetilde{\Lambda}{\rightarrow}{\Lambda}$ such that the new family $\widetilde{f}(\widetilde{\lambda})$ has all critical points marked. We claim that for every $1\leq k\leq \dim({\Lambda})$, (with obvious notation) $$\pi^{-1}\operatorname{Supp}({T_{\mathrm{bif}}}^k) = \operatorname{Supp}(\widetilde T_{\rm bif}^k).$$ Indeed the Lyapunov exponent function in $\widetilde{\Lambda}$ is $\widetilde{\lambda}\mapsto \chi(\pi(\widetilde{\lambda}))$. In particular in any open subset $U\subset\widetilde{{\Lambda}}$ where $\pi$ is a biholomorphism, for every $k$, we have that $\widetilde{\lambda}\in \operatorname{Supp}(\widetilde T_{\rm bif}^k)$ iff ${\lambda}=\pi(\widetilde {\lambda})\in\operatorname{Supp}({T_{\mathrm{bif}}}^k)$. Let now $\widetilde B$ denotes the branching locus of $\pi$. Since $\chi$ is continuous, ${T_{\mathrm{bif}}}^k$ (resp. $\widetilde T_{\rm bif}^k$) gives no mass to analytic subsets, so we infer that $$\operatorname{Supp}(\widetilde T_{\rm bif}^k) = \overline{\operatorname{Supp}(\widetilde T_{\rm bif}^k)\setminus \widetilde B} \text{ (resp. }
\operatorname{Supp}( {T_{\mathrm{bif}}}^k) = \overline{\operatorname{Supp}( {T_{\mathrm{bif}}}^k)\setminus \pi(\widetilde B)} \text{ )}.$$ Thus our claim follows.
Therefore we assume that critical points are marked as $(c_1({\lambda}), \ldots, c_{2d-2}({\lambda}))$, and denote by $T_1, \ldots , T_{2d-2}$ the respective bifurcation currents. Recall that $\operatorname{Supp}(T_i)$ is the activity locus of $c_i$ and ${T_{\mathrm{bif}}}$ equals $\sum T_i$ by Proposition \[prop:sum ci\].
Since the $T_i$ are (1,1) positive closed currents with local continuous potentials, it is possible to wedge them (see [@demailly]). Here is a first observation.
For every $1\leq i\leq 2d-2$, $T_i\wedge T_i =0$.
Assume that the convergence theorem \[thm:cvg preper\] holds in ${\Lambda}$ (e.g. if (H) holds). Then $T_i = \lim_{n{\rightarrow}\infty} d^{-n} [{\mathrm{PerCrit}}_i(n,0)]$ where of course ${\mathrm{PerCrit}}_i(n,k)$ is the subvariety of parameters such that $f^n_{\lambda}(c_i({\lambda})) = f^k_{\lambda}(c_i({\lambda}))$. Since $T_i$ has continuous potentials, we infer that $T_i\wedge T_i = \lim_{n{\rightarrow}\infty} d^{-n} [{\mathrm{PerCrit}}_i(n,0)]\wedge T_i$. Now for every $n$, $[{\mathrm{PerCrit}}_i(n,0)]\wedge T_i = T_i{ \arrowvert_{{\mathrm{PerCrit}}_i(n,0)}}$ vanishes since $c_i$ is passive on ${\mathrm{PerCrit}}_i(n,0)$, and we conclude that $T_i\wedge T_i=0$.
Without assuming Theorem \[thm:cvg preper\], the proof is more involved and due to Gauthier [@gauthier].
As a consequence of this proposition, we infer that for every $1\leq k\leq \dim({\Lambda})$ $$\label{eq:tbifk}
{T_{\mathrm{bif}}}^k = k! \sum_{1\leq i_1<\cdots <i_k\leq {2d-2}} T_{i_1}\wedge \cdots \wedge T_{i_k}.$$ In particular $$\label{eq:inclusion}
\operatorname{Supp}({T_{\mathrm{bif}}}^k)\subset {\left\{{\lambda}, \ k \text{ critical points are active at }{\lambda}\right\}}.$$ As we will see in a moment (Example \[ex:douady\] below), this inclusion is in general [*not*]{} an equality. It thus becomes an interesting problem, still open in general, to characterize $\operatorname{Supp}({T_{\mathrm{bif}}}^k)$. We will try to develop the idea that $\operatorname{Supp}({T_{\mathrm{bif}}}^k)$ is the set of parameters where $k$ critical points are active and “behave independently".
The following result follows from Theorem \[thm:basber\].
\[thm:tbifk per\] Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d\geq 2$. Then for every $k\leq \dim({\Lambda})$, $$\operatorname{Supp}({T_{\mathrm{bif}}}^k)\subset \overline{{\left\{{\lambda},\ f_{\lambda}\text{ admits }k\text{ periodic critical points}\right\}}}.$$
We argue by decreasing induction on $k$, by using the following principle: if $V\subset{\Lambda}$ is a smooth analytic hypersurface and $T$ a positive closed (1,1) current with continuous potential, then $T^k\wedge [V]= (T{ \arrowvert_{V}})^k$. Here as usual the current $T{ \arrowvert_{V}}$ is defined by restricting the potential of $T$ to $V$ and taking $dd^c$.
Under the assumptions of the theorem, let ${\lambda}_0\in \operatorname{Supp}({T_{\mathrm{bif}}}^k)$. Since $${T_{\mathrm{bif}}}= \lim_{n{\rightarrow}\infty} {\frac{1}{d^n}}[{\mathrm{Per}}(n,0)]$$ and ${T_{\mathrm{bif}}}$ has continuous potential, we infer that $${T_{\mathrm{bif}}}^k = \lim_{n{\rightarrow}\infty} {\frac{1}{d^n}} [{\mathrm{Per}}(n,0)]\wedge {T_{\mathrm{bif}}}^{k-1}.$$ In particular, ${\lambda}_0$ is approximated by parameters belonging to ${\left({T_{\mathrm{bif}}}{ \arrowvert_{{\mathrm{Per}}(n,0)}}\right)}^{k-1}$ (moving slightly if necessary we can always assume that these belong to the smooth part of ${\mathrm{Per}}(n,0)$).
We can now put ${\Lambda}_1= {\mathrm{Per}}(n,0)$ and repeat the argument to find a nearby parameter belonging to $\operatorname{Supp}[{\mathrm{Per}}({n_1},0)]\wedge {T_{\mathrm{bif}}}^{k-2}\subset {\Lambda}_1$ for some (possibly much larger) $n_1$, etc.
We see that it is important in this argument that no special assumption on ${\Lambda}$ is needed in Theorem \[thm:basber\]. For instance if one were to replace “periodic" by “strictly preperiodic" in this theorem, and try to use Theorem \[thm:cvg preper\], one would have to check the validity of (H) in the restricted submanifolds, which needn’t be satisfied (see however Theorem \[thm:tbifk preper\] below).
We can now explain how the inclusion in can be strict.
\[ex:douady\] In the two dimensional space of cubic polynomials with marked critical points, let $P_0(z) = z+\frac{z^2}{2} + z^3$. We claim that the two critical points are active at $P_0$, but $P_0$ does not belong to $\operatorname{Supp}({T_{\mathrm{bif}}}^2)$.
Indeed since $P_0$ is real and the critical points are not real, by symmetry, both critical points are attracted to the parabolic fixed point at the origin. Since the fixed point can be perturbed to become repelling (hence does not attract any critical point), Theorem \[thm:mss\][*v.*]{} implies that at least one critical point must be active. Hence by symmetry again, both critical points are active. On the other hand it can be proven (see [@preper Example 6.13] for details) that any nearby parameter admits an attracting (and not superattracting) fixed point. In particular $P_0$ cannot be perturbed to make both critical points periodic, therefore $P_0\notin \operatorname{Supp}({T_{\mathrm{bif}}}^2)$.
Denote by $(P_{{\lambda}})_{{\lambda}\in {\Lambda}\simeq{{{\mathbb{C}}^2}}}$ the family of cubic polynomials with marked critical points $c_1$ and $c_2$. We see that if $N$ is a small neighborhood of the above parameter 0, the values of $(P_{{\lambda}}^n(c_1({\lambda})), P_{{\lambda}}^n(c_2({\lambda})))$ for ${\lambda}\in N$ avoid an open set in ${{{\mathbb{C}}^2}}$. Indeed for ${\lambda}\in N$, either $c_1$ or $c_2$ must be attracted by an attracting cycle, so for large $n$, $P_{{\lambda}}^n(c_1({\lambda}))$ and $P_{{\lambda}}^n(c_2({\lambda}))$ cannot be simultaneously large. This is a manifestation of the Fatou-Bieberbach phenomenon (failure of Montel’s theorem in higher dimension).
One can also approximate $\operatorname{Supp}({T_{\mathrm{bif}}}^k)$ by parameters possessing $k$ indifferent periodic cycles. For $N_k=(n_1, \ldots n_k)\in {\mathbb{N}}^k$ and $\Theta_k=(\theta_1,\cdots ,\theta_k) \in ({\mathbb{R}}/{\mathbb{Z}})^k$, we denote by ${\mathrm{Per}}_k(N_k, e^{i\Theta_k})$ the union of codimension $k$ irreducible components of $${\mathrm{Per}}(n_1, e^{i\theta_1})\cap {\mathrm{Per}}(n_2, e^{i\theta_2})\cap \cdots \cap{\mathrm{Per}}(n_k, e^{i\theta_k}),$$ and if $E\subset {\mathbb{R}}/{\mathbb{Z}}$, we let $$\mathcal{Z}_k(E) = \bigcup_{N_k\in {\mathbb{N}}^k, \Theta_k\in E^k} {\mathrm{Per}}_k(N_k, e^{i\Theta_k}),$$ which is the (codimension $p$ part of the) set of parameters possessing $k$ neutral cycles with respective multipliers in $E$.
The following result is due to Bassanelli and Berteloot [@bas-ber1].
\[thm:indifferent\] Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d\geq 2$. If $E\subset {\mathbb{R}}/{\mathbb{Z}}$ is any dense subset, then for every $k\leq \dim({\Lambda})$, $$\operatorname{Supp}({T_{\mathrm{bif}}}^k)\subset \overline{\mathcal{Z}_k(E)}.$$
We argue by induction on $k$. For $k=1$, the result follows from Theorem \[thm:mss\]. The main idea of the induction step is as follows: assume ${\lambda}_0\in \operatorname{Supp}({T_{\mathrm{bif}}}^k)$. Then by the induction hypothesis, ${\lambda}\in \overline{\mathcal{Z}_{k-1}(E)}$, so there are plenty of $(k-1)$-dimensional disks near ${\lambda}_0$ along which $f_{\lambda}$ possesses $(k-1)$ neutral cycles with multipliers in $E$. Assume that the dynamics is $J$-stable along these disks. Then the Lyapunov exponent function $\chi$ is pluriharmonic there and it follows from a general pluripotential theoretic lemma (see [@fs2 Lemma 6.10]) that $(dd^c\chi)^k=0$ in the neighborhood of ${\lambda}_0$, a contradiction. So the dynamics is not $J$-stable along the disks of $\mathcal{Z}_{k-1}(E)$, hence Theorem \[thm:mss\] produces one more neutral cycle, thereby proving the result.
We see that in Example \[ex:douady\], the two critical points are related in a rather subtle way. On the other hand, under a certain transversality assumption, a clever argument of similarity between parameter and dynamical spaces, due to Buff and Epstein, shows that certain parameters belong to $\operatorname{Supp}({T_{\mathrm{bif}}}^k)$ [@buff; @epstein].
Let ${\lambda}_0\in {\Lambda}$ be a parameter where $c_1({\lambda}_0), \ldots , c_k({\lambda}_0)$ fall onto repelling cycles. More precisely we assume that there exist repelling periodic points $p_1({\lambda}_0), \ldots p_k({\lambda}_0)$ and integers $n_1,\ldots, n_k$ such that for $1\leq j\leq k$, $f^{n_j}_{{\lambda}_0}(c_j({\lambda}_0)) = p_j({\lambda}_0)$. The repelling orbits $p_j({\lambda}_0)$ can be uniquely continued to repelling periodic orbits $p_j({\lambda})$ for ${\lambda}$ in some neighborhood of ${\lambda}_0$. Fix for each $j$ a coordinate chart on ${{\mathbb{P}^1}}$ containing $p_j({\lambda}_0)$, so that for nearby ${\lambda}$, the function $\chi_j : {\lambda}\mapsto f^{n_j}_{{\lambda}_0}(c_j({\lambda})) - p_j({\lambda})$ is well defined. We say that the critical points $c_j({\lambda}_0)$ [*fall transversely*]{} onto the respective repelling points $p_j({\lambda}_0)$ if the mapping $\chi: {\Lambda}{\rightarrow}{\mathbb{C}}^k$ defined in the neighborhood of ${\lambda}_0$ by $\chi =(\chi_1, \ldots , \chi_k)$ has rank $k$ at ${\lambda}_0$. Of course this notion does not depend on the choice of coordinate charts.
\[thm:buff epstein\] Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d\geq 2$ with marked critical points $c_1,\ldots, c_k$, and associated bifurcation currents $T_1,\ldots , T_k$. Let ${\lambda}_0$ be a parameter at which $c_1({\lambda}_0), \ldots , c_k({\lambda}_0)$ fall transversely onto repelling cycles. Then ${\lambda}_0\in \operatorname{Supp}(T_1\wedge \cdots \wedge T_k)$.
Notice that this theorem does not appear in this form in [@buff; @epstein]. Our presentation borrows from [@berteloot; @survey]. The validity of the transversality assumption will be discussed in various situations in §\[subs:poly\] and \[subs:rat\] below.
First, a slicing argument shows that it is enough to prove that for a generic $k$-dimensional subspace ${\Lambda}'\ni{\lambda}_0$ (relative to some coordinate chart in ${\Lambda}$), the result holds by restricting to ${\Lambda}'$. Thus we can assume that $\dim({\Lambda})=k$ and that $\chi$ is a local biholomorphism on ${\Lambda}$. To simplify notation, we will assume that the $f_j$ are polynomials, $k=2$, $n_j=1$ and that the $p_j$ are fixed points.
Taking adapted coordinates $({\lambda}_1, {\lambda}_2)$ in ${\Lambda}$ (in which the initial parameter ${\lambda}_0$ is 0) we can assume that $\chi_1({\lambda}) = {\lambda}_1+ {h.o.t.}$ and $\chi_2({\lambda}) = {\lambda}_2+ {h.o.t.}$ The proof, based on a renormalization argument, consists in estimating the mass, relative to the measure $T_1\wedge T_2$, of small bidisks about 0 of carefully chosed size. Specifically, we will show that $$\liminf d^n (T_1\wedge T_2) {\left(D {\left(0, \frac{\delta}{m_1^n} \right)}\times D {\left(0, \frac{\delta}{m_2^n} \right)}\right)} >0,$$ where the $m_j$ are the respective multipliers of the $p_j$. Let $\delta_n$ be the scaling map defined by $$\delta_n({\lambda}) =\left(\frac{{\lambda}_1}{m_1^n}, \frac{{\lambda}_2}{m_2^n}\right).$$ An easy computation based on transversality and the fact that $f_{\lambda}$ is linearizable near $p_j({\lambda})$ shows that for $j=1,2$, $f_{\delta_n({\lambda})}^{n+1}(c_j(\delta_n({\lambda})))$ converges as $n{\rightarrow}\infty$ to a non-constant map $\psi_j$ [*depending only on*]{} ${\lambda}_j$, on some disk $D(0, \delta)$, with $\psi_j(0) =p_j$. Hence, since $G_{\lambda}$ depends continuously on ${\lambda}$ we get that $G_{\delta_n({\lambda})}(f_{\delta_n({\lambda})}^{n+1}(c_j(\delta_n({\lambda}))))$ converges to $G_0\circ \psi_j({\lambda}_j)$. Using the invariance relation for the Green’s function, we conclude that $$\label{eq:delta n}
d^{n+1} G_{\delta_n({\lambda})}( c_j(\delta_n({\lambda}))) =
G_{\delta_n({\lambda})}\left(f_{\delta_n({\lambda})}^{n+1}(c_j(\delta_n({\lambda})))\right) \underset{n{\rightarrow}\infty}\longrightarrow
G_0(\psi_j({\lambda}_j)),$$ hence $$\begin{aligned}
&(T_1\wedge T_2)(\delta_n (D(0, \delta)^2)) \simeq d^{-2(n+1)} \int_{D(0, \delta)^2} dd^c\left(G_0\circ \psi_1 ({\lambda}_1)\right)\wedge dd^c \left(G_0\circ \psi_2({\lambda}_2)\right) \\ &= d^{-2(n+1)}
{\left(\int_{D(0, \delta)} dd^c\left(G_0\circ \psi_1 ({\lambda}_1)\right) \right)}{\left(\int_{D(0, \delta)} dd^c\left(G_0\circ \psi_2 ({\lambda}_2)\right) \right)} \text{ by Fubini's theorem}.\end{aligned}$$ Note that the first line of this equation is justified by the local uniform convergence in . Finally, the integrals on the second line are positive since $G_0$ is not harmonic near $p_j$, so $G_0\circ \psi_j$ is not harmonic near the origin.
Building on similar ideas, Gauthier [@gauthier] relaxed the transversality assumption in Theorem \[thm:buff epstein\] as follows. Assume as before that ${\lambda}_0\in {\Lambda}$ is a parameter where $c_1({\lambda}_0), \ldots , c_k({\lambda}_0)$ fall onto respective repelling periodic points $p_1({\lambda}_0), \ldots p_k({\lambda}_0)$. Define $\chi:{\Lambda}{\rightarrow}{\mathbb{C}}^k$ as before Theorem \[thm:buff epstein\]. We say that the critical points $c_j({\lambda}_0)$ [*fall properly*]{} onto the respective repelling points $p_j({\lambda}_0)$ if $\chi^{-1}(0)$ has codimension $k$ at ${\lambda}_0$. To say it differently, we are requesting that in ${\Lambda}\times ({{\mathbb{P}^1}})^k$ the graphs of the two mappings ${\lambda}\mapsto (p_1({\lambda}), \ldots p_k({\lambda}))$ and ${\lambda}\mapsto (f^{n_1}_{\lambda}(c_1({\lambda})), \ldots ,
f^{n_k}_{\lambda}(c_k({\lambda})))$ intersect properly[^5] at $({\lambda}_0, p_1({\lambda}_0), \ldots p_k({\lambda}_0))$.
\[thm:gauthier proper\] Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps of degree $d\geq 2$ with marked critical points $c_1,\ldots, c_k$, and associated bifurcation currents $T_1,\ldots , T_k$. Let ${\lambda}_0$ be a parameter at which $c_1({\lambda}_0), \ldots , c_k({\lambda}_0)$ fall properly onto repelling cycles. Then ${\lambda}_0\in \operatorname{Supp}(T_1\wedge \cdots \wedge T_k)$.
Actually, it is enough that $c_1({\lambda}_0), \ldots , c_k({\lambda}_0)$ fall properly into an arbitrary hyperbolic set. We refer to [@gauthier] for details.
The results in this section show that for $1\leq k\leq \dim({\Lambda})$, $\operatorname{Supp}({T_{\mathrm{bif}}}^k)$ is a reasonable candidate for the locus of “bifurcations of order $k$". We will see in the next sections that when ${\Lambda}$ is the space of all polynomials or rational maps and $k$ is maximal, $\operatorname{Supp}({T_{\mathrm{bif}}}^k)$ can be characterized precisely. The picture is not yet complete in intermediate codimensions. In this respect let us state a few open questions.
\[q:tbifk\] Let $(f_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of rational maps on ${{\mathbb{P}^1}}$ of degree $d\geq 2$, with marked critical points $c_1, \ldots, c_k$.
1. Is it true that $$\operatorname{Supp}(T_1\wedge \cdots \wedge T_k) = \overline{{\left\{{\lambda}_0, \ c_1({\lambda}_0), \cdots ,c_k({\lambda}_0) \text{ fall transversely onto repelling cycles}\right\}}}$$ (by Theorem \[thm:buff epstein\] only the inclusion $\subset$ needs to be established).
2. More generally, do the codimension $k$ subvarieties $${\mathrm{PerCrit}}_1(n, k(n))\cap \cdots \cap {\mathrm{PerCrit}}_k(n, k(n))$$ equidistribute towards $T_1\wedge \cdots \wedge T_k$? Arithmetic methods could help here, especially when ${\Lambda}$ is the space of polynomials and $k$ is maximal (see the next paragraph).
3. Is the following characterization of $\operatorname{Supp}(T_1\wedge \cdots \wedge T_k)$ true: ${\lambda}_0\in \operatorname{Supp}(T_1\wedge \cdots \wedge T_k)$ iff for every neighborhood $U$ of ${\lambda}_0$, there exists a pluripolar subset $\mathcal{E}\subset({{\mathbb{P}^1}})^k$ such that the values of $f^n_{\lambda}(c_j({\lambda}))$ for $n\in {\mathbb{N}}$ and ${\lambda}\in U$ cover $ ({{\mathbb{P}^1}})^k\setminus \mathcal{E}$?
Work in progress of the author indicates that the answer to Question \[q:tbifk\].[1.]{} should be “yes”.
The space of polynomials {#subs:poly}
------------------------
In this paragraph we specialize the discussion to the case where ${\Lambda}$ is the space of polynomials of degree $d$, and will mostly concentrate on the maximal exterior power of the bifurcation current (the [*bifurcation measure*]{}). Our purpose is to show that it is in many respected the right analogue in higher degree of the harmonic measure of the Mandelbrot set. All results except Theorem \[thm:HD polynomial\] come from [@preper].
The space $\mathcal{P}_d$ of polynomials of degree $d$ with marked critical points is a singular affine algebraic variety. To work on this space, in practice we use an “orbifold parameterization" (not injective) $\pi:{\mathbb{C}}^{d-1}{\rightarrow}\mathcal{P}
_d$, defined as follows: $\pi$ maps $(c_1, \cdots, c_{d-2},a)
\in\mathbb{C}^{d-1}$ to the primitive of $z\prod_1^{d-2} (z-c_i)$ whose value at $0$ is $a^d$. In coordinates, denoting $c=(c_1, \ldots , c_{d-2})$, we get $$\label{e:111}
\pi(c,a) = {P}_{c,a} (z)
= \frac1d\, z^d+\sum_{j=2}^{d-1}(-1)^{d-j}\, \sigma_{d-j}(c)\,
\frac{z^j}{j} + a^d~,$$ where $\sigma_i(c)$ is the elementary symmetric polynomial in the $\{c_j\}_1^{d-2}$ of degree $i$. The critical points of $P_{c,a}$ are $\{0, c_1, \cdots , c_{d-2}\}$. We put $c_0 = 0$.
The choice of this parameterization (inspired from that used by Branner and Hubbard in [@branner-hubbard1]) is motivated by the fact that the bifurcation currents $T_i$ associated to the $c_i$ have the same projective mass. Furthermore, it is well suited in order to understand the behavior at infinity of certain parameter space subsets (this will be used to check condition (H) of Theorem \[thm:cvg preper\]).
We let $\mathcal{C}$ be the connectedness locus, which is compact in ${\mathbb{C}}^{d-1}$ by [@branner-hubbard1]. For $1\leq i\leq d-2$ we also define the closed subsets $\mathcal{C}_i$ by $$\mathcal{C}_i
= {\left\{(c,a), \ c_i\text{ has bounded orbit}\right\}}.$$ It is clear that $\mathcal{C} = \bigcap_i\mathcal
{C}_i$ and that ${\partial}\mathcal{C}_i$ is the activity locus of $c_i$.
We also let $g_{c,a}$ be the Green’s function of the polynomial $P_{c,a}$. Then $T_i = dd^cg_i$, where $g_i = g_{c,a}(c_i)$. Recall that the Manning-Przytycki formula asserts that $\chi = \log d+ \sum_{i=0}^{d-2} g_i$, hence ${T_{\mathrm{bif}}}= \sum_{i=0}^{d-2} T_i$.
In this specific situation we are able to solve the problem raised after Theorem \[thm:tbifk per\].
\[thm:tbifk preper\] In $\mathcal{P}_d$, for every $1\leq k\leq d-1$, $$\operatorname{Supp}({T_{\mathrm{bif}}}^k)\subset \overline{{\left\{{\lambda},\ f_{\lambda}\text{ admits }k\text{ strictly
preperiodic critical points}\right\}}}.$$
More precisely, for any collection of integers $i_1, \cdots , i_k \in \{ 0,
\cdots , d-2\}$, the analytic subset $W_{n_1, \ldots, n_k} =
\bigcap_{j=1}^k{\mathrm{PreperCrit}}_{i_j}(n_j,n_j-1)$ is of pure codimension $k$ and $$\label{eq:induc}
\lim_{n_k {\rightarrow}\infty} \cdots \lim_{n_1{\rightarrow}\infty} \,
\frac1{d^{n_k+ \cdots +n_1 }}\, [W_{n_1,
\ldots, n_k}] =T_{i_1} \wedge \cdots \wedge
T_{i_k}~.$$
Of course in we can replace ${\mathrm{PreperCrit}}(n,n-1)$ by ${\mathrm{PerCrit}}(n,k(n))$ for any sequence $k(n)$. As already observed, to prove this theorem one cannot simply take wedge products in Corollary \[cor:strpreper\] (resp. Theorem \[thm:cvg preper\]). The proof goes by induction on $\ell\leq k$, by successively applying this corollary to the parameter space ${\Lambda}= W_{n_1, \cdots, n_\ell}$. It is not obvious to check that assumption (H) is satisfied, and at this point the particular choice of the parameterization is useful (see also [@berteloot; @survey] for neat computations). It is unknown whether in one can take $n_1= \cdots = n_k=n$ (see Question \[q:tbifk\].2.)
Let us now focus on the maximal codimension case $k = d-1$. We set $\mu_{\rm bif} =
T_0\wedge \cdots \wedge T_{d-2} = {\frac{1}{(d-1)!}}{T_{\mathrm{bif}}}^{d-1}$, which is a probability measure supported on the boundary of the connectedness locus.
\[prop:equilibrium\] The bifurcation measure is the pluripotential equilibrium measure of the connectedness locus. In particular $\operatorname{Supp}(\mu_{\rm bif})$ is the Shiloff boundary of $\mathcal{C}$.
It is important to understand that when $d\geq 3$, $\operatorname{Supp}({\mu_{\mathrm{bif}}})$ is a proper subset of ${\partial}{\mathcal{C}}$. To get a (crude but instructive!) mental picture of the situation, think about the boundary of a polydisk in ${\mathbb{C}}^{d-1}$. This boundary can be decomposed into foliated pieces of varying dimension between $1$ and $d-2$, together with the unit torus, the unit torus $\mathbb{T}^{d-1}$ being the Shiloff boundary. The structure of ${\partial}{\mathcal{C}}$ should be somehow similar to this, with foliated pieces of dimension $j$ corresponding to parts of ${\partial}{\mathcal{C}}$ where $j$ critical points are passive. The precise picture is far from being understood, except for cubic polynomials (see below §\[subs:laminarity\]).
It is a well-known open question whether in higher dimension the connectedness locus is the closure of its interior. Theorem \[thm:tbifk per\] provides a partial answer to this question: if ${\lambda}_0\in {\partial}{\mathcal{C}}\cap \operatorname{Supp}({T_{\mathrm{bif}}}^j)$ and $j$ critical points are active at ${\lambda}_0$, then ${\lambda}_0\in \overline{\operatorname{Int}({\mathcal{C}})}$. Indeed, there exists a neighborhood $U\ni{\lambda}_0$ where $d-1-j$ critical points are passive, hence persistently do not escape, and by Theorem \[thm:tbifk per\] there is a sequence of parameters ${\lambda}_n{\rightarrow}{\lambda}_0$ for which the $j$ remaining critical points are periodic. Hence ${\lambda}_n\in \operatorname{Int}({\mathcal{C}})$ and we are done.
We see that the answer to the problem lies in the set of parameters in ${\partial}{\mathcal{C}}\cap \operatorname{Supp}({T_{\mathrm{bif}}}^j)$ with more than $j$ active critical points (like in Example \[ex:douady\]). So far there does not seem to be any reasonable (even conjectural) understanding of the structure of this set of parameters.
We can give a satisfactory dynamical characterization of $\operatorname{Supp}({\mu_{\mathrm{bif}}})$. A polynomial is said to be [*Misiurewicz*]{} if all critical points fall onto repelling cycles.
\[thm:mis\] $\operatorname{Supp}(\mu_{\rm bif})$ is the closure of the set of Misiurewicz parameters.
The fact that $\operatorname{Supp}({\mu_{\mathrm{bif}}})$ is contained in the closure of Misiurewicz polynomials follows from Theorem \[thm:tbifk preper\], since a strictly postcritically finite polynomial is automatically Misiurewicz. So the point here is to prove the converse. There are actually several proofs of this. The original one in [@preper] uses landing of external rays (see below). Another proof goes by observing that the properness assumption of Theorem \[thm:gauthier proper\] is satisfied at every Misiurewicz parameter. Indeed, as already observed in Theorem \[thm:tbifk preper\], for every $(n_j)_{j=1, \ldots , d-1}$ and $(n_j)_{j=1, \ldots , d-1}$ with $k_j<n_j$, $\bigcap_{j=1}^{d-1} {\mathrm{PreperCrit}}_j(n_j, k_j)$ is of dimension 0. Indeed, otherwise, this intersection would contain an analytic set contained in the connectedness locus, contradicting its compactness. Therefore, at a Misiurewicz point, critical points fall properly on the corresponding repelling points, and Theorem \[thm:gauthier proper\] applies.
Notice that the work of Buff and Epstein [@buff; @epstein] implies that these intersections are actually transverse.
Theorem \[thm:gauthier proper\] in its general form [@gauthier] shows that one can alternately characterize the support of ${\mu_{\mathrm{bif}}}$ as being the closure of [*generalized Misiurewicz*]{} polynomials, where generalized here means that critical points fall into a hyperbolic set disjoint from the critical set. These considerations lead to a generalization in higher degree the well-known theorem of Shishikura on the Hausdorff dimension of the boundary of the Mandelbrot set [@shishikura]. Notice that Tan Lei [@tan; @lei] extended Shishikura’s theorem by showing that the bifurcation locus has maximal Hausdorff dimension in any family of rational maps.
\[thm:HD polynomial\] If $U$ is any open set such that $U\cap \operatorname{Supp}({\mu_{\mathrm{bif}}})\neq \emptyset$, then $\operatorname{Supp}({\mu_{\mathrm{bif}}})\cap U$ has maximal Hausdorff dimension $2(d-1)$.
We now discuss external rays, following [@preper]. Let $P\in \mathcal{P}_d$ be a polynomial for which the Green’s function takes the same value $r>0$ at all critical points ($J(P)$ is then a Cantor set). The set $\Theta$ of external arguments of external rays landing at the critical points enables to describe in a natural fashion the combinatorics of $P$. This set of angles is known as the [*critical portrait*]{} of $P$. Now we can deform $P$ by keeping $\Theta$ constant and let $r$ vary in $\mathbb{R}_+^*$ (this is the “stretching" operation of [@branner-hubbard1]). This defines a ray in parameter space corresponding to the critical portrait $\Theta$.
The set $\mathrm{Cb}$ of combinatorics/critical portraits is endowed with a natural Lebesgue measure $
\mu_{\mathrm{Cb}}$ coming from ${\mathbb{R}}/{\mathbb{Z}}$. It is easy to show that $\mu_
{\mathrm{Cb}}$-a.e. ray lands as $r{\rightarrow}0$ (this follows from Fatou’s theorem on the existence of radial limits of bounded holomorphic functions). We thus obtain a measurable landing map $e: \mathrm{Cb} {\rightarrow}\mathcal{C}$. The measures $\mu_{\rm Cb}$ and $\mu_{\rm bif}$ are related as follows:
\[thm:landing\] ${\mu_{\mathrm{bif}}}$ is the landing measure, i.e. $e_*\mu_{\mathrm{Cb}} = \mu_{\rm bif}$.
The proof of Theorem \[thm:mis\] given in [@preper] relies on a more precise landing theorem for “Misiurewicz combinatorics”. A critical portrait $\Theta\in \mathrm{Cb}$ is said to be of Misiurewicz type if the external angles it contains are strictly preperiodic under multiplication by $d$. A combination of results due to Bielefeld, Fisher and Hubbard [@bfh] and Kiwi [@kiwi-portrait] asserts that the landing map $e$ is continuous at Misiurewicz combinatorics and that the landing point is a Misiurewicz point. It then follows from Theorem \[thm:landing\] that Misiurewicz points belong to $\operatorname{Supp}({\mu_{\mathrm{bif}}})$.
The description of $\mu_{\rm bif}$ in terms of external rays allows to generalize to higher dimensions a result of Graczyk-Światek [@gs] and Smirnov [@smirnov].
\[thm:ce\] The Topological Collet-Eckmann property holds for $\mu_{\rm
bif}$-almost every polynomial $P$.
In particular for a $\mu_{\rm bif}$-a.e. $P$, we have that:
all cycles are repelling;
the orbit of each critical point is dense in the Julia set;
$K_P=J_P$ is locally connected and its Hausdorff dimension is smaller than 2.
Notice that Gauthier [@gauthier] shows that for a topologically generic polynomial $P\in \operatorname{Supp}({\mu_{\mathrm{bif}}})$, $\mathrm{HD}(J_P)=2$. So the topologically and metrically generic pictures differ.
The connectedness of $\mathrm{Cb}$ naturally suggests the following generalization of the connectedness of the boundary of the Mandelbrot set.
Is $\operatorname{Supp}({\mu_{\mathrm{bif}}})$ connected?
The space of rational maps {#subs:rat}
--------------------------
The space $\mathrm{Rat}_d$ of rational maps of degree $d$ is a smooth complex manifold of dimension $2d+1$, actually a Zariski open set of $\mathbb{P}^{2d+1}$. For $d=2$ it is isomorphic to ${{{\mathbb{C}}^2}}$ [@milnor; @quadratic], and it was recently proven to be rational for all $d$ by Levy [@levy]. Automorphisms of ${{\mathbb{P}^1}}$ act by conjugation of rational maps, and the moduli space $\mathcal{M}_d$ is the quotient $\mathrm{Rat}_d/{\mathrm{PSL}(2,\mathbb C)}$. Even if the action is not free it can be shown that $\mathcal{M}_d$ is a normal quasiprojective variety of dimension $2(d-1)$ [@silverman; @book]. Exactly as we did for polynomials, to work on this space, it is usually convenient to work on a smooth family ${\Lambda}$ which is transverse to the fibers of the projection $\mathrm{Rat}_d {\rightarrow}\mathcal{M}_d$. Through every point of $\mathrm{Rat}_d$ there exists such a family. Working in such a family, we can consider elements of $\mathcal{M}_d$ as rational maps rather than conjugacy classes.
In this paragraph we briefly give some properties of the bifurcation measure ${\mu_{\mathrm{bif}}}= {T_{\mathrm{bif}}}^{2(d-1)}$ on $\mathcal{M}_d$. The first basic result was obtained in [@bas-ber1].
The bifurcation measure has positive and finite mass on $\mathcal{M}_d$.
These authors also give a nice argument showing that all isolated Lattès examples belong to $\operatorname{Supp}({\mu_{\mathrm{bif}}})$: it is known that a rational map is a Lattès example if and only if its Lyapunov exponent is minimal, that is, equal to $\log d/2$ [@ledrappier; @cras; @zdunik]. On the other hand if $u$ is a continuous psh function on a complex manifold of dimension $k$ with a local minimum at $x_0$, then $x_0\in \operatorname{Supp}((dd^cu)^k)$ (this follows from the so-called comparison principle [@bt]). The result follows.
The precise characterization of $\operatorname{Supp}({\mu_{\mathrm{bif}}})$ is due to Buff and Epstein [@buff; @epstein] and Buff and Gauthier [@buff; @gauthier]. Let
$\mathrm{SPCF}_d$ be the set of (conjugacy classes of) strictly post-critically finite rational maps;
$\mathcal{Z}_d$ be the set of rational maps possessing $2d-2$ indifferent cycles, without counting multiplicities (this is the maximal possible number).
In $\mathcal{M}_d$, $\operatorname{Supp}({\mu_{\mathrm{bif}}}) = \overline{\mathrm{SPCF}_d} = \overline{\mathcal{Z}_d}$.
It is no loss of generality to assume that all critical points are marked as ${\left\{c_1,\ldots , c_{2d-2}\right\}}$. Let $\mathrm{SPCF}^*_d$ be the set of (conjugacy classes of) strictly post-critically finite rational maps, which are not flexible Lattès examples[^6]. We already know from Theorem \[thm:indifferent\] that $\operatorname{Supp}({\mu_{\mathrm{bif}}})\subset \overline{\mathcal{Z}_d}$. It should be possible to prove that $\operatorname{Supp}({\mu_{\mathrm{bif}}})\subset \overline{\mathrm{SPCF}_d}$ in the spirit of Theorem \[thm:tbifk preper\], but again it is not easy to check assumption (H). Instead Buff and Epstein prove directly that $\overline{\mathrm{SPCF}^*_d} = \overline{\mathcal{Z}_d}$ using the Mañé-Sad-Sullivan theorem.
Let us show that $\mathrm{SPCF}^*_d \subset\operatorname{Supp}({\mu_{\mathrm{bif}}})$. If $f\in \mathrm{SPCF}^*_d$, it satisfies a relation of the form $f^{n_j+p_j}(c_j) = f^{n_j}(c_j)$, $j=1,\ldots ,2d-2$. Consider the subvariety of $\mathcal{M}_d$ defined by these algebraic equations. We claim that this subvariety is of dimension 0 at $f$. Indeed a theorem of McMullen [@mcm; @algo] asserts that any stable algebraic family of rational maps is a family of flexible Lattès maps, hence the result. It then follows from Theorem \[thm:gauthier proper\] that $f\in \operatorname{Supp}({\mu_{\mathrm{bif}}})$.
The original argument of [@buff; @epstein] was (under some mild restrictions on $f$) to check the transversality assumption of Theorem \[thm:buff epstein\], using Teichmüller-theoretic techniques. Another proof of this fact was given by Van Strien [@van; @strien].
Finally, by using an explicit deformation, Buff and Gauthier [@buff; @gauthier] recently proved that every flexible Lattès map can be approximated by strictly post-critically finite rational maps which are not Lattès examples. Therefore $\overline{\mathrm{SPCF}^*_d} = \overline{\mathrm{SPCF}_d}$.
As in the polynomial case, in the previous result one may relax the critical finiteness assumption by only requiring that the critical points map to a hyperbolic set disjoint from the critical set. Similarly to Theorem \[thm:HD polynomial\] one thus gets [@gauthier]:
The Hausdorff dimension of $\operatorname{Supp}({\mu_{\mathrm{bif}}})$ equals $2(2d-2)$.
A famous theorem of Rees [@rees] asserts that the set of rational maps that are ergodic with respect to Lebesgue measure is of positive measure in parameter space. It is then natural to ask: are these parameters inside $\operatorname{Supp}({\mu_{\mathrm{bif}}})$?
Laminarity {#subs:laminarity}
----------
A positive current $T$ of bidegree $(q,q)$ in a complex manifold $M$ of dimension $k$ is said to be *locally uniformly laminar* if in the neighborhood of every point of $\operatorname{Supp}(T)$ there exists a lamination by $q$ dimensional disks embedded in $M$ such that $T$ is an average of integration currents over the leaves. More precisely, the restriction of $T$ to a flow box of this lamination is of the form $\int_\tau [\Delta_t] dm(t)$, where $\tau$ is a local transversal to the lamination, $m$ is a positive measure on $\tau$, and $\Delta_t$ is the plaque through $t$.
A current $T$ of bidegree $(q,q)$ in $M$ is said to be *laminar* if there exists a sequence of open subsets ${\Omega}_i\subset M$ and a sequence of currents $T_i$, respectively locally uniformly laminar in ${\Omega}_i$, such that $T_i$ increases to $T$ as $i{\rightarrow}\infty$. Equivalently, $T$ is laminar iff there exists a measured family $((\Delta_a)_{a\in \mathcal{A}}, m)$ of compatible holomorphic disks of dimension $q$ in $M$ such that $T = \int_\mathcal{A} [\Delta_a]dm(a)$. Here compatible means that the intersection of two disks in the family is relatively open (possibly empty) in each of the disks (i.e. the disks are analytic continuations of each other). It is important to note that for laminar currents there is no control on the geometry of the disks (even locally). These geometric currents appear rather frequently in holomorphic dynamics. The reader is referred to [@bls] for a general account on this notion (see also [@maximal; @cantat; @survey]).
Why should we wonder about the laminarity of the bifurcation currents? We have been emphasizing the fact that $\operatorname{Supp}({T_{\mathrm{bif}}}^k)$ is in a sense the locus of “bifurcations of order $k$”. If true, this would mean that on $\operatorname{Supp}({T_{\mathrm{bif}}}^k)\setminus \operatorname{Supp}({T_{\mathrm{bif}}}^{k+1})$ we should see “stability in codimension $k$”, that is, we should expect $\operatorname{Supp}({T_{\mathrm{bif}}}^k)\setminus \operatorname{Supp}({T_{\mathrm{bif}}}^{k+1})$ to be filled with submanifolds of codimension $k$ where the dynamics is stable. There is a natural stratification of parameter space according to the dimension of the space of deformations, and our purpose is to compare this stratification with that of the supports of the successive bifurcation currents. Laminarity is the precise way to formulate this problem.
Let us be more specific. Throughout this paragraph ${\Lambda}$ is either the space of polynomials or the space of rational maps of degree $d$, with marked critical points if needed. We let $D = \dim({\Lambda})$. We say that two rational maps are deformations of each other if there is a $J$-stable family connecting them[^7].
In [@mcms] McMullen and Sullivan ask for a general description of the way the deformation space of a given rational map embeds in parameter space. Our thesis is that these submanifolds tend to be organized into laminar currents[^8].
We start with a few general facts. The first easy proposition asserts that if ${T_{\mathrm{bif}}}^k$ is laminar on some set of positive measure outside $\operatorname{Supp}({T_{\mathrm{bif}}}^{k+1})$, then the corresponding disks are indeed disks of deformations.
Assume that $S$ is a laminar current of bidegree $(k,k)$ such that $S\leq {T_{\mathrm{bif}}}^k$ and $\operatorname{Supp}(S)\cap \operatorname{Supp}({T_{\mathrm{bif}}}^{k+1})=\emptyset$. Then the disks subordinate to $S$ are disks of deformations.
By definition, a holomorphic disk of codimension $k$ is said to be subordinate to $S$ if there exists a non zero locally uniformly laminar current $U\leq S$ such that $\Delta$ is contained in a leaf of $U$. Observe that since $T_{\rm bif}^{k+1}$ gives no mass to analytic sets, there are no isolated leaves in $\operatorname{Supp}(U)$. Since $U\leq {T_{\mathrm{bif}}}^k$ and $\operatorname{Supp}(U)$ is disjoint from $\operatorname{Supp}({T_{\mathrm{bif}}}^{k+1})$, we see that $U\wedge {T_{\mathrm{bif}}}=0$. Now if $\chi$ was not pluriharmonic along $\Delta$, then by continuity, it wouldn’t be harmonic on the nearby leaves of $U$, implying that $U\wedge {T_{\mathrm{bif}}}$ would be non-zero, a contradiction.
For almost every parameter relative to the trace measure of ${T_{\mathrm{bif}}}^k$, the dimension of the deformation space is at most $D-k$.
In particular a ${\mu_{\mathrm{bif}}}$ generic parameter is rigid, that is it admits no deformations.
We should expect the codimension to be a.e. equal to $k$ on $\operatorname{Supp}({T_{\mathrm{bif}}}^k)\setminus \operatorname{Supp}({T_{\mathrm{bif}}}^{k+1})$.
Similarly to Theorem \[thm:buff epstein\], a slicing argument shows that is enough to consider the case of maximal codimension, that is, to show that for ${T_{\mathrm{bif}}}^D ={\mu_{\mathrm{bif}}}$-a.e. parameter the deformation space is zero dimensional.
If $f_0\in \operatorname{Supp}({\mu_{\mathrm{bif}}})$ possesses a disk $\Delta$ of deformations, then $\chi{ \arrowvert_{\Delta}}$ is harmonic. An already mentioned pluripotential theoretic lemma [@fs2 Lemma 6.10] asserts that if $E$ is a measurable set with the property that through every point of $E$ there exists a holomorphic disk along which $\chi$ is harmonic, then $(dd^c\chi)^{D} (E)= {\mu_{\mathrm{bif}}}(E) =0$. The result follows.
Let now ${\lambda}_0\in \operatorname{Supp}({T_{\mathrm{bif}}}^k)$. There are at least $k$ active critical points at $f_{{\lambda}_0}$. Assume the number is exactly $k$, say $c_1, \ldots, c_k$, so the remaining $D-k$ are passive ones. One might expect each of these passive critical points to give rise to a modulus of deformations of $f_{{\lambda}_0}$, but there is no general construction for this.
If these $D-k$ passive critical points lie in attracting basins, the existence of $D-k$ moduli of deformation for $f_{{\lambda}_0}$ should follow from classical quasi-conformal surgery techniques.
Observe that if the hyperbolicity conjecture holds, then $c_{k+1}, \ldots, c_D$ must be attracted by cycles. Indeed let $U$ be an open set where these points are passive. By Theorem \[thm:tbifk per\], there exists ${\lambda}\in U$ such that $c_1({\lambda}), \ldots, c_k({\lambda})$ are periodic. Hence there is an open set $U'\subset U$ where all critical points are passive. Assuming the hyperbolicity conjecture, in $U'$ all critical points lie in attracting basins. Thus this property persists for $c_{k+1}, \ldots, c_D$ throughout $U$.
When $D=2$ and $k=1$, one can indeed construct these deformations and relate them to the geometry of ${T_{\mathrm{bif}}}$. The following is a combination of results of Bassanelli-Berteloot [@bas-ber2] and the author [@cubic].
\[thm:unif lamin\] If ${\Lambda}= \mathcal{P}_3$ or $\mathcal{M}_2$ and if $U$ is an open set where one critical point is attracted by a cycle, then ${T_{\mathrm{bif}}}$ is locally uniformly laminar in $U$.
In the general case one is led to the following picture.
\[conj:lamin\] If $U\subset {\Lambda}$ is an open set where $D-k$ critical points (counted with multiplicity) lie in attracting basins, then ${T_{\mathrm{bif}}}^k$ is a laminar current in $U$, locally uniformly laminar outside a closed analytic set.
The necessity of an analytic subset where the uniform laminar structure might have singularities is due to the possibility of (exceptional) critical orbit relations
To address the question of laminarity of ${T_{\mathrm{bif}}}^k$ outside $\operatorname{Supp}({T_{\mathrm{bif}}}^{k+1})$, we also need to analyze the structure of ${T_{\mathrm{bif}}}^k$ in the neighborhood of the parameters lying outside $\operatorname{Supp}({T_{\mathrm{bif}}}^{k+1})$ but having more than $k$ active critical points. As Example \[ex:douady\] shows, these parameters can admit deformations. There does not seem to be any reasonable understanding of the bifurcation current near these parameters, even for cubic polynomials. Another interesting situation is that of cubic polynomials with a Siegel disk $\Delta$, such that a critical point falls in $\Delta$ after iteration. Of course in $\mathcal{P}_3$ both critical points are active. These parameters can be deformed by moving the critical point in the Siegel disk [@zakeri]. We do not know whether these parameters belong to $\operatorname{Supp}({\mu_{\mathrm{bif}}})$.
We also need to consider the possibility of “queer" passive critical points (whose existence contradicts the hyperbolicity conjecture, as seen above). In this case we have a positive result [@cubic].
\[thm:queer lamin\] If ${\Lambda}= \mathcal{P}_3$ and $U$ is an open set where one critical point is passive, then ${T_{\mathrm{bif}}}$ is laminar in $U$.
Perhaps unexpectedly, this result does not follow from the construction of some explicit deformation for a cubic polynomial with one active and one (queer) passive critical point. Instead we use a general laminarity criterion due to De Thélin [@dt-boule]: if a closed positive current $T$ in $U\subset {{{\mathbb{C}}^2}}$ is the limit in the sense of currents of a sequence of integration currents $T = \lim {d_n}{^{-1}}[C_n]$ with $\mathrm{genus}(C_n) =O(d_n)$, then $T$ is laminar. We refer to [@cubic] for the construction of the curves $C_n$.
Thus in the space of cubic polynomials, Theorems \[thm:unif lamin\] and \[thm:queer lamin\] show that ${T_{\mathrm{bif}}}$ is laminar outside the locus where two critical points are active (which is slightly larger than $\operatorname{Supp}({\mu_{\mathrm{bif}}})$). In the escape locus ${{{\mathbb{C}}^2}}\setminus \mathcal{C}$, where ${T_{\mathrm{bif}}}$ is uniformly laminar by Theorem \[thm:unif lamin\], we can actually give a rather precise description of ${T_{\mathrm{bif}}}$, which nicely complements the topological description given by Branner and Hubbard [@branner-hubbard1; @branner-hubbard2].
We are also able to show that this laminar structure really degenerates when approaching $\operatorname{Supp}({\mu_{\mathrm{bif}}})$, in the sense that there cannot exist a set of positive transverse measure of deformation disks “passing through" $\operatorname{Supp}({\mu_{\mathrm{bif}}})$[^9].
A consequence of this is that the genera of the curves ${\mathrm{PerCrit}}(n,k)$ must be asymptotically larger than $3^n$ near $\operatorname{Supp}({\mu_{\mathrm{bif}}})$. We refer to [@cubic] for details. The geometry of these curves was studied by Bonifant, Kiwi and Milnor [@milnor; @smooth; @bkm] in a series of papers (see also the figures in [@milnor; @cubic §2] for some visual evidence of the complexity of the ${\mathrm{PerCrit}}(n,k)$ curves).
The results in this paragraph suggest the following alternate characterization of the support of the bifurcation measure.
If ${\Lambda}$ is the moduli space of polynomials or rational maps of degree $d\geq 2$, is $\operatorname{Supp}({\mu_{\mathrm{bif}}})$ equal to the closure of the set of rigid parameters?
Bifurcation currents for families of Möbius subgroups {#sec:kleinbif}
=====================================================
The famous Sullivan dictionary provides a deep and fruitful analogy between the dynamics of rational maps on ${{\mathbb{P}^1}}$ and that of Kleinian groups [@sullivan1]. In this section we explain that bifurcation currents make also sense on the Kleinian group side, leading to interesting new results. Unless otherwise stated, all results are due to Deroin and the author [@kleinbif; @bers; @slice].
Holomorphic families of subgroups of ${\mathrm{PSL}(2,\mathbb C)}$
------------------------------------------------------------------
Here we gather some preliminary material and present the basic bifurcation theory of Möbius subgroups. We refer to the monographs of Beardon [@beardon] and Kapovich [@kapovich] for basics on the theory of Kleinian groups. Throughout this section, $G$ is a finitely generated group and ${\Lambda}$ a connected complex manifold. We consider a holomorphic family of representations of $G$ into ${\mathrm{PSL}(2,\mathbb C)}$, that is a mapping $\rho:{\Lambda}\times G{\rightarrow}{\mathrm{PSL}(2,\mathbb C)}$, such that for fixed ${\lambda}\in {\Lambda}$, $\rho({\lambda}, \cdot)$ is a group homomorphism, and for fixed $g\in G$ $\rho(\cdot, g)$ is holomorphic. The family will generally be denoted by $(\rho_{\lambda})_{{\lambda}\in {\Lambda}}$.
We make three standing assumptions:
1. the family is non-trivial, in the sense that there exists ${\lambda}_1$, ${\lambda}_2$ such that the representations $\rho_{{\lambda}_i}$, $i=1,2$ are not conjugate in ${\mathrm{PSL}(2,\mathbb C)}$;
2. there exists ${\lambda}_0\in {\Lambda}$ such that $\rho_{{\lambda}_0}$ is faithful;
3. for every ${\lambda}\in {\Lambda}$, $\rho_{\lambda}$ is non-elementary.
Assumptions (R1) and (R2) do not really restrict our scope: this is obvious for (R1), and for (R2) is suffices to take a quotient of $G$. Notice that under (R2), the representations $\rho_{\lambda}$ are *generally faithful*, that is, the set of of non faithful representations is a union of Zariski closed sets.
Recall that a representation is said to be [*elementary*]{} when it admits a finite orbit on $\mathbb{H}^3 \cup{{\mathbb{P}^1}}({\mathbb{C}})$ ($\mathbb{H}^3$ is the 3-dimensional hyperbolic space). Then, either $\Gamma$ fixes a point in $\mathbb{H}^3$ and is conjugate to a subgroup of $\mathrm{PSU}(2)$[^10] and in particular it contains only elliptic elements, or it has a finite orbit (with one or two elements) on ${{\mathbb{P}^1}}$. It can easily be proved that the subset of elementary representations of a given family $(\rho_{{\lambda}})$ is a real analytic subvariety $E$ of ${\Lambda}$. Hence (R3) will be satisfied upon restriction to ${\Lambda}\setminus E$ [^11].
We identify ${\mathrm{PSL}(2,\mathbb C)}$ with the group of transformations of the form $\gamma(z)=\frac{az
+b}{cz+d}$, with ${\left(\begin{smallmatrix} a& b \\ c&d\end{smallmatrix}\right)}\in {\mathrm{SL}(2,\mathbb C)}$ and let $${\left\Vert\gamma\right\Vert} = \sigma(A^*A)^{1/2} \text{ and }
\operatorname{tr}^2{\gamma} = (\operatorname{tr}A)^2,$$ where $A= {\left(\begin{smallmatrix} a& b \\ c&d\end{smallmatrix}\right)}$ is a lift of $\gamma$ to ${\mathrm{SL}(2,\mathbb C)}$, and $\sigma(\cdot)$ is the spectral radius. Of course these quantities do not depend on the choice of the lift.
As it is well-known, Möbius transformations are classified into three types according to the value of their trace:
[*parabolic*]{}
if $\operatorname{tr}^2(\gamma)=4$ and $\gamma \neq\mathrm{id}$; it is then conjugate to $z\mapsto z+1$;
[*elliptic*]{}
if $\operatorname{tr}^2(\gamma)\in [0,4)$, it is then conjugate to $z\mapsto e^{i\theta}z$ for some real number $\theta$, and $\operatorname{tr}^2{\gamma} = 2+2\cos(\theta)$.
[*loxodromic*]{}
if $\operatorname{tr}^2(\gamma)\notin [0,4]$, it is then conjugate to $z\mapsto kz$, with ${\left\vertk\right\vert}\neq 1$.
There is a well-established notion of bifurcation for a family of Möbius subgroups, which is the translation in the Sullivan dictionary of Theorem \[thm:mss\]. It has its roots in the work of Bers, Kra, Marden, Maskit, Thurston, and others on deformations of Kleinian groups.
\[Sullivan [@sullivan] (see also Bers [@bers])\]\[thm:sullivan\] Let $(\rho_\lambda)_{{\lambda\in {\Lambda}}}$ be a holomorphic family of representations of $G$ into ${\mathrm{PSL}(2,\mathbb C)}$ satisfying (R1-3), and let ${\Omega}\subset\Lambda$ be a connected open set. Then the following assertions are equivalent:
1. for every $\lambda\in {\Omega}$, $\rho_{\lambda}(G) $ is discrete;
2. for every $\lambda\in {\Omega}$, $\rho_\lambda$ is faithful;
3. for every $g$ in $G$, if for some $\lambda_0\in {\Omega}$, $\rho_{\lambda_0}(g)$ is loxodromic (resp. parabolic, elliptic), then $\rho_{\lambda}(g)$ is loxodromic (resp. parabolic, elliptic) throughout ${\Omega}$;
4. for any $\lambda_0$, $\lambda_1$ in ${\Omega}$ the representations $\rho_{\lambda_0}$ and $\rho_{\lambda_1}$ are quasi-conformally conjugate on ${{\mathbb{P}^1}}$.
If one of these conditions is satisfied, we say that the family is *stable* in ${\Omega}$. We define $\operatorname{Stab}$ to be the maximal such open set, and $\operatorname{Bif}$ to be its complement, so that ${\Lambda}= \operatorname{Stab}\cup\operatorname{Bif}$.
Theorem \[thm:sullivan\] shows that $\operatorname{Stab}= \operatorname{Int}(\mathrm{DF})$ is the interior of the set of discrete and faithful representations. One main difference with rational dynamics is that, as a consequence of the celebrated Jørgensen-Kazhdan-Margulis-Zassenhauss theorem, $\mathrm{DF}$ is a closed subset of parameters space. This is also referred to as Chuckrow’s theorem, see [@kapovich p. 170]. This implies that, whenever non-empty, $\operatorname{Bif}$ has non-empty interior, which is in contrast with Theorem \[thm:dense\].
The following corollary is immediate:
\[cor:accidental\] For every $t\in [0,4]$, the set of such parameters $\lambda_0$ at which there exists $g\in G$ such that $\operatorname{tr}^2{\rho_{\lambda_0}(g)}=t$ and $\lambda \mapsto \operatorname{tr}^2{\rho_{\lambda}(g)}$ is not locally constant, is dense in $\operatorname{Bif}$.
Again, a basic motivation for the introduction of bifurcation currents is the study of the asymptotic distribution of such parameters. The most emblematic value of $t$ is $t=4$. In this case one either gets “accidental” new relations or new parabolic elements. Notice that when $t = 4\cos^2(\theta)$ with $\theta\in \pi\mathbb{Q}$ (e.g. $t=0$), then if $\operatorname{tr}^2{\rho_{\lambda_0}(g)}=t$, $g$ is of finite order, so these parameters also correspond to accidental new relations in $\rho_{\lambda}(G)$.
A famous result in this area of research is a theorem by McMullen [@mcm; @cusps] which asserts that accidental parabolics are dense [*in the boundary*]{} of certain components of stability. One might also wonder what happens of Corollary \[cor:accidental\] when additional assumptions are imposed on $g$. Here is a question (certainly well-known to the experts) which was communicated to us by McMullen: if $G = \pi_1(S, \ast)$ is the fundamental group of a surface of finite type, does Corollary \[cor:accidental\] remain true when restricting to the elements $g\in G$ corresponding to simple closed curves on $S$?
Another important feature of the space of all representations of $G$ into ${\mathrm{PSL}(2,\mathbb C)}$ (resp. modulo conjugacy) is that it admits a natural action of the automorphism group of $\mathrm{Aut}(G)$ (resp. the outer automorphism group $\mathrm{Out}(G)$). Despite recent advances, the dynamics of this action is not well understood (see the expository papers of Goldman [@goldman] and Lubotzky [@lubotzky] for an account on this topic). There is a promising interplay between these issues and holomorphic dynamics, which was recently illustrated by the work of Cantat [@cantat; @bers].
Products of random matrices
---------------------------
To define a bifurcation current we use a notion of Lyapunov exponent of a representation, arising from a random walk on $G$. The properties of this Lyapunov exponent will be studied using the theory of random walks on groups and random products of matrices (good references on these topics are [@bougerol-lacroix; @furman]).
Let us fix a probability measure $\mu$ on $G$, satisfying the following assumptions:
1. $\operatorname{Supp}(\mu)$ generates $G$ as a semi-group;
2. there exists $s>0$ such that $\int_G \exp(s\; \operatorname{length}(g)) d\mu(g) <\infty$.
The length in (A2) is relative to the choice of any finite system of generators of $G$; of course the validity of (A2) does not depend on this choice. A typical case where these assumptions are satisfied is that of a finitely supported measure on a symmetric set of generators (like in the case of the simple random walk on the associated Cayley graph).
Loosely speaking, the choice of such a measure a measure on $G$ is somehow similar to the choice of a time parameterization for a flow, or more generally of a Riemannian metric along the leaves of a foliation.
We denote by $\mu^n$ the $n^{\rm th}$ convolution power of $\mu$, that is the image of $\mu^
{\otimes n}$ under $(g_1, \ldots ,g_n)\mapsto g_1\cdots g_n$. This is the law of the $n^{\rm th}$ step of the (left- or right-) random walk on $G$ with transition probabilites given by $\mu$.
If $\mathbf{g} = (g_n)_{n\geq 1}\in G^{\mathbb{N}}$, we let $l_n(\mathbf{g}) = g_n\cdots g_1$ be the product on the left of the $g_k$ (resp. $r_n(\mathbf{g}) = g_1\cdots g_n$ the product on the right). We denote by $\mu^{\mathbb{N}}$ the product measure on $G^{\mathbb{N}}$ so that $\mu^n = (l_n)_*\mu^{\mathbb{N}}= (r_n)_*
\mu^{\mathbb{N}}$.
The [*Lyapunov exponent*]{} of a representation $\rho
: G \rightarrow {\mathrm{PSL}(2,\mathbb C)}$ is defined by the formula $$\label{def:lyapunov exponent} \chi (\rho ) : = \lim _{n\rightarrow \infty} \frac{1}{n}\int _G
\log {\left\Vert\rho(g)\right\Vert} d\mu^{n}(g) = \lim _{n\rightarrow \infty} \frac{1}{n}\int \log {\left\Vert\rho(g_1\cdots
g_n)\right\Vert} d\mu(g_1)\cdots d\mu(g_n),$$ in which the limit exists by sub-addivity. It immediately follows from Kingman’s sub-additive ergodic theorem that $$\label{eq: lyapunov a.e.}
\text{for }\mu^{\mathbb{N}}\text{-a.e. }\mathbf{g}, \ \lim _{n\rightarrow \infty} \frac{1}{n} \log {\left\Vert\rho (l_n({\bf g}))\right\Vert} =\chi (\rho)$$ (this was before Kingman a theorem due to Furstenberg and Kesten [@furstenberg; @kesten]).
The following fundamental theorem is due to Furstenberg [@furstenberg]:
\[thm:furstenberg\] Let $G$ be a finitely generated group and $\mu$ a probability measure on $G$ satisfying (A1-2). Let $\rho:G{\rightarrow}{\mathrm{PSL}(2,\mathbb C)}$ be a non-elementary representation. Then the Lyapunov exponent $\chi(\rho)$ is positive and depends continuously on $\rho$.
The next result we need is due to Guivarc’h [@guivarch]:
\[thm:guivarch\] Let $G$ be a finitely generated group and $\mu$ a probability measure on $G$ satisfying (A1-2). Let $\rho:G{\rightarrow}{\mathrm{PSL}(2,\mathbb C)}$ be a non-elementary representation. Then for $\mu^{\mathbb{N}}$-a.e $\mathbf{g}$, we have that $$\label{eq:guivarch}
{\frac{1}{n}}\log {\left\vert\operatorname{tr}(\rho( l_n(\mathbf{g})))\right\vert} = {\frac{1}{n}}\log{\left\vert \operatorname{tr}(\rho(g_n\cdots g_1))\right\vert} \underset{n
{\rightarrow}\infty}\longrightarrow
\chi(\rho).$$
A trivial remark which turns out to be a source of technical difficulties, is that, as opposite to , one cannot in general integrate with respect to $\mu^{\mathbb{N}}$ in . The reason of course is that some words can have zero or very small trace. Conversely, if $h$ is a function on ${\mathrm{PSL}(2,\mathbb C)}$, which is bounded below and equivalent to $\log{\left\vert\operatorname{tr}(\cdot)\right\vert}$ as the trace tends to infinity, then one can integrate with respect to $\mu^{\mathbb{N}}$.
An example of such a function is given by the spectral radius, and under the assumptions of Theorem \[thm:guivarch\] we obtain that $$\label{eq:lambdamax}
\chi(\rho) = \lim_{n{\rightarrow}\infty}{\frac{1}{n}}\int \log{\left\vert\sigma(\rho( g))\right\vert} d\mu^n(g).$$
The bifurcation current {#the-bifurcation-current}
-----------------------
We are now ready for the introduction of the bifurcation current, following [@kleinbif]. Let $G$ be a finitely generated group, and $(\rho_{\lambda})_{{\lambda}\in {\Lambda}}$ be a holomorphic family of representations satisfying (R1-3). Fix a probability measure $\mu$ on $G$ satisfying (A1-2). It follows immediately from that $\chi:{\lambda}\mapsto\chi(\rho_{\lambda})$ is a psh function on ${\Lambda}$. Motivated by the analogy with rational dynamics, it is natural to suggest the following definition.
\[def:bifcur\] Let $(G,\rho,\mu)$ be as above. The bifurcation current associated to $(G,\mu,\rho)$ is ${T_{\mathrm{bif}}}= dd^c\chi$.
At this point it is still unclear whether this gives rise to a meaningful concept. This will be justified by the following results.
First, it is easy to see that $\operatorname{Supp}({T_{\mathrm{bif}}})$ is contained in the bifurcation locus. Indeed, if $(\rho_{\lambda})$ is stable on ${\Omega}$, then the Möbius transformations $\rho_{\lambda}(g)$, $g\in G$, do not change type throughout ${\Omega}$. In particular for every $g\in G$, ${\Omega}\ni{\lambda}\mapsto \log{\left\vert\sigma(\rho_{\lambda}(g))\right\vert}$ is pluriharmonic. We thus infer from that $\chi$ is pluriharmonic on ${\Omega}$.
It is a remarkable fact that conversely, pluriharmonicity of $\chi$ characterizes stability:
\[thm:support\] Let $(G, \rho, \mu)$ be a holomorphic family of representations of $G$, satisfying (R1-3), endowed with a measure $\mu$ satisfying (A1-2). Then $\operatorname{Supp}({T_{\mathrm{bif}}})$ is equal to the bifurcation locus.
Here is a sketch of the proof, which consists in several steps, and involves already encountered arguments. It is no loss of generality to assume that $\dim({\Lambda}) =1$. We need to show that if ${\Omega}$ is an open set disjoint from $\operatorname{Supp}({T_{\mathrm{bif}}})$, then $(\rho_{\lambda})$ is stable in ${\Omega}$. The main idea is to look for a geometric interpretation of ${T_{\mathrm{bif}}}$, in the spirit of what we did in §\[subs:marked\].
For this we need a substitute for the equilibrium measure of a rational map: this will be the unique stationary measure under the action of $\rho_{\lambda}(G)$ on ${{\mathbb{P}^1}}$. Let us be more specific. For every representation, $(G,\mu)$ acts by convolution on the set of probability measures on ${{\mathbb{P}^1}}$ by the assignment $$\nu\longmapsto \int \rho(g)_* \nu \; d\mu(g).$$ Any fixed point of this action is called a [*stationary measure*]{}. The following theorem is intimately related to Theorem \[thm:furstenberg\]. It is in a sense the analogue of the Brolin-Lyubich theorem in this context.
Let $G$ be a finitely generated group and $\mu$ a probability measure on $G$ satisfying (A1-2). Let $\rho:G{\rightarrow}{\mathrm{PSL}(2,\mathbb C)}$ be a non-elementary representation. Then there exists a unique stationary probability measure $\nu$ on ${{\mathbb{P}^1}}$. Furthermore, for any $z_0\in {{\mathbb{P}^1}}$, the sequence $\int ({\rho_{{\lambda}_0}(g)})_*\delta_{z_0} d\mu^n(g)$ converges to $\nu$.
In analogy with §\[subs:marked\], let us now work in ${\Lambda}\times {{\mathbb{P}^1}}$, and consider the fibered action $\widehat g$ on ${\Lambda}\times {{\mathbb{P}^1}}$ defined by $\widehat g:({\lambda},z)\mapsto ({\lambda}, \rho_{\lambda}(g)(z))$. We seek for a current $\widehat T$ on ${\Lambda}\times {{\mathbb{P}^1}}$ “interpolating” the stationary measures. Given $z_0\in {{\mathbb{P}^1}}$ , we then introduce the sequence of positive closed currents $ \widehat{T}_n$ defined by $$\label{eq:Tn hat}
\widehat{T}_n = {\frac{1}{ n}} \int \left[\widehat{g}\left({\Lambda}\times{\left\{z_0\right\}}\right) \right] d\mu^n(g).$$
To understand why it is natural to have a linear normalization in , think of a polynomial family of representations. Precisely, assume that ${\Lambda}={\mathbb{C}}$ and that for a set of generators $g^1, \ldots, g^k$ of $G$, the matrices $\rho_{\lambda}(g^j)$ are polynomial in ${\lambda}$. Then if $w$ is a word in $n$ letters in the generators, $\rho_{\lambda}(w)$ has degree $O(n)$ in ${\lambda}$, hence in ${\mathbb{C}}\times {{\mathbb{P}^1}}\subset{{\mathbb{P}^1}}\times {{\mathbb{P}^1}}$, the degree of the graph $\widehat{w}({\Lambda}\times {\left\{z_0\right\}})$ is $O(n)$.
It turns out that this sequence does [*not*]{} converge to a current interpolating the stationary measures (since e.g. it vanishes above the stability locus) nevertheless it gives another crucial information.
The sequence of currents $\widehat{T}_n$ converges to $\pi_1^*{T_{\mathrm{bif}}}$, where $
\pi_1:{\Lambda}\times{{\mathbb{P}^1}}{\rightarrow}{\Lambda}$ is the natural projection.
Figure \[fig:hat\] is a visual interpretation of this result.
\[fig:hat\]
This proposition implies that if ${\Omega}$ is disjoint from $\operatorname{Supp}({T_{\mathrm{bif}}})$, the average growth of the area of the sequence of graphs $\left(\widehat{g_1\cdots g_n}\right)\left({\Lambda}\times{\left\{z_0\right\}}\right)$ over ${\Omega}$ is sublinear in $n$. The attentive reader will have noticed that the situation is similar to that of Section \[sec:prologue\], except that the sequence of graphs over ${\Lambda}$ is replaced by an average of graphs. The next result is in the spirit of Proposition \[prop:support\].
If ${\Omega}$ is an open subset of ${\Lambda}$ such that $\overline{\Omega}$ is compact and disjoint from $\operatorname{Supp}({T_{\mathrm{bif}}})$, then the mass of $\widehat{T}_n$ in ${\Omega}$ is $O{\left({\frac{1}{n}}\right)}$. In other words: $$\int \mathrm{Area}{\left( \widehat{g}\left({\Lambda}\times{\left\{z_0\right\}}\right) \cap \pi_1^{-1}({\Omega}) \right)}d\mu^n(g) = O
(1).$$
To prove this estimate we give an analytic expression of this area: $$\begin{aligned}
\int \mathrm{Area}\big( \widehat{g}\left({\Lambda}\times{\left\{z_0\right\}}\right) & \cap \pi_1^{-1}({\Omega}) \big) d\mu^n(g) =
\int_{\pi_1^{-1}({\Omega})}\big[\widehat{g}\left({\Lambda}\times{\left\{z_0\right\}}\right)\big]\wedge (\pi_1^*\omega_{\Lambda}+
\pi_2^*\omega_{{\mathbb{P}^1}})d\mu^n(g) \\
&=\mathrm{Area}({\Omega}) + \int_{\Omega}(\pi_1)_* \left(\pi_2^*\omega_{{{\mathbb{P}^1}}}{ \arrowvert_{\widehat{g}\left({\Lambda}\times{\left\{z_0\right\}}\right)}}\right) d\mu^n(g) \\
&= \mathrm{Area}({\Omega})+n \int_{\Omega}dd^c\chi_n \text{ , where } \chi_n = {\frac{1}{n}} \int\log\frac{{\left\Vert\rho_{\lambda}(g)
(Z_0)\right\Vert}}{{\left\VertZ_0\right\Vert}} d\mu^n(g).\end{aligned}$$ Here $Z_0$ is a lift of $z_0$ to ${{{\mathbb{C}}^2}}$, ${\left\Vert\cdot\right\Vert}$ is the Hermitian norm, and the $dd^c$ is taken w.r.t. ${\lambda}$. Therefore, exactly as in Theorem \[thm:activity support\] we are left to prove that $\chi_n-\chi = O{\left({\frac{1}{n}}\right)}$. This estimate in turn follows from ergodic theoretic properties of random matrix products –specifically, from the “exponential convergence of the transition operator”, a result due to Le Page [@lepage].
The next step is to combine this estimate on the average area of the graphs over ${\Omega}$ with another result of Furstenberg which asserts that for each fixed ${\lambda}$, for $\mu^{\mathbb{N}}$-a.e. sequence $\mathbf{g}=(g_n)$, $\rho_{\lambda}(g_1\cdots g_n)({z_0})$ converges to a point $z_{\mathbf{g}}$ (a consequence of the martingale convergence theorem). From this, as in Lemma \[lem:dim1\], we infer that for a.e. $\mathbf{g} = (g_n)$, the sequence of graphs $\left(\widehat
{g_1\cdots g_n}\right)\left({\Lambda}\times{\left\{z_0\right\}}\right)$ converges to a limiting graph $\Gamma_{\mathbf
{g}}$, outside a finite set of vertical bubbles. We then obtain a measurable and $G$-equivariant family of limiting graphs over ${\Omega}$. Formally, this family is parameterized by the *Poisson boundary $P(G, \mu)$* of $(G,\mu)$.
To show that the family of representations is stable over ${\Omega}$, we need to upgrade this family of graphs into a holomorphic motion of the Poisson boundary, that is, we need to show that two distinct graphs are disjoint. For this, we use the fact that the number of intersection points between two graphs over some domain $D\subset {\Omega}$ is a function on $P(G, \mu)\times P(G,\mu)$ satisfying certain invariance properties. By an ergodic theorem due to Kaimanovich [@kaimanovich], this function is a.e. equal[^12] to a constant $\iota_D$ on $P(G, \mu)\times P(G,\mu)$, depending only on $D$ . The assignement $D\mapsto \iota_D$ being integer valued, we infer that there exists a locally finite set of points $F$ such that if $D\cap F = \emptyset$, then $\iota_D=0$. On such a $D$, it follows that the family of graphs is a holomorphic motion, and ultimately, that the family is stable. Finally, to show that the exceptional set $F$ is empty and conclude that the family is stable on all $D$, we use the fact that the set of discrete faithful representations is closed, which implies that $\operatorname{Bif}$ cannot have isolated points.
Equidistribution of representations with an element of a given trace {#subs:trace}
--------------------------------------------------------------------
Another important aspect of the bifurcation currents is that they enable to obtain equidistribution results in Corollary \[cor:accidental\]. This is the analogue in our context of the results of §\[subs:multiplier\]. For $t\in {\mathbb{C}}$ denote by $Z(g,t)$ the analytic subset of ${\Lambda}$ defined by $Z(g,t) = {\left\{{\lambda}, \ \operatorname{tr}^2(\rho_{\lambda}(g)) = t\right\}}$. We study the asymptotic properties of the integration currents $[Z(g,t)]$. Note that if $\operatorname{tr}^2(\rho_{\lambda}(g))\equiv t$, then $Z(g,t)={\Lambda}$, and by convention, $[{\Lambda}]=0$.
The first equidistribution result is the following.
\[thm:equidist1\] Let $(G,\rho,\mu)$ be a holomorphic family of representations of $G$, satisfying (R1-3), endowed with a measure $\mu$ satisfying (A1-2). Fix $t\in{\mathbb{C}}$. Then for $\mu^{\mathbb{N}}$ a.e. sequence $(g_n)$, we have that $${\frac{1}{2n}}\left[Z(g_n\cdots g_1, t )\right] \underset{n{\rightarrow}\infty}\longrightarrow
T_{\rm bif}.$$
Notice that, if instead of considering a random sequence in the group, we take a word obtained by applying to $g\in G$ an iterated element of $\mathrm{Aut}(G)$, then similar equidistribution results were obtained by Cantat in [@cantat; @bers]
The following “deterministic” corollary makes Corollary \[cor:accidental\] more precise. It seems difficult to prove it without using probabilistic methods.
Under the assumptions of Theorem \[thm:equidist1\], let ${\varepsilon}>0$ and ${\Lambda}'\Subset {\Lambda}$. Then there exists $g\in G$ such that ${\lambda}\mapsto \operatorname{tr}^2(\rho_{\lambda}(g))$ is non-constant and ${\left\{{\lambda}, \operatorname{tr}^2(\rho_{\lambda}(g))=t\right\}}$ is ${\varepsilon}$-dense in $\operatorname{Bif}\cap {\Lambda}'$.
For the value $t=4$, it is unclear which of accidental relations or parabolics prevail in $\left[Z(g_n\cdots g_1, 4)\right ]$.
Theorem \[thm:equidist1\] is actually a consequence of a more general theorem [@kleinbif Thm 4.1], which gives rise to several other random equidistribution results.
As opposed to the case of rational maps, we are able to estimate the speed of convergence, up to some averaging on $g$ and a global assumption on ${\Lambda}$.
\[thm:equidist speed\] Let $(G,\rho,\mu)$ be a holomorphic family of representations of $G$, satisfying (R1-3), endowed with a measure $\mu$ satisfying (A1-2). Fix $t\in{\mathbb{C}}$. Assume further that one of the following hypotheses is satisfied:
1. ${\Lambda}$ is an algebraic family of representations defined over $\overline{\mathbb{Q}}$.
2. There exists a geometrically finite representation in ${\Lambda}$.
Then there exists a constant $C$ such that for every test form $\phi$, $${\left\langle {\frac{1}{2n}} \int \left[Z(g,t)\right] d\mu^n(g) -T_{\rm bif}, \phi\right\rangle}\leq
C \frac{\log n }{n} {\left\Vert\phi\right\Vert}_{C^2}.$$
The meaning of the notion of an algebraic family of representations is the following: the space $\mathrm{Hom}(G,{\mathrm{PSL}(2,\mathbb C)})$ of representations of $G$ in ${\mathrm{PSL}(2,\mathbb C)}$ admits a natural structure of an affine algebraic variety over $\mathbb{Q}$, simply by describing it as a set of matrices satisfying certain polynomial relations[^13]. Changing this set of generators amounts to performing algebraic changes of coordinates, so that this structure of algebraic variety is well-defined. We say that an arbitrary family of representations, viewed as a holomorphic mapping $\rho:{\Lambda}{\rightarrow}\mathrm{Hom}(G,{\mathrm{PSL}(2,\mathbb C)})$ is algebraic (resp. algebraic over $K$) if $\rho{ \arrowvert_{{\Lambda}}}$ is a dominating map to some algebraic subvariety (resp. over $K$) of $\mathrm{Hom}(G,{\mathrm{PSL}(2,\mathbb C)})$. To say it differently, there exists an open subset ${\Omega}\subset {\Lambda}$ such that $\rho_{{\Omega}}$ is an open subset of an algebraic subvariety of $\mathrm{Hom}(G,{\mathrm{PSL}(2,\mathbb C)})$. .
These results parallel those of §\[subs:multiplier\], and exactly as in Theorem \[thm:basber\], they become much easier after some averaging with respect to the multiplier. Let us illustrate this by proving the following result:
Let $m$ be the normalized Lebesgue measure on $[0,4]$. Then under the assumptions of Theorems \[thm:equidist1\], we have that $${\frac{1}{2n}} \int \left[Z(g,t)\right] d\mu^n(g)dm(t) \underset{n{\rightarrow}\infty}\longrightarrow
T_{\rm bif}.$$
We prove the $L^1_{\rm loc}$ convergence of the potentials. Let $u(g,\cdot)$ be the psh function on ${\Lambda}$ defined by $$u(g,{\lambda}) = \int \log {\left\vert\operatorname{tr}^2(\rho_{\lambda}(g))-t\right\vert} dm(t) = v(\operatorname{tr}^2(\rho_{\lambda}(g))),$$ where $v$ is the logarithmic potential of $m$ in ${\mathbb{C}}$. The function $u$ is bounded below and $u(g, {\lambda}) \sim \log{\left\vert\operatorname{tr}^2 (\rho_{\lambda}(g))\right\vert}$ as $\operatorname{tr}^2 (\rho_{\lambda}(g))$ tends to infinity.
If ${\lambda}$ is fixed, then by Theorem \[thm:guivarch\], for $\mu^{\mathbb{N}}$ a.e. sequence $(g_n)$, ${\frac{1}{2n}} u(g_1\cdots g_n, {\lambda}){\rightarrow}\chi({\lambda})$. Since $u$ is bounded below, we can apply the dominated convergence theorem and integrate with respect to $g_1,\ldots ,g_n$. We conclude that $${\frac{1}{2n}} \int u(g,{\lambda}) d\mu^n(g)\underset{n{\rightarrow}\infty}\longrightarrow \chi({\lambda})$$ for all ${\lambda}\in {\Lambda}$, which, by taking the $dd^c$ in ${\lambda}$, implies the desired statement.
We need to show that for a.e. sequence $(g_n)$, the sequence of psh functions ${\frac{1}{2n}}\log{\left\vert\operatorname{tr}^2(\rho_{\lambda}(g_1\cdots g_n)) -t\right\vert}$ converges to $\chi$. The point is to find a choice of random sequence $(g_n)$ which does not depend on ${\lambda}$. For this, a kind of sub-additive ergodic theorem with values in the space of psh functions shows that for a.e. $(g_n)$, ${\frac{1}{n}} \log {\left\Vert\rho_{\lambda}(g_1\cdots g_n)\right\Vert}$ converges in $L^1_{\rm loc}$ to $\chi$. By Theorem \[thm:guivarch\], it is possible to choose the sequence $(g_n)$ so that for ${\lambda}$ belonging to a countable dense sequence of parameters, ${\frac{1}{2n}}\log{\left\vert\operatorname{tr}^2(\rho_{\lambda}(g_1\cdots g_n)) -t\right\vert}$ converges to $\chi({\lambda})$. On the other hand, $${\frac{1}{2n}}\log{\left\vert\operatorname{tr}^2(\rho_{\lambda}(g_1\cdots g_n)) -t\right\vert}\leq {\frac{1}{n}} \log {\left\Vert\rho_{\lambda}(g_1\cdots g_n)\right\Vert} + o(1).$$ Using the continuity of $\chi$ and the Hartogs lemma, we conclude that ${\frac{1}{2n}}\log{\left\vert\operatorname{tr}^2(\rho_{\lambda}(g_1\cdots g_n)) -t\right\vert}$ converges to $\chi$ in $L^1_{\rm loc}$.
For Theorem \[thm:equidist speed\], the main difficulty is that for a given parameter we cannot in general integrate with respect to $g_1, \cdots, g_n$ in the almost sure convergence $${\frac{1}{2n}}\log {\left\vert\operatorname{tr}^2(\rho_{\lambda}(g_1\cdots g_n)) - t\right\vert} {\rightarrow}\chi({\lambda}),$$ due to the possibility of elements with trace very close to $t$. This is exactly similar to the difficulty encountered in Theorem \[thm:basber\].[*iii*]{}. We estimate the size of the set of parameters where this exceptional phenomenon happens by using volume estimates for sub-level sets of psh functions and the global assumption [*i.*]{} or [*ii.*]{} In both cases this global assumption is used to show the existence of a parameter at which $ {\left\vert\operatorname{tr}^2(\rho_{\lambda}(g_1\cdots g_n)) - t\right\vert}$ is not super-exponentially small in $n$. Under [*i.*]{}, this follows from a nice number-theoretic lemma (a generalization of the so-called [*Liouville inequality*]{}), which was communicated to us by P. Philippon. Another key ingredient is a large deviations estimate in Theorem \[thm:guivarch\], which was obtained independently by Aoun [@aoun].
Canonical bifurcation currents
------------------------------
One might object that our definition of bifurcation currents in spaces of representations lacks of naturality, for it depends on the choice of a measure $\mu$ on $G$ –recall however from Theorem \[thm:support\] that the support of the bifurcation current is independent of $\mu$. In this paragraph, following [@bers; @slice], we briefly explain how a canonical bifurcation current can be constructed under natural assumptions.
Let $X$ be a compact Riemann surface of genus $g\geq 2$, and $G = \pi_1(X, \ast)$ be its fundamental group. Let $(\rho_{\lambda})$ be a holomorphic family of representations of $G$ into ${\mathrm{PSL}(2,\mathbb C)}$ satisfying (R1-3). We claim that there is a Lyapunov exponent function on ${\Lambda}$ which is canonically associated to the Riemann surface structure of $X$ (up to a multiplicative constant).
For this, let $\widetilde{X}$ be the universal cover of $X$ (i.e. the unit disk). $G$ embeds naturally as a subgroup of $\mathrm{Aut}(\widetilde X)$. For any representation $\rho\in {\Lambda}$, consider its suspension $X_\rho$, that is the quotient of $\widetilde{X}\times {{\mathbb{P}^1}}$ by the diagonal action of $G$. The suspension is a fiber bundle over $X$, with ${{\mathbb{P}^1}}$ fibers, and admits a holomorphic foliation transverse to the fibers whose holonomy is given by $\rho$. If $\gamma$ is any path on $X$, we denote by $h_\gamma$ its holonomy ${{\mathbb{P}^1}}_{\gamma(0)}{\rightarrow}{{\mathbb{P}^1}}_{\gamma(1)}$.
The Poincaré metric endows $X$ with a natural Riemannian structure, so we can consider the Brownian motion on $X$. It follows from the sub-additive ergodic theorem that for a.e. Brownian path ${\Omega}$ (relative to the Wiener measure), the limit $$\chi (\omega) = \lim_{t{\rightarrow}\infty} {\frac{1}{t}}\log{\left\Verth_{\omega(0), \omega(t)}\right\Vert}$$ (where ${\left\Vert\cdot\right\Vert}$ is any smoothly varying spherical metric on the fibers) exists, and does not depend on ${\Omega}$.
We define $\chi_{\rm Brownian}(\rho)$ to be this number, and introduce a natural bifurcation current on ${\Lambda}$ by putting ${T_{\mathrm{bif}}}= dd^c\chi_{\rm Brownian}$. We have the following theorem.
Let as above $(\rho_{\lambda})$ be a holomorphic family of non-elementary representations of the fundamental group of a compact Riemann surface, satisfying (R1-3).
Then the function $\chi_{\rm Brownian}$ is psh on ${\Lambda}$ and the support of ${T_{\mathrm{bif}}}= dd^c\chi_{\rm Brownian} $ is the bifurcation locus.
To prove this theorem, it is enough to exhibit a measure $\mu$ on $G$ satisfying (A1-2) and such that for every $\rho$, $\chi_\mu(\rho) = \chi_{\rm Brownian}(\rho)$ (up to a multiplicative constant). Such a measure actually exists and was constructed using a discretization procedure by Furstenberg [@furstenberg; @discretisation]. It is non-trivial to check that $\mu$ satisfies the exponential moment condition (A2) (for instance this measure can never be of finite support).
There is another natural family of paths on $X$: the geodesic trajectories. An argument similar to the previous one shows that if $(x,v)\in S^1(X)$ (unit tangent bundle) is generic relative to the Liouville measure, and if $\gamma_{(x,v)}$ denotes the unit speed geodesic stemming from $(x,v)$, then the limit $\lim_{t{\rightarrow}\infty} {\frac{1}{t}}\log{\left\Verth_{\gamma
(0), \gamma(t)}\right\Vert}$ exists and does not depend on (generic) $(x,v)$. We denote by $\chi_{\mathrm{geodesic}}(\rho)$ this number. It follows from the elementary properties of the Brownian motion on the hyperbolic disk that there exists a constant $v$ depending only on $X$ such that $\chi_{\rm Brownian} = v \chi_{\mathrm{geodesic}}$. Therefore the associated bifurcation current is the same.
Here is a situation where these ideas naturally apply: consider the set $\mathcal{P}(X)$ of complex projective structures over a Riemann surface $X$, compatible with its complex structure (see [@dumas] for a nice introductory text on projective structures). This is a complex affine space of dimension $3g-3$, admitting a distinguished point, the “standard Fuchsian structure”, namely the projective structure obtained by viewing $X$ as a quotient of the unit disk. A projective structure induces a [*holonomy representation*]{} (which is always non-elementary and defined only up to conjugacy) so the above discussion applies. We conclude that [*the space of projective structures on $X$ admits a natural bifurcation current*]{}.
From the standard Fuchsian structure, one classically constructs an embedding of the Teichmüller space of $X$ as a bounded open subset of $\mathcal{P}(X)$, known as the [*Bers embedding*]{} (or [*Bers slice*]{}). This open set can be defined for instance as the component of the distinguished point in the stability locus.
In [@mcm; @book], McMullen suggests the Bers slice as the analogue of the Mandelbrot set through the Sullivan dictionary. From this perspective, an interesting result in [@bers; @slice] is that the canonical Lyapunov exponent function $\chi_{\rm Brownian}$ is [*constant*]{} on the Bers embedding, so the analogy also holds at the level of Lyapunov exponents.
Open problems
-------------
There are many interesting open questions in this area, some of them stated in [@kleinbif §5.2]. In the spirit of this survey, let us only state two problems related to the exterior powers of ${T_{\mathrm{bif}}}$.
As opposite to the case of rational maps, we believe that the supports of ${T_{\mathrm{bif}}}^k$ for $k\geq 2$ do not give rise to “higher bifurcation loci”.
Let $(G, \rho, \mu)$ be a family of representations satisfying (R1-3) and (A1-2), and assume further that two representations in ${\Lambda}$ are never conjugate in ${\mathrm{PSL}(2,\mathbb C)}$ (that is, ${\Lambda}$ is a subset of the character variety). Then for every $k\leq \dim({\Lambda})$, $\operatorname{Supp}({T_{\mathrm{bif}}}^k) = \operatorname{Bif}$.
Here is some evidence for this conjecture: let $\theta \in {\mathbb{R}}\setminus\pi\mathbb{Q}$, $t=4\cos^2(\theta)$ and consider the varieties $Z(g,t)$. Since for ${\lambda}\in Z(g,t)$, $\rho_{\lambda}$ is not discrete, the bifurcation locus of ${\left\{\rho_{\lambda}, {\lambda}\in Z(g,t)\right\}}$ is equal to $Z(g,t)$. Hence $\operatorname{Supp}({T_{\mathrm{bif}}}\wedge [Z(g,t)]) = Z(g,t)$, which by equidistribution of $Z(g,t)$ makes the equality $\operatorname{Supp}({T_{\mathrm{bif}}}^2) = \operatorname{Supp}({T_{\mathrm{bif}}})$ reasonable.
It is also natural to look for equidistribution in higher codimension. Here is a specific question:
Let $(G, \rho, \mu)$ be a family of representations satisfying (R1-3) and (A1-2). Assume that $\dim({\Lambda})\geq 3$. Given a generic element $h\in {\mathrm{PSL}(2,\mathbb C)}$, and a $\mu^{\mathbb{N}}$ generic sequence $(g_n)$, do the solutions of the equation $g_1\cdots g_n = h$ equidistribute (after convenient normalization) towards ${T_{\mathrm{bif}}}^3$?
Further settings, final remarks {#sec:further}
===============================
In this section we gather some speculations about possible extensions of the results presented in the paper.
Holomorphic dynamics in higher dimension {#subs:higher}
----------------------------------------
It is likely that a substantial part of the theory of bifurcation currents for rational maps on ${{\mathbb{P}^1}}$ should remain true in higher dimension, nevertheless little has been done so far.
Let us first discuss the case of polynomial automorphisms of ${{{\mathbb{C}}^2}}$. A polynomial automorphism $f$ of degree $d$ of ${{{\mathbb{C}}^2}}$ with non-trivial dynamics admits a unique measure of maximal entropy, which has two (complex) Lyapunov exponents of opposite sign $\chi^+(f)>0>\chi^-(f)$ and describes the asymptotic distribution of saddle periodic orbits [@bls; @bls2]. See [@cantat; @survey] in this volume for a presentation of these results for automorphisms of compact complex surfaces. Notice that a polynomial automorphism has constant jacobian, so $\chi^+ (f)+ \chi^-(f) = \log{\left\vert\mathrm{Jac}(f)\right\vert}$ is a pluriharmonic function on parameter space. It is not difficult to see that the function $f\mapsto \chi^+(f)$ is psh (in particular upper semi-continuous), and it was shown in [@lyapunov] that is actually continuous (even for families degenerating to a one-dimensional map).
Since the Lyapunov exponents are well approximated by the multipliers of saddle orbits [@bls2], it follows that near any point in parameter space where $f\mapsto \chi^+(f)$ is not pluriharmonic, complicated bifurcations of saddle points occur. In the dissipative case they must become attracting. The main idea of Theorem \[thm:basber\] seems robust enough to enable some generalization to this setting.
On the other hand, a basic understanding of the phenomena responsible for the bifurcations of a family of polynomial automorphisms of ${{{\mathbb{C}}^2}}$ –e.g. the role of homoclinic tangencies– is still missing (see [@bsR] for some results in a particular case). In particular no reasonable analogue of Theorem \[thm:mss\] is available for the moment. Therefore it seems a bit premature to hope for a characterization of the support of the bifurcation current $dd^c\chi^+$, let alone $(dd^c\chi^+)^p$.
The situation is analogous in the case of families of holomorphic endomorphisms of ${{\mathbb{P}^k}}$ (and more generally for families of polynomial-like mappings in higher dimension). The regularity properties of the Lyapunov exponent(s) function(s) are rather well understood, due to the work of Dinh-Sibony [@ds-pl] and Pham [@pham] (a good account on this is in [@ds-survey §2.5]). In particular it is known that the sum $L_p(f)$ of the $p$ largest Lyapunov exponents of the maximal entropy measure is psh for $1\leq p\leq k$ and the sum of all Lyapunov exponents is Hölder continuous. It is also known [@bdm] that $L_p(f)$ is well approximated by the corresponding quantity evaluated at repelling periodic cycles, so that any point in parameter space where $L_p$ is not pluriharmonic is accumulated by bifurcations of periodic points. Notice that the relationship between the currents $dd^cL_p$ is unclear.
Again, one may reasonably hope for equidistribution results in the spirit of Theorem \[thm:basber\].
Another interesting point is a formula, given in [@bas-ber1], for the sum $L_k$ of Lyapunov exponents of endomorphisms of ${{\mathbb{P}^k}}$ which generalizes Przytycki’s formula . From this formula one may expect to reach some understanding on the role of the critical locus towards bifurcations.
Cocycles
--------
Yoccoz suggests in [@yoccoz] to study the geography of the (finite dimensional) space of locally constant $\mathrm{SL}(2,{\mathbb{R}})$ cocycles over a transitive subshift of finite type in the same way as spaces of one-dimensional holomorphic dynamical systems, with some emphasis on the description of hyperbolic components and their boundaries. For ${\mathrm{SL}(2,\mathbb C)}$ cocycles (and more generally for any cocycle with values in a complex Lie subgroup of $\mathrm{GL}(n,{\mathbb{C}})$) we have an explicit connection with holomorphic dynamics given by the bifurcation currents. Indeed, locally constant cocycles over a subshift are generalizations of random products of matrices, which correspond to cocycles over the full shift. In this situation we can define a Lyapunov exponent function relative to a fixed measure on the base dynamical system (the Parry measure is a natural candidate), and construct a bifurcation current by taking the $dd^c$.
Notice that the subharmonicity properties of Lyapunov exponents are frequently used in this area of research (an early example is [@herman]).
For a general holomorphic family of (say, locally constant) ${\mathrm{SL}(2,\mathbb C)}$ cocycles over a fixed subshift of finite type, one may ask the same questions as in Section \[sec:kleinbif\]: characterize the support of the bifurcation current, prove equidistribution theorems. Another interest of considering this setting is that it is somehow a simplified model of the tangent dynamics of 2-dimensional diffeomorphisms, so it might provide some insight on the bifurcation theory of those. In particular there is an analogue of heteroclinic tangencies in this setting (“heteroclinic connexions"), and it might be interesting to study the distribution of the corresponding parameters.
Random walks on other groups
----------------------------
Another obvious possible generalization of Section \[sec:kleinbif\] is the study of bifurcation currents associated to holomorphic families of finitely generated subgroups of $\mathrm{SL}(n,{\mathbb{C}})$. Again, if $(\rho_{\lambda})$ is a holomorphic family of strongly irreducible representations (see [@furman] for the definition) of a finitely generated group $G$ endowed with a probability measure satisfying (A1-2), then Definition \[def:bifcur\] makes sense, with $\chi$ being the top Lyapunov exponent. It is likely that equidistribution theorems for representations possessing an element of given trace should follow as in §\[subs:trace\]. More generally, one may investigate the distribution of representations with an element belonging to a given hypersurface of $\mathrm{SL}(n,{\mathbb{C}})$, in the spirit of Section \[sec:prologue\].
On the other hand, for the same reasons as in §\[subs:higher\], the characterization of the support of the bifurcation current is certainly a more challenging problem.
Non-archimedian dynamics
------------------------
It is a standard fact in algebraic geometry that studying families $(X_{\lambda})_{{\lambda}\in {\Lambda}}$ of complex algebraic varieties often amounts to studying varieties over a field extension of ${\mathbb{C}}$, that is a function field in the variable ${\lambda}$. The same idea applies in the dynamical context and was explored by several authors. This fact was used notably by Culler, Morgan and Shalen [@culler; @shalen; @morgan; @shalen] to construct compactifications of spaces of representations into ${\mathrm{PSL}(2,\mathbb C)}$ and obtain new results on the geometry of 3-manifolds. In rational dynamics Kiwi [@kiwi; @puiseux; @kiwi; @rational] used a similar construction to study the behaviour at infinity of families of cubic polynomials or quadratic rational maps (see [@demarco; @mcmullen] for a different approach to this problem).
It would be natural to explore the interaction of bifurcation currents with these compactifications, as well as the general bifurcation theory of non-archimedian rational dynamical systems.
[\[ABCD\]]{} Aoun, Richard. [*Random subgroups of linear groups are free.*]{} Duke Math. J. 160 (2011), 117–173. Baker, Matthew; Hsia, Liang-Chung. *Canonical heights, transfinite diameters, and polynomial dynamics.* J. Reine Angew. Math. 585 (2005), 61–92.
Baker, Matthew; DeMarco, Laura. [*Preperiodic points and unlikely intersections.*]{} Duke Math. J. 159 (2011), no. 1, 1–29.
Bassanelli, Giovanni; Berteloot, Fran[ç]{}ois. *Bifurcation currents in holomorphic dynamics on ${\bf
P}^k$.* J. Reine Angew. Math. 608 (2007), 201–235.
Bassanelli, Giovanni; Berteloot, Fran[ç]{}ois. *Bifurcation currents and holomorphic motions in bifurcation loci.* Math. Ann. 345 (2009), 1–23.
Bassanelli, Giovanni; Berteloot, Fran[ç]{}ois. *Distribution of polynomials with cycles of a given multiplier.* Nagoya Math. J. 201 (2011), 23–43.
Beardon, Alan F.[*The geometry of discrete groups.*]{} Graduate Texts in Mathematics, 91. Springer-Verlag, New York, 1983.
Bedford, Eric; Lyubich, Mikhail; Smillie, John. [*Polynomial diffeomorphisms of ${\mathbb{C}}^ 2$. IV. The measure of maximal entropy and laminar currents.*]{} Invent. Math. 112 (1993), 77-125.
Bedford, Eric; Lyubich, Mikhail; Smillie, John. [*Distribution of periodic points of polynomial diffeomorphisms of ${{{\mathbb{C}}^2}}$.*]{} Invent. Math. 114 (1993), 277-288.
Bedford, Eric; Smillie, John. [*Real polynomial diffeomorphisms with maximal entropy: Tangencies.*]{} Ann. of Math. (2) 160 (2004), no. 1, 1–26.
Bedford, Eric; Taylor, B. Alan. *A new capacity for plurisubharmonic functions.* Acta Math. 149 (1982), no. 1-2, 1–40.
Bers, Lipman. [*Holomorphic families of isomorphisms of Möbius groups.*]{} J. Math. Kyoto Univ. 26 (1986), no. 1, 73–76.
Berteloot, François. [*Lyapunov exponent of a rational map and multipliers of repelling cycles.*]{} Riv. Math. Univ. Parma, 1(2) (2010), 263-269.
Berteloot, François. [*Bifurcation currents in one-dimensional holomorphic dynamics.*]{} CIME Lecture notes, 2011.
Berteloot, François; Dupont, Christophe; Molino, Laura. [*Normalization of bundle holomorphic contractions and applications to dynamics.*]{} Ann. Inst. Fourier (Grenoble) 58 (2008), no. 6, 2137–2168.
Bielefeld, Ben; Fisher, Yuval; Hubbard, John H. *The classification of critically preperiodic polynomials as dynamical systems.* J. Amer. Math. Soc. 5 (1992), no. 4, 721–762.
Bonifant, Araceli; Kiwi, Jan; Milnor, John. [*Cubic polynomial maps with periodic critical orbit. II. Escape regions.*]{} Conform. Geom. Dyn. 14 (2010), 68–112.
Bougerol, Philippe; Lacroix, Jean. [*Products of random matrices with applications to Schrödinger operators.*]{} Progress in Probability and Statistics, 8. Birkhäuser Boston, Inc., Boston, MA, 1985.
Branner, Bodil. in [*Topological methods in modern mathematics (Stony Brook, NY, 1991)*]{}, 391–427, Publish or Perish, Houston, TX, 1993.
Branner, Bodil; Hubbard, John H. *The iteration of cubic polynomials. I. The global topology of parameter space.* Acta Math. 160 (1988), no. 3-4, 143–206.
Branner, Bobil; Hubbard, John H. *The iteration of cubic polynomials. II. Patterns and parapatterns.* Acta Math. 169 (1992), no. 3-4, 229–325.
Briend, Jean-Yves; Duval, Julien. [*Exposants de Liapounoff et distribution des points périodiques d’un endomorphisme de $\mathbb{C}\mathbb{P}^{k}$.* ]{} Acta Math. 182, 143–157, (1999).
Brolin, Hans. *Invariant sets under iteration of rational functions.* Ark. Mat. 6 1965 103–144 (1965).
Buff, Xavier; Epstein, Adam. [*Bifurcation measure and postcritically finite rational maps.*]{} Complex dynamics, 491–512, A K Peters, Wellesley, MA, 2009.
Buff, Xavier; Gauthier, Thomas. [*Perturbations of flexible Lattès maps.*]{} Preprint, arxiv:1111.5451.
Cantat, Serge. [*Bers and Hénon, Painlevé and Schrödinger.*]{} Duke Math. J. 149 (2009), 411–460.
Cantat, Serge. [*Dynamics of automorphisms of compact complex surfaces*]{}. Preprint (2011).
Chirka, Evgueny M. [*Complex analytic sets.*]{} Mathematics and its Applications (Soviet Series), 46. Kluwer Academic Publishers Group, Dordrecht, 1989.
Culler, Marc; Shalen, Peter B. [*Varieties of group representations and splittings of $3$-manifolds.*]{} Ann. of Math. (2) 117 (1983), no. 1, 109–146.
Demailly, Jean-Pierre. [*Complex analytic and differential geometry, Chap. III.*]{} Book available online at [http://www-fourier.ujf-grenoble.fr/ demailly/manuscripts/agbook.pdf]{}.
DeMarco, Laura. *Dynamics of rational maps: a current on the bifurcation locus.* Math. Res. Lett. 8 (2001), no. 1-2, 57–66.
DeMarco, Laura. *Dynamics of rational maps: Lyapunov exponents, bifurcations, and capacity.* Math. Ann. 326 (2003), no. 1, 43–73.
DeMarco, Laura; McMullen, Curtis.[*Trees and the dynamics of polynomials.*]{} Ann. Sci. Éc. Norm. Supér. (4) 41 (2008), no. 3, 337–382..
Deroin, Bertrand; Dujardin, Romain. [*Random walks, Kleinian groups, and bifurcation currents.*]{} Invent. Math. (to appear).
Deroin, Bertrand; Dujardin, Romain. [*Lyapunov exponents for surface group representations.* ]{} In preparation (2011).
De Th[é]{}lin, Henry. Ann. Sci. [É]{}cole Norm. Sup. (4) 37 (2004), 304–311.
Dinh, Tien Cuong; Sibony, Nessim. [*Dynamique des applications d’allure polynomiale.* ]{} J. Math. Pures Appl. (9) 82 (2003), 367-423.
Dinh, Tien Cuong; Sibony, Nessim. [*Distribution des valeurs de transformations méromorphes et applications.*]{} Comment. Math. Helv. 81 (2006), no. 1, 221–258.
Dinh, Tien Cuong; Sibony, Nessim. [*Super-potentials for currents on compact Kähler manifolds and dynamics of automorphisms.*]{} J. Algebraic Geom. 19 (2010), no. 3, 473–529.
Douady, Adrien; Hubbard, John H. Publications Mathématiques d’Orsay. Université de Paris-Sud, Orsay, 1985.
Dujardin, Romain. [*Continuity of Lyapunov exponents for polynomial automorphisms of $\mathbb{C}^2$*]{}. Ergodic Theory and Dynamical Systems 27 (2007), 1111-1133.
Dujardin, Romain. [*Cubic polynomials: a measurable view on parameter space.*]{} [*in*]{} Complex Dynamics and Friends, A K Peters, Wellesley, MA, 2009.
Dujardin, Romain. [*Fatou directions along the Julia set for endomorphisms of $\mathbf{CP}^k$.*]{} Preprint (2010).
Dujardin, Romain; Favre, Charles. [*Distribution of rational maps with a preperiodic critical point*]{}. Amer. J. Math. 130 (2008), 979-1032.
Dujardin, Romain; Guedj, Vincent. [*Homogeneous Monge-Ampère equations and foliations by holomorphic disks*]{}. in [*Complex Monge-Ampère equations and geodesics in the space of Kähler metrics*]{}, Lecture Notes in Mathematics (2038), to appear.
Dumas, David. [*Complex projective structures.*]{} Handbook of Teichmüller theory. Vol. II, 455–508, Eur. Math. Soc., Zürich, 2009.
Favre Charles; Rivera-Letelier, Juan. *[É]{}quidistribution quantitative des points de petite hauteur sur la droite projective.* Math. Ann. 335 (2006), 311–361.
Fornaess, John Erik; Sibony, Nessim [*Complex dynamics in higher dimension. [II]{}*]{}, Modern methods in complex analysis, Ann. of Math. Studies 135–182. Princeton Univ. Press.
Freire, Alexandre; Lopes, Artur; Mañé, Ricardo [*An invariant measure for rational maps.*]{} Bol. Soc. Brasil. Mat. 14 (1983), 45–62.
Fulton, William, [*Intersection theory.*]{} Second edition. Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics. Springer-Verlag, Berlin, 1998.
Furman, Alex. [*Random walks on groups and random transformations.*]{} Handbook of dynamical systems, Vol. 1A, 931–1014, North-Holland, Amsterdam, 2002.
Furstenberg, Hillel [*Noncommuting random products.*]{} Trans. Amer. Math. Soc. 108 (1963) 377–428.
Furstenberg, Hillel. [*Random walks and discrete subgroups of Lie groups.*]{} 1971 Advances in Probability and Related Topics, Vol. 1 pp. 1–63 Dekker, New York
Furstenberg, Hillel; Kesten, Harry. [*Products of random matrices.*]{} Ann. Math. Statist. 31 (1960), 457–469.
Gauthier, Thomas. Preprint, arxiv:1103.2656.
Goldman, William M. [*Mapping class group dynamics on surface group representations.*]{} Problems on mapping class groups and related topics, 189–214, Proc. Sympos. Pure Math., 74, Amer. Math. Soc., Providence, RI, 2006
Graczyk, Jacek; Światek, Gregorz. *Harmonic measure and expansion on the boundary of the connectedness locus.* Invent. Math. 142 (2000), no. 3, 605–629.
Guivarc’h, Yves. [*Produits de matrices aléatoires et applications aux propriétés géométriques des sous-groupes du groupe linéaire.*]{} Ergodic Theory Dynam. Systems 10 (1990), no. 3, 483–512.
Herman, Michael. [*Une méthode pour minorer les exposants de Lyapounov et quelques exemples montrant le caractère local d’un théorème d’Arnol’d et de Moser sur le tore de dimension $2$*]{} Comment. Math. Helv. 58 (1983), no. 3, 453–502.
H[ö]{}rmander, Lars. [*Notions of convexity.*]{} Progress in Math 127. Birkh[ä]{}user, Boston, MA, 1994.
Ivashkovich, Sergey; Neji, Fethi. [*Weak normality of families of meromorphic mappings and bubbling in higher dimensions.*]{} Preprint (2011), arxiv:1104:3973.
Kaimanovich, Vadim A. [*Double ergodicity of the Poisson boundary and applications to bounded cohomology.*]{} Geom. Funct. Anal. 13 (2003), no. 4, 852–861.
Kapovich, Michael [*Hyperbolic manifolds and discrete groups.*]{} Progress in Mathematics, 183. Birkhäuser Boston, Inc., Boston, MA, 2001.
Kiwi, Jan. *Combinatorial continuity in complex polynomial dynamics.* Proc. London Math. Soc. (3) 91 (2005), no. 1, 215–248.
Kiwi, Jan. *Puiseux series, polynomial dynamics, and iteration of complex cubic polynomials* Ann. Inst. Fourier (Grenoble) 56 (2006), no. 5, 1337–1404.
Kiwi, Jan. *Puiseux series dynamics of quadratic rational maps.* Preprint (2011) arXiv:1106.0059.
Ledrappier, François. [*Quelques propriétés ergodiques des applications rationnelles.*]{} C. R. Acad. Sci. Paris Sér. I Math. 299 (1) (1984), 37–40.
Le Page, Émile. [*Théorèmes limites pour les produits de matrices aléatoires.*]{} In “Probability measures on groups”, ed. H. Heyer. Lecture Notes in Math. no. 928, (1982), 258–303.
Levin, Guennadi. Russian Math. Surveys 36:6 (1981), 189–190.
Levin, Gennadi M. *On the theory of iterations of polynomial families in the complex plane.* J. Soviet Math. 52 (1990), no. 6, 3512–3522.
Levy, Alon. [*The space of morphisms on projective space.*]{} Preprint, arxiv:0903.1318.
Lubotzky, Alex. [*Dynamics of $\mathrm{Aut}(F_n)$ actions on group presentations and representations.*]{} to appear in [ Geometry, Rigidity, and Group Actions. A Festschrift in honor of Robert Zimmer’s 60th birthday.]{} (edited by B. Farb and D. Fisher) Chicago Univ. Press (2011)
Lyubich, Mikhail [*Entropy properties of rational endomorphisms of the Riemann sphere.*]{} Ergodic Theory Dynam. Systems 3 (1983), 351-385.
Lyubich, Mikhail Yu. Russian Math. Surveys 38:5 (1983), 154–155.
Ma[ñ]{}[é]{}, Ricardo. Dynamical systems, Valparaiso 1986, 86-117, Lecture Notes in Math., 1331, Springer, Berlin, 1988.
Manning, Anthony. Ann. of Math. (2) 119 (1984), 425-430.
Ma[ñ]{}[é]{}, Ricardo; Sad, Paulo; Sullivan, Dennis. *On the dynamics of rational maps.* Ann. Sci. [É]{}cole Norm. Sup. (4) 16 (1983), 193–217.
McMullen, Curtis T. Prépublication (1985).
C.T. McMullen. [*Families of rational maps and iterative root-finding algorithms.*]{} Ann. of Math. 125 (1987), 467–493.
McMullen, Curtis T. Ann. of Math. (2) 133 (1991), 217–247.
McMullen, Curtis T. [*Renormalization and 3-manifolds which fiber over the circle.*]{} Annals of Mathematics Studies, 142. Princeton University Press, Princeton, NJ, 1996.
McMullen, Curtis T.; Sullivan, Dennis P. Adv. Math. 135 (1998), 351–395.
Milnor, John. [*Remarks on iterated cubic maps.*]{} Experiment. Math. 1 (1992), no. 1, 5–24.
Milnor, John. [*Geometry and dynamics of quadratic rational maps.*]{} With an appendix by the author and Lei Tan. Experiment. Math. 2 (1993), no. 1, 37–83.
Milnor, John. [ *Hyperbolic components in spaces of polynomial maps.*]{} With an appendix by Alfredo Poirier. Stony Brook preprint, 1992.
Milnor, John. [*On rational maps with two critical points.*]{} Experiment. Math. 9 (2000), no. 4, 481–522.
Milnor, John. [*Cubic polynomial maps with periodic critical orbit. I.*]{} Complex dynamics, 333–411, A K Peters, Wellesley, MA, 2009.
Morgan, John W.; Shalen, Peter B. [ *Valuations, trees, and degenerations of hyperbolic structures. I.*]{} Ann. of Math. (2) 120 (1984), no. 3, 401–476.
Pham, Ngoc-Mai. *Lyapunov exponents and bifurcation current for polynomial-like maps.* Preprint (2005), arxiv math:0512557.
Przytycki, Feliks. Invent. Math. 80 (1985), 161-179.
Rees, Mary. [*Positive measure sets of ergodic rational maps.* ]{} Ann. Sci. École Norm. Sup. (4) 19 (1986), 383–407.
Russakovskii, Alexander; Shiffman, Bernard. [*Value distribution for sequences of rational mappings and complex dynamics.*]{} Indiana Univ. Math. J. 46 (1997), 897-932.
Russakovskii, Alexander; Sodin, Mikhail. [ *Equidistribution for sequences of polynomial mappings.*]{} Indiana Univ. Math. J. 44 (1995), no. 3, 851–882.
Shishikura, Mitsuhiro. [*The Hausdorff dimension of the boundary of the Mandelbrot set and Julia sets.*]{} Ann. of Math. (2) 147 (1998), no. 2, 225–267.
Sibony, Nessim. Lecture notes (unpublished), UCLA, 1984.
Sibony, Nessim, Wong, Pit Mann. [*Some results on global analytic sets.*]{} Séminaire Lelong-Skoda (Analyse). Années 1978/79, pp. 221–237, Lecture Notes in Math., 822, Springer, Berlin, 1980.
Silverman, Joseph H. [*The arithmetic of dynamical systems.*]{} Graduate Texts in Mathematics, 241. Springer, New York, 2007.
Smirnov, Stanislav. *Symbolic dynamics and Collet-Eckmann conditions.* Internat. Math. Res. Notices 2000, no. 7, 333–351.
Sullivan, Dennis. [*Quasiconformal homeomorphisms and dynamics. I. Solution of the Fatou-Julia problem on wandering domains.*]{} Ann. of Math. (2) 122 (1985), no. 3, 401–418.
Sullivan, Dennis. [*Quasiconformal homeomorphisms and dynamics. II. Structural stability implies hyperbolicity for Kleinian groups.* ]{} Acta Math. 155 (1985), no. 3-4, 243–260.
Tan, Lei. [*Hausdorff dimension of subsets of the parameter space for families of rational maps. (A generalization of Shishikura’s result).*]{} Nonlinearity 11 (1998), no. 2, 233–246.
Ueda, Tetsuo. [*Fatou sets in complex dynamics on projective spaces.*]{} J. Math. Soc. Japan 46 (1994), 545–555.
van Strien, Sebastian [*Misiurewicz maps unfold generically (even if they are critically non-finite).* ]{} Fund. Math. 163 (2000), no. 1, 39–54.
Yoccoz, Jean-Christophe [*Some questions and remarks about ${\rm SL}(2,{\mathbb{R}})$ cocycles.*]{} Modern dynamical systems and applications, 447–458, Cambridge Univ. Press, Cambridge, 2004.
Zakeri, Saeed. *Dynamics of cubic Siegel polynomials.* Comm. Math. Phys. 206 (1999), 185–233.
Zdunik, Anna. [*Parabolic orbifolds and the dimension of the maximal measure for rational maps.*]{} Invent. Math. 99 (3) (1990), 627–649.
[^1]: In this respect it is instructive to compare this with the more topological point of view of Branner and Hubbard, which was summarized 20 years ago by Branner in [@branner-turning].
[^2]: This is where we use the assumption that $\dim({\Lambda})= 1$. In higher dimension, to estimate this volume one needs to integrate the exterior power $(\pi_1^*\beta + \pi_2^*\omega)^{d}$ where $d=\dim({\Lambda})$.
[^3]: The expression of DeMarco’s formula for a rational map $f$ is of the form $\chi(f) = \sum_{i=1}^{2d-2} G_F(c_j) + H(f)$, where $G_F$ is the dynamical Green’s function of a homogeneous lift $F:{{{\mathbb{C}}^2}}{\rightarrow}{{{\mathbb{C}}^2}}$ of $f$, the $c_j$ are certain lifts of the critical points, and $H(f)$ depends pluriharmonically on $f$
[^4]: The additional ${\frac{1}{n}}$ in that formula follows from the fact that in the sum is over periodic points while here we sum over periodic cycles.
[^5]: The is the usual terminology in intersection theory, see [@fulton; @chirka].
[^6]: A [*flexible Lattès map*]{} is a rational mapping descending from an integer multiplication on elliptic curve. These can be deformed with the elliptic curve, hence the terminology.
[^7]: Notice that this is weaker than the notion considered in [@mcms] (stability over the whole Riemann sphere), which introduces some distinctions which are not relevant from our point of view, like distinguishing the center from the other parameters in a hyperbolic component.
[^8]: We do not address the problem of the global holonomy of these laminations, which gives rise to interesting phenomena [@branner-turning].
[^9]: Milnor discusses in [@milnor; @cubic §3] the possibility of so-called “product configurations" in the connectedness locus of real cubic polynomials. Our result actually asserts that in the complex setting such configurations cannot be of positive ${\mu_{\mathrm{bif}}}$ measure.
[^10]: or $\mathrm{SO}(3,{\mathbb{R}})$ if we view $\mathbb{H}^3$ in its ball model.
[^11]: In order to study the space of all representations of $G$ to ${\mathrm{PSL}(2,\mathbb C)}$, it is nevertheless interesting to understand which results remain true when allowing a proper subset of elementary representations. This issue is considered in [@kleinbif], but here for simplicity we only work with non-elementary representations.
[^12]: As stated here, the result is true only when $\mu$ is invariant under $g\mapsto g{^{-1}}$. The general case needs some adaptations. Notice also that the Kaimanovich theorem can be viewed as a far reaching generalization of the ergodicity of the geodesic flow for manifolds of constant negative curvature.
[^13]: To view ${\mathrm{PSL}(2,\mathbb C)}$ as a set of matrices, observe that ${\mathrm{PSL}(2,\mathbb C)}$ is isomorphic to $\mathrm{SO}(3,{\mathbb{C}})$ by the adjoint representation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In Nevzorov’s $F^\alpha$-scheme, one deals with a sequence of independent random variables whose distribution functions are all powers of a common continuous distribution function. A key property of the $F^\alpha$-scheme is that the record indicators for such a sequence are independent. This allows one to obtain several important limit theorems for the total number of records in the sequence up to time $n\to\infty$. We extend these theorems to a much more general class of sequences of random variables obeying a “threshold $F^\alpha$-scheme" in which the distribution functions of the variables are close to the powers of a common $F$ only in their right tails, above certain non-random non-decreasing threshold levels. Of independent interest is the characterization of the growth rate for extremal processes that we derived in order to be able to verify the conditions of our main theorem. We also establish the asymptotic pair-wise independence of record indicators in a special case of threshold $F^\alpha$-schemes.
records, maxima of random variables, extremal process, growth rate, $F^\alpha$-scheme, almost sure behavior.
60G70, 60F05, 60F20.
author:
- Patrick He$^1$ and Konstantin Borovkov$^2$
title: 'Limit theorems for record indicators in threshold $F^{\alpha}$-schemes'
---
Introduction and main results
=============================
Let $\boldsymbol{X}:=\{X_n\}_{n\ge 1}$ be a sequence of random variables (r.v.’s) on a common probability space, $M_n:= \bigvee_{1\le k\le n}X_k,$ $n\ge 1,$ be the sequence of the partial maxima of these r.v.’s, and $I_1:=1,$ $$\quad I_n:= {{\bf 1}}(X_n> M_{n-1}), \quad n\ge 2,$$ the (upper) record indicators for $\boldsymbol{X}$. Denote by $N_n:=\sum_{k=1}^n I_k$ the total number of records in $\boldsymbol{X}$ up to time $n\ge 1.$ Apart from the natural motivation related to the theory of records, the study of the distribution of $N_n$ is also of interest for other applications, e.g. in connection with the secretary problem [@Pf89] or for the linear search problem of the maximum element in a field of $n$ entries, where $N_n$ denotes the number of re-storages during the procedure (for details see e.g. [@Ke85; @Pf91]). Another field of relevance one might mention is the average-case analysis of the simplex method in linear programming [@DePf87; @Ro82].
For an outline of the history of the theory of records, we refer the reader to Appendix 1 in [@Ne00]. For further detail, see also [@Ne88; @NeAh15] and the bibliography therein. The most studied case is, of course, when the $X_n$’s are independent and identically distributed (i.i.d.) with a continuous distribution function (d.f.) $F$. By the Dwass–Rényi theorem (see e.g. Ch. 3 in [@NeAh15]), the record indicators for such sequences $\boldsymbol{X}$ are jointly independent with ${{\bf P}}(I_n=1)= n^{-1},$ $n\ge 1.$ The independence property enables one to establish a number of limit theorems for the distribution of $N_n$, including the Poisson and normal approximations, the respective convergence rates in the uniform norm for the d.f.’s being the rather slowly decaying $O(\ln^{-3/2}n)$ and $O(\ln^{-1/2}n)$ as $n\to\infty$. One should note, however, that the former approximation can be dramatically improved by a remarkably simple “adjusted Poisson approximation" from [@BoPf96] with a convergence rate of the form $O(n^{-2})$ (with an explicit bound for the constant).
It was discovered in [@Ne81; @Ne85] that the important independence property for the record indicators also holds for a special class of scenarios nowadays referred to as the (Nevzorov) “$F^{\alpha}$-scheme". In that scheme, the $X_n$’s are independent r.v.’s following the respective distribution functions $F^{{\alpha}_n}$, $n\ge 1,$ with a common continuous d.f. $F,$ ${\boldsymbol{\alpha}}:=\{{\alpha}_n \}_{n\ge 1}$ being an arbitrary positive sequence (a special case of the $F^{\alpha}$-scheme where the ${\alpha}_n$’s are integers had been earlier analyzed in [@Ya75]). This scheme plays an important role in the present paper, so it will be convenient for us to adopt a special notation for the related r.v.’s to distinguish them from the ones for the original $\boldsymbol{X}$: we will use ${\mathcal{X}}_n, {\mathcal{M}}_n, $ ${\mathcal{I}}_n$ and ${\mathcal{N}}_n,$ respectively, for the independent r.v.’s in the $F^{\alpha}$-scheme with some $F$ and ${\boldsymbol{\alpha}}$ (setting $\boldsymbol{{\mathcal{X}}}:=\{{\mathcal{X}}_n\}_{n\ge 1}$), their partial maxima, record indicators and record counts. It turned out that the record indicators ${\mathcal{I}}_1,{\mathcal{I}}_2,\ldots $ in the case of the $F^{\alpha}$-scheme form a sequence of independent r.v.’s with $$\label{a/s}
p_n:={{\bf P}}({\mathcal{I}}_n=1)=\frac{{\alpha}_n}{s_n },\quad s_n :=\sum_{k=1}^n {\alpha}_k,\quad n\ge 1$$ (see e.g. p. 217 in [@NeAh15] and [@BoPf95]). This result was also demonstrated in [@BaRe87] using a natural embedding of the sequence of partial maxima ${\mathcal{M}}_n$ in the so-called extremal process, yielding the additional fact that $$\label{I_M}
\mbox{$ {\mathcal{M}}_n $ is independent of ${\mathcal{I}}_1,\ldots,{\mathcal{I}}_n$, \ $n\ge 1.$ }$$
Moreover, it turned out that the $F^{\alpha}$-scheme is basically the only situation with the original $X_n$’s being independent where the r.v.’s $I_n$ and $M_n$ are independent of each other for any $n\ge 1$ (Theorem 3 in [@BoPf95]).
The independence of the record indicators in the $F^{\alpha}$-scheme enables one to establish a number of asymptotic results for the behavior of ${\mathcal{N}}_n$ as $n\to\infty.$ We will summarize the key ones. Note that $E_n:={{\bf E}}{\mathcal{N}}_n=\sum_{k=1}^n p_k,$ $V_n:={\mbox{\rm Var}\, }({\mathcal{N}}_n)=\sum_{k=1}^n p_k (1-p_k),$ $n\ge 1.$
1. First of all, ${\mathcal{N}}_n\to \infty $ a.s. as $n\to \infty $ iff condition
[[**\[C$_{\boldsymbol{1}}$\]**]{}]{} $\lim_{n\to \infty }s_n =\infty$
is met [@Ne85].
This condition will be assumed to be satisfied throughout this paper (and, in particular, in assertions (A2)–(A4) in this list). Note that under condition [[**\[C$_{\boldsymbol{1}}$\]**]{}]{} both $E_n$ and $V_n$ tend to infinity as $n\to\infty$.
2. $\lim_{n\to\infty}{\mathcal{N}}_n/E_n=1$ a.s. (Theorem 1 in [@DoKlPaSt13]).
3. $\lim_{n\to\infty}({\mathcal{N}}_n-E_n)/\ln s_n =0$ a.s. If $p_n\to 0 $ then $\lim_{n\to\infty} {\mathcal{N}}_n /\ln s_n =1$ a.s., while if $p_n\to 1 $ then $\lim_{n\to\infty} {\mathcal{N}}_n /\ln s_n =0$ a.s. [@DoKlSt15].
4. If $\lim_{n\to \infty } V_n<\infty$ then ${\mathcal{N}}_n-E_n$ converges a.s. to a proper r.v. as $n\to\infty,$ while if $\lim_{n\to \infty } V_n=\infty$ then $V_n^{-1/2} ({\mathcal{N}}_n-E_n) {\stackrel{d}{\longrightarrow}}Z\sim N(0,1)$ (Theorem 7 in [@DoKlPaSt13]).
In the general case, even when the $X_n$’s are independent, the nature of dependence between record indicators is very complicated and difficult to describe, so obtaining results similar to (A1)–(A4) appears to be a rather hard task. We will note, however, the remarkable results on coupling of the record times and values for a class of strictly stationary sequences $\boldsymbol{X}$ with “time-shifted" record times and values, respectively, for sequences of i.i.d. r.v.’s with the same univariate marginal distributions as for $\boldsymbol{X}$, see [@Ha87; @HaMaNePu98] and the references in the latter paper. Such couplings imply, in particular, that the above assertions (A1)–(A.4) will hold for such stationary sequences as well, with the quantities $E_n,$ $V_n$ corresponding to the i.i.d. case. Any advances extending the limit theory for records beyond such special cases remain to be highly desirable.
The key observation that led us to writing this paper was, roughly speaking, that, for large $n$ values, for the observation $X_n$ to be a record it needs to be “large" as it has to exceed the previous partial maximum value $M_{n-1}$ that is likely to already be “large". Therefore, if, for $n\ge 1,$ the d.f. of $X_n$ is equal (or close) to the respective $F^{{\alpha}_n }$ from an $F^{{\alpha}}$-scheme $\boldsymbol{{\mathcal{X}}}$ in its “right tail" only, i.e. above a certain non-random threshold $\ell_n$, then one can still expect the record indicators to display an asymptotic behavior close to that of the record indicators for $\boldsymbol{{\mathcal{X}}}$. Moreover, one can relax the independence assumption as well, since the nature of dependence between the $X_n$’s on the event where the observations are unlikely to be records would be of little relevance. The respective result, stated as Theorem \[th1\] below, is the first main contribution of this paper. The key condition in the theorem is that eventually ${\mathcal{M}}_n> \ell_n$ (meaning that ${\mathcal{M}}_n > \ell_n$ for all sufficiently large $n$ and abbreviated as “${\mathcal{M}}_n> \ell_n$ ev.") a.s.
Our second contribution concerns the question of when that key condition is met. We establish the a.s. rate of growth of the sequence $\{{\mathcal{M}}_n\}$, providing criteria for ${{\bf P}}({\mathcal{M}}_n > \ell_n \ {\rm ev.})=1$. This result is stated in Corollary \[co2\] of Theorem \[th2\], the latter dealing with a similar question for the extremal process. This result is an extension of the work in [@Kl84; @Kl85] on such criteria in the case of i.i.d. sequences $\boldsymbol{X}$ and is of independent interest.
The third main contribution of this paper is Theorem \[th3\] below, which establishes the uniform asymptotic pairwise independence of the record indicators in the special case of the threshold $F^{\alpha}$-scheme where $\boldsymbol{X}$ consists of independent r.v.’s and there is a common threshold $\ell_n\equiv \ell$.
Now we will give a formal definition of the above-mentioned threshold $F^{\alpha}$-scheme and state our main results.
Denote by $F_n$ the d.f. of $X_n $ and by $$F_{n|x_1, \ldots, x_{n-1}}(x_n):=
{{\bf P}}(X_n\le x_n|X_1=x_1, \ldots, X_{n-1}= x_{n-1}), \quad x_j\in {\mathbb{R}}, \quad 1\le j\le n,$$ the conditional d.f. of $X_n$ given the values of the $n-1$ “preceding observations". We will use the standard notation $$G^\leftarrow (u):= \inf\{x\in{\mathbb{R}}: G(x)\ge u\}, \quad u\in (0,1),$$ for the generalized inverse function of the d.f. $G.$ Finally, we will denote by $G|_x$ the restriction of the distribution $G$ to the half-line $(x,\infty)$: $dG|_x(y):={{\bf 1}}(y>x)\, dG(y),$ $y\in {\mathbb{R}}.$
By a threshold $F^{\alpha}$-scheme we will mean any sequence $\boldsymbol{X}$ of r.v.’s satisfying the following condition:
[[**\[C$_{\boldsymbol{2}}$\]**]{}]{}
*There exist a continuous d.f. $F$, a positive sequence ${\boldsymbol{\alpha}}$ satisfying [[**\[C$_{\boldsymbol{1}}$\]**]{}]{} and a non-decreasing real sequence $\{\ell_n\}_{n\ge 1}$ such that$:$*
1. $F_n^{\leftarrow} (1-)\le F^{\leftarrow} (1-) ,$ $n\ge 1;$
2. $
F_{n|x_1, \ldots, x_{n-1}}(x_n)
= F_{n }(x_n) $ for $x_n \ge \ell_n,$ $ x_j\in {\mathbb{R}},$ $ 1\le j< n,$ $n\ge 2;$
3. $\sum_{n\ge 1}{\delta}_n< \infty,$ where $${\delta}_n:= d_{TV} (F_n|_{ \ell_n }, F^{{\alpha}_n}|_{ \ell_n }):=
\int_{(\ell_n,\infty)} |d(F_n- F^{{\alpha}_n})|$$ denotes the total variation distance between the restrictions of the distributions $F_n$ and $F^{{\alpha}_n}$ to the half-line $(\ell_n,\infty)$, and
4. ${{\bf P}}({\mathcal{M}}_n>\ell_n{\rm \ ev.})=1$ for the sequence $\{{\mathcal{M}}_n\}_{n\ge 1}$ of the partial maxima in the $F^{\alpha}$-scheme $\boldsymbol{{\mathcal{X}}}= \{{\mathcal{X}}_n\}_{n\ge 1}$ specified by $F$ and ${\boldsymbol{\alpha}}.$
That is, for any $n\ge 2,$ on the event $\{X_n>\ell_n\}$ the r.v. $X_n$ is independent of the observations $X_1, \ldots, X_{n-1}$ and the restriction of the distribution of $X_n$ to the half-line $(\ell_n,\infty)$ is close (in the total variation sense) to the restriction to $(\ell_n, \infty)$ of the law of the $n$th element of an $F^{{\alpha}}$-scheme that has the property that, with probability 1, its partial maxima ${\mathcal{M}}_n$ will eventually lie above the threshold values $\ell_n$.
Parts (ii) and (iii) of the above condition may seem quite strong. In fact, the distributional properties of the sequence of record indicators are very sensitive to changes in the distribution of the original sequence. Therefore it should not be surprising that, to obtain results at the level of (A1)–(A4), one would need to make relatively strong assumptions about $\boldsymbol{X}.$ What we would like to stress, though, is that those assumptions only need to be made about the right tails of the (conditional) distributions of the elements of $\boldsymbol{X}.$
Note that the assumption that $\{\ell_n\}_{n\ge 1}$ is a non-decreasing sequence does not actually restrict the generality, cf. Remark \[re1\] below. The purpose of part (i) of condition [[**\[C$_{\boldsymbol{2}}$\]**]{}]{} is to ensure that the partial maxima $M_n$ of the r.v.’s in the original sequence $\boldsymbol{X}$ cannot take values “beyond the reach" of the maxima in the $F^{\alpha}$-scheme $\boldsymbol{\mathcal{X}}$. Necessary and sufficient conditions for part (iv) to hold are established in our Corollary \[co2\] below.
A simple special case where condition [[**\[C$_{\boldsymbol{2}}$\]**]{}]{} is satisfied is a sequence $\boldsymbol{X}$ of independent r.v.’s such that $F_n(x) = F^{{\alpha}_n}(x)$ for $x> \ell_n$ (so that ${\delta}_n\equiv 0$), $n\ge 1$, for some continuous d.f. $F$ and positive sequence ${\boldsymbol{\alpha}}.$ A more interesting example of a situation where that condition can be satisfied is when $X_n= V_n \vee Y_n,$ $n\ge 1,$ under the assumptions that $ \{V_n\}_{n\ge 1}$ is an arbitrary sequence of bounded from above r.v.’s (${{\bf P}}(V_n\le \ell_n)=1$ for all $n\ge 1$), whereas the r.v.’s $Y_n$ are independent of each other and of $ \{V_n\}_{n\ge 1}$. As for further conditions on the $Y_n$’s, it suffices to assume that there exists an $F^{\alpha}$-scheme satisfying conditions [[**\[C$_{\boldsymbol{1}}$\]**]{}]{}, [[**\[C$_{\boldsymbol{2}}$\]**]{}]{}(iv) such that the total variation distances ${\Delta}_n$ between the laws of $Y_n,$ $n\ge 1,$ and the respective $F^{{\alpha}_n}$ are such that $\sum_{n\ge 1}{\Delta}_n<\infty.$
Our first result establishes the existence of a coupling of the sequences $\{(M_n, I_n)\}_{n\ge 1}$ and $\{({\mathcal{M}}_n, {\mathcal{I}}_n)\}_{n\ge 1},$ the latter corresponding to the $F^{\alpha}$-scheme $\boldsymbol{{\mathcal{X}}}$ from [[**\[C$_{\boldsymbol{2}}$\]**]{}]{}.
\[th1\] If $\boldsymbol{X}$ satisfies condition [[**\[C$_{\boldsymbol{2}}$\]**]{}]{} then one can construct the sequences $\boldsymbol{X}$ and $\boldsymbol{{\mathcal{X}}}$ on a common probability space so that there exists a random time $T<\infty$ a.s. such that $(M_n, I_n)=({\mathcal{M}}_n, {\mathcal{I}}_n)$ for all $n\ge T.$
The following corollary is a straightforward consequence of the coupling established in Theorem \[th1\] and the observation that $E_n,$ $V_n\to\infty$ as $n\to \infty$ under condition [[**\[C$_{\boldsymbol{1}}$\]**]{}]{}.
\[co1\] Under the condtions of Theorem \[th1\], the above assertions [(A1)–(A4)]{} concerning the limiting behavior of ${\mathcal{N}}_n$ as $n\to\infty$ remain true if we replace in them ${\mathcal{N}}_n$ with $N_n$, the definitions of $E_n, V_n$ staying unchanged.
One of the key components of [[**\[C$_{\boldsymbol{2}}$\]**]{}]{} is the condition that ${{\bf P}}({\mathcal{M}}_n >\ell_n\ {\rm ev.})=1.$ We will obtain a criterion for that relation as a consequence of our Theorem \[th2\] on the growth rate of the extremal process. The assertion that theorem builds on is the criterion for the partial maxima of i.i.d. r.v.’s which was derived in [@Kl84; @Kl85] and is stated in below. For comments on the history of the problem on the growth rate for the partial maxima in the i.i.d. case see [@Kl84]. For an alternative martingale-based proof of criterion see [@Go87].
First we need to recall a constructive definition of the extremal process associated with the d.f. $F$ (cf. Ch. 4 in [@Re87]). Let $\mathscr{P}$ be a Possion point process on ${\mathbb{R}}^2$ with the intensity measure ${\lambda}$ specified by $${\lambda}\bigl((a,b)\times (x,\infty)\bigr)=(a-b)\ln F(x), \quad a<b,\ x>x_0,$$ and ${\lambda}\bigl({\mathbb{R}}\times (-\infty,x_0]\bigr)=0,$ where $x_0:=\inf\{x\in{\mathbb{R}}: F(x) >0\}$ and we assume for simplicity that $F(x_0)=0$ if $x_0>-\infty$ (which is no loss of generality as we are interested in the behavior of the upper records only). Introduce the following notation: $$\label{L}
L(B):= \inf\bigl\{x\in {\mathbb{R}}: \mathscr{P}\bigl(B\times (x,\infty)\bigr)=0\bigr\}, \quad B\subset {\mathbb{R}}.$$ The (continuous–time) extremal process $\{\mathscr M_t\}_{t> 0}$ is then defined by $
{\mathscr M_t}:= L((0,t]), $ $t> 0.$ Clearly, $$\label{embed}
{\mathscr M_{s_n}}=\bigvee_{j=1}^n {\mathcal{X}}_j^*, \quad n\ge 1,$$ where $ {\mathscr {\mathcal{X}}_j^*}:=L\bigl( (s_{j-1}, s_j] \bigr) $, $ j\ge 1,$ are independent r.v.’s, $${{\bf P}}({\mathcal{X}}_j^*\le x) = {{\bf P}}\bigl(\mathscr{P} \bigl((s_{j-1},s_j]\times (x,\infty)\bigr)=0\bigr)=
e^{ {\alpha}_j\ln F(x)}=F^{{\alpha}_j} (x), \quad x\in{\mathbb{R}},$$ for $j\ge 1$ (and likewise ${{\bf P}}(\mathscr{M}_t\le x)=F^t (x),$ $t>0$), so that $\{{\mathcal{X}}_n^*\}_{n\ge 1} {\stackrel{d}{=}}\{{\mathcal{X}}_n\}_{n\ge 1}$ and hence $\{{\mathcal{M}}_n\}_{n\ge 1} {\stackrel{d}{=}}\{\mathscr M_{s_n}\}_{n\ge 1}.$ This remarkable embedding of the partial maxima sequence for an $F^{\alpha}$-scheme into a continuous–time extremal process is one of the powerful tools for solving problems related to such schemes, see e.g. [@BaRe87]. It proved to be very handy in our case as well.
Let $b_t,$ $t\ge 0,$ be a non-decreasing right-continuous real-valued function. As in the discrete-time case, we will use $
\{\mathscr{M}_t > b_t {\rm\ ev.}\}
$ to denote the event that $\mathscr{M}_t>b_t$ eventually, i.e. that $\sup\{t>0: \mathscr{M}_t \le b_t \}<\infty$. We do not assume in the following theorem that $F$ is continuous.
\[th2\]
1. One always has ${{\bf P}}(\mathscr{M}_t > b_t {\rm \ ev.})=0$ or $1. $
2. If $g_t:=1-F(b_t)\to c >0$ as $t\to\infty$ then ${{\bf P}}(\mathscr{M}_t > b_t {\rm\ ev.})=1.$
3. If $\lim_{t\to 0}g_t = 0$ and $\liminf_{t\to\infty} tg_t <\infty$ then ${{\bf P}}(\mathscr{M}_t > b_t {\rm\ ev.})=0.$
4. If $\lim_{t\to 0}g_t = 0$ and $\lim_{t\to\infty} tg_t =\infty$ then $$\begin{aligned}
\label{J}
{{\bf P}}(\mathscr{M}_t > b_t {\rm\ ev.})=
\left\{
\begin{array}{ll}
0 & \mbox{if}\ \ J(b) =\infty,\\
1 & \mbox{if}\ \ J(b) <\infty,
\end{array}
\right.
\quad \mbox{where}\ J(b):= \int_0^\infty g_t e^{-t g_t}dt.\end{aligned}$$
\[re1\] Note that the assumption that $b_t$ is non-decreasing does not actually restrict the generality. Thus, for part (iv), arguing as on p. 382 in [@Kl84], it is not hard to verify that if $b_t$ is a general real-valued function such that $ g_t\to 0,$ $tg_t\to\infty$ as $t\to\infty,$ then, for some $t_0<\infty$, the non-decreasing function $$\overline{b}_t:=
\left\{
\begin{array}{ll}
b_{t_0}, & t\in (0,t_0),\\
\sup_{t_0\le u\le t} b_u,& t\ge t_0,
\end{array}
\right.$$ has the property that ${{\bf P}}(\mathscr{M}_t > b_t {\rm\ ev.})={{\bf P}}(\mathscr{M}_t > \overline{b}_t {\rm\ ev.}).$
Theorem \[th2\] enables us to give a complete characterization of the situations where the condition ${{\bf P}}({\mathcal{M}}_n > \ell_n {\rm \ ev.})=1$ from Theorem 1 is satisfied.
\[co2\] The following assertions hold true for an $F^{\alpha}$-scheme under condition [[**\[C$_{\boldsymbol{1}}$\]**]{}]{} and a non-decreasing sequence $\{\ell_n\}$.
1. One always has ${{\bf P}}({\mathcal{M}}_n > \ell_n {\rm \ ev.})=0$ or $1. $
2. If $q_n:=1-F(\ell_n)\to c>0$ as $n\to\infty$ then ${{\bf P}}({\mathcal{M}}_n > \ell_n {\rm \ ev.})=1. $
3. If $\lim_{n\to\infty}q_n= 0$ and $\liminf_{n\to \infty} s_n q_n< \infty$ then ${{\bf P}}({\mathcal{M}}_n > \ell_n {\rm \ ev.})=0. $
4. If $\lim_{n\to\infty}q_n= 0$ and $\lim_{n\to \infty} s_n q_n=\infty$ then
$${{\bf P}}({\mathcal{M}}_n > \ell_n {\rm \ ev.})=
\left\{
\begin{array}{ll}
0 & \mbox{if}\ \ K(\ell) =\infty,\\
1 & \mbox{if}\ \ K(\ell) <\infty,
\end{array}
\right.
\quad \mbox{where}\ K(\ell):=
\sum_{n\ge 1} e^{-s_nq_n} (1- e^{-{\alpha}_{n+1}q_n}).$$
[Note that without condition [[**\[C$_{\boldsymbol{1}}$\]**]{}]{} the assertions of the theorem do not need to hold. Indeed, if $s_\infty:=\lim_{n\to\infty} s_n<\infty $ then there are only finitely many records in the sequence $\boldsymbol{{\mathcal{X}}}$ and ${\mathcal{M}}_\infty:=\lim_{n\to\infty} {\mathcal{M}}_n{\stackrel{d}{=}}\mathscr{M}_{ s_\infty }$ is a proper non-degenerate r.v. Hence for, say, a constant sequence $\ell_n\equiv \ell$ one would have ${{\bf P}}({\mathcal{M}}_n > \ell_n {\rm \ ev.})= {{\bf P}}(\mathscr{M}_{ s_\infty }>\ell )=F^{s_\infty} (\ell), $ which can be neither 0 nor 1 etc.]{}
Finally, we will address the question concerning the dependence of the record indicators in threshold $F^{\alpha}$-schemes. As was pointed out earlier, in the general case, even when the $X_n$’s are independent, the nature of dependence between record indicators is very complicated and difficult to describe. However, we were able to obtain the following asymptotic pair-wise independence result in the case of independent $X_n$’s in a threshold $F^{\alpha}$-scheme with a flat threshold $\ell_n=\ell ,$ $n\ge 1.$ It is actually possible to extend the result of Theorem \[th3\] to assert that, say, ${{\bf P}}(I_n=1|I_{m_1}=1, I_{m_2}=1)$ and $ {{\bf P}}(I_n =1)$ are asymptotically equivalent as $n>m_2>m_1\ge k\to\infty$ etc, but the set of conditions for such an assertion will already be quite cumbersome.
\[th3\] Assume that the r.v.’s in the sequence $\boldsymbol{X}$ are independent, condition [[**\[C$_{\boldsymbol{2}}$\]**]{}]{} is satisfied for a constant threshold sequence $\ell_n=\ell $ and ${\delta}_n=0,$ $n\ge 1 $. If $$\label{con_s_al}
\lim_{k\to \infty}\frac{ s^2_k}{{\alpha}_k }({\alpha}_k \vee 1) h^{s_k }= 0, \quad\mbox{where}\quad h:=F(\ell),$$ then $$\label{to_1}
\sup_{n>m\ge k}\biggl|
\frac{{{\bf P}}(I_n=1|I_m=1)}{ {{\bf P}}(I_n =1)}-1\biggr|\to 0\quad as \quad k\to \infty.$$
[We can somewhat extend the conditions of the last theorem by assuming that we have a sequence $\boldsymbol{X}^{(k)}=\{X_n^{(k)}\}_{n\ge 1},$ $k\ge 1,$ of threshold $F^{\alpha}$-schemes indexed by the parameter $k\to\infty$ and that these schemes share a common d.f. $F$ and sequence ${\boldsymbol{\alpha}},$ but have growing thresholds $\ell=\ell^{ (k)}$ such that $h=h^{ (k)}:=F(\ell^{ (k)})\to 1$. Then relation will still hold true under assumption with $h=h^{(k)}.$ We just note here that the upper bound for the last term in in the proof of Lemma \[le2\] will be valid when $(h^{(k)})^{s_k }<e^{-2}$ and that (with $h=h^{(k)}$) ensures that $(h^{(k)})^{s_k }\to 0 $ as $k\to\infty.$ ]{}
Proofs
======
To construct the desired coupling, we will start with a sequence of i.i.d. uniform-$(0,1)$ r.v.’s $U_1, U_2, \ldots$ given on some probability space, set $X_1:=F_1^{\leftarrow} (U_1),$ $$\label{Xn}
X_n:=F_{n|X_1,\ldots,X_{n-1}}^\leftarrow (U_{n })\quad\mbox{for}\quad n\ge 2,$$ and let ${\mathcal{X}}_n :=(F^{{\alpha}_n})^\leftarrow (U_n)=F^\leftarrow (U_n^{1/{\alpha}_n}) ,$ $n\ge 1.$ These sequences of r.v.’s will clearly have the desired distributions.
Next introduce the events $$\begin{aligned}
A_n
& := \{X_n>\ell_n\}= \{U_n>F_{n|X_1,\ldots,X_{n-1}}(\ell_n)\}
=\{U_n>F_n (\ell_n)\},
\\
{\mathcal{A}}_n &
:= \{{\mathcal{X}}_n>\ell_n\}= \{U_n>F^{{\alpha}_n}(\ell_n)\}\end{aligned}$$ for $n\ge 2,$ where we used condition [[**\[C$_{\boldsymbol{2}}$\]**]{}]{}(ii) for the last equality in the first line. Observe that, for the symmetric difference of these events, one has $${{\bf P}}(A_n\triangle {\mathcal{A}}_n)
= |F_n (\ell_n)- F^{{\alpha}_n}(\ell_n)|\le {\delta}_n.$$ Therefore, by the Borel–Cantelli lemma and condition [[**\[C$_{\boldsymbol{2}}$\]**]{}]{}(iii), these events occur finitely often a.s., so that $$T_1: = \inf\Bigl\{n\ge 1: \sum_{k\ge n} {{\bf 1}}( A_k \triangle {\mathcal{A}}_k)=0 \Bigr\}
<\infty \quad\mbox{a.s.}$$ Now introduce the events $B_n:= A_n\cap {\mathcal{A}}_n=\{U_n>F_n (\ell_n) \vee F^{{\alpha}_n} (\ell_n)=:l_n\},$ $n\ge 1.$ Clearly, $$d_{TV} \bigl(F_n |_{F_n^{\leftarrow}(l_n)}, F^{{\alpha}_n}|_{(F^{{\alpha}_n})^{\leftarrow}(l_n)}\bigr) \le {\delta}_n.$$ Therefore, using Dobrushin’s maximal coupling theorem [@Do70], we can recursively re-define for $n\ge 2$ the r.v.’s $X_n$ and ${\mathcal{X}}_n$ on the set $B_n$ only (extending the underlying probability space if necessary and updating at each step the definitions for the “later" observations $X_{n+1}, X_{n+2},\ldots$ accordingly) in such a way that $${{\bf P}}(X_n\neq {\mathcal{X}}_n; B_n)\le {\delta}_n.$$ Again by the Borel–Cantelli lemma and conditions [[**\[C$_{\boldsymbol{2}}$\]**]{}]{}(iii),(iv), one has $$T_2: = \inf\Bigl\{n \ge T_1: \sum_{k\ge n} \bigl( {{\bf 1}}(X_k\neq {\mathcal{X}}_k; B_k)+ {{\bf 1}}({\mathcal{M}}_k\le \ell_k)\bigr)=0 \Bigr\}
<\infty \quad\mbox{a.s.}$$ Now for $n\ge T_2$ we always have $ {\mathcal{M}}_n > \ell_n$ and whenever one of the events $A_n, {\mathcal{A}}_n$ occurs, the other one occurs as well and $X_n={\mathcal{X}}_n$ for that $n$.
In view of condition [[**\[C$_{\boldsymbol{2}}$\]**]{}]{}(i), one has $F (M_{T_2})<1$ a.s. Therefore, by virtue of [[**\[C$_{\boldsymbol{1}}$\]**]{}]{}, with probability 1 there exists an $n\ge T_2$ such that ${\mathcal{X}}_n={\mathcal{M}}_n> M_{T_2}$. From the definition of $T_2$, one then also has ${\mathcal{X}}_n >\ell_n$ and $X_n= {\mathcal{X}}_n,$ so that $M_n> M_{T_2}$. We conclude that $$T_3:= \inf\{n> T_2: M_n> M_{T_2}, {\mathcal{M}}_n> {\mathcal{M}}_{T_2}, {\mathcal{I}}_n=1 \}<\infty \quad\mbox{a.s.}$$
We claim that $(M_n, I_n)=({\mathcal{M}}_n, {\mathcal{I}}_n)$ for all $n\ge T_3.$ Indeed, assume that ${\mathcal{I}}_n=1$ for an $n\ge T_3$. Then ${\mathcal{X}}_n={\mathcal{M}}_n> \ell_n $ and so also $X_n={\mathcal{X}}_n> \ell_n .$ Now if $I_n=0$ then there exists a $k\in [T_2, n] $ such that $X_k>X_n>\ell_n\ge \ell_k$, which implies that $X_k={\mathcal{X}}_k $ (by the definition of $T_2$) and so ${\mathcal{X}}_k>{\mathcal{X}}_n$, which contradicts the assumption that ${\mathcal{I}}_n=1$. A symmetric argument shows that if $I_n=1$ then also ${\mathcal{I}}_n=1$, thus proving that $I_n= {\mathcal{I}}_n $ for all $n\ge T_3.$ Now, whenever ${\mathcal{X}}_n$ with $n\ge T_2$ is a record we must have ${\mathcal{X}}_n={\mathcal{M}}_n>\ell_n,$ which ensures that $X_n={\mathcal{X}}_n$. Since we already know that at such times $n\ge T_3$ there will be a record in $\boldsymbol{X}$ as well, we obtain that $M_n=X_n={\mathcal{X}}_n={\mathcal{M}}_n$. Theorem \[th1\] is proved.
(i) Recalling notation , set $$Y_n:=L((n-1,n]), \quad n\in {\mathbb{Z}}, \qquad M^Y_m:=\bigvee_{k=1}^m Y_k, \quad m\ge 1.$$ By construction, the $Y_n$’s are i.i.d. with d.f. $F$. As both $\mathscr{M}$ and $b$ are non-decreasing, we have (using $\lfloor t\rfloor$ for the integer part of $t$) $$\begin{aligned}
\{M^Y_n > b_{n+1} {\rm\ ev.}\}
= \{\mathscr{M}_t > b_{\lfloor t\rfloor +1} {\rm\ ev.}\}
& \subset
\{\mathscr{M}_t > b_t {\rm\ ev.}\}
\notag \\
&
\subset
\{\mathscr{M}_t > b_{\lfloor t\rfloor} {\rm\ ev.}\}
= \{M^Y_n > b_{n} {\rm\ ev.}\}.
\label{MM}\end{aligned}$$
We can assume w.l.o.g. that there is no $x_0$ such that $F(x_0-)<F(x_0)=1,$ as if such a point existed then one would have $M^Y_n=x_0$ ev. (and hence $\mathscr{M}_t=x_0$ ev.) a.s., making assertion (i) obvious. With that assumption, clearly ${{\bf P}}(A)=1$ for $A:=\bigl\{\sum_{k=1}^\infty {{\bf 1}}(Y_{k+1} >M_k)=\infty\bigr\}$ and, for any $k\ge 1,$ $$\begin{aligned}
A\{M^Y_n > b_{n} {\rm\ ev.}\}
& = A \bigl\{ \mbox{$\bigvee_{j=k}^n$} Y_j > b_n {\rm\ ev.}\bigr\}
\\
& = \biggl\{\sum_{m=k+1}^\infty {{\bf 1}}\bigl(Y_m > \mbox{$\bigvee_{j=k}^{m-1}$}Y_j \bigr)=\infty;
\ \mbox{$\bigvee_{j=k}^n$} Y_j > b_n {\rm\ ev.} \biggr\}.\end{aligned}$$ Therefore the event on the left-hand side belongs to the tail ${\sigma}$-algebra for $\{Y_n\}_{n\ge 1}.$ We conclude that, by Kolmogorov’s 0–1 law, its probability must be either 0 or 1 and that the same applies to ${{\bf P}}(M^Y_n > b_{n} {\rm\ ev.})$ as well since ${{\bf P}}(A)=1$. The same argument is valid for ${{\bf P}}(M^Y_n > b_{n+1} {\rm\ ev.})$.
It remains to observe that $\{M_n\}_{n\ge 2} {\stackrel{d}{=}}\{Y_0\vee M_{n-1}\}_{n\ge 2}$ and so $$\begin{aligned}
{{\bf P}}(M^Y_n > b_{n} {\rm\ ev.})
& = {{\bf P}}(Y_0\vee M^Y_{n-1} > b_{n} {\rm\ ev.})
= {{\bf P}}\bigl(A\{ Y_0\vee M^Y_{n-1} > b_{n} {\rm\ ev.}\}\bigr)
\notag
\\
&
=
{{\bf P}}\bigl( A\{ M^Y_{n-1} > b_{n} {\rm\ ev.}\}\bigr)
= {{\bf P}}( M^Y_{n} > b_{n+1} {\rm\ ev.} ).
\label{n_n+1}\end{aligned}$$ Assertion (i) is proved, as the probabilities of the left-hand and right-hand sides in are equal to each other, their only possible values being 0 and 1.
(ii) This assertion is obvious as $\mathscr{M}_t$ is non-decreasing, $\mathscr{M}_n \ge Y_n,$ $n\ge 1,$ and so, setting $B:=\bigcup_{n\ge 1} (-\infty, b_n] $ (which can be an open or closed half-line, ${{\bf P}}(Y_j\in B)=1-c<1$), one has $
{{\bf P}}(\mathscr{M}_t > b_t {\rm\ ev.}) \ge
{{\bf P}}\bigl( \bigcup_{n\ge 1} \{Y_n\in B^c\}\bigr)
= 1- {{\bf P}}\bigl( \bigcap_{n\ge 1} \{Y_n\in B\}\bigr)=1.
$
(iii) In this case, there exists a sequence $t_n\to\infty $ as $n\to \infty$ such that $t_n g_{t_n}<c<\infty$, $n\ge 1.$ One has $$\begin{aligned}
{{\bf P}}(\mathscr{M}_{t_n} \le b_{t_n} {\rm\ i.o.})
& = \lim_{k\to\infty}
{{\bf P}}\bigl( \mbox{$\bigcup_{n\ge k}$ } \{ \mathscr{M}_{t_n} \le b_{t_n}\} \bigr)
\\
& \ge \liminf_{n\ge 1} {{\bf P}}\bigl( \mathscr{M}_{t_n} \le b_{t_n} \bigr)
=
\liminf_{n\ge 1} F^{t_n}(b_{t_n})
\\
&
= \liminf_{n\ge 1} e^{- t_n g_{t_n}(1+o(1))}
\ge e^{-c} >0.\end{aligned}$$ Hence ${{\bf P}}(\mathscr{M}_t > b_t {\rm\ ev.})<1.$ In view of (i), that probability must be 0.
(iv) We will make use of the following criterion from [@Kl84; @Kl85] derived for an i.i.d. sequence $\{Y_n\}$ in the case a non-decreasing sequence $\{b_n\}$ when $g_n= 1- F(b_n)\to 0,$ $n g_n\to\infty$ as $n\to\infty$: $$\label{Klass}
{{\bf P}}(M^Y_n > b_n {\rm\ ev.})=
\left\{
\begin{array}{ll}
0 & \mbox{if}\ \ \Sigma (b) =\infty,\\
1 & \mbox{if}\ \ \Sigma(b) <\infty,
\end{array}
\right.
\quad \mbox{where}\quad \Sigma (b):= \sum_{n\ge 1} g_n e^{-n g_n}.$$ As the function $f(x)= x e^{-ax}$ with $a>0$ is decreasing for $x\ge a^{-1}$, $g_n=o(1)$ as $n\to\infty,$ and the function $g_t$ is non-increasing, we have, for $t\in [n,n+1]$ and all sufficiently large $n$, $$\begin{aligned}
g_n e^{- n g_n}(1+o(1))& = g_n e^{-(n+1) g_n} \le g_t e^{-(n+1) g_t}
\le g_t e^{-t g_t}
\\
& \le g_t e^{-n g_t} \le g_{n+1} e^{-n g_{n+1}}
=
g_{n+1} e^{-(n+1) g_{n+1}} (1+o(1)) .\end{aligned}$$ Therefore the integral $J(b)$ in and the sum $\Sigma (b)$ in converge (diverge) simultaneously. It remains to make use of relations and . Theorem \[th2\] is proved.
Setting $s_0:=0,$ introduce the non-decreasing function $$\begin{aligned}
\label{bt}
b_t:= \sum_{n\ge 1} \ell_n {{\bf 1}}(t\in [s_{n}, s_{n+1})), \quad t>0.\end{aligned}$$ In view of the embedding of $\{{\mathcal{M}}_n\}$ in the extremal process $\{\mathscr{M}_t\}$ associated with the d.f. $F$ and the special form of the boundary $b_t,$ we clearly have ${{\bf P}}({\mathcal{M}}_n >\ell_n {\rm \ ev.})={{\bf P}}(\mathscr{M}_t > b_t {\rm \ ev.})$. Therefore part (i) follows from Theorem \[th2\](i). It remains to verify that the conditions from parts (ii)–(iv) of the corollary are equivalent to the conditions from the respective parts (ii)–(iv) of Theorem \[th2\] in the case of our boundary .
Part (ii) is obvious. Part (iii) is also obvious since, for the boundary , one has $t g_t = s_n q_n$ for $t=s_n$ and $s_n\to \infty$ as $n\to\infty $ by [[**\[C$_{\boldsymbol{1}}$\]**]{}]{}. Finally, in part (iv) one has $$\begin{aligned}
J(b) = \sum_{n\ge 1} \int_{s_{n }}^{s_{n+1}}
q_n e^{-t q_n} dt
=\sum_{n\ge 1} e^{-s_nq_n} (1- e^{-{\alpha}_{n+1}q_n}).\end{aligned}$$ Corollary \[co2\] is proved.
We will split the proof of the theorem into three lemmata.
\[le0\] Under conditions of Theorem \[th3\], one can construct $\boldsymbol{X}$ and $\boldsymbol{{\mathcal{X}}}$ so that, for any $m\ge 1,$ on the event $\{M_m>\ell\}$ one has $$\begin{aligned}
( M_n, I_n) =({\mathcal{M}}_n, {\mathcal{I}}_n), \quad \mbox{for}\ n\ge m.\end{aligned}$$
Under conditions of Theorem \[th3\], we can construct $\boldsymbol{X}$ and $\boldsymbol{{\mathcal{X}}}$ by letting $
X_n:=F_n^{\leftarrow} (U_n),$ ${\mathcal{X}}_n:=(F^{{\alpha}_n})^{\leftarrow} (U_n),$ $n\ge 1
$, where $\{U_n\}_{n\ge 1}$ is a sequence of independent uniform-$(0.1)$ r.v.’s. Then clearly $$\begin{aligned}
\label{XXX}
\{X_n>\ell\}= \{{\mathcal{X}}_n>\ell\},\quad
X_n {{\bf 1}}(X_n>\ell) = {\mathcal{X}}_n {{\bf 1}}({\mathcal{X}}_n>\ell), \quad n\ge 1.
\end{aligned}$$ Hence, for $n\ge m\ge 1,$ $$\begin{aligned}
\{I_n=1, M_m>\ell\}
& =\{I_n=1, M_m>\ell, X_n >\ell\}
\\
& = \{ X_n >\ell\}\cap \biggl[\bigcup_{k\le m}\{ X_k>\ell\}\biggr] \bigcap_{r<n}
\bigl[ \{X_r \le \ell\}\cup \{\ell< X_r <X_n\}\bigr]
\\
& = \{ {\mathcal{X}}_n >\ell\}\cap \biggl[\bigcup_{k\le m}\{ {\mathcal{X}}_k>\ell\}\biggr] \bigcap_{r<n}
\bigl[ \{{\mathcal{X}}_r \le \ell\}\cup \{\ell< {\mathcal{X}}_r < {\mathcal{X}}_n\}\bigr]
\\
&=\{{\mathcal{I}}_n=1, {\mathcal{M}}_m>\ell, {\mathcal{X}}_n >\ell\}
=\{{\mathcal{I}}_n=1, {\mathcal{M}}_m>\ell \}.
\end{aligned}$$ The claim concerning the equality of $M_n$ and ${\mathcal{M}}_n$ is next to obvious from : setting $\tau (m):=\min\{k\ge 1: X_k=M_m\}$ (which is the same as $\min\{k\ge 1: {\mathcal{X}}_k={\mathcal{M}}_m\}$ on the event $\{M_m>\ell\}=\bigcup_{k\le m} \{X_k>\ell\}=\bigcup_{k\le m} \{{\mathcal{X}}_k>\ell\}=\{{\mathcal{M}}_m>\ell\}$), one has $$\begin{aligned}
M_n {{\bf 1}}(M_m>\ell)
& = \bigvee_{k=1}^n X_k {{\bf 1}}(X_{\tau (m)} >\ell)
= \bigvee_{k=1}^n X_k {{\bf 1}}(X_{k} >\ell) {{\bf 1}}(X_{\tau (m)} >\ell)
\\
& = \bigvee_{k=1}^n {\mathcal{X}}_k {{\bf 1}}({\mathcal{X}}_{k} >\ell) {{\bf 1}}({\mathcal{X}}_{\tau (m)} >\ell)
= {\mathcal{M}}_n {{\bf 1}}({\mathcal{M}}_m>\ell).
\end{aligned}$$ Lemma \[le0\] is proved.
Next observe that one clearly has $$\begin{aligned}
b_m & := {{\bf P}}(I_m=1, M_m\le \ell)\le {{\bf P}}(M_m\le \ell) = h^{s_m },\quad m\ge 1,
\label{b_for_b}\\
c_{m,n} & :={{\bf P}}(I_m=I_n=1, M_n\le \ell)
\le {{\bf P}}(M_n\le \ell) = h^{s_n },\quad n>m\ge 1.
\label{b_for_c}
\end{aligned}$$
\[le1\] For $ n>m\ge 1,$ $$\begin{aligned}
{{\bf P}}(I_m =1)&= \frac{{\alpha}_m }{s_m }(1-h^{s_m })+b_m
= \frac{{\alpha}_m }{s_m }+\theta_m,\quad |\theta_m|\le h^{s_m } ,
\label{b_for_I}
\\
{{\bf P}}(I_m =I_n=1) & =
\frac{{\alpha}_m {\alpha}_n }{s_n -s_m }
\biggl(\frac{1-h^{ s_m }}{s_m }
- \frac{1-h^{ s_n }}{s_n }\biggr)
\notag\\
&\hskip 2 cm + b_m {\alpha}_n \frac{1-h^{ s_n -s_m }}{s_n -s_m }+ c_{m,n}.
\label{b_for_II}\end{aligned}$$
By Lemma \[le0\], and , $$\begin{aligned}
{{\bf P}}(I_m =1)&= {{\bf P}}(I_m =1, M_m >\ell) +b_m
={{\bf P}}({\mathcal{I}}_m =1, {\mathcal{M}}_m >\ell) +b_m\notag\\
& ={{\bf P}}({\mathcal{I}}_m =1){{\bf P}}({\mathcal{M}}_m >\ell) +b_m
=\frac{{\alpha}_m }{s_m } (1-h^{s_m }) +b_m,
\label{***_2}\end{aligned}$$ which is the first equality in . The second one is obvious.
To establish , we set $A_{u,v}(x):= \bigcap_{u<r<v}\{{\mathcal{X}}_r<x \} $ for $v>u\ge 0$ and observe that, in view of Lemma \[le0\] and independence of the ${\mathcal{X}}_r$’s, $$\begin{aligned}
P_1&:= {{\bf P}}(I_m =I_n=1, M_m>\ell) = {{\bf P}}({\mathcal{I}}_m ={\mathcal{I}}_n=1, {\mathcal{M}}_m>\ell)\\
& = \int_{\ell <x_1 <x_2}
{{\bf P}}\bigl(A_{0,m}(x_1)A_{m,n}(x_2)
\big|{\mathcal{X}}_m=x_1, {\mathcal{X}}_n=x_2\bigr)dF^{{\alpha}_m }(x_1)dF^{{\alpha}_n}(x_2)\\
& = \int_{\ell <x_1 <x_2}
\biggl(\prod_{u=1}^{m-1}F^{{\alpha}_u }(x_1)\biggr)
\biggl(\prod_{v=m+1}^{n-1}F^{{\alpha}_v }(x_2)\biggr)
dF^{{\alpha}_m }(x_1)dF^{{\alpha}_n}(x_2)
\\
& = \int_{h <y_1 <y_2\le 1}
y_1^{s(m-1)} y_2^{s_n - s_m }
dy_1^{{\alpha}_m } dy_2^{{\alpha}_n}
\\
& = \frac{{\alpha}_m {\alpha}_n }{s_n -s_m }
\biggl(\frac{1-h^{s_m }}{s_m }
-\frac{1-h^{s_n }}{s_n }\biggr).\end{aligned}$$ Next, by the independence of the $X_r$’s, one has $$\begin{aligned}
P_2&:= {{\bf P}}(I_m =I_n=1, M_m\le \ell <M_n)\notag\\
& = {{\bf P}}\bigl(\{I_m = 1, M_m\le \ell\}\cap \{ X_r< X_n, m<r<n; X_n>\ell\}\bigr)\notag
\\
& = {{\bf P}}\bigl( I_m = 1, M_m\le \ell\bigr)\,
{{\bf P}}\bigl( X_r< X_n, m<r<n; X_n>\ell\bigr).
\label{P2}
\end{aligned}$$ The first factor of the right-hand side is $b_m,$ whereas the second one is of the form ${{\bf P}}(I'_{n-m}=1, M'_{n-m}>\ell)$ for the shifted sequence $\{X'_j:= X_{j+m}\}_{j \ge 1},$ which satisfies the same assumptions as the original $\boldsymbol{X}$. Therefore, similarly to the computation of the term ${{\bf P}}(I_m =1, M_m >\ell)$ in and using the self-explanatory notations ${\mathcal{I}}'_j ,$ ${\mathcal{M}}'_j$, we obtain that the second factor on the right-hand side of is equal to $${{\bf P}}({\mathcal{I}}'_{n-m} =1){{\bf P}}({\mathcal{M}}'_{n-m} >\ell)
= \frac{{\alpha}_n}{s_n -s_m } (1-h^{s_n -s_m }).$$ We conclude that $
P_2 = b_m{\alpha}_n(1-h^{s_n -s_m })/(s_n -s_m ).
$ Now immediately follows from the obvious representation $ {{\bf P}}(I_m =I_n=1) =P_1+P_2+c_{m,n}.$
\[le2\] As $k\to \infty,$ $$\begin{aligned}
\frac{s_m s_n }{{\alpha}_m {\alpha}_n }{{\bf P}}(I_m =1){{\bf P}}(I_n =1)
\to 1,
\quad \frac{s_m s_n }{{\alpha}_m {\alpha}_n }{{\bf P}}(I_m =1,I_n =1) \to 1
\label{XX}\end{aligned}$$ uniformly in $n>m\ge k.$
The first relation in is obvious from as the left-hand side of the former relation equals $$\frac{s_m s_n }{{\alpha}_m {\alpha}_n }
\biggl(\frac{{\alpha}_m }{s_m }+\theta_m\biggr)
\biggl(\frac{{\alpha}_n }{s_n }+\theta_n\biggr)
= \biggl(1+ \theta_m\frac{s_m }{{\alpha}_m } \biggr)
\biggl(1+ \theta_n\frac{s_n }{{\alpha}_n } \biggr)$$ and $|\theta_r|\le h^{s_r }=o({\alpha}_r /s_r )$ as $r\to \infty$ by virtue of and [[**\[C$_{\boldsymbol{1}}$\]**]{}]{}.
To establish the second relation in , we will use . Recall that the first two terms on the latter relation’s right-hand side we denoted by $P_i,$ $i=1,2.$ First note that, by the mean value theorem for the function $f(x):=xh^x,$ $h=\rm const$, one has $$\begin{aligned}
\frac{s_m s_n }{{\alpha}_m {\alpha}_n }P_1
& =
\frac{s_m s_n }{s_n -s_m }
\biggl(\frac{1-h^{s_m }}{s_m }
-\frac{1-h^{s_n }}{s_n }\biggr)
\notag\\
&
= 1- h^{s_m }-h^{s_n } +\frac{s_n h^{s_n }-s_m h^{s_m } }{s_n -s_m }
\notag\\
&
= 1- h^{s_m }-h^{s_n } +(1 +\ln h^{s})h^s
\label{ssaa1}\end{aligned}$$ for some $s\in [s_m , s_n ].$ Clearly, the expression in the last line in tends to 1 uniformly in $n>m\ge k\to\infty.$
Next, in view of and again by the mean value theorem (this time for the function $f(x) =h^{1/x},$ $h=\rm const \in (0,1)$), one has $$\begin{aligned}
0 &\le \frac{s_m s_n }{{\alpha}_m {\alpha}_n }P_2
=
\frac{b_m s_n s_m \bigl(1-h^{s_n -s_m }\bigr)}{{\alpha}_m (s_n -s_m )}
\notag \\
& \le
\frac{h^{ s_m }-h^{s_n }}{{\alpha}_m (1/s_m -1/s_n )}
=\frac{|\ln h |}{{\alpha}_m } s^2 h^s
\label{ssaa2}\end{aligned}$$ for some $s\in [s_m , s_n ].$ As the function $s^2 h^s$ is decreasing for $s$ such that $h^s<e^{-2},$ the right-hand side of is bounded by $|\ln h |s^2_m h^{s_m }/ {\alpha}_m $ for all sufficiently large $m$. That expression clearly vanishes uniformly in $m\ge k\to\infty$ in view of condition .
Finally, $$\begin{aligned}
\frac{s_m s_n }{{\alpha}_m {\alpha}_n } c_{m,n}
& \le
\frac{s_m s_n }{{\alpha}_m {\alpha}_n } h^{s_n }
\le \frac{s_m }{{\alpha}_m } h^{s_m /2}
\frac{ s_n }{ {\alpha}_n } h^{s_n /2} \to 0\end{aligned}$$ uniformly in $n> m\ge k\to\infty$, again due to . Lemma \[le2\] is proved.
It remains to observe that immediately follows from relations . Theorem \[th3\] is proved.
Acknowledgments {#acknowledgments .unnumbered}
===============
A major part of the research presented in this paper was done when K. Borovkov was visiting the School of Informatics and Data Science, Hiroshima University, whose hospitality and support is gratefully acknowledged. P. He’s work on this research was supported by the Maurice H. Belz scholarship.
[99]{}
J. Appl. Prob., 24:4 (1987), pp. 827–837.
J. Statist. Plan. Inf. 45:1 (1995), pp. 65–79.
J. Appl. Prob., 33:1 (1996), pp. 146–155.
In: M. Puri et al., eds., [*New Perspectives in Theoretical and Applied Statistics.*]{} Wiley, New York (1987), pp. 439–448.
, 40:4 (1988), pp. 671–681.
Theor. Prob. Appl., 15:3 (1970), pp. 458–486.
16:2 (2013), pp. 147–171.
In: M.Hallin et al., eds, [*Mathematical Statistics and Limit Theorems: Festschrift in Honour of Paul Deheuvels.*]{} Springer, Cham, 2015, pp.287–303.
Statist. Probab. Letters, 5:3 (1987), pp. 235–237.
Ann. Inst. H. Poincaré Probab. Statist. 23:3 (1987), pp. 425–457.
Ann. Inst. H. Poincaré Probab. Statist. 34:4 (1998), pp. 481–503.
Wiley-Teubner, New York, 1985, 242 pp.
, 12:2 (1984), pp. 380–389.
, 13:4 (1985), pp. 1369–1370.
In: [*Abstracts of the III Vilnius Conference on Probability Theory and Mathematical Statistics*]{}, Vol. 2. Lithuanian Academy of Sciences Institute of Mathematics and Cybernetics, Vilnius, 1981, pp. 86–87.
Zap. Nauchn. Sem. LOMI, 142 (1985), pp. 109–118. \[English translation: [*J. Soviet Math.,*]{} 36:4 (1987), pp. 510–516.\]
Theor. Prob. Appl. 32:2 (1988), pp. 201–228.
AMS, Providence, 2000, 255 pp.
Springer, New York, 2015, 164 pp.
J. Appl. Prob., 26:4 (1989), pp. 722–733.
Adv. Appl. Prob., 23:4 (1991), pp. 823–834.
Springer, New York, 1987, 320 pp.
Eur. J. Operat. Res. 9:4 (1982), pp. 344–346.
J. Appl. Probab., 12:1 (1975), pp. 148–154.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the problem of segmenting a biomedical image into anatomical regions of interest. We specifically address the frequent scenario where we have no paired training data that contains images and their manual segmentations. Instead, we employ unpaired segmentation images to build an anatomical prior. Critically these segmentations can be derived from imaging data from a different dataset and imaging modality than the current task. We introduce a generative probabilistic model that employs the learned prior through a convolutional neural network to compute segmentations in an unsupervised setting. We conducted an empirical analysis of the proposed approach in the context of structural brain MRI segmentation, using a multi-study dataset of more than 14,000 scans. Our results show that an anatomical prior can enable fast unsupervised segmentation which is typically not possible using standard convolutional networks. The integration of anatomical priors can facilitate CNN-based anatomical segmentation in a range of novel clinical problems, where few or no annotations are available and thus standard networks are not trainable. The code is freely available at <http://github.com/adalca/neuron>.'
author:
- |
Adrian V. Dalca\
MIT and MGH\
[[email protected]]{}
- |
John Guttag\
MIT\
[[email protected]]{}
- |
Mert R. Sabuncu\
Cornell University\
[[email protected]]{}
bibliography:
- 'bibliography.bib'
title: |
Anatomical Priors in Convolutional Networks for\
Unsupervised Biomedical Segmentation
---
Conclusion
==========
In this paper, we introduced a generative probabilistic model that employs a prior model learned through a convolutional neural network to compute segmentations in an unsupervised setting. We can interpret the anatomical prior as encouraging the neural network to predicting segmentation maps that come from a known distribution characterized by ${\boldsymbol{z}}$ while simultaneously producing images that agree with the observed scan. We demonstrate that our model enables segmentation using convolutional networks leading to rapid inference in a setting where segmentation is traditionally not possible, or takes hours to obtain for a single scan. The integration of priors promises to facilitate accurate anatomical segmentation in a variety of novel clinical problems with limited dataset availability.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we consider the monodomain model of cardiac electrophysiology. After an analysis of the well-posedness of the model we determine an asymptotic expansion of the perturbed potential due to the presence of small conductivity inhomogeneities (modelling small ischemic regions in the cardiac tissue) and use it to detect the anomalies from partial boundary measurements. This is done by determining the topological gradient of a suitable boundary misfit functional. The robustness of the algorithm is confirmed by several numerical experiments.'
author:
- 'Elena Beretta, Cecilia Cavaterra, Luca Ratti'
title: On the determination of ischemic regions in the monodomain model of cardiac electrophysiology from boundary measurements
---
Introduction {#intro}
============
Cardiac ischemia consists in a restriction of blood supply to the heart tissue usually caused by atherosclerosis or coronary syndrome. The shortage of oxygen may lead to dysfunction of the cell metabolism and eventually to their death. The possible outcomes range from ventricular arrhythmia, fibrillation and ultimately to myocardial infarction.
The ischemic heart syndrome is the most common cardiovascular disease, and the most common cause of death. Hence, the detection of ischemic regions at early stage of their development is of primary importance. This is usually performed by imaging techniques such as echocardiography, gamma ray scintigraphy or magnetic resonance imaging. Nevertheless, the most common test for patients not exhibiting evident symptoms is the electrocardiogram (ECG), which consists in recording electrical impulses across the thorax by means of a set of electrodes. Physicians are often able to identify myocardial ischemia by analysing the evolution of the voltage recorded in the ECG leads although with several technical difficulties (see the mathematical studies in [@art:ECG] and [@art:gerbeau]).
The approach we propose in this paper is to obtain information regarding the electrical functioning of the tissue (and ultimately the presence of ischemic regions) from the knowledge of the electrical potential on the surface of the heart. When employing the potential on the epicardium, i.e., the external boundary of the heart, this can be considered as a natural extension of the well-known mathematical problem often referred to as *the inverse problem of electrocardiography*, which consists in using the ECG recordings to compute the potential distribution on the epicardium. Alternatively, we remark that it is also possible to obtain electrical measurements on the endocardium, namely the internal boundaries of the heart cavities. Although much more invasive than the ECG, intracardiac ECG (iECG) has become a standard of care in patients with symptoms of heart failure, and allows to get a map of the endocardiac potential by means of non-contact electrodes carried by a catheter inside a heart cavity. We therefore assume that the distribution of the electrical potential is available on some portion of the heart boundary (on the epicardium, in the case of ECG data, or on the endocardium, in the case of iECG measurements). After a reliable model is introduced for the description of the evolution of the electrical potential within the heart, the problem of detecting ischemic regions is hence formulated as an inverse boundary value problem of identifying parameters in a nonlinear reaction diffusion system from boundary data.
In this paper we focus on the determination of small ischemias from boundary measurements of the electrical potential, generalizing the results obtained in previous papers (see [@art:BCMP],[@art:BMR],[@art:bccmr]) to the monodomain time-dependent model of cardiac electrophysiology.
In general, this model is described by a system consisting in a semilinear parabolic equation coupled with a system of nonlinear ordinary differential equations, where the state variables are the transmembrane potential across the cell membrane, the concentrations of ionic species and the gating variables describing the activity of ionic channels in the membrane. It is well known (see [@book:pavarino] and [@art:boulakia]) that the monodomain model can be derived from the more complex bidomain one, introduced for the first time in [@phd:tung], for instance by assuming proportionality between intracellular and extracellular conductivity tensors. In particular, in this case, the corresponding conductivity tensor becomes the harmonic mean of the extra and intracellular conductivities. On the other hand, this choice of the conductivity tensor turns out to be the best one to approximate the electrical propagation described by the bidomain model (see, e.g., [@monovsbi1],[@monovsbi2]).
As a first attempt, we limit ourselves to analyze a coupled system of two equations in the potential $u$ and the recovery variable $w$. This corresponds to a large class of phenomenological models, which are characterized by the choice of the nonlinear terms appearing in the partial differential equation and in the ordinary differential equation. Among the most important two-equation systems (see [@book:pavarino] for a general overview) we mention FitzHugh-Nagumo, Rogers-McCulloch and Aliev-Panfilov models. Throughout this paper, we focus on the commonly used version of the Aliev-Panfilov model (originally introduced in [@art:alpanf]), even though the analysis could be extended also to the other two.
In order to detect ischemic regions, we extend the approach of [@art:bccmr], determining a rigorous asymptotic expansion for the perturbed boundary potential due to small conductivity anomalies. To accomplish this task, we need an accurate analysis of the well-posedness of the direct problem for the coupled system in the case of discontinuous anisotropic coefficients and suitable regularity estimates for solutions. In particular, we establish a comparison principle for this class of systems, which to our knowledge was not present in the literature. Here, we consider the case of an insulated heart and we assume to have measurements of the potential on a portion of the boundary.
The theory of detection of small conductivity inhomogeneities from boundary measurements via asymptotic techniques has been developed in the last three decades in the framework of Electrical Impedance Tomography (see, e.g., [@book:ammari-kang],[@art:cfmv],[@art:ammari2012]). A similar approach has also been used in Thermal Imaging (see, e.g., [@art:aikk]). Here, we are able to extend in a non trivial way the results obtained previously in [@art:BMR] and [@art:bccmr] for simplified versions of the monodomain model making use of fine regularity for the solutions to nonlinear reaction diffusion systems. In particular, we establish a rigorous expansion for the perturbed transmembrane potential and use this to implement an effective reconstruction algorithm.
The paper is organized as follows. In Section \[direct\] we analyze the well-posedness of the forward problem in the unperturbed and perturbed case. Section \[inverse\] is devoted to obtaining energy estimates on the difference between the perturbed and unperturbed electrical potential and an asymptotic expansion of suitable integral terms involving such difference on the boundary. In Section \[algorithm\] we describe our topological-based optimization algorithm and derive rigorously the topological gradient of a suitable mismatch functional. This is obtained by using the results of Section \[inverse\] and some interior regularity results for the solution of parabolic systems. Finally, in Section \[results\] we outline the numerical implementation of the proposed algorithm, relying on the Finite Element Method for the discrete formulation of the problem. A significant set of numerical experiments is provided in order to assess the effectiveness of the reconstruction, even in presence of data noise.
Analysis of the direct problem {#direct}
==============================
The well-posedness analysis of the monodomain system has been object of several studies. We refer to [@book:pavarino Chapter 3] for a general overview. In [@art:BCP] a result of existence and uniqueness of weak solution is proved for the FitzHugh-Nagumo, the Aliev-Panfilov and the Rogers-MacCulloch models by means of a Faedo-Galerkin procedure. A result of existence of strong solutions, local in time, is also derived. In [@art:veneroni], instead, results of well-posedness are obtained for a wider range of models, on the base of a fixed point argument.
Regarding the regularity of the solutions of the monodomain system, we report a result in [@pierrecoudiere] for FitzHugh-Nagumo, Aliev-Panfilov and Rogers-MacCulloch models: if no discontinuities are present in the coefficients of the system, existence and uniqueness of strong solutions is guaranteed, locally in time (see for instance [@book:smoller] and [@book:henry]). A comparison principle is also provided, by means of the tool of invariant sets, allowing to prove existence of global solutions. We also report a result of local existence of classical solutions for the bidomain model, recently obtained in [@giga].
In this section we focus on the monodomain system in the case of smooth diffusion coefficient and reaction term, corresponding to the case of the healthy tissue (unperturbed case) and in the case of discontinuities in the diffusion coefficient and in the reaction term, corresponding to the presence of an ischemia in the heart tissue (perturbed case). We state an existence, uniqueness and comparison result for classical solutions of the unperturbed case (a proof of which, alternative to the approach of [@pierrecoudiere], is proposed in [@phd:Luca Chapter 6]) and prove a result regarding existence, uniqueness and regularity of weak solutions in the perturbed case.
In particular, the initial and boundary value problem associated to the unperturbed monodomain system is the following one $$\left\{
\begin{aligned}
\partial_t u - div(K_0\nabla u) + f(u,w) &= 0 \qquad &\text{in } Q_T, \\
K_0 \nabla u \cdot \normal &= 0 \qquad &\text{on } \Gamma_T, \\
\partial_t w + g(u,w) &= 0 \qquad &\text{in } Q_T, \\
u(\cdot,0) = u_0 \qquad w(\cdot,0) &= w_0 \qquad &\text{in } \Omega,
\end{aligned}
\right.
\label{eq:formaforte}$$ where $\Omega$ is the region occupied by the hearth tissue, $\nu$ is the outward unit normal vector to the boundary $\partial\Omega$, $Q_T := \Omega \times (0,T)$ and $\Gamma_T: \partial\Omega \times (0,T)$. A slightly different formulation of the monodomain model (see, e.g., [@book:pavarino]) involves the presence of a source term in the right-hand side of the first equation in , representing an applied current, during a limited time window (the initial activation of the tissue). We replace the effect of such a current with the presence of a non-null initial value $u_0$. Since the modeling differences between the two formulations are negligible after the first instants, we choose the one proposed in more suited for the mathematical analysis of the problem.
In presence of an ischemia $\omega\subset\Omega$, the perturbed case is described by the model $$\left\{
\begin{aligned}
\partial_t \uo - div(K_\omega \nabla \uo) + (1-\chi_\omega)f(\uo,\wo) &= 0 \qquad &\text{in } Q_T, \\
K_\omega \nabla \uo \cdot \normal &= 0 \qquad &\text{on } \Gamma_T, \\
\partial_t \wo + g(\uo,\wo) &= 0 \qquad &\text{in } Q_T, \\
\uo(\cdot,0) = u_0 \qquad \wo(\cdot,0) &= w_0 \qquad &\text{in } \Omega,
\end{aligned}
\right.
\label{eq:perturbed}$$ where $\chi_\omega$ is the indicator function of $\omega$ and $K_\omega = K_0 - (K_0-K_1)\chi_\omega$. We now specify the requirements on the domain, the coefficients and the source terms.
1. $\Omega \subset \R^d$ bounded domain, $d=2,3$, and $\partial \Omega \in C^{2+\alpha}$;
2. the inclusion $\omega\subset\Omega$ is well separated from the boundary, i.e., $$\vspace{-0.25cm}
\exists \ \mathcal{C}_0 \text{ compact subset of } \Omega \text{ s.t. } \omega \subset \mathcal{C}_0 \text{ and } dist(\mathcal{C}_0,\partial\Omega) \geq d_0 >0.
\vspace{-0.25cm}
\label{eq:separ}$$
3. $K_0, K_1 \in C^{2}(\overline{\Omega};\R^{d \times d})$ are symmetric matrix-valued functions in $\Omega$; $\forall x \in \overline{\Omega}$, the matrices $K_0(x)$ and $K_1(x)$ admit $d$ positive eigenvalues $k_{0,1} \leq \ldots \leq k_{0,d}$ and $k_{1,1} \leq \ldots \leq k_{1,d}$ respectively, associated to the same eigenvectors $\vec{e}_1(x),\ldots \vec{e}_d(x)$ such that $k_{1,i} \leq k_{0,i}$ $\forall i = 1,\ldots,d$;
4. $u_0 \in C^{2+\alpha}(\overline{\Omega})$, $w_0 \in C^2(\overline{\Omega})$, and $K_0 \nabla u_0 \cdot \normal = 0$ on $\partial\Omega$;
5. we assume $f,g$ as in the Aliev-Panfilov model, namely: $$\vspace{-0.25cm}
f(u,w) = Au(u-a)(u-1) + uw \qquad g(u,w) = \epsilon (Au(u-1-a) + w),
\label{eq:AP}$$ being $A, \epsilon > 0$ and $0<a<1$.
\[ass:1\]
The requirements on the matrix-valued functions $K_0$ and $K_1$ are satisfied by the conductivity tensors prescribed in the model under consideration (see [@book:pavarino; @book:sundes-lines]). According to experimental evidence, the cardiac tissue can be modeled as an orthotropic material, characterized by the presence of fibers and sheets, which define the conductivity eigenvectors. Moreover, the presence of an ischemia does not affect the direction of the fibers, but reduces the value of all the associated eigenvalues. From now on, we will indicate by $k_{min}$ the minimum eigenvalue of $K_1$ and by $k_{max}$ the maximum eigenvalue of $K_0$.
As an immediate consequence of , $f$ and $g$ satisfy the *Tangency condition* on the rectangle $S \defeq [0,1]\times [0,\frac{A(a+1)^2}{4}]$, (see [@amann]), i.e., indicating by $\vec{p}$ a generalized outward normal on $\partial S$ ( for all $(\xi_1,\xi_2) \in \partial S$, $\vec{p}(\xi_1,\xi_2)$ is such that $\vec{p}(\xi_1,\xi_2)\cdot (\xi_1,\xi_2) \geq \vec{p}(\xi_1,\xi_2)\cdot (\eta_1,\eta_2)$ $\forall (\eta_1,\eta_2) \in S$) then, $$\vec{p}(\xi_1,\xi_2) \cdot \left( \begin{aligned} -f(\xi_1,\xi_2) \\ -g(\xi_1,\xi_2)\end{aligned}\ \right) \leq 0 \qquad \forall (\xi_1,\xi_2) \in \partial S.
\label{eq:nagumo}$$ Moreover, the functions $f,g$ are Lipschitz continuous on $S$ with constants $L_f, L_g \leq L$.
For the sake of brevity, in all the formulas we avoid to indicate time and space integration variables with respect to the classical Lebesgue measure, unless it is necessary.
We now outline the main results regarding the well-posedness of the problems and .
Let Assumption \[ass:1\] holds, and suppose that, $\forall x \in \overline{\Omega},$ $(u_0(x),w_0(x)) \in S$. Then, problem admits a unique classical solution $(u,w)$, namely $u \in C^{2+\alpha,1+\alpha/2}(\overline{Q_T})$, $w \in C^{\alpha,1+\alpha/2}(\overline{Q_T})$. Moreover, $(u(x,t),w(x,t)) \in S$, for each $(x,t) \in \overline{Q_T}$. \[th:1\]
The proof of Theorem \[th:1\] is derived by means of classical fixed point argument (see [@book:pao Chapter 8, Sections 9 and 11]). A detailed proof is reported in [@phd:Luca Theorem 6.1].
Regarding the perturbed problem , we note that, although the conductivity tensor and the nonlinear term are discontinuous, we can extend the results obtained in [@art:BCP] thanks to the uniform ellipticity to the boundedness of the conductivity tensors and to the form of the reaction term, deriving the following existence and uniqueness result.
Let the Assumption \[ass:1\] holds, and suppose that, $\forall x \in \overline{\Omega},$ $(u_0(x),w_0(x)) \in S$. Then, problem admits a unique weak solution, i.e., a couple $(\uo,\wo)$ such that $\uo \in L^2(0,T;H^1(\Omega))\cap L^\infty(0,T;L^2(\Omega))$, $\partial_t \uo \in L^2(0,T;H^*)$, $\wo \in L^\infty(0,T;L^2(\Omega))$, $\partial_t \wo \in L^2(Q_T)$ and satisfying, for a.e. $t \in (0,T)$, $$\begin{aligned}
\langle \partial_t \uo, \varphi \rangle_* + \int_\Omega K_\omega \nabla \uo \cdot \nabla \varphi + \int_\Omega (1-\chi_\omega)f(\uo,\wo) \varphi &= 0
\qquad &\forall \varphi \in H^1(\Omega),\\
\int_\Omega \partial_t \wo \psi + \int_\Omega g(\uo,\wo)\psi &= 0 \qquad &\forall \psi \in L^2(\Omega).
\end{aligned}
\label{eq:weak}$$ Moreover, $\uo \in C^{\alpha,\alpha/2}(\overline{Q_T})$, $\wo \in C^{\alpha,\alpha/2}(\overline{Q_T})$ and $(\uo(x,t),\wo(x,t)) \in S$, $\forall \, (x,t) \in \overline{Q_T}$.
\[th:2\]
To start with, note that uniqueness of the weak solution of has been shown in the case of Aliev-Panfilov model in [@art:kunisch-wagner Theorem 1.3] as a byproduct of a stability result obtained by exploiting the specific nonlinear expression of $f$ and $g$.
We proceed now introducing a sequence of regularized problems $(P_k), \, k\in\N,$ of and showing that the sequence of their solutions converges to a weak solution of . We then exploit additional properties inherited by the approximation process to conclude the stated regularity results. The uniqueness argument is briefly sketched.
Since $\chi_\omega$ is an indicator function, surely $\chi_\omega \in L^2(\Omega)$; by density arguments and according to , $$\exists \{\phi_k\}\subset C^2_c(\Omega): \quad 0 \leq \phi_k(x) \leq 1 \quad \forall x \in \Omega, \quad \phi_k \rightarrow \chi_\omega \text{ in $L^2(\Omega)$ and a.e.}, \label{eq:ass2}$$ being $C^2_c(\Omega)$ the space of $C^2$ functions with compact support in $\Omega$. Define $K_k = K_0 + (K_1-K_0)\phi_k$ and let $(u_k,w_k)$ be the solution of the following problem $$\left\{
\begin{aligned}
\partial_t u_k - div(K_k \nabla u_k) + (1-\phi_k)f(u_k,w_k) &= 0 \qquad &\text{in } Q_T, \\
K_0 \nabla u_k \cdot \normal &= 0 \qquad &\text{on } \Gamma_T, \\
\partial_t w_k + g(u_k,w_k) &= 0 \qquad &\text{in } Q_T, \\
u_k(\cdot,0) = u_0 \qquad w_k(\cdot,0) &= w_0 \qquad &\text{in } \Omega,
\end{aligned}
\right.
\label{eq:regularized}$$ where we have used the fact that $K_k = K_0$ on $\partial \Omega$. We observe that, for any fixed $k$, an application of Theorem \[th:1\] ensures the existence and uniqueness of a classical solution of Problem . Moreover by we have that $(1-\phi_k)f$ and $g$ satisfy the Tangency condition on $S$. Hence, from Theorem \[th:1\] we deduce $(u_k,w_k)\in S$, for all $k$.
We now prove that from $\phi_k \xrightarrow{L^2} \chi_\omega$ the convergence of $(u_k,w_k)$ to a weak solution $(u,w)$ of holds. We start by proving some *a priori* estimates. Consider the weak form of the problem solved by $(u_k,w_k)$ and take the classical solutions $u_k$, $w_k$ as test functions. Then, we get $$\begin{aligned}
&\half \frac{d}{dt}\left(\norm{L^2(\Omega)}{u_k(\cdot,t)}^2 + \norm{L^2(\Omega)}{w_k(\cdot,t)}^2 \right)
+ \int_\Omega K_k\nabla u_k(\cdot,t)\cdot \nabla u_k(\cdot,t) \\
& \quad = -\int_\Omega (1-\phi_k)f(u_k(\cdot,t),w_k(\cdot,t))u_k(\cdot,t)- \int_\Omega g(u_k(\cdot,t),w_k(\cdot,t))w_k(\cdot,t).
\end{aligned}$$ Recall now that $k_{min}$ is the minimum eigenvalue of $K_1$, whereas $k_{max}$ is the maximum eigenvalue of $K_0$. Moreover, since $\phi_k$, $u_k$, $w_k$ are uniformly bounded independently of $k$ (indeed, $\phi_k(x)\in [0,1]$ and $(u_k,w_k)\in S$) and $f,g$ are continuous, we can introduce $M_f := max_{(x,t)\in Q_T} \left|(1-\phi_k)f(u_k,w_k)\right|$ and $M_g:= max_{(x,t)\in Q_T} \left|g(u_k,w_k)\right|$, which are independent of $k$. Hence, by Young inequality, $$\label{Young}
\begin{aligned}
&\half \frac{d}{dt}\left(\norm{L^2(\Omega)}{u_k(\cdot,t)}^2 + \norm{L^2(\Omega)}{w_k(\cdot,t)}^2 \right)
+ k_{min} \norm{L^2(\Omega)}{\nabla u_k(\cdot,t)}^2 \\
&\leq \half \left(\norm{L^2(\Omega)}{u_k(\cdot,t)}^2 + \norm{L^2(\Omega)}{w_k(\cdot,t)}^2 \right)
+ \half |\Omega| (M^2_f + M^2_g).
\end{aligned}$$ Using Gronwall’s inequality, we get $$\begin{aligned}
&\left(\norm{L^2(\Omega)}{u_k(\cdot,t)}^2 + \norm{L^2(\Omega)}{w_k(\cdot,t)}^2 \right) \\
&\quad \leq \left( \norm{L^2(\Omega)}{u_0}^2 + \norm{L^2(\Omega)}{w_0}^2 + |\Omega|(M^2_f+M^2_g)t \right)e^{t},
\end{aligned}$$ which implies $$\begin{aligned}
& \norm{L^\infty(0,T;L^2(\Omega))}{u_k}^2+ \norm{L^\infty(0,T;L^2(\Omega))}{w_k}^2 \\
&\leq \left( \norm{L^2(\Omega)}{u_0}^2 + \norm{L^2(\Omega)}{w_0}^2 + |\Omega|(M^2_f+M^2_g)T \right)e^{T} \defeq c_1^2.
\end{aligned}$$ Integrating from $0$ to $t$ (\[Young\]) and using last estimate it also follows that $$\norm{L^2(0,T,H^1(\Omega))}{u_k}^2 \leq c_1^2T + \frac{1}{2 k_{min}}\left((|\Omega|(M^2_f+M^2_g)+c_1^2)T+\norm{L^2(\Omega)}{u_0}^2 + \norm{L^2(\Omega)}{w_0}^2\right) \eqdef c_2^2.$$ A bound for the $H^*$ norm of $\partial_t u$ can be found by considering that, for each $\varphi \in H^1(\Omega)$, $$\begin{aligned}
\left|\langle \partial_t u_k(\cdot,t), \varphi \rangle_*\right| &\leq k_{max}\norm{L^2(\Omega)}{\nabla u_k(\cdot,t)}\norm{L^2(\Omega)}{\nabla \varphi}
+ M_f |\Omega|^\half \norm{L^2(\Omega)}{\varphi} \\
%&\leq max\{k_{max},M_f |\Omega|^\half\}\norm{H^1(\Omega)}{u_k(\cdot,t)} \norm{H^1(\Omega)}{\varphi}
\end{aligned}$$ and computing the $L^2$ norm in time $$\norm{L^2(0,T,H^*)}{\partial_t u_k} \leq (k_{max}c_2+M_f |\Omega|^\half) \eqdef c_3.$$ Analogously, one proves that $\norm{L^2(Q_T)}{\partial_t w_k}\leq c_4$, with $c_4$ independent of $k$.
As a consequence of the uniform bounds we can ensure that (up to a subsequence) $\exists u \in L^2(0,T;H^1(\Omega)) \cap L^\infty(0,T;L^2(\Omega))$, $\exists w \in L^\infty(0,T;L^2(\Omega))$, $\exists u^* \in L^2(0,T;H^*)$, $\exists w^* \in L^2(Q_T)$ such that $$u_k \xrightharpoonup{L^2(0,T,H^1)} u, \quad \partial_t u_k \xrightharpoonup{L^2(0,T,H^*)} u^*, \quad w_k \xrightharpoonup{L^2(Q_T)} w,
\quad \partial_t w_k \xrightharpoonup{L^2(Q_T)} w^*.$$ Using the result contained in [@book:robinson Theorem 8.1] and the uniqueness of the weak solution of the perturbed problem we have that $u_k \xrightarrow{L^2(Q_T)} \uo$ (see [@book:robinson Theorem 8.1]), hence $u_k \rightarrow \uo$ a.e. in $Q_T$. Moreover, also $w_k \rightarrow \wo$ a.e. in $Q_T$ as it can be straightforwardly obtained by the expression $$w_k(x,t) = e^{-\epsilon t}w_0(x) + \epsilon A e^{-\epsilon t}\int_0^t ((1+a)u_k - u_k^2)e^{\epsilon s} ds.
\label{eq:solwk}$$ Consider now the weak form of the problem solved by $(u_k,w_k)$: $\forall \varphi \in H^1(\Omega)$, $\psi \in L^2(\Omega)$. $$\begin{aligned}
& \langle \partial_t u_k, \varphi \rangle_* + \int_\Omega K_k \nabla u_k \cdot \nabla \varphi + \int_\Omega(1-\phi_k) Au_k(u_k-a)(u_k-1) \varphi \\
& \quad + \int_\Omega (1-\phi_k)u_k w_k \varphi + \int_\Omega \partial_t w_k \psi
+ \int_\Omega A\epsilon(u_k^2 - (1+a)u_k) \psi + \int_\Omega \epsilon w_k \psi = 0.
\label{eq:regularizedweak}
\end{aligned}$$ Taking the limit in $k$, exploiting the weak convergence of $u_k$, $w_k$, $\partial_t u_k$, $\partial_t w_k$, we obtain that $u^* = \partial_t \uo$, $w^* = \partial_t \wo$ in $\mathcal{D}'(0,T)$ and that the limit $(\uo,\wo)$ satisfies, $\forall \varphi \in H^1(\Omega)$, $\psi \in L^2(\Omega)$, $$\begin{aligned}
& \langle \partial_t \uo, \varphi \rangle_* + \int_\Omega K_\omega \nabla \uo \cdot \nabla \varphi + \int_\Omega(1-\chi_\omega) f(\uo,\wo) \varphi \\
& \quad + \int_\Omega \partial_t \wo \psi + \int_\Omega g(\uo,\wo) \psi = 0.
\label{eq:weakuw}
\end{aligned}$$ Indeed, the convergence of the terms involving the time derivatives is a direct consequence of the weak convergence of $\partial_t u_k$, $\partial_t w_k$ and of the definition of distributional derivative. The limit of the nonlinear reaction terms can be proved by taking advantage of the dominated convergence theorem and of the pointwise (a.e.) convergence of $u_k$ and $\phi_k$. The convergence of the diffusion term is obtained by combining the weak convergence of $u_k$ in $H^1(\Omega)$, the pointwise (a.e.) convergence of $\phi_k$ and the uniform $L^\infty(\Omega)$ bound on $\phi_k$. According to , we can ensure that the limit $(\uo,\wo)$ is the weak solution of .
The weak solution $(\uo,\wo)$ is moreover a pointwise (a.e.) limit of the regularized solutions $(u_k,w_k)$. As a consequence, the uniform bound on $(u_k,w_k)$ is valid also for the limit: $(\uo(x,t),\wo(x,t)) \in S$ a.e. in $Q_T$. This allows to prove the additional Hölder regularity of $\uo$. Indeed, since $K_\omega \in L^{\infty}(\Omega)$, $f(\uo,\wo) \in L^{\infty}(Q_T)$, we can apply Theorem 10.1 of [@book:lady Chapter 3] on the first equation in $$\partial_t \uo - div(K_\omega \nabla \uo) = -(1-\chi_\omega) f(\uo,\wo)$$ to get $\uo \in L^{\infty}(Q_T)$. Now, we can extend the regularity result up to the boundary due to the hypothesis on $\partial \Omega$ and on $u_0$ contained in Assumption \[ass:1\], and conclude $\uo \in C^{\alpha,\alpha/2}(\overline{Q_T})$. Using the analytic expression of $\wo$ that can be obtained by and the regularity of $\uo$, we can also deduce that $\wo \in C^{\alpha,1+\alpha/2}(\overline{Q_T})$.
Analysis of the inverse problem {#inverse}
===============================
We now tackle the inverse problem of identifying a perturbation $\omega$ from boundary measurements. Suppose to know $u_{meas}$, the trace of the solution of in presence of an unknown inclusion. The inverse problem reads $$\textit{find $\omega \subset \Omega$ s.t. } u_\omega|_{\partial \Omega} = u_{meas}.
\label{eq:inv}$$ Although the analysis of the direct problem has been performed in presence of an arbitrary inclusion $\omega \subset \Omega$, for the purpose of solving the inverse problem and derive a reconstruction algorithm, we limit ourselves at considering the case of inclusions of small size, in analogy to what done in [@art:BCMP], [@art:bccmr]. In particular, we consider a family of inclusions $\omega_\varepsilon$ satisfying for each $\varepsilon$ and such that $$|\omega_\varepsilon| \rightarrow 0 \text{ as } \varepsilon \rightarrow 0.
\label{eq:small}$$
We define $\chi_\varepsilon$ the indicator function of $\omegae$, $K_\varepsilon :=K_0 - (K_0-K_1)\chie$ and $(\ue,\we)$ the solution of problem with $\omega = \omegae$, in the sense of Theorem \[th:2\].
Energy estimates
----------------
We now derive some *energy estimates* for the difference between $(\ue,\we)$ and $(u,w)$, the solution of the unperturbed problem in terms of $|\omega_\varepsilon|$ when $\varepsilon \rightarrow 0$.
From now on we will indicate by $C$ a positive constant depending on the data, independent of $\varepsilon$ and that may vary also in the same line.
Under Assumption 2.1 the following inequalities hold
$\displaystyle \norm{L^\infty(0,T,L^2(\Omega))}{\ue- u} + \norm{L^\infty(0,T,L^2(\Omega))}{\we- w} \leq C |\omegae|^{\half}$, 0.2truecm $\displaystyle \norm{L^2(0,T,H^1(\Omega))}{\ue- u} \leq C |\omegae|^{\half}$, 0.2truecm $\displaystyle \norm{L^2(Q_T)}{\ue - u} \leq C |\omegae|^{\half + \eta}, \quad \text{ for some } \eta > 0$. \[prop:energy\]
We consider with $\omega=\omegae$ and the weak formulation of the unperturbed problem . Subtracting term by term and defining $\Ue = \ue - u$ and $\We = \we - w$, we have that $\Ue(\cdot,0)= 0$, $\We(\cdot,0)=0$ and, for almost every $t\in (0,T)$, $$\begin{aligned}
&\int_\Omega \partial_t \Ue \varphi + \int_\Omega K_\varepsilon \nabla \Ue \cdot \nabla \varphi + \int_\Omega (1-\chi_\varepsilon)(f(\ue,\we)-f(u,w))\varphi
+ \int_\Omega \partial_t \We \psi \\
&+ \int_\Omega (g(\ue,\we)-g(u,w))\psi = \int_{\omegae} (K_0- K_1) \nabla u\cdot \nabla \varphi + \int_{\omegae} f(u,w) \varphi.
\end{aligned}
\label{eq:mn}$$ Let $\varphi = \Ue$, $\psi = \We$ then, according to Theorem \[th:1\] and Theorem \[th:2\], both $(u,w)$ and $(\ue,\we)$ range within the rectangle $S$, on which the functions $f$ and $g$ are Lipschitz continuous with constants less or equal than $L$. Hence $$\begin{aligned}
& \half \frac{d}{dt}\left(\norm{L^2(\Omega)}{\Ue}^2 + \norm{L^2(\Omega)}{\We}^2\right)
+ k_{min} \norm{L^2(\Omega)}{\nabla \Ue}^2 \leq 2L\left( \norm{L^2(\Omega)}{\Ue}^2 + \norm{L^2(\Omega)}{\We}^2\right) \\
& \quad + \int_{\omegae} (K_0- K_1) \nabla u\cdot \nabla \Ue + \int_{\omegae} f(u,w) \Ue.
\end{aligned}$$ Via an application of Schwarz and Young inequalities on the last two terms in the right-hand side of the previous inequality, and from the regularity of the solution $(u,w)$, we can deduce $$\begin{aligned}
\half \frac{d}{dt}\left(\norm{L^2(\Omega)}{\Ue}^2 + \norm{L^2(\Omega)}{\We}^2\right) &+\frac{k_{min}}{2} \norm{L^2(\Omega)}{\nabla \Ue}^2\\
&\leq C \left( \norm{L^2(\Omega)}{\Ue}^2 + \norm{L^2(\Omega)}{\We}^2\right) + C |\omegae|.
\label{eq:ener}
\end{aligned}$$ An application of Gronwall’s lemma to entails that $\norm{L^\infty(0,T;L^2(\Omega))}{\Ue}^2$, $\norm{L^\infty(0,T;L^2(\Omega))}{\We}^2$ and ultimately $\norm{L^2(0,T;H^1(\Omega))}{\Ue}^2$ can be bounded by $C |\omegae|$. This allows to conclude the first two statements.
Observe now that the pair $(\Ue,\We)$ is also the solution of $$\left\{
\begin{aligned}
&\partial_t \Ue - div(K_0 \nabla \Ue) + (1-\chi_\varepsilon)(f(\ue,\we)-f(u,w)) =\\
& \qquad -div((K_0-K_1)\chi_\varepsilon \nabla \ue) + \chi_\varepsilon f(u,w) \\
&\partial_t \We + g(\ue,\we) - g(u,w) = 0.
\end{aligned}
\right.
\label{eq:mn2}$$ By the mean value theorem, there exist two pairs of functions $(u_{\xi_1}, w_{\xi_1})$, $(u_{\xi_2}, w_{\xi_2})$ s.t. [^1] $$\begin{aligned}
\int_\Omega (1-\chi_\varepsilon)(f(\ue,\we)-f(u,w))\Ue &= \int_\Omega (1-\chi_\varepsilon)f_u(u_{\xi_1}, w_{\xi_1})\Ue^2
+ \int_\Omega (1-\chi_\varepsilon)f_w(u_{\xi_1}, w_{\xi_1})\Ue\We \\
\int_\Omega (g(\ue,\we)-g(u,w))\We &= \int_\Omega g_u(u_{\xi_2}, w_{\xi_2})\Ue\We + \int_\Omega g_w(u_{\xi_2}, w_{\xi_2})\We^2.
\end{aligned}$$ By definition, $(u_{\xi_1}, w_{\xi_1})$ and $(u_{\xi_2}, w_{\xi_2})$ are convex combinations of $(u,w)$ and $(\ue,\we)$, thus they assume values in the rectangle $S$. Let now $(\overline{U}_\varepsilon,\overline{W}_\varepsilon)$ be the solution of the adjoint problem $$\left\{
\begin{aligned}
\partial_t \overline{U}_\varepsilon + div(K_0 \nabla \overline{U}_\varepsilon)
- (1-\chi_\varepsilon)f_u(u_{\xi_1},w_{\xi_1})\overline{U}_\varepsilon - g_u(u_{\xi_2},w_{\xi_2}) \overline{W}_\varepsilon&= - \Ue \\
\partial_t \overline{W}_\varepsilon - (1-\chi_\varepsilon)f_w(u_{\xi_1},w_{\xi_1})\overline{U}_\varepsilon - g_w(u_{\xi_2},w_{\xi_2})\overline{W}_\varepsilon
&= -\We,
\end{aligned}
\right.
\label{eq:mn3}$$ with initial conditions $\overline{U}_\varepsilon(\cdot,T)=0$, $\overline{W}_\varepsilon(\cdot,T)=0$ and homogeneous Neumann boundary condition. Consider the change of variable: $z(\cdot,t) = \overline{U}_\varepsilon(\cdot,T-t)$, $y(\cdot,t) = \overline{W}_\varepsilon(\cdot,T-t)$ and define $\widehat{U}_\varepsilon(\cdot,t) = \Ue(\cdot,T-t)$, $\widehat{W}_\varepsilon(\cdot,t) = \We(\cdot,T-t)$, $\hat{f}_u(\cdot,t) = f_u(u_{\xi_1}(\cdot,T-t),w_{\xi_1}(\cdot,T-t))$, and analogously for $\hat{f}_w,\hat{g}_u,\hat{g}_w$. Hence $z$ and $y$ solve $$\left\{
\begin{aligned}
\partial_t z - div(K_0 \nabla z) + (1-\chi_\varepsilon)\hat{f}_u z + \hat{g}_u y&= \widehat{U}_\varepsilon\quad \text{ in }Q_T, \\
\partial_t y + (1-\chi_\varepsilon)\hat{f}_w z + \hat{g}_w y &= \widehat{W}_\varepsilon\quad\text{ in }Q_T,\\
z(0)=0,\quad y(0)&=0\quad \text{ in }\Omega,\\
K_0\nabla z\cdot{\nu}&=0\quad\text{ on }\Gamma_T.
\end{aligned}
\right.
\label{eq:mn4}$$ Since $\widehat{U}_\varepsilon,\widehat{W}_\varepsilon \in L^2(Q_T)$ and $\hat{f}_u, \hat{f}_w,\hat{g}_u,\hat{g}_w$ are bounded in $\overline{Q}_T$, by standard Faedo-Galerkin technique we obtain that the solution $(z,y)$ of exists and is unique with the properties $z \in L^2(0,T;H^1(\Omega)) \cap L^\infty(0,T;L^2(\Omega))$, $\partial_t z \in L^2(0,T;H^*)$, $y \in L^\infty(0,T;L^2(\Omega))$, $\partial_t y \in L^2(Q_T)$. Moreover $$\begin{aligned}
\norm{L^\infty(0,T;L^2(\Omega))}{z} &+ \norm{L^2(0,T;H^1(\Omega))}{z} + \norm{L^2(0,T;H^*)}{\partial_t z} + \norm{L^\infty(0,T;L^2(\Omega))}{y} + \norm{L^2(Q_T)}{\partial_t y}
\\ &\leq C \left(\norm{L^2(Q_T)}{\widehat{U}_\varepsilon} + \norm{L^2(Q_T)}{\widehat{W}_\varepsilon}\right) = C \left(\norm{L^2(Q_T)}{\Ue} + \norm{L^2(Q_T)}{\We}\right).
\end{aligned}$$ Additional regularity of $z$ can be proved with an analogous argument as in [@book:lady Chapter 4, Theorem 9.1], applied on the first equation in . Indeed, by the regularity of $K_0$ and the square integrability of $\widehat{U}_\varepsilon - \hat{g}_u y$, we can conclude that $z \in L^2(0,T;H^2(\Omega))$, $\partial_t z \in L^2(Q_T)$ and $$\norm{L^2(0,T;H^2(\Omega))}{z} + \norm{L^2(Q_T)}{\partial_t z} \leq C \norm{L^2(Q_T)}{\widehat{U}_\varepsilon - \hat{g}_u y}
\leq C\left( \norm{L^2(Q_T)}{\Ue} + \norm{L^2(Q_T)}{\We}\right).$$ Moreover, multiplying the first two equations in respectively by $\partial_t z$ and $\partial_t y$ and integrating on $\Omega$, straightforward computations allow to conclude that $z \in L^\infty(0,T;H^1(\Omega))$ with $$\norm{L^\infty(0,T;H^1(\Omega))}{z} \leq C \left(\norm{L^2(Q_T)}{\Ue} + \norm{L^2(Q_T)}{\We}\right).$$ By Sobolev inequality we obtain $$\begin{aligned}
\nabla z \in (L^2(0,T;L^6(\Omega)))^d,\,\, z \in L^6(Q_T),
\end{aligned}$$ (all the norms being bounded by $\norm{L^2(Q_T)}{\Ue} + \norm{L^2(Q_T)}{\We}$; the same bounds hold on $\overline{U}_\varepsilon$). Thus, via Nirenberg interpolation estimates (see [@nirenberg]), we get $$\begin{aligned}
\norm{L^{10/3}(Q_T)}{\nabla \overline{U}_\varepsilon} \leq C (\norm{L^2(Q_T)}{\Ue} + \norm{L^2(Q_T)}{\We}).
\end{aligned}$$ Recalling also the previous results, this allows to conclude that, taking $p \in \left(2,\frac{10}{3}\right]$, $$\norm{L^p(Q_T)}{\overline{U}_\varepsilon} + \norm{L^p(Q_T)}{\nabla \overline{U}_\varepsilon} \leq C \left( \norm{L^2(Q_T)}{\Ue} + \norm{L^2(Q_T)}{\We}\right).
\label{eq:Lp}$$ Let us multiply the equations of respectively by $\overline{U}_\varepsilon,\overline{W}_\varepsilon$ and the first two equations of respectively by $\Ue,\We$. Integrating on $(0,T)$ and summing the resulting identities it is easy to see that $$\int_0^T \int_\Omega (\Ue^2+ \We^2) = \int_0^T \int_{\omegae} (K_0 - K_1) \nabla \ue \cdot \nabla \overline{U}_\varepsilon
+ \int_0^T \int_{\omegae} f(u,w)\overline{U}_\varepsilon.
\label{eq:aux1}$$ Thanks to and Hölder inequality the first term of the right-hand side can be bounded as follows $$\begin{aligned}
\int_0^T \int_{\omegae} (K_0 - K_1) \nabla \ue \cdot \nabla \overline{U}_\varepsilon
&\leq C \norm{L^q(\omegae \times (0,T))}{\nabla \ue} \norm{L^p(\omegae \times (0,T))}{\nabla \overline{U}_\varepsilon} \\
&\leq C \norm{L^q(\omegae \times (0,T))}{\nabla \ue} \left( \norm{L^2(Q_T)}{\Ue} + \norm{L^2(Q_T)}{\We} \right),
\end{aligned}$$ where $p \in \left(2,\frac{10}{3}\right]$ and $q = \frac{p}{p-1} \in \left[\frac{10}{7}, 2\right)$. In addition, again by Hölder inequality, $$\begin{aligned}
\norm{L^q(\omegae \times (0,T))}{\nabla \ue} &\leq \norm{L^q(\omegae \times (0,T))}{\nabla u}
+ |\omegae|^{\frac{2-q}{2q}}\norm{L^q(\Omega \times (0,T))}{\nabla \Ue} \leq c |\omegae|^\frac{1}{q}.
\end{aligned}$$ Furthermore, it is straightforward to see that the last term in can be bounded by $c|\omegae|^\frac{1}{q} \norm{L^2(Q_T)}{\overline{U}_\varepsilon}$. Finally, from we conclude that, for $q \in \left[\frac{10}{7},2\right)$, $$\norm{L^2(Q_T)}{\Ue}^2 + \norm{L^2(Q_T)}{\We}^2 \leq c |\omegae|^{\frac{1}{q}}\left( \norm{L^2(Q_T)}{\Ue} + \norm{L^2(Q_T)}{\We}\right).$$ Hence $$\norm{L^2(Q_T)}{\Ue} + \norm{L^2(Q_T)}{\We} \leq c|\omegae|^\frac{1}{q} = c|\omegae|^{\frac{1}{2}+\beta},$$ where $\beta \in \left(0,\frac{1}{5}\right]$.
Asymptotic expansion of boundary voltage
----------------------------------------
In this section we prove the following result.
For every sequence $\{\omega_{\varepsilon_n}\}$ with $|\omega_{\varepsilon_n}|>0$, $|\omega_{\varepsilon_n}| \rightarrow 0$, there exist a subsequence (still denoted by $\omega_{\varepsilon_n}$), a Radon measure $\mu$ and a matrix-valued function $\mathcal{M} \in L^2(\Omega,d\mu;\R^{d\times d})$ such that, for every pair $(\Phi,\Psi)\in (C^1(\overline{Q}_T),C(\overline{Q}_T))$ satisfying $$\left\{
\begin{aligned}
\partial_t \Phi + div(K_0 \nabla \Phi) - f_u(u,w) \Phi - g_u(u,w)\Psi &= 0 \\
\partial_t \Psi - f_w(u,w)\Phi - g_w(u,w)\Psi &= 0 \\
\end{aligned}
\right.
\label{eq:PhiPsi}$$ with the final conditions $\Phi(\cdot,T)=0$, $\Psi(\cdot,T)=0$, it holds $$\int_0^T\!\!\!\!\int_{\partial \Omega} \!\!\!\!K_0 \nabla \Phi \cdot \normal (\uen - u)d\sigma
= |\omegaen| \int_0^T\!\!\!\!\int_\Omega \!\!\left[ \mathcal{M}(K_0-K_1)\nabla u \cdot \nabla \!\Phi + f(u,w)\Phi \right]d\mu + o(|\omegaen|).
\label{eq:expansion}$$ \[th:3\]
In order to prove Theorem \[th:3\], we need to introduce the auxiliary functions $\vj$, $\venj$ solving $$\left\{ \begin{aligned}
div(K_0 \nabla \vj) &= F^{(j)} \quad &\textit{ in } \Omega ,\\
K_0\nabla \vj \cdot \normal &= f^{(j)}\quad &\textit{ on } \partial \Omega, \\
\int_{\partial \Omega} \vj &= 0, &
\end{aligned}\right.
\qquad \quad
\left\{ \begin{aligned}
div(\Ken \nabla \venj) &= F^{(j)} \quad &\textit{ in } \Omega, \\
K_0\nabla \venj \cdot \normal &= f^{(j)} \quad &\textit{ on } \partial \Omega, \\
\int_{\partial \Omega} \venj &= 0, &
\end{aligned}\right.
\label{eq:vjvenj}$$ with $F^{(j)} = \sum_{i} \frac{\partial[K_0]_{ij}}{\partial x_i}$ and $f^{(j)} = \sum_{i} [K_0]_{ij}\normal_i$. By the choice of $F^{(j)}$ and $f^{(j)}$ it holds $\vj = x_j - \int_{\partial \Omega} x_j$. Moreover, $\vj$ and $\venj$ satisfy the following energy estimates (see [@art:capdebosq]) $$\norm{H^1(\Omega)}{\venj - \vj} \leq c |\omegaen|^\half, \qquad \norm{L^2(\Omega)}{\venj - \vj} \leq c |\omegaen|^{\half + \gamma}, \,\, \gamma >0.
\label{eq:energyv}$$ We need also a preliminary lemma.
For each $\phi \in C^1(\overline{Q}_T)$ such that $\phi(x,T)=0$, we have, as $|\omegaen| \rightarrow 0$, $$\int_0^T \int_\Omega \frac{\chi_\omegaen}{|\omegaen|}(K_0-K_1)\nabla u \cdot \nabla \venj \phi
= \int_0^T \int_\Omega \frac{\chi_\omegaen}{|\omegaen|}(K_0 - K_1)\nabla \uen \cdot \nabla \vj \phi + o(1).
\label{eq:lemma4}$$ \[lem:asym\]
Since $\vj$ and $\venj$ are the solutions of problems , we obtain $$\int_\Omega \Ken \nabla \venj \cdot \nabla \varphi = \int_\Omega K_0 \nabla \vj \cdot \nabla \varphi \qquad \forall \varphi \in H^1(\Omega).$$ Take $\varphi = \Uen \phi$, being $\Uen = \uen - u$. By computation we get $$\int_\Omega \Ken \nabla \venj \cdot \nabla \Uen \phi + \int_\Omega \Ken \nabla \venj \cdot \nabla \phi \Uen =
\int_\Omega K_0 \nabla \vj \cdot \nabla \Uen\phi + \int_\Omega K_0 \nabla \vj \cdot \nabla \phi \Uen$$ that we can rewrite as $$\begin{aligned}
&\int_\Omega \Ken \nabla (\venj\phi) \cdot \nabla \Uen - \int_\Omega \Ken \venj \nabla \phi \cdot \nabla \Uen
+ \int_\Omega \Ken \nabla \venj \cdot \nabla \phi \Uen \\
&\quad = \int_\Omega K_0 \nabla (\vj\phi) \cdot \nabla \Uen - \int_\Omega K_0 \vj \nabla \phi \cdot \nabla \Uen
+ \int_\Omega K_0\nabla \vj \cdot \nabla \phi \Uen.
\end{aligned}
\label{eq:bob}$$ Proposition \[prop:energy\] and lead to $$\begin{aligned}
&\int_0^T\int_\Omega \Ken \venj \nabla \phi \cdot \nabla \Uen - \int_0^T\int_\Omega K_0 \vj \nabla \phi \cdot \nabla \Uen = \int_0^T\int_\Omega (\Ken - K_0) \nabla \uen \cdot \nabla \phi \vj\\
& \quad - \int_0^T\int_\Omega (\Ken - K_0) \nabla u \cdot \nabla \phi \venj + o(|\omegaen|),
\end{aligned}$$ whereas, $$\begin{aligned}
&\int_0^T\int_\Omega \Ken \nabla \venj \cdot \nabla \phi \Uen - \int_0^T\int_\Omega K_0\nabla \vj \cdot \nabla \phi \Uen \\
&\quad = \int_0^T\int_\Omega (\Ken - K_0) \nabla \venj \cdot \nabla \phi \Uen + \int_0^T\int_\Omega K_0 \nabla (\venj - \vj) \cdot \nabla \phi \Uen = o(|\omegaen|).
\end{aligned}$$ Collecting the previous two relations in , we conclude $$\begin{aligned}
& \int_0^T\int_\Omega \Ken \nabla (\venj\phi) \cdot \nabla \Uen - \int_0^T\int_\Omega K_0 \nabla (\vj\phi) \cdot \nabla \Uen \\
& \quad = \int_0^T\int_\Omega (\Ken - K_0) \nabla \uen \cdot \nabla \phi \vj - \int_0^T\int_\Omega (\Ken - K_0) \nabla u \cdot \nabla \phi \venj + o(|\omegaen|).
\end{aligned}
\label{eq:vpart}$$ It can be easily verified that, for every $\varphi \in H^1(\Omega)$, the following identities hold $$\begin{aligned}
&\int_\Omega \partial_t \Uen \varphi + \int_\Omega \Ken \nabla \Uen \cdot \nabla \varphi + \int_\Omega (1-\chi_{\omegaen})(f(\uen,\wen)-f(u,w))\varphi \\
& \quad = \int_\Omega (K_0 - \Ken)\nabla u \cdot \nabla \varphi + \int_\Omega \chi_{\omegaen}f(u,w) \varphi ,\\
&\int_\Omega \partial_t \Uen \varphi + \int_\Omega K_0 \nabla \Uen \cdot \nabla \varphi + \int_\Omega (1-\chi_{\omegaen})(f(\uen,\wen)-f(u,w))\varphi \\
& \quad = \int_\Omega (K_0 - \Ken)\nabla \uen \cdot \nabla \varphi + \int_\Omega \chi_{\omegaen}f(u,w) \varphi. \\
\end{aligned}$$ Taking $\varphi = \venj \phi$ in the first identity and $\varphi = \vj \phi$ in the second one, integrating in time and subtracting we obtain $$\begin{aligned}
&\int_0^T \!\!\int_\Omega \partial_t \Uen(\venj-\vj)\phi + \int_0^T\!\!\int_\Omega \Ken\nabla\Uen\cdot\nabla(\venj\phi) - \int_0^T\!\!\int_\Omega K_0\nabla\Uen\cdot\nabla(\vj\phi)\\
& + \!\! \int_0^T\!\!\!\!\int_\Omega (1-\!\chi_{\omegaen})(f(\uen,\wen)-\!f(u,w))(\venj-\vj)\phi = \!\! \int_0^T \!\!\!\!\int_\Omega \chi_{\omegaen}f(u,w) (\venj-\!\vj)\phi \\
& +\int_0^T\!\!\int_\Omega (K_0 - \Ken)\nabla u \cdot \nabla (\venj\phi) - \int_0^T\!\!\int_\Omega (K_0 - \Ken)\nabla \uen \cdot \nabla (\vj\phi).
\end{aligned}
\label{eq:aux43}$$ Via integration by parts, and taking advantage of the homogeneous initial and final conditions satisfied respectively by $U_{\varepsilon_n}$ and $\phi$ it follows $$\begin{aligned}
&\left | \int_0^T\int_\Omega \partial_t \Uen(\venj-\vj)\phi \right | \leq \norm{C^1(Q_T)}{\phi}\norm{L^2(Q_T)}{\Uen}\norm{L^2(\Omega)}{\venj-\vj} = o(|\omegaen|).
\end{aligned}$$ Analogous bounds can be proved for the last term in the left-hand side and the first term in the right-hand side of , thanks to the energy estimates of $\venj - \vj, \, \uen - u,\, \wen - w$ and the regularity of $f$. As a consequence, it holds that $$\begin{aligned}
& \int_0^T\int_\Omega \Ken\nabla\Uen\cdot\nabla(\venj\phi) - \int_0^T\int_\Omega K_0\nabla\Uen\cdot\nabla(\vj\phi) \\
& = \quad \int_0^T\int_\Omega (K_0 - \Ken)\nabla u \cdot \nabla (\venj\phi) - \int_0^T\int_\Omega (K_0 - \Ken)\nabla \uen \cdot \nabla (\vj\phi) + o(|\omegaen|).
\end{aligned}
\label{eq:upart}$$ In conclusion, by a combination of and , $$\begin{aligned}
&\int_0^T\int_\Omega (K_0 - \Ken)\nabla u \cdot \nabla (\venj\phi) - \int_0^T\int_\Omega (K_0 - \Ken)\nabla \uen \cdot \nabla (\vj\phi) \\
& \quad = \int_0^T\int_\Omega (\Ken - K_0) \nabla \uen \cdot \nabla \phi \vj - \int_0^T\int_\Omega (\Ken - K_0) \nabla u \cdot \nabla \phi \venj + o(|\omegaen|),
\end{aligned}$$ which immediately entails the thesis.
We are finally ready to prove the main result of this section.
Arguing as in [@art:capdebosq], we deduce that there exist a Radon measure $\mu \in C(\overline{\Omega})^*$, a symmetric matrix $\mathcal{M} \in L^2(\Omega,d\mu;\R^{d\times d})$ and a sequence $\{\omegaen\}$ such that $$\frac{\chi_\omegaen}{|\omegaen|} \dx \rightarrow \dmu
\quad \text{and} \quad
\frac{\chi_\omegaen}{|\omegaen|}\frac{\partial \venj}{\partial x_i}\dx \rightarrow \mathcal{M}_{ij}\dmu$$ in the weak\* topology of $C(\overline{\Omega})$. Moreover, thanks to the regularity estimates for $u$ and the symmetries of the matrices $K_0$, $K_1$ and $\mathcal{M}$, we have, for every $j=1,\ldots,d$, $$\frac{\chi_\omegaen}{|\omegaen|} (K_0-K_1)_{ik}\frac{\partial u}{\partial x_k} \frac{\partial \venj}{\partial x_i}\dx \rightarrow \mathcal{M}_{ji}(K_0-K_1)_{ik}\frac{\partial u}{\partial x_k}\dmu, \qquad \forall t \in (0,T),
\label{eq:Mu}$$ which also implies $$\frac{\chi_\omegaen}{|\omegaen|} (K_0-K_1)_{ik}\frac{\partial u}{\partial x_k} \frac{\partial \venj}{\partial x_i}\dx \dt \rightarrow \mathcal{M}_{ji}(K_0-K_1)_{ik}\frac{\partial u}{\partial x_k}\dmu dt
\label{eq:Mut}$$ in the weak$^*$ topology of $C(\overline{Q}_T)$.
On account of the energy estimates , straightforward computations show that for every $i,j=1,\ldots,d$ there exists a constant $C$, depending on the data and on $T$ but independent of $\varepsilon_n$, such that $$\left|\int_0^T \int_\Omega \frac{\chi_\omegaen}{|\omegaen|} (K_0 - K_1)_{ik} \frac{\partial \uen}{\partial x_k} \frac{\partial \vj}{\partial x_i} \dx \dt \right |\leq C.$$ In particular, we define the limit measure $\nu_j$ in the weak$^*$ topology of $C(\overline{Q}_T)$) as follows $$\frac{\chi_\omegaen}{|\omegaen|}\left[ (K_0 - K_1) \nabla \uen \right]_i \frac{\partial \vj}{\partial x_i} \dx \dt \rightarrow \dnuj
\label{eq:nuj}$$ Exploiting and , and recalling Lemma 3.3, we deduce $$\int_0^T\int_{\Omega}\phi\dnuj = \int_0^T\int_{\Omega} \mathcal{M}_{ji}(K_0-K_1)_{ik}\frac{\partial u}{\partial x_k}\phi \dmu \qquad \forall \phi\in C^1(\overline{Q}_T)
\label{eq:nuj2}$$ and by the density of $C^1(\overline{Q}_T)$ in $C^0(\overline{Q}_T)$ we derive $$\dnuj=\mathcal{M}_{ji}(K_0-K_1)_{ik}\frac{\partial u}{\partial x_k} \dmu.$$ Now, setting $\Uen = \uen-u$ and $\Wen = \wen-w$, and selecting $\Phi$ and $\Psi$ as test functions, we have $$\begin{aligned}
&\int_\Omega \partial_t \Uen \Phi + \int_\Omega K_0 \nabla \Uen\cdot \nabla \Phi + \int_\Omega (1 - \chi_\omegaen)(f(\ue,\we)-f(u,w))\Phi
+ \int_\Omega \partial_t \Wen \Psi \\&+ \int_\Omega (g(\ue,\we)-g(u,w))\Psi = \int_\omegaen (K_0 - K_1)\nabla \uen \cdot \nabla \Phi
+ \int_\omegaen f(u,w)\Phi.
\end{aligned}$$ Furthermore, since $(\Phi,\Psi)$ is a solution to , it also holds: $$\begin{aligned}
&\int_\Omega \partial_t \Phi \Uen - \int_\Omega K_0 \nabla \Phi \cdot \nabla \Uen - \int_\Omega (f_u(u,w)\Phi + g_u(u,w)\Psi)\Uen
+ \int_\Omega \partial_t \Psi\Wen \\&- \int_\Omega (f_w(u,w)\Phi + g_w(u,w)\Psi)\Wen = -\int_{\partial \Omega}K_0 \nabla \Phi \cdot \normal \Uen.
\end{aligned}$$ Summing up the last two identities and integrating in time we observe that the terms involving the time derivatives vanish due to the initial and final conditions on $\Uen,\Wen,\Phi,\Psi$. Consider now the non-linear terms containing $f$ $$\begin{aligned}
& \int_0^T \int_\Omega (1-\chi_\omegaen)(f(\uen,\wen)-f(u,w))\Phi - \int_0^T\int_\Omega (f_u(u,w)\Uen + f_w(u,w)\Wen)\Phi \\
& = \int_0^T \int_\Omega (1-\chi_\omegaen)(f(\uen,\wen) - f(u,w) - f_u(u,w)\Uen - f_w(u,w)\Wen) \Phi \\
&+ \int_0^T \int_\omegaen (f_u(u,w)\Uen + f_w(u,w)\Wen)\Phi.
\end{aligned}$$ Observe that by means of the Lagrange’s mean value theorem and the Lipschitz-continuity of the functions $f_u,f_w$ with constants $L_{f_u}, L_{f_w}$ we have $$\begin{aligned}
&\int_0^T \int_\Omega (1-\chi_\omegaen)\left[ \left(f_u(u_{\xi_1}, w_{\xi_1})-f_u(u,w)\right) \Uen + \left(f_w(u_{\xi_1}, w_{\xi_1})-f_w(u,w)\right)\Wen \right] \Phi \\
& \quad \leq \int_0^T \int_\Omega (1-\chi_\omegaen)\left[L_{f_u} U_{\varepsilon_n}^2 + L_{f_u} \Uen\Wen + L_{f_w}\Uen\Wen + L_{f_w} W_{\varepsilon_n}^2 \right] \Phi \\
& \quad \leq c (\norm{L^2(Q_T)}{\Uen}^2 + \norm{L^2(Q_T)}{\Wen}^2) = o(|\omegaen|),
\end{aligned}$$ whereas $$\begin{aligned}
& \int_0^T \int_\omegaen (f_u(u,w)\Uen + f_w(u,w)\Wen)\Phi \leq c \int_0^T \int_\Omega \chi_\omegaen (\Uen+\Wen) \\
& \quad \leq c |\omegaen|^\half (\norm{L^2(Q_T)}{\Uen} + \norm{L^2(Q_T)}{\Wen}) = o(|\omegaen|).
\end{aligned}$$ Analogous estimates can be obtained for the terms containing $g$. Finally, $$\begin{aligned}
\int_0^T \int_{\partial \Omega} K_0 \nabla \Phi \cdot \normal \Uen &= |\omegaen| \int_0^T \int_\Omega \frac{\chi_\omegaen}{|\omegaen|}\left[(K_0 - K_1)\nabla \uen \cdot \nabla \Phi + f(u,w)\Phi \right] + o(|\omegaen|) \\
&= |\omegaen| \sum_{i,j} \int_0^T \int_\Omega \frac{\chi_\omegaen}{|\omegaen|} [K_0 - K_1]_{ji} \frac{\partial \uen}{\partial x_i} \frac{\partial \Phi}{\partial x_j} \\
& \quad + \int_0^T \int_\Omega \frac{\chi_\omegaen}{|\omegaen|}f(u,w)\Phi + o(|\omegaen|).
\end{aligned}$$ Thanks to the regularity of $\Phi$, and employing the computed weak\* limits, we derive $$\begin{aligned}
&\int_0^T \int_{\partial \Omega} K_0 \nabla \Phi \cdot \normal (\uen-u) \\
& \quad = |\omegaen| \left( \sum_{i,j,k} \int_0^T\int_\Omega \mathcal{M}_{kj} [K_0 - K_1]_{ki}\frac{\partial u}{\partial x_i} \frac{\partial \Phi}{\partial x_j} d\mu\ dt + \int_0^T\int_\Omega f(u,w)\Phi d\mu \ dt \right)+ o(|\omegaen|) \\
& \quad = |\omegaen| \int_0^T\int_\Omega \left[ \mathcal{M}(K_0-K_1) \nabla u \cdot \nabla \Phi + f(u,w)\Phi\right] d\mu\ dt + o(|\omegaen|).
\end{aligned}$$ which concludes the proof.
Reconstruction algorithm: a topology optimization approach {#algorithm}
==========================================================
The asymptotic expansion provided in Theorem \[th:3\] allows to describe the perturbation of the electrical potential on the boundary of the domain due to the presence of a small conductivity inhomogeneity $\omega_{\varepsilon}$. In order to derive a reconstruction algorithm for problem , we introduce the mismatch functional $J$ $$J(\omega) = \half \int_0^T \int_{\partial \Omega} (u_\varepsilon - u_{meas})^2,
\label{eq:J}$$ where $u_\varepsilon=u_{\omega_\varepsilon}$ solves the perturbed problem in the presence of an ischemic region $\omega_\varepsilon$. We now prove that the functional $J$ restricted to the class of inclusions satisfying , admits an asymptotic expansion with respect to the size of the inclusion. Moreover, as it is shown in Theorem \[th:4\], the first-order term of the expansion (which will be denoted as $G$, the topological gradient of $J$) can be computed by solving the unperturbed problem and a suitable adjoint problem. In particular, we restrict ourselves to inclusions $\omegae$ satisfying and of the form $$\omegae = \{z+ \varepsilon B\},
\label{eq:omegaez}$$ where $B$ is a bounded, smooth set containing the origin.
We have the following
Consider a family $\{\omegae\}$ satisfying and . Then, there exists a matrix $\mathcal{M}$ (which may depend on $z$, $B$, $K_0$ and $K_1$) such that, as $\varepsilon\to 0$, $$\begin{aligned}
J(\omegae) = &J(0) + |\omegae| \int_0^T \left[ \mathcal{M}(K_0(z)-K_1(z))\nabla u(z,t) \nabla \Phi(z,t) + f(u(z,t),w(z,t))\Phi(z,t) \right]\dt \\
&+ o(|\omegae|),
\end{aligned}$$ where $(u,w)$ solves and $(\Phi,\Psi)$ is the solution of the *adjoint problem*: $$\left\{
\begin{aligned}
\partial_t \Phi + div(K_0 \nabla \Phi) - f_u(u,w)\Phi - g_u(u,w)\Psi &= 0 \quad &\textit{ in } Q_T,\\
K_0 \nabla \Phi \cdot \normal &= u - u_{meas} \quad &\textit{ on } \Gamma_T,\\
\partial_t \Psi - f_w(u,w)\Phi - g_w(u,w)\Psi &= 0 \quad &\textit{ in } Q_T,\\
\Phi(\cdot,T) = 0 \qquad \Psi(\cdot,T) &= 0 \quad &\textit{ in } \Omega.
\end{aligned}
\right.
\label{eq:adjoint}$$ \[th:4\]
First of all, we note that, since Lemma 3.3 holds under the weaker assumption on the test functions, $\phi\in W^{1,\infty}({Q}_T)$, then the validity of Theorem \[th:3\] can be easily extended when $(\Phi,\Psi)\in (W^{1,\infty}({Q}_T), C(\overline{Q}_T))$. We now show that the solution to (\[eq:adjoint\]) enjoys this regularity. In fact this can be ensured by a lifting argument, since the boundary datum $u-u_{meas} \in C^{2+\alpha,1+\alpha/2}(\Gamma_T)$. Estimates of the $W^{1,\infty}(Q_T)$ norm of $\Phi$ can be obtained by flattening the boundary and by reflection arguments, see for example [@book:lady Chapter 7, Theorem 2.2] for a two-dimensional case.
It is straightforward to verify that $$\begin{aligned}
J(\omegae) - J(0) &= \int_0^T\int_{\partial \Omega} (\ue - u)(u - u_{meas}) + \half \int_0^T\int_{\partial \Omega} (\ue - u)^2.
\end{aligned}
\label{eq:solito}$$ The first term in the right-hand side of involves the boundary condition of the adjoint problem, and in particular it can be written as $\int_0^T\int_{\partial \Omega} K_0 \nabla \Phi \cdot \normal (\ue - u)$. We now apply Theorem \[th:3\] in order to conclude that $$\int_0^T\!\!\int_{\partial \Omega} (\ue - u)(u - u_{meas}) = |\omegae| \int_0^T \!\!\int_\Omega \left[ \mathcal{M}(K_0-K_1)\nabla u \cdot \nabla \Phi + f(u,w)\Phi \right]d\mu + o(|\omegae|).$$ According to assumption , the limit measure $\mu$ is independent of the choice of the subsequence and is equal to $\delta_z$, the Dirac measure centered in $z$. Finally, the second term in the right-hand side of is $o(|\omegae|)$ by means of Lemma \[lem:boundary\] below.
By Assumption 2.1 and by the homogeneous boundary conditions on $\ue$ we can conclude, using standard local regularity results on parabolic equations applied in a neighbourhood of $\partial\Omega$, that $\ue \in C^{2+\alpha,1+\alpha/2}(\Gamma_T)$
Let $\{\omegae\}$ satisfy and . Then, $$\int_0^T\int_{\partial \Omega} (\ue - u)^2 = o(|\omegae|).$$ \[lem:boundary\]
By the same argument as in the proof of Theorem \[th:4\] and by Remark 4.2, $$\int_0^T\int_{\partial \Omega} (\ue - u)^2 = |\omegae|\int_0^T\int_\Omega \left[\mathcal{M}(K_0 - K_1)\nabla u\cdot \nabla \Theta + f(u,w)\Theta \right]d\mu + o(|\omegae|),$$ where $(\Theta,\Xi)$ satisfies $$\left\{
\begin{aligned}
\partial_t \Theta + div(K_0 \nabla \Theta) - f_u(u,w)\Theta - g_u(u,w)\Xi &= 0 \quad &\textit{ in } Q_T,\\
K_0 \nabla \Theta \cdot \normal &= \ue - u \quad &\textit{ on } \Gamma_T,\\
\partial_t \Xi - f_w(u,w)\Theta - g_w(u,w)\Xi &= 0 \quad &\textit{ in } Q_T,\\
\Theta(\cdot,T) = 0 \qquad \Xi(\cdot,T) &= 0 \quad &\textit{ in } \Omega.
\end{aligned}
\right.
\label{eq:test}$$ and $(\Theta,\Xi)\in(W^{1,\infty}({Q}_T), C(\overline{Q}_T)) $.
We now focus on proving that for a compact set $K$ such that $\omegae\subset K$ and satisfying $dist(K,\partial \Omega) \geq d_0$ one has $$\norm{L^2(0,T;H^3(K))}{\Theta} \leq \norm{L^2(0,T;H^1(\Omega))}{\ue-u},
\label{eq:H3}$$ and, as a consequence of a Sobolev immersion, (\[eq:H3\]) implies (here $d=2,3$) $$\norm{L^2(0,T;W^{1,\infty}(K))}{\Theta} \leq \norm{L^2(0,T;H^1(\Omega))}{\ue-u} \leq |\omegae|^\half.$$ By a combination of and we can ensure that the support of the measure $\mu = \delta_z$ is contained in $K$. This would entail that $$\int_0^T\int_\Omega \left[\mathcal{M}(K_0 - K_1)\nabla u\cdot \nabla \Theta + f(u,w)\Theta \right]d\mu \leq c |\omegae|^\half,$$ hence $\int_0^T\int_{\partial \Omega} (\ue - u)^2 = O(|\omegae|^{3/2}) = o(|\omegae|)$.
In order to prove , by standard Faedo-Galerkin technique applied to the linear system , one has $$\norm{L^\infty(0,T;L^2(\Omega))}{\Theta} + \norm{L^\infty(0,T;L^2(\Omega))}{\Xi} + \norm{L^2(0,T;H^1(\Omega))}{\Theta} \leq c \norm{L^2(0,T;L^2(\partial \Omega))}{\ue - u}.
\label{eq:L2bound}$$ Moreover, according to , $\Theta$ is the solution of the following problem $$\left\{
\begin{aligned}
\partial_t \Theta + div(K_0 \nabla \Theta) - f_u(u,w)\Theta &= g_u(u,w)\Xi \quad &\textit{ in } Q_T,\\
K_0 \nabla \Theta \cdot \normal &= \ue - u \quad &\textit{ on } \Gamma_T,\\
\Theta(\cdot,T) &= 0 \quad &\textit{ in } \Omega.
\end{aligned}
\right.$$ Analogously to what stated in [@book:lady Chapter 4, Theorem 9.1], it holds that $$\norm{L^2(0,T;H^2(\Omega))}{\Theta} \leq C \left(\norm{L^2(Q_T)}{g_u(u,w)\Xi} + \norm{L^2(0,T;H^{1/2}(\partial \Omega))}{\ue - u}\right)$$ and in conclusion, by the regularity of $g, u, w$ and by $$\norm{L^2(0,T;H^2(\Omega))}{\Theta} \leq C \norm{L^2(0,T;H^{1/2}(\partial \Omega))}{\ue - u}.
\label{eq:H2bound}$$ Similar estimates can be derived also for $\Xi$ since $f,g \in C^3(\R^2)$ and $u,w \in C^{2,1}(\overline{Q_T})$. This last property is guaranteed on $u$ by Theorem \[th:1\], whereas it can be extended to $w$ since we consider $w_0 \in C^2(\overline\Omega)$. Now, by computing the solution of the third equation in in closed form, we can easily verify that $$\norm{L^2(0,T;H^2(\Omega))}{\Xi} \leq C \norm{L^2(0,T;H^2(\Omega))}{\Theta} \leq C \norm{L^2(0,T;H^{1/2}(\partial \Omega))}{\ue - u}.
\label{eq:Xi}$$ Consider the change of variable $s = T-t$ and denote by $\widehat{\Theta}(\cdot,s) = \Theta(\cdot,T-t), \widehat{\Xi}(\cdot,s) = \Xi(\cdot,T-t)$. We now focus on the first equation in and compute the second derivatives of each term. Then, $V \defeq \frac{\partial^2 \widehat{\Theta}}{\partial x_i \partial x_j}$ satisfies the following equation in a weak sense: $$\partial_{s} V -div(K_0 \nabla V) + f_u V = R,
\label{eq:VR}$$ being $$\begin{aligned}
&R = div\left(\frac{\partial K_0}{\partial x_i} \nabla \frac{\partial \widehat{\Theta}}{\partial x_j} \right) + div\left(\frac{\partial K_0}{\partial x_j} \nabla \frac{\partial \widehat{\Theta}}{\partial x_i} \right) + div\left(\frac{\partial^2 K_0}{\partial x_i \partial x_j} \nabla \widehat{\Theta} \right)\\
& -\left(f_{uuu}\frac{\partial u}{\partial x_i}\frac{\partial u}{\partial x_j} + f_{uuw}\frac{\partial u}{\partial x_i} \frac{\partial w}{\partial x_j} + f_{uu} \frac{\partial^2 u}{\partial x_i \partial x_j} + f_{uwu}\frac{\partial w}{\partial x_i}\frac{\partial u}{\partial x_j} + f_{uww}\frac{\partial w}{\partial x_i} \frac{\partial w}{\partial x_j} + f_{uw} \frac{\partial^2 w}{\partial x_i \partial x_j} \right)\widehat{\Theta} \\
&-\left( f_{uu}\frac{\partial u}{\partial x_i} + f_{uw}\frac{\partial w}{\partial x_i}\right)\frac{\partial \widehat{\Theta}}{\partial x_j} - \left( f_{uu}\frac{\partial u}{\partial x_j} + f_{uw}\frac{\partial w}{\partial x_j}\right)\frac{\partial \widehat{\Theta}}{\partial x_i} \\
& - \left(g_{uuu}\frac{\partial u}{\partial x_i}\frac{\partial u}{\partial x_j} + g_{uuw}\frac{\partial u}{\partial x_i} \frac{\partial w}{\partial x_j} + g_{uu} \frac{\partial^2 u}{\partial x_i \partial x_j} + g_{uwu}\frac{\partial w}{\partial x_i}\frac{\partial u}{\partial x_j} + g_{uww}\frac{\partial w}{\partial x_i} \frac{\partial w}{\partial x_j} + g_{uw} \frac{\partial^2 w}{\partial x_i \partial x_j} \right)\widehat{\Xi} \\
& - \left( g_{uu}\frac{\partial u}{\partial x_i} + g_{uw}\frac{\partial w}{\partial x_i}\right)\frac{\partial \widehat{\Xi}}{\partial x_j} - \left( g_{uu}\frac{\partial u}{\partial x_j} + g_{uw}\frac{\partial w}{\partial x_j}\right)\frac{\partial \widehat{\Xi}}{\partial x_i} - g_u \frac{\partial^2 \widehat{\Xi}}{\partial x_i \partial x_j}.
\end{aligned}$$ Let $\Omega_1$ be an open subset of $\Omega$, such that $\omegae\subset K\subset \Omega_1 \subset \subset \Omega$. By interior regularity results it holds that $(\widehat{\Theta},\widehat{\Xi})$ is smooth on $\Omega_1\times[0,T]$, which entails by the initial conditions that $$\frac{\partial^2\widehat{\Theta}}{\partial x_i \partial x_j}(x,0)=0,\quad\forall x\in \Omega_1.
\label{eq:Theta0}$$ According to the smoothness of $f,g,u,w$ and to the $H^2$ bounds and , we get $$\norm{L^2(0,T;H^{-1}(\Omega_1))}{R} \leq C \norm{L^2(0,T;H^{1/2}(\partial \Omega))}{\ue - u}.
\label{eq:Rbound}$$ In fact, all the terms in $R$ except the first three belong to $L^2(Q_T)$, with norm bounded by $\norm{L^2(0,T;H^2(\Omega))}{\Theta}$ or by $\norm{L^2(0,T;H^2(\Omega))}{\Xi}$. Set $H:=div\left(\frac{\partial K_0}{\partial x_i} \nabla \frac{\partial \widehat{\Theta}}{\partial x_j} \right)$, then, for any $v \in H^1_0(\Omega_1)$, $$|\langle H,v\rangle_{\star}|= \left|\int_{\Omega_1} \frac{\partial K_0}{\partial x_i} \nabla \frac{\partial \widehat{\Theta}}{\partial x_j} \cdot \nabla v \right|\leq C\norm{H^2(\Omega)}{\widehat{\Theta}}\norm{H^1(\Omega_1)}{v}.$$ where $\langle \cdot,\cdot \rangle_{\star}$ indicates the duality pairing between the involved spaces. Integrating in time last relation we finally get $$\norm{L^2(0,T;H^{-1}(\Omega_1))}{H} \leq C\norm{L^2(0,T;H^{1/2}(\partial \Omega))}{\ue - u}.$$ With similar arguments we can estimate also the second and third term in the definition of the function $R$ leading finally to (\[eq:Rbound\]).
Consider now a test function $\zeta \in C^\infty_C(\Omega_1)$, $0 \leq \zeta \leq 1$, such that $\zeta = 1$ in $K$. According to , by simple computations we verify that $\widetilde{V} = V\zeta$ satisfies $$\partial_s \widetilde{V} - div(K_0 \nabla \widetilde{V}) + f_u \widetilde{V} = \widetilde{R},
\label{eq:VRt}$$ being $\widetilde{R} = \zeta R - 2 div(V K_0 \nabla \zeta) + V div(K_0 \nabla \zeta)$. It holds that $div(V K_0 \nabla \zeta) \in L^2(0,T;H^{-1}(\Omega_1))$, and $$\norm{L^2(0,T,H^{-1})}{div(V K_0 \nabla \zeta)} \leq C \norm{L^2(Q_T)}{V},$$ hence we observe that $$\norm{L^2(0,T,H^{-1}(\Omega_1))}{\widetilde{R}} \leq C ( \norm{L^2(Q_T)}{V} + \norm{L^2(0,T,H^{-1}(\Omega_1))}{R}) \leq C \norm{L^2(0,T;H^{1/2}(\partial \Omega))}{\ue - u}
\label{eq:Rtilde}$$ By standard Faedo-Galerkin argument (see, e.g., [@book:dautraylions Chap. 18, Par. 3, Theorem 3], we verify that $$\norm{L^2(0,T;H^1_0(\Omega_1))}{\widetilde{V}}\leq C \norm{L^2(0,T,H^{-1}(\Omega_1))}{\widetilde{R}}.
\label{eq:FG}$$ Finally, by the definition of $\zeta$, combining and we conclude that $$\norm{L^2(0,T;H^1(K))}{V} \leq \norm{L^2(0,T;H^1_0(\Omega_1))}{\widetilde{V}}\leq C \norm{L^2(0,T;H^{1/2}(\partial \Omega))}{\ue - u}.$$
The thesis is now proved thanks to the trace inequality and the energy estimates.
Theorem \[th:4\] gives a representation formula for the topological gradient $G$, the first order term in the expansion of the mismatch functional $J$. In analogy to [@art:BMR; @art:bccmr], we propose a *one-step* reconstruction Algorithm \[al:top\] for the identification of small inclusions satisfying .
$u_{meas}(x,t) \ \forall \, x \in \partial \Omega$, $t \in (0,T)$. approximated centre of the inclusion, $\bar{z}$
- compute $(u,w)$ by solving ;
- compute $(\Phi,\Psi)$ by solving ;
- determine the topological gradient $G$ of $J$ according to Theorem \[th:4\];
- find $\bar{z}$ s.t. $G(\bar{z}) \leq G(z) \quad \forall \, z \in \Omega$.
The polarization tensor $\mathcal{M}$ can be computed in an explicit form, e.g., when the shape $B$ of the inclusion is a disk, even though it is not straightforward in the current anisotropic case. Since $\mathcal{M}$ depends on $K_0$ and $K_1$ and on the shape of the set $B$, which we assume to be a disk, it varies according to the space variable $z$. In order to compute $\mathcal{M}(z) = \mathcal{M}(K_0(z),K_1(z),B)$, we first use [@book:ammari-kang Lemma 4.30] to transform $K_0$ and $K_1$ in diagonal matrices. By a rescaling of the spatial variables, we can reduce to the case in which one of the two diagonalized tensors is the identity, which allows to apply [@book:ammari-kang Proposition 4.31]. We remark that the anisotropic rescaling entails that the original circle $B$ is transformed in an ellipse of known semiaxes.
Numerical results {#results}
=================
In this section we describe the implementation of Algorithm \[al:top\] and report some numerical results in order to show its effectiveness. In particular, we set our experiments in a two-dimensional idealized geometry, representing a horizontal section of the ventricles. The application of the model on a three-dimensional geometry would be equivalently possible, and will be object of future studies, employing advanced numerical analysis techniques in order to tackle the increased computational effort. We rely on synthetic data, i.e., we solve the monodomain system in presence of a prescribed ischemic region and then use the value of the solution on the boundary (or a portion of it) as an input for the reconstruction algorithm. In order to prevent *inverse crimes*, we employ different numerical settings for the synthetic data generation and for the solution of the unperturbed and adjoint problems, required for the reconstruction algorithm. In particular, we adopt a much more refined discretization, both in space and in time, for the simulation of the synthetic data.
In the following experiments we assume to measure the voltage only on a portion of the heart. As outlined in Section \[intro\], measurements of the voltage can be acquired on the inner surface of a ventricle by intracavitary measurements, or, alternatively, we might be able to compute a map of the electrical potential on the epicardium starting from ECG data. This does not affect the reconstruction procedure described in Algorithm \[al:top\], apart from the definition of the adjoint problem , which now prescribes oblique boundary conditions involving $u_{meas}-u$ on the portion of the boundary where the measurements are acquired, and homogeneous Neumann conditions elsewhere.
Finite Element approximation
----------------------------
In order to numerically approximate the solution of the monodomain model, we rely on a Galerkin Finite Element scheme, introducing a tessellation of the domain consisting of triangular elements. In particular, we adopt two different meshes for the solution of the unperturbed (and adjoint) problem and for the generation of the synthetic measurements, the latter being more refined especially close to the boundary of the prescribed ischemic region.
Moreover, in order to reproduce the anisotropic behavior of the conductivity coefficients $K_0$ and $K_1$, we consider the presence of fibers within the domain. We adopt a procedure analogous to the one reported in [@Quarteroni2016] for the generation of the fiber directions, resorting on the solution of a Laplace problem with suitable boundary conditions. Once the direction of the fibers $\vec{e_f}(x) = \vec{e}_1(x)$ is defined within $\Omega$, as well as the transmural vector field $\vec{e_n}(x) = \vec{e}_2(x)$ (obtained by a clockwise rotation of $90^\circ$ of $\vec{e_f}(x)$), the definition of the conductivity tensors is
$$K_0(x) = k_{0,1} \vec{e_f}(x) \otimes \vec{e_f}(x) + k_{0,2} \vec{e_n}(x) \otimes \vec{e_n}(x), \quad
K_1(x) = k_{1,1} \vec{e_f}(x) \otimes \vec{e_f}(x) + k_{1,2} \vec{e_n}(x) \otimes \vec{e_n}(x).$$ The numerical mesh for the solution of the background problem and the directions of the fibers are reported in Figure \[fig:setup\].
As already mentioned in the introduction, the best choice of the conductivity tensor $K_0$ to be employed in the numerical simulations of the monodomain model is given by the harmonic mean of the intracellular and extracellular conductivity tensors $D_i$ and $D_e$ appearing in the bidomain model that can be expressed by $$K_0 = D_e (D_e + D_i)^{-1} D_i.
\label{eq:}$$ Exploiting the fact that $D_e$ and $D_i$ have the same eigenvectors (again defined by the fiber and transmural directions), we can therefore compute the eigenvalues of $K_0$ (and analogously of $K_1$) as the harmonic mean of the ones of $D_i$ and $D_e$ as reported, e.g., in [@book:pavarino Table 8.1]. The values of the main parameters involved in the numerical simulations are reported in Table \[tab:param\].
$k_{0,1}$ $k_{0,2}$ $k_{1,1}$ $k_{1,2}$ $A$ $a$ $\epsilon$
----------- ----------- ----------- ----------- ----- -------- ------------
$1.200$ $0.2538$ $0.2308$ $0.0062$ $8$ $0.15$ $0.05$
: Values of the main parameters[]{data-label="tab:param"}
The numerical solution of the background problem relies on a Newton-Galerkin scheme. The spatial discretization is performed thanks to the P1-finite element space, i.e., the space of the continuous functions over $\Omega$ which are linear polynomials when restricted on each element of the mesh. The temporal discretization is done via an implicit Euler scheme. This leads to solve a nonlinear problem at each timestep, which is performed via a Newton iterative algorithm. More details on the solver can be found in [@art:RV], where a thorough convergence analysis is performed.
Reconstruction of small inclusions
----------------------------------
We study the effectiveness of the reconstruction algorithm in identifying the position of small ischemic regions within the cardiac tissue. For, we employ synthetic measurements of the perturbed boundary voltage in presence of ischemias located in different sections of the cardiac tissue: in the left ventricle, in the septum, or in the right ventricle. In each simulation, we consider circular ischemic regions of radius ranging from $1mm$ to $2mm$. In Figure \[fig:recon\] we report the contour plot of the topological gradient $G$ in four different cases, and superimpose a black line representing the boundary of the ischemia that we aim to reconstruct.
\
From an analysis of Figure \[fig:recon\] we can deduce that the topological gradient always attains its negative minimum value close to the real position of the ischemia, and the accuracy of the localization may vary according to the position of the ischemia. We achieve significant results also in the case of multiple inclusions. When more than one ischemic region is present, the topological gradient shows a local minimum close to each region. As depicted in the captions, some experiments employ the knowledge of the voltage on the epicardium and some on one endocardiac surface. The reconstruction appears fairly accurate in both cases. We remark that the algorithm is always effective when using epicardiac measurements, but fails in detecting ischemic regions located in a ventricle whenever the measurements are acquired on the inner surface of the other one. Nevertheless, even in those cases the technique avoids false positive detection, i.e., the topological gradient does not show significant local minima in wrong locations.
Reconstruction from noisy measurements
--------------------------------------
We now focus on the performance of the algorithm in presence of noisy measurements, in order to assess the stability of the algorithm with respect to small perturbations of the boundary data of the form $$\widetilde{u}_{meas}(x,t) = u_{meas}(x,t) + \rho \eta(x,t),
\label{eq:noise}$$ where $\eta(x,t)$ is a standard Gaussian random variable for each point $x$ and instant $t$, and $\rho \in [0,1]$ is the noise level.
\
In Figure \[fig:noise\] we report the contour plot of the topological gradient computed with noiseless measurements and compare it with the ones obtained with growing levels of noise. The algorithm shows to be stable even under significantly corrupted measurements: up to the level $\rho = 0.15$, the negative region in correspondence to the exact position is clearly identifiable.
Reconstruction of large inclusions
----------------------------------
We eventually remark that the proposed algorithm produces significant results also when applied to measurements associated to large ischemic regions. The identification of larger regions could be performed by means of iterative reconstruction algorithms, as the one proposed in [@art:BRV] for a semilinear elliptic problem. Nevertheless, as it is shown in Figure \[fig:big\], the information coming from the topological gradient could be a suitable initial guess for such iterative algorithms.
Conclusions
===========
In this paper, we have studied the identification of small ischemic regions in the cardiac tissue, represented by discontinuous alterations in the conductivity and in the nonlinear reaction term of the monodomain model, taking advantage of the measurement of the voltage on the boundary of the heart. We have extended the existing results regarding the well-posedness of the direct problem and derived a rigorous asymptotic expansion of the perturbed boundary voltage, which allows to formulate an effective reconstruction algorithm based on topological optimization.
We foresee several significant extensions of the presented analysis, both from an analytical and a numerical viewpoint. The coupling of the monodomain model of the heart with a passive conductor model for the surrounding torso (see [@art:boulakia]) would enable us to make use of ECG data for reconstruction purposes, possibly comparing the results with the ones obtained in [@art:lyka],[@wang2013inverse] with a stationary version of the bidomain model.
The effectiveness of the reconstruction algorithm should also be tested in a three-dimensional setting; in such a context, due to the high computational cost of the numerical simulation, a Reduced Order Modeling approach could be considered, as recently analysed in [@art:pagani].
Finally, the detection of arbitrarily large ischemic regions can be tackled, exploiting an analogous strategy as the one adopted in [@art:BMR] (possibly reducing the computational effort by employing an adaptive solver of the monodomain system, as proposed in [@art:RV]).
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors are grateful to Roberto Gianni for his useful suggestions. The authors thank the New York University of Abu Dhabi for its kind hospitality. The work of the authors has been supported by GNAMPA (Gruppo Nazionale per l’Analisi Matematica, la Probabilità e le loro Applicazioni). The numerical simulations presented in this work have been performed thanks to the MATLAB library redbKIT, [@redbKIT].
[^1]: it follows by an application of the Lagrange’s mean value theorem on the real-valued function $h(\tau) = \int_\Omega (1-\chi_\varepsilon)(f(u+\tau(\ue - u), w+\tau(\we - w))-f(u,w))\Ue$ on the interval $[0,1]$; the same holds for $g$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper aims at settling the issue of the validity of the de Jonquières formulas. We consider the space of divisors with prescribed multiplicity, or de Jonquières divisors, contained in a linear series on a smooth projective curve. Assuming zero expected dimension of this space, the de Jonquières formulas compute the virtual number of de Jonquières divisors. Using degenerations to nodal curves we show that for a general curve equipped with a general complete linear series, the space is of expected dimension, which shows that the counts are in fact true. This implies that in the case of negative expected dimension a general linear series on a general curve does not admit de Jonquières divisors of the expected type.'
author:
- Mara Ungureanu
bibliography:
- 'dejonq2.bib'
title: |
Dimension theory and degenerations of\
de Jonquières divisors
---
Introduction
============
In his 1866 memoir [@deJo], de Jonquières computed the number of divisors with prescribed multiplicities that are contained in a fixed linear series on a given plane algebraic curve.
Almost a century later, and using modern techniques of topology and intersection theory of cycles on the symmetric product on a curve, MacDonald [@Mac] and Mattuck [@Mat] recovered the original formula in characteristic zero and arbitrary characteristic, respectively, and generalised it to linear series of any dimension. However, their work does not address the vagueness of the classical statements, assuming either that the linear series in question is sufficiently generic, or that the multiplicities are counted correctly. To address this issue, Vainsencher [@Va] described the locus of divisors with prescribed multiplicities as the vanishing locus of a section of a bundle of the appropriate rank. Using a natural filtration of this bundle, he computed its Chern classes without making use of the Grothendieck-Riemann-Roch theorem, and established the enumerative validity of the de Jonquières formula for plane curves and for some higher dimensional cases.
The aim of this paper is twofold. On the one hand, we settle the issue of the validity of the de Jonquières formula for linear series of arbitrary degree and dimension on a general curve by studying the geometry of the respective moduli space. On the other hand, we develop a theory of degenerations for de Jonquières divisors to nodal curves, which plays a central role as the main tool of in the study of the aforementioned moduli space.
Aside from being interesting objects in their own right, de Jonquières divisors and their degenerations are natural generalisations of the concept of strata of abelian differentials which were first introduced in the context of Teichmüller dynamics and flat surfaces - see the works of Masur [@Mas] and Veech [@Ve], and more recently, of Bainbridge, Chen, Gendron, Grushevsky, and Möller [@BCGGM]. These strata are, however, interesting objects also from the point of view of algebraic geometry, as can be seen in the work of Farkas and Pandharipande [@FP], Chen and Tarasca [@CT], or Mullane [@Mu]. In fact, the result of Polishchuk [@Po] concerning the dimension of the strata in $\mathcal{M}_{g,n}$ provides an important clue towards the validity of the de Jonquières formulas.
In what follows we fix the notation and describe the objects of interest. Let $C$ be a smooth projective curve of genus $g$ and denote by $C_d$ its $d$-th symmetric product. Furthermore, let ${G^r_d(C)}$ parametrise linear series of type ${g^r_d}$, i.e. $${G^r_d(C)}:=\{ l=(L,V) \mid L\in\operatorname{\text{Pic}}^d(C), V\in G(r+1,H^0(C,L)) \}.$$ In this paper we focus on the case of Brill-Noether general curves, meaning that ${G^r_d(C)}$ is a smooth variety and its dimension is given by the Brill-Noether number $$\label{eq:brillnoether}
\rho(g,r,d)=g-(r+1)(g-d+r)\geq 0.$$ Moreover it follows that the Hilbert scheme $\operatorname{\text{Hilb}_{d,g,r}}$ parametrising curves in $\operatorname{\mathbb{P}}^r$ of degree $d$ and (arithmetic) genus $g$ has a unique component $\mathcal{H}_{d,g,r}$, whose general point corresponds to a smooth curve and which maps dominantly onto $\mathcal{M}_g$. Thus, when we talk about a general curve $C$ with a general linear series $l\in{G^r_d(C)}$ we refer to a general point $[C \xrightarrow{l} \operatorname{\mathbb{P}}^r]\in\mathcal{H}_{d,g,r}$.
We now define the main object of interest in this paper: for a fixed smooth curve $C$ of genus $g$ with a fixed linear series $l=(L,V)\in{G^r_d(C)}$, a *de Jonquières divisor of length* $N$ is a divisor $a_1 D_1 + \ldots + a_k D_k \in C_d$ such that $$a_1 D_1 + \ldots + a_k D_k \in\operatorname{\mathbb{P}}V,$$ where $k\leq d$ is a positive integer and the $D_i$ are effective divisors of degree $d_i$ for $i=1,\ldots,k$ such that $N=\sum_{i=1}^k d_i$. If $l$ is complete (i.e. such that $g-d+r\geq 0$), the definition of a de Jonquières divisor is equivalent to $$L\simeq\operatorname{\mathcal{O}}_C(a_1 D_1 + \ldots + a_k D_k).$$ Furthermore, if we let $\mu_1=(a_1,\ldots,a_k)$ and $\mu_2=(d_1,\ldots,d_k)$ be two positive partitions such that $\sum_{i=1}^k a_i d_i = d$, then we denote the set of de Jonquières divisors of length $N$ determined by $\mu_1$ and $\mu_2$ by $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$.
In the particular case when $d_i=1$ for all $i=1,\ldots,k$, let $n:=N=k$ and the de Jonquières divisor is of the form $$a_1 p_1 + \ldots + a_n p_n,$$ for some distinct points $p_1,\ldots,p_n \in C$. Here we simplify the notation to $$DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)=DJ^{r,d}_n(\mu_1,C,l).$$
It turns out (see Section \[sec:deglocus\]) that the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ has the structure of a determinantal variety and its expected dimension (or, equivalently, lower bound for its dimension) is $$\text{exp}\dim DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l) = N-d+r.$$ The de Jonquières formula (cf. [@Mat] §5) states that, if we expect there to be a finite number of de Jonquières divisors of length $N$ (so if $N-d+r=0$), then this virtual number is given by the coefficient of the monomial $t_1^{d_1}\cdot\ldots\cdot t_k^{d_k}$ in $$\label{eq:dejformula}
(1+a_1^2 t_1 + \ldots + a_k^2 t_k)^{g}(1+a_1 t_1 + \ldots + a_k t_k)^{d-r-g}.$$ Substituting $r=1$ and $d_1=\ldots=d_k=1$ in formula (\[eq:dejformula\]) recovers the number of ramification points of a Hurwitz cover of $C$ obtained from the Plücker formula. In addition, if $C$ is the plane cubic, then $g=1$, $d-r-g=1$ and we recover its 9 flex points. Lastly, taking the linear series to be the canonical one, we recover the number of odd theta characteristics on a smooth general curve. Hence we expect these counts to be true. To settle the issue we must however study the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$, establish whether it is empty when the expected dimension is negative, and when non-empty whether it is smooth, reduced, and of expected dimension.
Luckily, we are able to settle these questions in the affirmative. In fact, the non-emptiness of the space of de Jonquières when the expected dimension is non-negative follows from an easy class computation, which we explain in Section \[sec:existencedej\].
The questions regarding the dimension of the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ and whether it is empty when the expected dimension is negative are less straightforward and require the degeneration techniques. Using limit linear series on nodal curves of compact type, we prove
\[dimension\] Fix a general curve $C$ of genus $g$ equipped with a general complete linear series $l\in{G^r_d(C)}$. If $N-d+r\geq 0$, the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ is of expected dimension, $$\dim DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l) = N-d+r.$$
A direct consequence of the dimension theorem is the non-existence statement for complete linear series:
Let $C$ be a general curve equipped with a general complete linear series $l\in{G^r_d(C)}$. If $N-d+r<0$, the variety $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ is empty.
The validity of de Jonquières’ counts is a direct consequence of Theorem \[dimension\], and of the determinantal variety structure of the space of de Jonquières divisors. The latter implies that $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ is in fact a Cohen-Macaulay variety (Proposition 4.1, Chapter II, [@ACGH]). As such, if it is zero-dimensional, it consists of a finite number of discrete closed points. This yields
\[cor:counts\] Let $C$ be a general curve equipped with a general complete linear series $l\in{G^r_d(C)}$. If $N-d+r=0$, the variety $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ is a finite collection of reduced points.
We address the issue of the smoothness of $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ by expressing the space as an intersection of subvarieties inside the symmetric product $C_d$ and obtaining a transversality condition from the study of the relevant tangent spaces.
\[thm:smooth\] Let $C$ be a smooth general curve of genus $g$. Then for any complete linear series $l\in G^r_d(C)$, the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ is smooth, whenever $N-d+r>0$.
The proof is also by degeneration to nodal curves and limit linear series, however this time using a strategy developed in [@Fa2] .
Finally, we prove the non-existence result for non-complete linear series using a different degeneration technique, namely compactified Picard schemes for moduli of stable pointed curves. We obtain
\[cor:nonex\] Let $C$ be a general curve equipped with a general linear series $l\in{G^r_d(C)}$ satisfying $g-d+r<0$ and let $\mu$ be a positive partition of $d$ length $n$. If $n-d+r<0$, the variety $DJ^{r,d}_n(\mu,C,l)$ is empty.
These degenerations are only suitable for treating the case of de Jonquières divisors satisfying $d_1=\ldots=d_k=1$, as they rely on the fact that the points in the support of the divisor are distinct.
The paper is organised as follows. In Section \[sec:geomdej\] we collect some preliminary results about the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ which will form the basis for the arguments in the remainder of the paper. In Section \[sec:deglocus\] the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ is endowed with the structure of a determinantal subvariety of $C_d$, while in Section \[sec:existencedej\] an easy argument shows that the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ is indeed non-empty when the expected dimension is positive. Furthemore, in Section \[sec:nonspecial\] we express the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ as an intersection of subvarieties of the symmetric product $C_d$ and establish the condition for this intersection to be transverse.
We then proceed in Section \[nodes\] to describe degenerations of de Jonquières divisors for families of smooth curves with a nodal central fibre. We approach the question from two perspectives:
1. In Section \[sec:lls\] we look at the theory of limit linear series, as developed by Eisenbud and Harris in [@EH86], which applies to the case when the central fibre is a curve of compact type. In particular, we address the question of what it means to say that a limit linear series admits a de Jonquières divisor. In this context we use a very simple degeneration and an induction argument to prove Theorem \[dimension\] in Section \[sec:prooflls\] and Theorem \[thm:smooth\] in Section \[sec:smooth\].
2. In Section \[sec:compactified\] we discuss compactifications of the Picard schemes over the moduli space of stable curves with marked points, following [@Ca] and [@Me], which come into play for central fibres that are stable curves. This approach represents our attempt at generalising the degenerations of abelian differentials, as they appear in [@Ch] or [@FP]. Of course, what distinguishes our case from the case of differentials is that we do not have a readily available equivalent of the relative dualising sheaf on the family of curves. These degenerations have as direct practical application the proof by induction of Theorem \[cor:nonex\] in Section \[sec:noncomplete\].
We conclude the paper in Section \[sec:lorentz\] with a discussion of the space de Jonquières divisors where the partition $\mu_1$ is allowed to have negative coefficients as well.
Acknowledgements {#acknowledgements .unnumbered}
================
This paper is part of my PhD thesis. I would like to thank my advisor Gavril Farkas for introducing me to this circle of ideas and for his suggestions to approaching this problem. I am also grateful to Irfan Kadiköylü and Nicola Pagani for useful discussions.
The geometry of the space of de Jonquières divisors {#sec:geomdej}
===================================================
In this section we extract as much information as possible about the geometry of the space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ of de Jonquières divisors for a fixed curve $C$ equipped with a linear series $l=(L,V)$ of type ${g^r_d}$, without using any degeneration techniques.
We recall some standard definitions: for a smooth curve $C$, let ${C^r_d}$ be the subvariety of $C_d$ parametrising effective divisors of degree $d$ on $C$ moving in a linear series of dimension at least $r$: $${C^r_d}:= \{ D\in C_d \mid \dim | D | \geq r \},$$ and ${W^r_d(C)}$ be the associated variety of complete linear series of degree $d$ and dimension at least $r$, i.e. $${W^r_d(C)}:= \{ L\in\operatorname{\text{Pic}}^d(C) \mid h^0(C,L)\geq r+1 \}\subseteq\operatorname{\text{Pic}}^d(C).$$ Allowing the curve to vary in the moduli space $\mathcal{M}_g$ of curves of genus $g$, we denote by $\mathcal{W}^r_d$ the relative counterpart of ${W^r_d(C)}$.
The space of de Jonquières divisors as degeneracy locus {#sec:deglocus}
-------------------------------------------------------
Fix an integer $k\leq d$ and two vectors of positive integers $\mu_1=(a_1,\ldots,a_k)$ and $\mu_2=(d_1,\ldots,d_k)$ such that $\sum_{i=1}^{k} a_i d_i = d$. The space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ can be described as a degeneracy locus of vector bundles over $C_d$ as follows: the idea is that the condition $$a_1 D_1 + \ldots + a_k D_k \in\operatorname{\mathbb{P}}V$$ is equivalent to the condition that the natural restriction map $$V\rightarrow H^0(C,L/L(-a_1 D_1-\ldots-a_k D_k))$$ has non-zero kernel. To reformulate this globally in terms of a morphism of two vector bundles over $C_d$, let the first bundle $\mathcal{E}=\operatorname{\mathcal{O}}_{C_d}\otimes V$ be the trivial bundle. As for the second bundle, consider the diagram
\(m) \[matrix of math nodes,row sep=2em,column sep=1em,minimum width=1em\] [ & C C\_d &\
C & & C\_d\
]{}; (m-1-2) edge node \[auto,swap\] [$\sigma$]{} (m-2-1) edge node \[auto\][$\tau$]{} (m-2-3);
where $\sigma$ and $\tau$ are the usual projections and $\mathcal{U}$ is the universal divisor defined as $$\mathcal{U} = \{ (p,D) \mid D\in C_d \text{ and }p\in D \} \subset C\times C_d.$$ Alternatively, identifying $C_d$ with the Hilbert scheme $C^{[d]}$ of $d$ points on $C$, one defines $\mathcal{U}$ as the universal family $\mathcal{U}\subset C\times C^{[d]}$. For the second bundle, consider the sheaf: $$\mathcal{F}_d(L) = \tau_*(\sigma^* L\otimes \operatorname{\mathcal{O}}_{\mathcal{U}}),$$ By cohomology and base change $\mathcal{F}_d(L)$ is indeed a vector bundle. The fibre of $\mathcal{F}_d(L)$ over any point $D\in C_d$ is given by the $d$-dimensional vector space $H^0(C,L/L(-D))$.
Finally, let $\Phi:\mathcal{E}\rightarrow \mathcal{F}_d(L)$ be the vector bundle morphism obtained by pushing down to $C_d$ the restriction $\sigma^*L\rightarrow\sigma^*L\otimes\operatorname{\mathcal{O}}_{\mathcal{U}}$. Moreover let $$\Sigma_{k,N}(\mu_1,\mu_2) = \Bigl\{ E \in C_d \mid E = \sum_{i=1}^k a_i D_i\text{ for some }D_1\in C_{d_1},\ldots,D_k\in C_{d_k}\}.$$
The space $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ is defined as the $r$*-th degeneracy locus* of $\Phi$, i.e. the locus in $\Sigma_{k,N}(\mu_1,\mu_2)$ where $\operatorname{rk}\Phi\leq r$.
\[lemma:dimbound\] For every point $D\in DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$, one has $$\dim_D DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l) \geq N-d+r.$$
From the description of $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ as a degeneracy locus, its codimension in $\Sigma_{k,N}(\mu)$ is at most $$(\operatorname{rk}\mathcal{E}-r)(\operatorname{rk}\mathcal{F}_d(L) -r)=(r+1-r)(d-r)=d-r.$$ Since $\dim\Sigma_{k,N}(\mu_1,\mu_2)=N$, the dimension estimate follows.
Finally, we record here an easy result that forms the base case for the induction argument in the proof of Theorem \[cor:nonex\].
\[lemma:nonspecial\] Let $C$ be any smooth curve of genus $g$ with a general linear series $l\in{G^r_d(C)}$. Fix an integer $k\leq d$ and two vectors of positive integers $\mu_1=(a_1,\ldots,a_k)$ and $\mu_2=(d_1,\ldots,d_k)$ such that $\sum_{i=1}^{k} a_i d_i = d$ and $N-d+r<0$.
1. If $g-d+r=0$, then $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)=\emptyset$.
2. If $g-d+r<0$ and $N<g$, then $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)=\emptyset$.
(1): Consider the following restriction of the Abel-Jacobi-type map: $$\begin{aligned}
u:\Sigma_{k,N}(\mu_1,\mu_2) &\rightarrow \operatorname{\text{Pic}}^d(C) \\
a_1 D_1+\ldots+a_k D_k &\mapsto \operatorname{\mathcal{O}}_C(a_1 D_1 + \ldots + a_k D_k).
\end{aligned}$$ In the non-special regime $\operatorname{\text{Pic}}^d(C)=W^r_d(C)$. Moreover the image of $\varphi$ is closed and $\dim\operatorname{\text{im}}u\leq n<g=\dim\operatorname{\text{Pic}}^d(C)$. Thus a general line bundle $L\in\operatorname{\text{Pic}}^d(C)$ is not contained in the image of $u$ whence we conclude that the divisor $a_1 D_1 + \ldots + a_k D_k$ is not contained in a general linear series $l$ of type ${g^r_d}$, i.e. $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)=\emptyset$. This is Corollary \[cor:nonex\] for non-special linear series.\
As a consequence, for all $r'<r$, a general linear series $l'$ in $\text{G}^{r'}_d(C)$ also has $$DJ_{k,N}^{r',d}(\mu_1,\mu_2,C,l')=\emptyset.$$ To see this, let $c:\text{G}^{r'}_d(C)\rightarrow\text{W}^{r'}_d\subset{W^r_d(C)}$ be the forgetful map $(L,V)\mapsto L$. Note that a line bundle $L\in{W^r_d(C)}\setminus\operatorname{\text{im}}u$ does not admit de Jonquières divisors of length $n$. Now, since $c$ is continuous (as the projection morphism from a Grassmann bundle), $c^{-1}(\text{W}^{r'}_d(C)\setminus\operatorname{\text{im}}u)$ is also open in $\text{G}^{r'}_d$ and nonempty. Hence no $l'\in c^{-1}(\text{W}^{r'}_d(C)\setminus\operatorname{\text{im}}u)$ admits a de Jonquières divisor of length $n$ and our claim is proved.\
(2): Set $r_1=d-g$ so that $g-d+r_1=0$ and $r<r_1$. We conclude from the discussion above that if $n<g$, then $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)=\emptyset$ for a general linear series $l$. The non-existence for $n\geq g$ for $DJ_n^{r,d}(\mu,C,l)$ follows by an induction argument explained in Section \[sec:compactified\].
Existence of de Jonquières divisors {#sec:existencedej}
-----------------------------------
Luckily, the question of existence is easily answered in a manner similar to that of the first proofs of the existence part of the Brill-Noether theorem ([@Ke] and [@KL]). The idea is to simply look at the class of $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$ and establish its positiveness. Consider the *diagonal mapping* for $C_d$: $$\begin{aligned}
\epsilon: C_{d_1}\times\ldots\times C_{d_k} &\rightarrow C_d \\
D_1+\ldots+D_k &\mapsto a_1 D_1+\ldots +a_k D_k.\end{aligned}$$ It is well-known (see for example chapter VIII §5 of [@ACGH]) that the image, via $\epsilon$ of the fundamental class of $\epsilon(C_{d_1}\times\ldots\times C_{d_k})$ is equal to the coefficient of the monomial $t_1^{d_1} \cdot\ldots \cdot t_k^{d_k}$ in $$\sum_{a\geq b} \frac{(-1)^{a+b}}{b!(a-b)!}\left(1+\sum_{i=1}^{k}a_it_i\right)^{N-g+b} \left( 1+\sum_{i=1}^{k}a_i^2t_i \right)^{g-b}x^{d-N-a}\theta^a,$$ where $\theta$ is the pullback of the fundamental class of the theta divisor to $C_d$ and $x$ the class of the divisor $q+C_{d-1}\subset C_d$. Evaluating this formula on a linear series $l$ of degree $d$ and dimension $r$, and using that $\theta\mid l=0$, we obtain the following expression for the class of $DJ^{r,d}_n(\mu,C,l)$: $$\left(1+\sum_{i=1}^{k}a_it_i\right)^{N-g} \left( 1+\sum_{i=1}^{k}a_i^2t_i \right)^{g}x^{d-N}[l].$$ If $N-d+r\geq 0$, this class is clearly positive and yields the non-emptiness of $DJ_{k,N}^{r,d}(\mu_1,\mu_2,C,l)$.
Transversality condition {#sec:nonspecial}
------------------------
We deal here only with the case of complete linear series $l\in{G^r_d(C)}$ such that $|D|=l$ for some $D\in C_d$. Consider the alternative description of the space $DJ^{r,d}_{k,N}(\mu_1\mu_2,C,l)$ as the intersection $$DJ^{r,d}_{k,N}(\mu_1,\mu_2,C,l) = \Sigma_{k,N}(\mu_1,\mu_2) \cap |D|.$$ The condition for transversality of intersection is: $$\label{eq:transvers}
T_D(C_d) = T_D(\Sigma_{k,N}(\mu_1,\mu_2)) + T_D(|D|),$$ for $D=\sum_{i=1}^k a_i D_i$, for divisors $D_i \in C_{d_i}$ and fixed vectors of strictly positive integers $\mu_1=(a_1,\ldots,a_k)$ and $\mu_2=(d_1,\ldots,d_k)$ satisfying $\sum_{i=1}^k a_k d_k=d$.
Recall that $T_D (C_d) = H^0(C,\operatorname{\mathcal{O}}_D (D))$, as shown for instance in [@ACGH] chapter IV, §1. Moreover its dual is $T_D^{\vee}(C_d)=H^0(K_C/K_C(-D))$ and the pairing between the tangent and cotangent space is given by the residue.
To compute $T_D(\Sigma_{k,N}(\mu_1,\mu_2))$, let $\mathcal{D}_i$ denote the diagonal in the $a_i$-th product $C_{d_i}\times\ldots \times C_{d_i}$ so that $\Sigma_{k,N}=\mathcal{D}_1 \times\ldots\times\mathcal{D}_k/S_d$. Hence $$T_D(\Sigma_{k,N}(\mu_1,\mu_2)) = T_{a_1 D_1}\mathcal{D}_1\oplus\ldots\oplus T_{a_k D_k}\mathcal{D}_k.$$ Since $\mathcal{T}\mathcal{D}_i=\mathcal{T}C_{d_i}$ for all $i=1,\ldots,k$, $$\begin{aligned}
T_D(\Sigma_{k,N}(\mu_1,\mu_2)) &= T_{D_1}C_{d_1} \oplus \ldots \oplus T_{D_k}C_{d_k} \\
&\simeq T_{(D_1,\ldots,D_k)} C_{d_1} \times \ldots \times C_{d_k} \\
&\simeq T_{D_1 + \ldots + D_k} C_N \\
&= H^0(C,\operatorname{\mathcal{O}}_C(D_1 + \ldots + D_k)/\operatorname{\mathcal{O}}_C).\end{aligned}$$ Its dual is given by $T_D(\Sigma_{k,N}(\mu_1,\mu_2))^{\vee}=H^0(K_C(-D_1-\ldots-D_k)/K_C(-D))^0$, where the superscript $^0$ denotes the annihilator of a vector space.
To determine $T_D |D|$, consider the following restriction of the Abel-Jacobi map: $$u:{C^r_d}\rightarrow {W^r_d(C)}$$ with differential given by $$\delta:\operatorname{\text{im}}(\alpha\mu_0)^0\rightarrow\operatorname{\text{im}}(\mu_0)^0,$$ where $\delta$ denotes the restriction of the coboundary map $$H^0(C,\operatorname{\mathcal{O}}_D(D))\rightarrow H^1(C,\operatorname{\mathcal{O}}_C)$$ of the Mittag-Leffler sequence to $T_D{C^r_d}=\operatorname{\text{im}}(\alpha\mu_0)^0$, while $$\alpha:H^0(C,K_C)\rightarrow H^0(C,K_C\otimes\operatorname{\mathcal{O}}_D)$$ is the restriction mapping and $$\mu_0:H^0(C,K_C-D)\otimes H^0(C,\operatorname{\mathcal{O}}_C(D))\rightarrow H^0(C,K_C)$$ the cup-product mapping (see Chapter IV of [@ACGH] for details).
Let $D\in{C^r_d}$. Then $\vert D \vert\subset{C^r_d}$ and $u(D)\in{W^r_d(C)}$ with $u^{-1}(u(D))=\vert D\vert$. Since $\delta$ is surjective by definition, $$T_D\vert D\vert = T_D (u^{-1}(u(D)))=\ker\delta=\operatorname{\text{im}}(\delta^{\vee})^0,$$ where the dual map $\delta^{\vee}$ is the restriction of $\alpha$ to $(\operatorname{\text{im}}(\mu_0)^0)^{\vee}=\operatorname{\text{coker}}(\mu_0)$.
The transversality condition (\[eq:transvers\]) translates to $$T_D C_d = H^0\Bigl(C,K_C\Bigl(-\sum_{i=1}^k D_i\Bigr)/K_C(-D)\Bigr)^0 + \operatorname{\text{im}}(\delta^{\vee})^0$$ which is equivalent to: $$H^0\Bigl(C,K_C\Bigl(-\sum_{i=1}^k D_i\Bigr)/K_C(-D)\Bigr)\cap\operatorname{\text{im}}(\delta^{\vee})=0.$$ If the Brill-Noether number $\rho(g,r,d)=0$, then $\mu_0$ is an isomorphism and this means that there are two possibilities for a differential in $H^0(C,K_C)$:
1. the differential is of the form $1\otimes\omega$ with $\omega\in H^0(C,K_C-D)$. Then clearly $\delta^{\vee}(\omega)=0$.
2. the differential is of the form $f\otimes\omega$, where $f$ is a meromorphic function with divisor of poles $D$. Assuming that $f\otimes\omega\in H^0(C,K_C-D_1-\ldots-D_k)$, we notice that it vanishes on $D_1 + \ldots + D_k$ if and only if $\omega\in H^0(C,K_C-D-D_1-\ldots-D_k)$.
We conclude that the transversality condition (\[eq:transvers\]) ca be reformulated as: $$\label{eq:transversefinal}
H^0(C,K_C-D-D_1-\ldots-D_k)=0.$$ If the Brill-Noether number $\rho(g,r,d)\geq 1$, then the transversality condition becomes $$H^0\Bigl(C,K_C-\sum_{i=1}^k D_i\Bigr)\cap\operatorname{\text{coker}}(\mu_0)=0.$$ Since $H^0(C,K_C)=\operatorname{\text{im}}(\mu_0)\oplus\operatorname{\text{coker}}(\mu_0)$, the transversality condition becomes $$H^0\Bigl(C,K_C-\sum_{i=1}^k D_i\Bigr)\subseteq\operatorname{\text{im}}(\mu_0).$$ Using the same argument as in the previous case, we obtain the same transversality condition (\[eq:transversefinal\]) as in the case $\rho(g,r,d)=0$.
Note that the condition (\[eq:transversefinal\]) is immediately satisfied by non-special and canonical linear series therefore proving Theorem \[dimension\] in these cases. There are actually a few more cases where transversality follows without using degenerations to nodal curves.
### The case $r=1$
The argument in this case is similar to the one in Section 5 of [@HM]. The idea is to consider the map $$\pi:C\rightarrow\operatorname{\mathbb{P}}^1$$ given by the de Jonquières divisor $D=\sum_{i=1}^k a_i D_i$ and its versal deformation space $\mathcal{V}$. Moreover, let $\mathcal{V}'\subset\mathcal{V}$ be the subvariety of maps given by divisors with the same coefficients as $D$. Then the tangent space to $\mathcal{V}$ at $\pi$ is $T_\pi \mathcal{V} = H^0(C,\mathcal{N})$, where $\mathcal{N}$ is the normal sheaf of $\pi$ defined by the exact sequence $$0 \rightarrow \mathcal{T}_C \rightarrow \pi^*\mathcal{T}_{\operatorname{\mathbb{P}}^1}\rightarrow\mathcal{N}\rightarrow 0,$$ where $\mathcal{T}_C$ is the tangent sheaf of $C$ and $\mathcal{T}_{\operatorname{\mathbb{P}}^1}$ the tangent sheaf of $\operatorname{\mathbb{P}}^1$.
Consider also the forgetful map $\beta:\mathcal{V}\rightarrow\mathcal{M}_g$, with differential $\beta_*$ given by the coboundary map $$H^0(C,\mathcal{N})\rightarrow H^1( C,\mathcal{T}_C)$$ of the exact sequence above. We now identify the tangent space to $\mathcal{V}'$ with the subspace of $H^0(C,\mathcal{N})$ of sections of $\mathcal{N}$ that vanish in a neighbourhood of the points in the support of $D_1 +\ldots+D_k$, i.e. the sections of the sheaf $\mathcal{N}'$ defined by the sequence $$0\rightarrow\mathcal{T}_C\rightarrow\pi^*\mathcal{T}_{\operatorname{\mathbb{P}}^1}(-(a_1-1)D_1-\ldots-(a_k-1)D_k)\rightarrow\mathcal{N}'\rightarrow 0.$$ Since $\pi$ is a point in the general fibre of $\beta|_{\mathcal{V}'}$, from Sard’s theorem it follows that the differential $\beta_*$ restricted to $\mathcal{T}_{\pi}(\mathcal{V})$ is surjective. This in turn means that the map $\beta'$ below is surjective: $$H^0(C,\mathcal{N}')\xrightarrow{\beta'}H^1(C,\mathcal{T}_C)\rightarrow H^1(C,\pi^*\mathcal{T}_{\operatorname{\mathbb{P}}^1}(-(a_1-1)D_1-\ldots-(a_k-1)D_k))\rightarrow 0.$$ Now, note that $\mathcal{T}_{\operatorname{\mathbb{P}}^1}\simeq\operatorname{\mathcal{O}}_{\operatorname{\mathbb{P}}^1}(2)$ and moreover $\pi^*\operatorname{\mathcal{O}}_{\operatorname{\mathbb{P}}^1}(1)=\operatorname{\mathcal{O}}_C(D)$. Therefore $$\begin{aligned}
0&=H^1(C,\operatorname{\mathcal{O}}_C(2D -(a_1-1)D_1-\ldots-(a_k-1)D_k))\\
&=H^0(C,K_C-D-D_1-\ldots-D_k)\end{aligned}$$ as desired.
### The case $r=2$
Denote by $\mathcal{Z}^N$ the $(N+2)$-dimensional subvariety of the $N$-th symmetric product of $\operatorname{\mathbb{P}}^2$ corresponding to collinear length $N$ zero-cycles in $\operatorname{\mathbb{P}}^2$. Further imposing on $\mathcal{Z}^n$ that the coefficients in these zero-cycles add up to $d$ (i.e. imposing $d$ independent conditions) yields indeed that $$\dim DJ^{2,d}_{k,N}(\mu_1,\mu_2,C,l) = N-d+2$$ for every linear series $l$.
### The case $g-d+r=1$
Let $D=\sum_{i=1}^k a_i D_i$ be a de Jonquières divisor such that $l=|D|$ is a ${g^r_d}$ with $g-d+r=1$ and, as usual, let $L$ denote the corresponding line bundle. This means that the residual linear series $K_C-l$ is an isolated divisor $E\in C_{2g-2-d}$ such that $K_C=\operatorname{\mathcal{O}}_C(D+E)$. Consider the subspace $$\mathcal{P}_d = \{ L\in\operatorname{\text{Pic}}^d(C) \mid h^1(C,L)=1 \}\subset\operatorname{\text{Pic}}^d(C),$$ which, by the previous observation, has dimension $2g-2-d$. Now consider the space $$\mathcal{Q}=\{ (E,D_1 + \ldots + D_k)\in C_{2g-d-2}\times C_N \mid \operatorname{\mathcal{O}}_C(E + a_1 D_1 + \ldots + a_k D_k) = K_C \}.$$ Polishchuk [@Po] shows that this space is smooth and such that $$\dim\mathcal{Q}=N+g-d-1.$$ Hence, for a general fixed isolated divisor $E\in C_{2g-d-2}$, the space $$\mathcal{Q}'=\{ D_1 +\ldots+ D_k\in C_N \mid \operatorname{\mathcal{O}}_C(a_1 D_1 + \ldots + a_k D_k) = \operatorname{\mathcal{O}}_C(K_C-E) \}$$ is also smooth and of dimension $$(N+g-d-1)-(2g-d-2) = N-g+1=N-d+r,$$ which immediately implies the same for the space $DJ^{r,d}_{k,N}(\mu_1,\mu_2,C,l)$ for a general linear series $l$ with $g-d+r=1$.
We can in fact do better than this and prove transversality for an arbitrary linear series $l$ with $g-d+r=1$. From Polishchuk’s result we have that the intersection $$\Sigma=\{ E+D \in C_{2g-2} \mid D\in \Sigma_{k,N}(\mu_1,\mu_2)\}\cap |K_C|$$ is transverse, i.e. $$T_{E+D} (\Sigma) + T_{E+D}|K_C| = T_{E+D} (C_{2g-2}).$$ Using the fact that $$\begin{aligned}
T_{E+D} (\Sigma) &= T_E (C_{2g-d-2}) \oplus T_{D} (\Sigma_{k,N})\\
T_{E+D} (C_{2g-2}) &= T_E C_{2g-d-2} \oplus T_D (C_d) \\
T_{E+D} |K_C| &= T_E |E| \oplus T_D |K_C - E| = T_D |L|,\end{aligned}$$ we obtain $$T_D \Sigma_{k,N}(\mu_1,\mu_2) + T_D |L| = T_D C_d,$$ which is the sought after transversality condition.
Therefore, in order to prove Theorem \[thm:smooth\], it remains to check the transversality condition (\[eq:transversefinal\]) for $r\geq 3$ and $g-d+r\geq 2$. We do this using degenerations in Section \[sec:smooth\].
De Jonquières divisors on nodal curves of compact type {#nodes}
======================================================
In the case of nodal curves, the usual correspondence between divisors and line bundles breaks down. Most significantly for our problem, the Abel-Jacobi map $$C_d \rightarrow \operatorname{\text{Pic}}^d(C)$$ does not make sense any more, even though the two spaces $C_d$ and $\operatorname{\text{Pic}}^d(C)$ are still defined. As a simple example of this failure, the sheaf of functions with one pole at one of the nodes is not a line bundle, while the sheaf of functions with two poles at the node has degree 3. We therefore first need to make sense of the statement that a linear series on a nodal curve admits a de Jonquières divisor. We do this in a variational setting, by considering families of smooth curves degenerating to a nodal curve and analysing what happens on the central fibre to limits of line bundles admitting de Jonquières divisors. As mentioned in the introduction, we approach this issue from two points of view: limit linear series for central fibres of compact type in Section \[sec:lls\] and compactified Picard schemes for stable central fibres in Section \[sec:compactified\].
Limit linear series approach {#sec:lls}
----------------------------
### Review of limit linear series {#sec:llsreview}
We recall a few well-known facts. Consider a smooth 1-parameter family $\pi:\mathscr{X}\rightarrow \Delta$ of curves of genus $g$ over the disk $\Delta$ such that the fibres over the punctured disk $\Delta^*=\Delta\setminus 0$ are smooth curves, while the special fibre is given by a nodal curve of compact type $\mathscr{X}_0$. Let $\mathscr{X}^*$ be the restriction of the family to $\Delta^*$. Suppose that $\mathscr{L}^*$ is a line bundle on $\mathscr{X}^*$ such that the restriction $\mathscr{L}_t$ to each fibre $\mathscr{X}_t$ is of degree $d$ for all $t\in\Delta^*$. Then we can extend $\mathscr{L}^*$ to a limit line bundle $\mathscr{L}$ over the whole family $\mathscr{X}$. If $Y$ is any irreducible component of $\mathscr{X}_0$, then $$\label{eq:twistingbycomponent}
\mathscr{L}\otimes\operatorname{\mathcal{O}}_{\mathscr{X}}(Y)$$ is also an extension of $\mathscr{L}^*$.
Given a line bundle $\mathscr{L}^*$ on $\mathscr{X}^*$, for each $t\in\Delta^*$ fix a non-zero subvector space $V_t\subseteq H^0(\mathscr{X}_t,\mathscr{L}_t)$ of dimension $r+1$. If $\mathscr{L}$ is an extension of $\mathscr{L}^*$ to the whole $\mathscr{X}$, let $\mathscr{V}$ be a free module of rank $r+1$ over $\Delta$ with $$\mathscr{V}_t:=V_t \cap H^0(\mathscr{X},\mathscr{L})\text{ with }t\in\Delta,$$ where the intersection is taken in $H^0(\mathscr{X}_t,\mathscr{L}_t)$. We denote by $\mathscr{V}^*$ the corresponding module over $\Delta^*$. Note that the induced homomorphism $$\mathscr{V}_0\rightarrow (\pi_* \mathscr{L})_0 \rightarrow H^0(\mathscr{X}_0, \mathscr{L}_0)$$ is injective. To summarise, we call $\mathcal{L}^*=(\mathscr{L}^*,\mathscr{V}^*)$ a *linear series on* $\mathscr{X}^*$. Given a limit line bundle $\mathscr{L}$ on $\mathscr{X}$, $\mathcal{L}^*$ extends to a linear series $\mathcal{L}:=(\mathscr{L},\mathscr{V})$ on $\mathscr{X}$. Its restriction $\mathcal{L}_0=(\mathscr{L}_0,\mathscr{V}_0)$ is a linear series of degree $d$ and dimension $r$ on $\mathscr{X}_0$. Unfortunately, twisting the bundle $\mathscr{L}$ by different components of $\mathscr{X}_0$ as in (\[eq:twistingbycomponent\]) yields infinitely many possible extensions of the linear series $\mathcal{L}^*$. Hence any geometric information that one may extract from the degeneration is lost.
In [@EH86], Eisenbud and Harris explain that in order to get the most information about the limiting behaviour of the linear series on the central fibre we should only focus on some particular extensions of the line bundle in question. Thus, for each component $Y$ of $\mathscr{X}_0$, we denote by $\mathscr{L}_Y$ the unique extension of the line bundle $\mathscr{L}^*$ that has degree $d$ on $Y$ and degree 0 on all other components of $\mathscr{X}_0$ and by $\mathscr{V}_Y$ the corresponding free module of rank $r+1$ over $\Delta$ defined as above. The advantage of this is that the sections belonging to $(\mathscr{V}_Y)_0$ vanish on all components of $\mathscr{X}_0$ except for $Y$. Hence each $\mathcal{L}_Y=(\mathscr{L}_Y,\mathscr{V}_Y)$ induces on the component $Y$ a linear series $l_Y$ of type ${g^r_d}$, which is called an *aspect* of $\mathcal{L}^*$.
The relationship between the various aspects of $\mathcal{L}^*$ is best described in terms of the *vanishing sequence* at the point $p\in Y$ $$0 \leq a_0(l_Y,p) < a_1(l_Y,p) < \cdots < a_r(l_Y,p)\leq d,$$ where the $a_i(l_Y,p)$ are the orders with which non-zero sections of $l_Y$ vanish at $p$. In the subsequent sections we shall also need the concept of a *ramification sequence* at a point $p\in Y$: $$0 \leq \alpha_0(l_Y,p) < \alpha_1(l_Y,p) < \cdots < \alpha_r(l_Y,p)\leq d,$$ where $\alpha_i(l_Y,p)=a_i(l_Y,p)-i$.
If $Z$ is another component of $\mathscr{X}_0$ such that $Y\cap Z = p$, then for all $i=0,\ldots,r$, $$\label{eq:vanish}
a_i(l_Y,p) + a_{r-i}(l_Z,p) \geq d-r .$$ To sum up, a collection $l$ of aspects of $\mathcal{L}^*$ satisfying (\[eq:vanish\]) is called a *crude limit linear series* and it was proved in [@EH86] that it indeed arises as a limit of ordinary linear series on smooth curves. If all inequalities in (\[eq:vanish\]) become equalities, then $l$ is called a *refined limit linear series*. Since refined limit linear series are in fact the ones playing the role of ordinary limit series on smooth curves, we shall usually drop the adjective “refined” unless necessary.
Unfortunately, it is not always true that a limit linear series on a nodal curve $\mathscr{X}_0$ occurs as the limit of linear series on a family $\mathscr{X}$ of smooth curves specialising to $\mathscr{X}_0$. While on the one hand there are examples of limit linear series that cannot be smoothed, on the other there are techniques for proving the smoothability of certain series, under some assumptions. For details, see Section 3 of [@EH86].
### De Jonquières divisors on a single nodal curve of compact type
In the framework of \[sec:llsreview\], fix $Y\subset\mathscr{X}_0$ an irreducible component of the central fibre. Let $\mathscr{D}^*=(\sigma)\in|\mathscr{L}^*|$ (and $\mathscr{D}^*_t=(\sigma)|_{\mathscr{X}^*_t}$) be a divisor on $\mathscr{X}^*$, where $\sigma$ is a section of $\mathscr{L}^*$. To find the limit of $\mathscr{D}^*$ on $\mathscr{X}_0$, we multiply $\sigma$ by the unique power of $t\in\Delta^*$ so that it extends to a holomorphic section $\sigma_Y$ of the extension $\mathscr{L}_Y$ on the whole of $\mathscr{X}$ and so that it does not vanish identically on $\mathscr{X}_0$. The limit of $\mathscr{D}^*$ is the divisor $(\sigma_Y|_Y)$.
\[def:comtype\] Let $\mathscr{X}_0$ be a nodal curve of compact type equipped with a smoothable limit linear series $l$ of type ${g^r_d}$. Fix an integer $k\leq d$ and two vectors of positive integers $\mu_1=(a_1,\ldots,a_k)$ and $\mu_2=(d_1,\ldots,d_k)$ such that $\sum_{i=1}^{k} a_i d_i = d$. The divisor $\sum_{i=1}^k a_i D_i$ with $D_i\in C_{d_i}$ on $\mathscr{X}_0$ is a *de Jonquières divisor for* $l$ if for each aspect $l_Y$ corresponding to an irreducible component $Y\subset X$, there is a section $\sigma_Y|_Y$ as above vanishing on $\sum_{D_{i,Y}\in Y} a_i D_{i,Y}$, where $D_{i,Y}$ is the specialization of the divisor $D_i$ on the component $Y$.
The section $\sigma_{Y}|_Y$ will also vanish at the nodes of $\mathscr{X}_0$ belonging to $Y$, and in such a way that (\[eq:vanish\]) is satisfied (we assume the limit series $l$ to be refined, so we have equality in (\[eq:vanish\])). Hence the series $l_{Y}$ of type ${g^r_d}$ on $Y$ admits the de Jonquières divisor $$\sum_{i=1}^k a_i D_{i,Y} + \sum_{q\in\text{Sing}(\mathscr{X}_0),q\in Y} \biggl(d-\sum_{i=1}^k a_i d_{i,Y}\biggr) q,$$ where the sum is over the preimages $q\in Y$ of the nodes of $\mathscr{X}_0$, and $d_{i,Y}=\deg D_{i,Y}$. We therefore have a way to go from de Jonquières divisors on a nodal curve of compact type to de Jonquières divisors on its smooth components, where the coefficients of the nodes must of course satisfy the equality in (\[eq:vanish\]). In what follows we construct the space of de Jonquières divisors on families of nodal curves of compact type and endow it once more with the structure of a degeneracy locus.
### Degenerations of de Jonquières divisors
Let $\mathscr{X}\rightarrow B$ be a flat, proper family of curves, where $B$ is a scheme. Fix a partition $\mu=(a_1,\ldots,a_n)$ of $d$ and let $p_1,\ldots,p_n:B\rightarrow\mathscr{X}$ be the sections corresponding to the markings on each fibre $\mathscr{X}_t$, for $t\in B$. By making the necessary base changes we ensure that the markings specialise on smooth points of the central fibre $\mathscr{X}_0$.
Before discussing degenerations of de Jonquières divisors, we recall some facts about moduli spaces of (limit) linear series, following the work of Osserman [@Os] and the exposition in [@Osbook].
To begin with, let $\pi:\mathscr{X}\rightarrow B$ be a proper family of smooth curves of genus $g$ with a section. Following Definition 4.2.1 in [@Osbook], the *functor* $\mathscr{G}^r_d(\mathscr{X}/B)$ *of linear series of type* ${g^r_d}$ is defined by associating to each $B$-scheme $T$ the set of equivalence classes of pairs $(\mathscr{L},\mathscr{V})$, where $\mathscr{L}$ is now a line bundle on $\mathscr{X} \times_B T$ with degree $d$ on all fibres, and $\mathscr{V}\subseteq\pi_{2*}\mathscr{L}$ is a subbundle of rank $r+1$, where $\pi_2$ denotes the second projection from the fibre product. For the precise definition of the equivalence relation, we refer the reader to [@Osbook]. This functor is represented by a scheme $G^r_d(\mathscr{X}/B)$ which is proper over $B$.
Assume now that the fibres of the family $\pi:\mathscr{X}\rightarrow B$ are nodal curves of genus $g$ of compact type such that no nodes are smoothed. Hence all fibres have the same dual graph $\Gamma$. For each vertex $v$ of $\Gamma$, let $Y^v_t$ denote the irreducible component of $\mathscr{X}_t$ corresponding to $v$. Thus for each $v$ we have a family $\mathscr{Y}^v$ of smooth curves over $B$ with fibres given by $Y^v_t$. In this case the functor $\mathscr{G}^r_d(\mathscr{X}/B)$ of linear series of type ${g^r_d}$ is defined as follows. Consider the product fibred over $B$ $$\prod_{v} \mathscr{G}^r_d(\mathscr{Y}^v/B).$$ Let $T$ be a scheme over $B$. A $T$-valued point of the above product consists of tuples of pairs $(\mathscr{L}^v,\mathscr{V}^v)$, where $\mathscr{L}^v$ is a vector bundle of degree $d$ on $\mathscr{Y}^v \times_B T$ and $\mathscr{V}\subseteq\pi_{2*}\mathscr{L}^v$ is a subbundle of rank $r+1$. Denote by $\mathscr{L}^{\vec{d}}$ the “canonical” line bundle of degree $d$ and multidegree $\vec{d}$ on $\mathscr{X}\times_B T$ obtained as in 4.4.2 of [@Osbook]. Moreover, a line bundle has multidegree $\vec{d}^v$ if it has degree $d$ on the component corresponding to the vertex $v$ and degree zero on all the other components. Note also that for two distinct multidegrees $\vec{d}$ and $\vec{d'}$, there is a unique twist map $f_{\vec{d},\vec{d'}}:\mathscr{L}^{\vec{d}}\rightarrow\mathscr{L}^{\vec{d'}}$ obtained by performing the unique minimal number of line bundle twists. According to Definition 4.4.7 in loc.cit., a $T$-valued point of $\prod_{v} \mathscr{G}^r_d(\mathscr{Y}^v/B)$ is in $\mathscr{G}^r_d(\mathscr{X}/B)(T)$ if, for all multidegrees $\vec{d}$ of $d$, the map $$\pi_{2*} \mathscr{L}^{\vec{d}} \rightarrow \bigoplus_v (\pi_{2*}\mathscr{L}^v)/\mathscr{V}^v$$ induced by the restriction to $\mathscr{Y}^v$ and $f_{\vec{d},\vec{d}^v}$ has its $(r+1)$st degeneracy locus equal to all of $T$. With this construction, $\mathscr{G}^r_d(\mathscr{X}/B)$ is also represented by a scheme $G^r_d(\mathscr{X}/B)$ proper over $B$.
Finally, if $\pi:\mathscr{X}\rightarrow B$ is a smoothing family (for details, see 4.5 of [@Osbook]), the irreducible components $Y^v_t$ may not exist for certain $t\in B$ and it follows that the dual graph of the fibres of the family is not constant. We assume from now on that there is a unique maximally degenerate fibre with dual graph $\Gamma_0$ (i.e. the family is locally smoothing). We fix a vertex $v_0 \in V(\Gamma_0)$ and set $\vec{d_0}:=\vec{d}^{v_0}$. We then replace the tuples of pairs $(\mathscr{L}^v,\mathscr{V}^v)$ with tuples $(\mathscr{L},(\mathscr{V}^v)_{v\in V(\Gamma_0)})$, where $\mathscr{L}$ is a line bundle of multidegree $\vec{d_0}$ on $\mathscr{X}\times_B T$, and for each $v\in V(\Gamma_0)$, the $\mathscr{V}^v$ are subbundles of rank $r+1$ of the twists $\pi_{2*}\mathscr{L}^{\vec{d}^v}$. Let $f:T\rightarrow B$ be a $B$-scheme. A $T$-valued point $(\mathscr{L},(\mathscr{V}^v)_{v\in V(\Gamma_0)})$ is in $\mathscr{G}^r_d(\mathscr{X}/B)(T)$ if for an open cover $\{U_m\}_{m\in I}$ of $B$ satisfying certain technical properties explained in 4.5.2 of [@Osbook], for all $m\in I$ and all multidegrees $\vec{d}$ of $d$, the map $$\pi_{2*}\mathscr{L}^{\vec{d}}|_{(f\circ \pi_2)^{-1}(U_m)} \rightarrow \bigoplus_v \left( \pi_{2*}\mathscr{L}^{\vec{d}^v}|_{(f\circ\pi_2)^{-1}(U_m)} \right)/\mathscr{V}^v|_{f^{-1}(U_m)},$$ induced by the appropriate (local) twist maps, has its $(r+1)$st degeneracy locus equal to the whole of $U_m$.
\[rem:functor\] The functor of linear series with points given by tuples $(\mathscr{L},(\mathscr{V}^v)_{v\in V(\Gamma_0)})$ is naturally isomorphic to the linear series functor with points given by tuples of pairs $(\mathscr{L}^v,\mathscr{V}^v)$ in the case of families where no nodes are smoothed (this is Proposition 4.5.5 in loc.cit.).
All the constructions can be shown to be independent of the choice of vertex $v_0$, twist maps, and open covers $\{U_m\}_{m\in I}$.
Note that all constructions are compatible with base change and moreover, the fibre over $t\in B$ is a limit linear series space when $\mathscr{X}_t$ is reducible, and a space of classical linear series when $\mathscr{X}_t$ is smooth. As a last remark, since working with (refined) limit linear series in the sense of Eisenbud and Harris is more convenient for practical purposes, we generally restrict to those (see Section 6 of [@Os] for the connection between the two approaches).
Denote by $\ell$ a $T$-valued point of $\mathscr{G}^r_d(\mathscr{X}/B)(T)$. In what follows, we construct a functor $\mathcal{DJ}^{r,d}_{k,N}(\mu_1,\mu_2,\mathscr{X},\ell)$, represented by a scheme which is projective over $B$, and which parametrises de Jonquières divisors for a family $\mathscr{X}\rightarrow B$ of curves of genus $g$ of compact type equipped with a linear series $\ell$.
\[prop:dejspacecomtype\] Fix a projective, flat family of curves $\mathscr{X}\rightarrow B$ over a scheme $B$ equipped with a linear series $\ell$ of type ${g^r_d}$. Let $\mu_1=(a_1,\ldots,a_k)$ and $\mu_2=(d_1.\ldots,d_k)$ be vectors of positive integers such that $\sum_{i=1}^k a_i d_i = d$. As usual, let $N=\sum_{i=1}^k d_i$. Consider also the relative divisors $\mathscr{D}_i\subset\mathscr{X}^{d_i}$. There exists a scheme $\mathcal{DJ}^{r,d}_{k,N}(\mu_1,\mu_2,\mathscr{X},\ell)$ projective over $B$, compatible with base change, whose point over every $t\in B$ parametrises pointed curves $[\mathscr{X}_t, \mathscr{D}_1(t),\ldots,\mathscr{D}_k(t)]$ such that $\sum_{i=1}^k a_i \mathscr{D}_i(t)$ is a de Jonquières divisor of $\ell_t$. Furthermore, every irreducible component of $\mathcal{DJ}^{r,d}_{k,N}(\mu_1,\mu_2,\mathscr{X},\ell)$ has dimension at least $\dim B - d+r$.
We construct the functor $\mathcal{DJ}^{r,d}_{k,N}(\mu_1,\mu_2,\mathscr{X},\ell)$ as a subfunctor of the functor of points of the fibre product $\mathscr{X}^N$ over $B$. We show that it is representable by a scheme that is projective over $B$ and which we also denote by $\mathcal{DJ}^{r,d}_{k,N}(\mu_1,\mu_2,\mathscr{X},\ell)$.
Let $T\rightarrow B$ be a scheme over $B$. Suppose first that all the fibres of the family are nonsingular. In this case, from the discussion above, a ${g^r_d}$ on $\mathscr{X}$ is given by a pair $(\mathscr{L},\mathscr{V})$, where $\mathscr{V}\subseteq\pi_{2*}\mathscr{L}$ is a vector bundle of rank $r+1$ on $B$. Then the $T$-valued point $[\mathscr{X},\mathscr{D}_1,\ldots,\mathscr{D}_k]$ belongs to $\mathcal{DJ}^{r,d}_{k,N}(\mu_1,\mu_2,\mathscr{X},\ell)(T)$ if the $r$-th degeneracy locus of the map $$\mathscr{V} \rightarrow \pi_{2*}\mathscr{L}|_{\sum_{i=1}^k a_i \mathscr{D}_i }$$ is the whole of $T$. By construction $\mathcal{DJ}^{r,d}_{k,N}(\mu_1,\mu_2,\mathscr{X},\ell)$ is compatible with base change, so it is a functor, and it has the structure of a closed subscheme, hence it is representable and the associated scheme is projective.
Alternatively, more explicitly, take the projective bundle $\operatorname{\mathbb{P}}\mathscr{V}$ corresponding to $\mathscr{V}$ which has rank $r$, with elements in its fibres given by sections $\sigma\in H^0(\mathscr{L}|_{\mathscr{X}_t})$ up to equivalence with respect to scalar multiplication. Consider the subscheme $\mathcal{DJ}'(\mathscr{X},\mathscr{V})$ in $\operatorname{\mathbb{P}}\mathscr{V}$ cut by the equations coming from the condition that the sections vanish on $\mathscr{D}_i$ with multiplicity at least $a_i$. This imposes in total $\sum_{i=1}^k a_i d_i = d$ conditions, so the dimension of every irreducible component of $\mathcal{DJ}'(\mathscr{X},\mathscr{V})$ is at least $\dim B - d + r$. Collecting all irreducible components of $\mathcal{DJ}'(\mathscr{X},\mathscr{V})$ such that the section $\sigma$ does not vanish on the whole underlying curve, we obtain the desired $\mathcal{DJ}^{r,d}_{k,N}(\mu_1,\mu_2,\mathscr{X},\ell)$.
Now suppose that some of the fibres have nodes (that may or may not be smoothed by $\mathscr{X}$ - see Remark \[rem:functor\]). From the discussion above, a ${g^r_d}$ on $\mathscr{X}$ is a tuple $(\mathscr{L},(\mathscr{V}^v)_{v\in V(\Gamma_0)})$. Let $v_j\in \Gamma_0$ be the vertex corresponding to an irreducible component. Denote by $\mathscr{D}_{i,j}$ the specialisation of $\mathscr{D}_i$ to $v_j$. Then the $T$-valued point $[\mathscr{X},\mathscr{D}_1,\ldots,\mathscr{D}_k]$ belongs to $\mathcal{DJ}_{k,N}^{r,d}(\mu_1,\mu_2,\mathscr{X},\ell)(T)$ if, for all vertices $v_j$, the $r$-th degeneracy locus of the map $$\mathscr{V}^{v_j} \rightarrow \pi_{2*}\mathscr{L}^{\vec{d^{v_j}}}|_{a_i \mathscr{D}_{i,j}}$$ is the whole of $T$. Checking for compatibility with base change (and hence functoriality) is more delicate than in the previous case because the base change may change the graph $\Gamma_0$. However, arguing like in the proof of Proposition 4.5.6 in loc.cit. yields the desired property. Representability and projectiveness then follow analogously.
Alternatively, if no nodes are smoothed in $\mathscr{X}$, for each vertex $v$ of the dual graph $\Gamma$ of the fibres, we have a family $\mathscr{Y}^v$ of smooth curves with the divisors $\mathscr{D}_i$ belonging to $\mathscr{Y}^v$ and additional sections $q_j$ corresponding to the preimages of the nodes. Consider now the space $\mathcal{DJ}'(\mathscr{Y}^v,\mathscr{V}^v)$ defined as in the case of families with smooth fibres by the vanishing at the $\mathscr{D}_i$. In addition, we cut $\mathcal{DJ}'(\mathscr{Y}^v,\mathscr{V}^v)$ with the equations corresponding to the vanishing of the sections at the points $q_j$, subject to the constraints explained in the discussion following Definition \[def:comtype\]. We denote the space thus obtained by $\mathcal{DJ}'(\mathscr{Y}^v,\mathscr{V}^v)$ as well. Finally, the desired space $\mathcal{DJ}_{k,N}^{r,d}(\mu_1,\mu_2,\mathscr{X},\ell)$ is obtained by taking the fibre product over $B$ of the $\mathcal{DJ}'(\mathscr{Y}^v,\mathscr{V}^v)$. The dimension estimate follows as in the case of smooth fibres. If there are smoothed nodes, for each $v\in V(\Gamma_0)$, consider the subscheme $\mathcal{DJ}'(\mathscr{X},\mathscr{V}^v)$ in $\operatorname{\mathbb{P}}\mathscr{V}^v$ cut by the vanishing conditions at the divisors $\mathscr{D}_i$ and at the nodes. Taking the fibre product over $B$ yields the space $\mathcal{DJ}_{k,N}^{r,d}(\mu_1,\mu_2,\mathscr{X},\ell)$ and the dimension bound.
\[rem:fibre\] Let $\phi:\mathcal{DJ}_{k,N}^{r,d}(\mu_1,\mu_2,\mathscr{X},\ell)\rightarrow\mathscr{X}$ be the forgetful map, which is projective by base change. Then the fibre of $\phi$ over a curve $\mathscr{X}_t$ is precisely $DJ_{k,N}^{r,d}(\mu_1,\mu_2,\mathscr{X},\ell)$.
To conclude the study of the space $\mathcal{DJ}_{k,N}^{r,d}(\mu_1,\mu_2,\mathscr{X},\ell)$ of de Jonquières divisors for a family of curves, we investigate their smoothability.
Suppose the pointed curve $[C,D_1,\ldots,D_k]\in B$ is contained in an irreducible component $U\subset \mathcal{DJ}_{k,N}^{r,d}(\mu_1,\mu_2,\mathscr{X}/B,\ell)$ with $\dim U = \dim B - d +r$. Then the general point of $U$ parametrises a de Jonquières divisor on a smooth curve.
We essentially follow the argument in the proof of Theorem 3.4 of [@EH86].
Let $\widetilde{\mathscr{X}}\rightarrow\widetilde{B}$ be the versal family of pointed curves around $[C,D_1,\ldots,D_k]$ and let $f:B\rightarrow\widetilde{B}$ be the map inducing $\pi:\mathscr{X}\rightarrow B$ with sections corresponding to the marked points. Moreover, let $\widetilde{\mathscr{L}}$ be the corresponding linear series on $\widetilde{\mathscr{X}}$. Let $\widetilde{U}\subset \mathcal{DJ}_{k,N}^{r,d}(\mu_1,\mu_2,\widetilde{\mathscr{X}}/\widetilde{B},\widetilde{\ell})$ be a component such that $U\subset f^*\widetilde{U}$ and denote by $\widetilde{C}$ the point of $\widetilde{U}$ corresponding to $C$. By Proposition \[prop:dejspacecomtype\], $\dim\widetilde{U}\geq \dim\widetilde{B} - d + r$. Hence, if $\widetilde{U}$ does not completely lie in the discriminant locus of $\widetilde{X}\rightarrow\widetilde{B}$ which parametrises nodal curves, then a general point of $\widetilde{U}$ corresponds to a de Jonquières divisor on a smooth curve. On the other hand, if $\widetilde{U}$ lies over a component $\widetilde{B}'$ of the discriminant locus, then $$\dim\widetilde{U}\geq \dim \widetilde{B} - d+r > \dim \widetilde{B}' - d+r,$$ since $\widetilde{B}'$ is a hypersurface in $\widetilde{B}$. Therefore every component of $f^*\widetilde{U}$ (hence also $U$) must have dimension strictly larger than $\dim B -d+r$ which contradicts the assumption on $\dim U$. Hence $\widetilde{U}$ cannot lie entirely in the discriminant locus, and we are done.
De Jonquières divisors on nodal stable curves {#sec:compactified}
---------------------------------------------
Consider now a smooth 1-parameter family $\pi:\mathscr{X}\rightarrow \Delta$ of curves of genus $g$ over the disk $\Delta$ such that the fibres over the punctured disk $\Delta^*=\Delta\setminus 0$ are smooth curves, while the special fibre is given by a stable nodal curve $\mathscr{X}_0$. Denote by $I(\mathscr{X}_0)$ the set of all irreducible components of the central fibre and by $N(\mathscr{X}_0)$ the set of nodes lying at the intersection of distinct irreducible components, together with their respective supports, i.e. $$N(\mathscr{X}_0) = \{ (q,C) \mid q\in C\cap C'\text{ where }C,C'\in I(\mathscr{X}_0) \}.$$ Suppose that $\mathscr{L}^*$ is a line bundle on $\mathscr{X}^*$ such that the restriction $\mathscr{L}_t$ to each fibre $\mathscr{X}_t$ is of degree $d$ for all $t\in\Delta^*$. Then, using Caporaso’s approach [@Ca] we can extend $\mathscr{L}^*$ over the central fibre $0\in\Delta$ such that the fibre $\mathscr{L}_0$ is a limit line bundle on $\mathscr{X}_0$ (or possibly a quasistable curve of $\mathscr{X}_0$) of degree $d$. As observed before, this limit is not unique because, for any $m_C\in\mathbb{Z}$, $$\label{def:twistbundle}
\mathscr{L}\otimes \operatorname{\mathcal{O}}_{\mathscr{X}}\Bigl(\sum_{C\in I(\mathscr{X}_0)} m_C C \Bigr)$$ is also an extension of $\mathscr{L}^*$ to $\Delta$. We call the new extension in (\[def:twistbundle\]) a *twisted line bundle*. Observe also the following “computation” rules $$\begin{aligned}
\label{eq:rules}
\begin{split}
\operatorname{\mathcal{O}}_{\mathscr{X}}&\simeq\operatorname{\mathcal{O}}_{\mathscr{X}}\Bigl(\sum_{C\in I(\mathscr{X}_0)} C\Bigr)\\
\operatorname{\mathcal{O}}_{\mathscr{X}}\Bigl(\sum_{C\in I(\mathscr{X}_0)} m_ C C\Bigr)\Bigl|_{C'}&\simeq\operatorname{\mathcal{O}}_{C'}\Bigl(\sum_{q\in C\cap C'} (m_C-m_{C'}) q\Bigr).
\end{split}\end{aligned}$$
We encode this information in a *twist function*: $$\begin{aligned}
T: N(\mathscr{X}_0) &\rightarrow \mathbb{Z} \\
(q,C) &\mapsto m_{C'} - m_C\end{aligned}$$ and introduce the following
\[def:twist\] A *twist* of the line bundle $\mathscr{L}$ is a function $T: N(\mathscr{X}_0) \rightarrow \mathbb{Z}$ satisfying the following properties
1. Given $C,C'\in I(\mathscr{X}_0)$ and $q\in C\cap C'$, then $T(q,C)=-T(q,C')$.
2. Given $C,C'\in I(\mathscr{X}_0)$ and $q_1,\ldots,q_n\in C\cap C'$, then $$T(q_1,C)=\ldots=T(q_n,C) = -T(q_1,C')=\ldots=-T(q_n,C).$$
3. Given $C,C',\widehat{C},\widehat{C'}\in I(\mathscr{X}_0)$, and points $q_C \in C\cap \widehat{C}$, $q_{C'}\in C'\cap\widehat{C'}$, $q\in C\cap C'$, and $\widehat{q}\in \widehat{C}\cap\widehat{C'}$, such that $$T(q_C,C)=T(q_{C'},C')=0,$$ we have that $$T(q,C)=T(\widehat{q},\widehat{C}).$$
The definition for the twist $T$ of a line bundle $L$ on a single curve $X$ is analogous.
Let $\underline{d}$ denote the *multidegree* $\underline{d}=(d_C)_{C\in I(\mathscr{X}_0)}$ of $\mathscr{L}_0$ assigning to each irreducible component $C \in I(\mathscr{X}_0)$ the degree of $\mathscr{L}_0|_{C_j}$.
Since we ultimately want to degenerate de Jonquières divisors, we introduce markings: let the sections $p_1,\ldots,p_n:\Delta\rightarrow\mathscr{X}$ correspond to the $n$ markings on each of the fibres $\mathscr{X}_t$.
We now specify what happens to the limit line bundle $\mathscr{L}_0$. We rely on Melo’s construction of the compactified Picard stack on the moduli stack of curves with marked points described in [@Me]. The Caporaso compactification emerges as a particular case (where no markings are present). We chose to work with this compactification (instead of using rank-1 torsion-free sheaves) because we want to use an induction procedure involving restrictions of line bundles on different irreducible components of the nodal curve, as in the arguments from \[sec:step1\] and \[sec:step2\]. Rank-1 torsion-free sheaves would not allow this, since their restrictions to subcurves are not necessarily torsion-free themselves.
We summarise here the most important aspects of this compactification that are relevant to us. Let $X$ be a semistable curve of genus $g\geq 2$ with $n$ marked points. For a subcurve $X'\subset X$, let $k_{X'} = \#\Bigl( X' \cap \overline{X\setminus X'} \Bigr)$. A *rational tail* $C$ of $X$ is a rational proper subcurve with $k_{X'} =1$, whereas a *rational bridge* is a rational proper subcurve $X'$ of $X$ satisfying $k_{X'} =2$. An *exceptional component* of $X$ is a destabilising component without marked points. Finally the semi-stable curve $X$ is called *quasi-stable* if the following conditions are satisfied:
- all destabilising components are exceptional;
- rational tails do not contain any exceptional components;
- each rational bridge contains at most one exceptional component.
Let $Y$ be a quasi-stable curve (obtained via semi-stable reduction) of the stable curve $X$ of genus $g\geq 2$ with $n$ marked points equipped with a line bunde $L$ of degree $d$. The multidegree of $L$ is *balanced* if
1. If $Y'\subset Y$ is an exceptional component, then $\deg_{Y'} L=1$.
2. If $Y'$ is a rational bridge, then $\deg_{Y'} L\in\{0,1\}$.
3. If $Y'$ is a rational tail, then $\deg_{Y'} L=-1$.
4. If $Y'$ is a proper subcurve whose irreducible components are not contained in any rational tail or bridge, then $\deg_{Y'} L$ must satisfy the following inequality: $$\label{ineq}\Bigl| \deg_{Y'} L - \frac{d(w_{Y'} - t_{Y'})}{2g-2} -t_{Y'} \Bigr| \leq \frac{k_{Y'} - t_{Y'} - 2b^L_{Y'}}{2},$$ where $w_{Y'} = 2(g_{Y'} - 2)$, $t_{Y'}$ is the number of rational tails meeting $Y'$, and $b^L_{Y'}$ is the number of rational bridges where the degree of $L$ vanishes and which meet $Y'$ in two points.
Denote by $\overline{P}^X_{d}$ the set of all the pairs $(Y,L)$ of quasi-stable curves $Y$ of $X$ equipped with a balanced line bundle $L$ of degree $d$. Let $\overline{W}^X_{r,d}\subset\overline{P}^X_{d}$ denote all those pairs where the line bundles satisfy $h^0(Y,L)\geq r+1$.
The compactification $\overline{P}_{d,g,n}$ of the Picard stack on the moduli stack of stable curves with marked points is given by the line bundles with balanced multidegrees on quasistable curves. More pecisely, $\overline{P}_{d,g,n}$ is a smooth and irreducible Artin stack of dimension $4g-3+n$ whose objects over a scheme $B$ are families $(\pi:\mathscr{X}\rightarrow B, p_i: B\rightarrow\mathscr{X}, \mathscr{L})$ of quasi-stable curves of genus $g$ with $n$ marked points equipped with a relative degree $d$ balanced line bundle $\mathscr{L}$. The stack $\overline{P}_{d,g,n}$ is endowed with a (forgetful) universally closed morphism $\Psi_{d,g,n}$ onto $\overline{\mathcal{M}}_{g,n}$. If moreover $(d-g+1,2g-2)=1$, the rigidification (in the sense of [@ACV]) of $\overline{P}_{d,g,n}$ is a Deligne-Mumford stack and the morphism $\Psi_{d,g,n}$ is proper. For more details, see [@Me] Definition 4.1, Theorem 4.2, and Section 7.
In what follows we also need the result below (for a proof, see [@Ra] Proposition 6.1.3).
\[lemma:raynaud\] Let $B$ be a smooth curve and let $f:\mathscr{X}\rightarrow B$ be a flat and proper morphism. Fix a point $b_0\in B$ and set $B^*=B\setminus b_0$. Let $\mathscr{L}$ and $\mathscr{M}$ be two line bundles on $\mathscr{X}$ such that $\mathscr{L}|_{f^{-1}(B^*)}\simeq\mathscr{M}_{f^{-1}(B^*)}$. Then $$\mathscr{L} = \mathscr{M} \otimes \operatorname{\mathcal{O}}_{\mathscr{X}}(\mathcal{C}),$$ where $\mathcal{C}$ is a Cartier divisor on $\mathscr{X}$ supported on $f^{-1}(b_0)$.
With all this in mind, we define the notion of de Jonquières divisors for quasi-stable nodal curves.
\[def:dejnodes\] Fix a quasi-stable curve $Y$ of a stable curve $X$ with $n$ marked points $p_1,\ldots,p_n$. The line bundle $L$ with balanced multidegree $\underline{d}$ on $Y$ admits a de Jonquières divisor $\sum_{i=1}^n a_ip_i$ if there exists a twist $T$ such that, for all $C\in I(Y)$, $$L|_C = \operatorname{\mathcal{O}}_{C}\Bigl(\sum_{p_i \in C} a_ip_i\Bigr) \otimes \operatorname{\mathcal{O}}_{C}\Bigl(\sum_{q\in C} T(q,C)q \Bigr).$$ In other words, each restriction of $L$ to the irreducible components $C$ of $Y$ admits the de Jonquières divisor $$\sum_{p_i\in C}a_i p_i + \sum_{q\in C} T(q,C)q.$$
If $C$ is an exceptional component, then the de Jonquières divisor has only the nodes $q$ in the support.
If any of the coefficients in the divisor above are negative, we find ourselves in the situation described in Section \[sec:lorentz\].
Here, our perspective on de Jonquières divisors on quasi-stable curves is naive in the sense that we ignore the precise vanishing or residue conditions at the nodes. In what follows we construct a space that not only contains the closure of the space of smooth curves with marked points and line bundles admitting de Jonquières divisors, but also some “virtual” components which we keep, in the same vein as the space of twisted canonical divisor of [@FP].
We now define the notion of de Jonquières divisors for a family of stable curves with $n$ marked points. We work locally so that a Poincaré bundle exists (otherwise we would have to assume that $(d-g+1,2g-2)=1$).
\[def:relativedej\] Let $(\pi:\mathscr{X}\rightarrow \Delta, p_i: \Delta\rightarrow\mathscr{X}, \mathscr{L})$ be a flat, proper family of quasi-stable curves of genus $g$ with $n$ marked points equipped with a relative degree $d$ balanced line bundle $\mathscr{L}$ such that $\mathscr{L}_t\in\overline{W}^{\mathscr{X}_t}_{r,d}$. For a fixed partition $\mu=(a_1,\ldots,a_n)$ of $d$ we say that $\mathscr{L}$ admits the de Jonquières divisor $\sum_{i=1}^n a_ip_i$ if for all $t\in \Delta$, $\mathscr{L}_t$ admits the de Jonquières divisor $\sum_{i=1}^n a_ip_i(t)$. Furthermore, define the locus $\mathcal{DJ}_{g,n,\mu}^{r,d}(\Delta)$ of de Jonquières divisors in $\overline{P}_{d,g,n}$ by $$\mathcal{DJ}_{g,n,\mu}^{r,d}(\Delta)=\Bigl\{ \bigl(\pi:\mathscr{X}\rightarrow \Delta, p_i: \Delta\rightarrow\mathscr{X}, \mathscr{L}\bigr) \mid \mathscr{L} \text{ admits the divisor }\sum_{i=1}^n a_i p_i\Bigr\}.$$
The content of the following proposition is that, for a one-parameter family of quasi-stable curves, the limit of de Jonquières divisors is itself a de Jonquières divisor.
\[prop:closed\] The locus $\mathcal{DJ}_{g,n,\mu}^{r,d}(\Delta)$ is closed in $\overline{P}_{d,g,n}$.
We use the valuative criterion. Take a map $\iota$ from the punctured disk $\Delta^*$ to $\mathcal{DJ}_{g,n,\mu}^{r,d}(\Delta)$. We must show that there exists a lift $\bar{\iota}$ of $\iota$ from $\Delta$, as shown in the commutative diagram below.
\(m) \[matrix of math nodes,row sep=2em,column sep=1em,minimum width=1em\] [ \^\* & \_[g,n,]{}\^[r,d]{}()\
&\
]{}; (m-1-1) edge \[right hook ->\] (m-2-1) (m-1-1) edge node \[above\] [$\iota$]{} (m-1-2) (m-2-1) edge \[dashed,->\] node \[above\] [$\bar{\iota}$]{} (m-1-2);
Since a map from $\Delta^*$ to $\mathcal{DJ}_{g,n,\mu}^{r,d}(\Delta)$ is the same as a family $(\pi:\mathscr{X}^*\rightarrow \Delta^*, p_i: \Delta^*\rightarrow\mathscr{X}^*, \mathscr{L}^*\bigr)$, we must show that we can extend this to a family $(\pi:\mathscr{X}\rightarrow \Delta, p_i: \Delta\rightarrow\mathscr{X}, \mathscr{L}\bigr)$ in $\mathcal{DJ}_{g,n,\mu}^{r,d}(\Delta)$. In other words, we must show that if the general fibre $(\mathscr{X}_t,p_i(t),\mathscr{L}_t)$, for $t\in\Delta^*$, is such that $\mathscr{L}_t$ admits the de Jonquières divisor $\sum_{i=1}^n a_i p_i(t)$, then the central fibre $(\mathscr{X}_0,p_i(0),\mathscr{L}_0)$ is such that $\mathscr{L}_0$ also admits the de Jonquières divisor $\sum_{i=0}^n a_i p_i(0)$.
From Definition \[def:dejnodes\], the family admits de Jonquières divisors if there exists a twist $T_t$ for each fibre $\mathscr{X}_t$, with $t\in\Delta^*$, such that, for all components $C\in\mathscr{X}_t$ and all nodes $q\in C$, $$\mathscr{L}_t|_C \simeq \operatorname{\mathcal{O}}_C \Bigl( \sum_{p_i(t)\in C} a_i p_i(t) \Bigr) \otimes \operatorname{\mathcal{O}}_C\Bigl( \sum_{q\in C}T_t(q,C)q \Bigr).$$ By shrinking the disk $\Delta$, and after possibly performing a base change, we may assume that the fibres of $\mathscr{X}$ are of constant topological type, the twist $T_t$ is the same twist $T$ over $\Delta^*$, and there is no monodromy in the components of the fibres over $\Delta^*$. We must now assign a twist $T_0$ to the central fibre $\mathscr{X}_0$ equipped with $\mathscr{L}_0$.
Recall that the twist $T_0$ is a function $T_0:N(\mathscr{X}_0)\rightarrow\mathbb{Z}$. There are two types of elements in $N(\mathscr{X}_0)$:
- $(q_0,C_0)$ where $q_0$ is a node not smoothed by the family $\mathscr{Y}$. Here $T_0(q_0,C_0)=T(q_t,C_t)$, where $q_t$ is the corresponding node in the component $C_t$ in the generic fibre over $t\in \Delta$.
- $(q_0,C_0)$ where $q_0$ is smoothed by the family $\mathscr{X}$. Here the twist $T_0$ must be assigned “by hand”.
To do so, note also that the component $C_0\in I(\mathscr{X}_0)$ belongs to a connected subcurve $X$ of $\mathscr{X}_0$ which consists of all components belonging to the same equivalence class with respect to twists at the non-smoothed nodes, i.e. $$C_0, C'_0\in X \Leftrightarrow C_0 \sim C'_0 \Leftrightarrow T(q,C_0) = T(q,C'_0) = 0, \forall q\in C_0\cap C'_0.$$ This yields a sub-family of $\mathscr{X}\rightarrow\Delta$, which we call $\mathscr{X}'$, whose central fibre is $X$ and whose generic fibre is given by the corresponding subcurves in $\mathscr{X}_t$. The markings $p_i$ which lie on the fibres of $\mathscr{X}'$ give sections which we rename $p'_i:\Delta\rightarrow\mathscr{X}'$, for $i=1,\ldots,n'$, where $n'\leq n$. The nodes connecting $\mathscr{X}'_t$ to its complement in $\mathscr{X}_t$ also yield sections $q_j:\Delta\rightarrow\mathscr{X}'$, for $q=1,\ldots,m$, for some $m\geq 1$; we emphasize here that the $q_j(t)$ are smooth points of $\mathscr{X}'$. Since the twist $T$ at the $q_j(t)$ is non-zero (by the definition of the equivalence classes), for $t\in\Delta^*$ and for any component $C_t\in I(\mathscr{X}'_t)$, $$\mathscr{L}_t|_{C_t} \simeq \operatorname{\mathcal{O}}_{C_t}\Bigl( \sum_{p'_i(t)\in C_t}a_i p'_i(t) + \sum_{q_j(t)\in C_t}T(q_j(t),C_t)q_j(t) \Bigr).$$
By our previous assumptions, $T(q_j(t),C_t)$ is constant for $t\in\Delta$, so in what follows we omit the terms in the bracket. Hence the line bundles $$\mathscr{L} \text{ and } \operatorname{\mathcal{O}}_{\mathscr{X}'}\biggl(\sum_{i=1}^{n'} a_i p_i + \sum_{j=1}^m Tq_j\biggr)$$ are isomorphic over $\Delta^*$ and they therefore differ by a Cartier divisor $\mathcal{C}$ on $\mathscr{X}'$, supported over $0\in\Delta$. This Cartier divisor is a sum of irreducible components of the fibre $\mathscr{X}'_0=X$, that is $$\mathcal{C}=\sum_{C\in I(X)} m_{C_0} C_0, \text{ with } m_{C_0}\in\mathbb{Z}.$$ Since the non-smoothed nodes of the family $\mathscr{X}'$ all have zero twist, this Cartier divisor yields in fact the definition of the twist $T_0:N(X)\rightarrow \mathbb{Z}$ for a node $q_0\in C_0\cap C'_0$ that is smoothed by $\mathscr{X}'$: $$(q_0,C_0) \mapsto m_{C'_0}-m_{C_0}.$$ Putting everything together, we obtain a twist $T_0:N(\mathscr{X}_0)\rightarrow\mathbb{Z}$ which by construction satisfies all the conditions of Definition \[def:twist\]. Moreover, for $t=0$,
$$\mathscr{L}_0|_{C_0} = \operatorname{\mathcal{O}}_{C_0}\biggl(\sum_{p_i \in C_0} a_ip_i(0)\biggr) \otimes \operatorname{\mathcal{O}}_{C_0}\biggl(\sum_{q_0\in C_0} T_0(q_0,C_0)q \biggr)$$ for each irreducible component $C_0$ of $\mathscr{X}_0$. By Definition \[def:relativedej\], $$(\mathscr{X}\rightarrow\Delta,p_i:\Delta\rightarrow\mathscr{X},\mathscr{L})\in\mathcal{DJ}_{g,n,\mu}^{r,d}(\Delta).$$ We conclude that $\mathcal{DJ}_{g,n,\mu}^{r,d}(\Delta)$ is closed.
Arguing like in the proof of Lemma 6 of [@FP], one can show that the line bundle associated to a de Jonquières divisor on a quasi-stable curve can be smoothed to a line bundle on a nonsingular curve. More precisely, let $(\mathscr{X}\rightarrow\Delta,p_i:\Delta\rightarrow\mathscr{X},\mathscr{L})$ be a smoothing of a quasi-stable curve with marked points $[X,p_1,\ldots,p_n]$ (so $\mathscr{X}_0=X$). Suppose also that for some $(X,L)\in\overline{P}^X_d$, $$\operatorname{\mathcal{O}}_X\biggl( \sum_{i=1}^n a_i p_i \biggr) = L \in \overline{P}^X_d,$$ Then there exists a line bundle $\mathscr{L}'\rightarrow\mathscr{X}$ and an isomorphism $\mathscr{L}'_0\simeq L$, which is constructed by twisting $\mathscr{L}$.
For the next two results, assume that $(d-g+1,2g-2)=1$ so that the definitions of de Jonquières divisors hold not just locally, but also for families over any scheme $B$. We give a lower bound on the dimension of irreducible components of $\mathcal{DJ}_{g,n,\mu}^{r,d}(\overline{P}_{d,g,n})$.
Every irreducible component of $\mathcal{DJ}_{g,n,\mu}^{r,d}(\overline{P}_{d,g,n})$ has dimension at least $3g-3+\rho(g,r,d)+n-d+r$.
The proof of this statement is the same as the one of Proposition 11 in [@FP]. The only difference is the dimension bound itself, which we explain below.
Let $[X,p_1,\ldots,p_n,L]\in\mathcal{DJ}_{g,n,\mu}^{r,d}(\overline{P}_{d,g,n})$ and $L=\operatorname{\mathcal{O}}_X\left( \sum_{i=1}^n a_i p_i \right)$ its associated twisted line bundle. We drop the markings $p_i$ without contracting the unstable components that we obtain. We then add $m$ new markings to $X$ to get rid of the automorphisms of the unstable components (see loc. cit. for details) and we obtain a stable pointed curve $[X,q_1,\ldots,q_m]$. Let $\mathcal{V}$ be its nonsingular versal deformation space. Hence $$\dim\mathcal{V} = \dim\text{Def}([X,q_1,\ldots,q_m]) = 3g-3+m.$$ Let $\pi:\mathcal{C}\rightarrow\mathcal{V}$ be the universal curve and consider the relative moduli space $\epsilon:\mathcal{B}\rightarrow\mathcal{V}$ of line bundles of degree $d$ on the fibres of $\pi$. Let $\mathcal{V}^*\subset\mathcal{V}$ be the locus of smooth curves and $\mathcal{B}^*\rightarrow\mathcal{V}^*$ the relative Picard scheme of degree $d$. Finally, let $\mathcal{W}^{r*}_d\subset\mathcal{B}^*$ be the codimension at most $(r+1)(g-d+r)$ locus of line bundles with dimension of the space of sections $r+1$. Let $\mathcal{W}^r_d$ be the closure of $\mathcal{W}^{r*}_d$ in $\mathcal{B}$. Then $$\dim\mathcal{W}^r_d \geq \dim\mathcal{B}-(r+1)(g-d+r)+m = 3g-3+\rho(g,r,d)+m.$$ This then contributes to the lower bound in the same way as in loc. cit.
Moreover, we also obtain an upper bound for the dimension of certain irreducible components of $\mathcal{DJ}_{g,n,\mu}^{r,d}(\overline{P}_{d,g,n})$ supported on the locus of marked quasi-stable curves with at least one node.
Every irreducible component of $\mathcal{DJ}_{g,n,\mu}^{r,d}(\overline{P}_{d,g,n})$ supported entirely on the locus of quasi-stable curves with $n$ marked points and at least one node has dimension at most $4g-4+n-d+r$.
Suppose $Z\subset \mathcal{DJ}_{g,n,\mu}^{r,d}(\overline{P}_{d,g,n})$ is an irreducible component supported entirely on the locus of quasi-stable curves with $n$ marked points and at least one node. Let $(X,p_1\ldots,p_n,L)\in Z$ be a generic element and denote by $\Gamma_Z$ the dual graph of the curve $X$. By the definition of $Z$, the set $E$ of edges of $\Gamma_Z$ has at least one element. Denote by $v$ the vertices of $\Gamma_Z$ and their set by $V$ (with $|V|\geq 1$). By definition, each $v$ corresponds to an irreducible component of $X$ whose genus we denote by $g_v$. Recall the genus formula: $$\label{eq:genusformula}
g-1 = \sum_{v\in V} (g_v-1) + |E|.$$ The strategy in what follows is to bound the dimension of the space of $(X,p_1,\ldots,p_n,L)\in Z$ with graph exactly $\Gamma_Z$. Now $X$ is equipped with a line bundle $L$ of degree $d$ with strictly balanced multidegree $\underline{d}=(d_v)_{v\in\Gamma_Z}$) and $h^0(X,L)=r+1$. Denote by $L_v$ the restriction of $L$ to the irreducible component corresponding to the vertex $v$ and by $n_v$ the number of the marked points on it. Thus, for a fixed vertex $v$ of $\Gamma_v$ and assuming the result of Theorem \[thm:smooth\], the dimension of the space of de Jonquières divisors of length $n_v$ on the component corresponding to $v$ is at most $3g_v-3+\rho(g_v,r_v,d_v)+n_v-d_v+r_v$, where $r_v:=h^0(L_v)-1$. The dimension bound is obtained by summing over the vertices $$\begin{aligned}
\dim Z &\leq \sum_{v\in V} \left( 3g_v - 3 + \rho(g_v,r_v,d_v) + n_v - d_v + r_v \right)\\
&\leq 3\sum_{v\in V} (g_v - 1) + n + 2|E| - d + \sum_{v\in V} r_v + \sum_{v\in V}\rho(g_v,r_v,d_v),\end{aligned}$$ where we used the fact that $\sum_{v\in V} n_v \leq n+2|E|$. The surplus of $2|E|$ comes from the preimages of the nodes on each component in case the twist from the definition of de Jonquières divisors is nonzero. From (\[eq:genusformula\]) we have $$\dim Z \leq 3g-3+n-d-|E|+\sum_{v\in V}r_v + \sum_{v\in V}\rho(g_v,r_v,d_v).$$ To estimate $\sum_{v\in V}r_v$, let $X_1$ and $X_2$ be two connected subcurves of $X$ intersecting each other at $k$ nodes. From the Mayer-Vietoris sequence $$0\rightarrow H^0(X,L)\rightarrow H^0(X_1,L|_{X_1}) \oplus H^0(X_2,L|_{X_2}) \rightarrow \mathbb{C}^k$$ we obtain $h^0(X_1,L|_{X_1}) + h^0(X_2,L|_{X_2})\leq r+1+k$. Consider in turn the same Mayer-Vietoris sequence for two connected subcurves of $X_1$ and of $X_2$, etc. , until we are left only with irreducible components. Working backwards and adding up the dimensions of the spaces of global sections for all irreducible components of $X$, we obtain $$\begin{aligned}
&\sum_{v\in V} h^0(L_v) = h^0(X,L) + |E|\Leftrightarrow \\
&\sum_{v\in V} (r_v + 1) = r+1 +|E|\Leftrightarrow\\
&\sum_{v\in V} r_v = r+1+|E|-|V|.\end{aligned}$$ For the sum of Brill-Noether numbers, we use the bound $\sum_v \rho(g_v,r_v,d_v)\leq\sum_{v\in V}g_v$, which in turn yields, using (\[eq:genusformula\]), $\sum_{v\in V}g_v=g-1-|E|+|V|$. Hence $\dim Z \leq 4g-3+n-d+r-|E|\leq 4g-4+n-d+r$.
The dimension theorem for complete linear series {#sec:prooflls}
================================================
We now give a proof of the dimension theorem (Theorem \[dimension\]) for complete linear series (i.e. those with $s=g-d+r\geq 0$) that makes use of the framework of limit linear series as discussed in Section \[sec:lls\].
We construct a nodal curve $X=C_1\cup_p C_2$ of genus $g$ out of two general pointed curves $(C_1,p)$ of genus $g_1$ and $(C_2,p)$ of genus $g_2$, where $g_1+g_2=g$. Furthermore, we equip $X$ with a limit linear series of type ${g^r_d}$ which we construct from the corresponding aspects ${g^{r}_{d_1}}(b_1p)$ on $C_1$ and ${g^{r}_{d_2}}(b_2p)$ on $C_2$, where $b_1,b_2\in\mathbb{Z}_{\geq 0}$. The genera $g_j$, the degrees $d_j$, and the multiplicities $b_j$ are chosen in such a way as to allow for a convenient induction step, where the induction hypothesis is the dimension theorem for ${g^{r}_{d_j}}$ on $C_j$ for $j=1,2$. We do this in two steps:
1. The proof for series with $s\geq 2$ and $\rho(g,r,d)=0$ works by induction on $s$ (while keeping $\rho(g,r,d)=0$ fixed), with base case given by the canonical linear series on a general smooth curve (which has $s=1$ and $\rho(g,r,d)=0$). This is done in Section \[sec:step1\].
2. The proof for linear series with $\rho(g,r,d)>0$ works by induction on $\rho(g,r,d)$ (and keeping $s$ constant), with base case given by the linear series with $\rho(g,r,d)=0$ from the previous step. This is done in Section \[sec:step2\].
In choosing the aspects ${g^{r}_{d_j}}(b_jp)$ on $C_j$ (with $j=1,2$), one has to take the following restrictions into consideration, which ensure that the limit we constructed exists and is smoothable:
- a general pointed curve $(C_j,p)\in\mathcal{M}_{g_j,1}$ may carry a ${g^{r}_{d_j}}(b_j p)$ with ramification sequence at least $(\alpha_0,\ldots,\alpha_r)$ at the point $p$ if and only if (cf. [@EH87], Proposition 1.2) $$\label{eq:ramif}
\sum_{i=0}^r(\alpha_i + g_j - d +r)_+ \leq g_j,$$ where $(x)_+=\max\{x,0\}$. In our case, the ramification sequence at $p$ is $(b_j,\ldots,b_j)$.
- the limit ${g^r_d}$ on $X$ must be refined in order to satisfy the hypotheses of the smoothability result of Eisenbud and Harris (Theorem 3.4 of [@EH83]). This means that the inequality in (\[eq:vanish\]) must be in fact an equality, thus further constraining the choice of $b_j$.
Combining these with Theorem 1.1 of [@EH87] and Corollary 3.7 of [@EH86] we obtain the smoothability of the limit ${g^r_d}$ on $X$. Assume that we are in the setting of Definition \[def:comtype\] with $\mathscr{X}_0=X$. If the limit ${g^r_d}$ on $X$ admits a de Jonquières divisor $\sum_{i=1}^n a_i D_i$, then each aspect ${g^{r}_{d_j}}(b_jp)$ on $C_j$ (with $j=1,2$) admits the de Jonquières divisor $$\sum_{i=1}^k a_i D_{i,C_j} + \Bigl(d- \sum_{i=1}^k a_i d_{i,C_j} \Bigr)p,$$ where the following inequality must hold in order to preserve the chosen ramification at $p$: $$\label{eq:ramifatp}
d- \sum_{i=1}^k a_i d_{i,C_j} \geq b_j.$$ Removing the base point $p$ from the series ${g^{r}_{d_j}}(b_j p)$, we are left with a general linear series $l_j:={g^{r}_{d_j}}$ on $C_j$ (for $j=1,2$), with simple ramification at $p$ and admitting a de Jonquières divisor $$\sum_{i=1}^k a_i D_{i,C_j} + \Bigl(d_j- \sum_{i=1}^k a_i d_{i,C_j} \Bigr)p.$$
The strategy is to prove that $\dim DJ_{k,N}^{r,d}(\mu_1,\mu_2,X,l)\leq N-d+r$ by using the dimension theorem for the spaces of de Jonquières divisors of the series $l_j$ on $C_j$. By the upper semicontinuity of fibre dimension applied to the map $\phi$ from Remark \[rem:fibre\] it follows that $$\dim DJ_{k,N}^{r,d}(\mu_1,\mu_2,X_t,l_t)\leq N-d+r$$ for a smoothing of $X$ to a general curve $X_t$ equipped with a general linear series $l_t$ of type ${g^r_d}$. Combining this with Lemma \[lemma:dimbound\], we obtain the statement of the dimension theorem for a general curve with a general linear series.
Step 1: proof for $\rho(g,r,d)=0$ {#sec:step1}
---------------------------------
Having fixed $r$ and $s=g-d+r\geq 2$, the proof in this case works by induction on $s$. The base case is given by the dimension theorem for the canonical linear series, (the unique linear series with index of speciality $s=1$ and vanishing Brill-Noether number), on a general smooth curve of any genus. This follows either from our discussion in Section \[sec:nonspecial\] or from Theorem 1.1 a) of Polishchuk [@Po] with $D=0$. The induction step constructs a curve $X$ of genus $g$ with a limit linear series $l$ of type ${g^r_d}$ with index of speciality $s$ and Brill-Noether number $\rho(g,r,d)=0$ from two irreducible components: $C_1$ equipped with a linear series $l_1$ with index of speciality $s_1=s-1$ and Brill-Noether number $\rho(l_1)=0$ and $C_2$ equipped with its canonical linear series (with index of speciality $s_2=1$). The induction hypothesis at each step is the dimension theorem for each of the components $C_1$ and $C_2$ equipped with their respective linear series $l_1$ and $l_2$.
We now show how to obtain the curve $X$. From the condition $\rho(g,r,d)=0$, we get $$\begin{aligned}
&g=s(r+1),\\
&d=g+r-s.\end{aligned}$$ We start with a general curve $C_1$ of genus $(s-1)(r+1)$ equipped with a general linear series $l_1$ of type ${g^{r}_{g-s}}$. Hence the index of speciality of $l_1$ is $$s_1=(s-1)(r+1)-g+s+r=(s-1)(r+1)-(s-1)(r+1)-r+s-1+r=s-1$$ and its Brill-Noether number is $$\begin{aligned}
\rho((s-1)(r+1),r,g-s)= (s-1)(r+1)-(r+1)(s-1)=0.\end{aligned}$$ We choose a general point $p\in C_1$ to which we attach another general curve $C_2$ of genus $r+1$ equipped with its canonical linear series $l_2={g^{r}_{2r}}$. This series has index of speciality $s_2=1$ and Brill-Noether number $$\rho(r+1,r,2r)=0.$$
Thus we obtained a curve $X=C_1\cup_p C_2$ of genus $g$. We construct on $X$ a refined limit linear series $l$ of type ${g^r_d}$ aspect by aspect using $l_1$ and $l_2$. On $C_1$ we take the aspect to be the series $l_1(rp)$, which therefore has the following vanishing sequence on $C_1$: $$(r,r+1,\ldots,2r).$$ Since the limit is refined, the vanishing sequence on $C_2$ must be $$(d-2r,\ldots,d-r),$$ so we take the aspect corresponding to $C_2$ to be the series $l_2((d-2r)p)$. Finally, we check that the limit series $$\{ (C_1,l_1(rp)),(C_2,l_2((d-2r)p) \}$$ satisfies (\[eq:ramif\]): $$\begin{aligned}
&\text{on }C_1: \sum_{i=0}^r(r+(s-1)(r+1)-d+r)_+=(r+1)(s-1)\leq (r+1)(s-1),\\
&\text{on }C_2: \sum_{i=0}^r(d-2r+r+1-d+r)_+=r+1\leq r+1.\end{aligned}$$ Hence the limit linear series $l$ on $X$ is smoothable.
We now prove that $\dim DJ_{k,N}^{r,d}(\mu_1,\mu_2,X,l)\leq N-d+r$. For $j=1,2$, let $N_j=\sum_{i=1}^k d_{i,C_j}$ and therefore $N_1 + N_2=N$. As seen in Section \[sec:lls\], $\sum_{i=1}^k a_i D_i \in DJ_{k,N}^{r,d}(\mu_1,\mu_2,X,l)$ if and only if $$\sum_{i=1} a_i D_{i,C_j} + \Bigl( d-\sum_{i=1}^k a_i d_{i,C_j} \Bigr)p$$ is a de Jonquières divisor of length (at most) $N_j+1$ of $d$ of the aspect of $l$ corresponding to $C_j$, where $j=1,2$. We distinguish a few possibilities:
1. If all points specialise on one of the $C_j$ (with $j=1,2$), then $d-\sum_{i=1}^k a_i d_{i,C_j} = 0$, contradicting inequality (\[eq:ramifatp\]). Hence we must have $$\begin{aligned}
&d-2r\leq\sum_{i=1}^k a_i d_{i,C_1}\leq d-r, \\
&2r\geq\sum_{i=1}^k a_i d_{i,C_2}\geq r.
\end{aligned}$$
2. If $\sum_{i=1}^k a_i d_{i,C_1}=d-r$, then $\sum_{i=1}^k a_i d_{i,C_2}=r$ and moreover $$\sum_{i=1}^k a_i D_{i,C_1} \in DJ_{k,N_1}^{r,d-r}(\mu'_1,\mu'_2,C_1,l_1)\text{ and } \sum_{i=1}^k a_i D_{i,C_2} + rp \in DJ_{k,N_2+1}^{r,2r}(\mu''_1,\mu''_2,C_2,l_2),$$ where $\mu'_1=(a_i)_{D_{i,C_1}>0}$, $\mu'_2=(d_{i,C_1})$ are the strictly positive vectors corresponding to the component $C_1$, while $\mu''_1=(a_i)_{D_{i,C_2}>0}$ and $\mu''_2=(d_{i,C_2},r)$ are the ones corresponding to $C_2$. By the induction hypothesis, the following inequalities must be satisfied $$\begin{aligned}
&\dim DJ_{k,N_1}^{r,d-r}(\mu'_1,\mu'_2,C_1,l_1)=N_1-d+2r=:x\geq 0 \\
&\dim DJ_{k,N_2+1}^{r,2r}(\mu''_1,\mu''_2,C_2,l_2) =N_2+1-r= (N-d+r) - x +1 \geq 0,\end{aligned}$$ where we used the fact that $N_1+N_2=N$. Furthermore, note that on $C_2$ we are actually only interested in the locus in $DJ_{k,N_2+1}^{r,2r}(\mu''_1,\mu''_2,C_2,l_2)$ consisting of divisors with $p$ in their support. More precisely, consider the incidence correspondence $$\Gamma = \{ (D,p) \mid p\in D \} \subset DJ_{k,N_2 +1}^{r,2r}(\mu''_1,\mu''_2,C_2,l_2) \times C$$ and let $\pi_1$, $\pi_2$ be the canonical projections. The locus we are after is $\pi_1(\pi_2^{-1}(p))$. By construction, $\pi_2$ is dominant and since $p$ is general, $$\dim\pi_1(\pi_2^{-1}(p))=\dim DJ_{k,N_2 +1}^{r,2r}(\mu''_1,\mu''_2,C_2,l_2)-1.$$ Therefore the dimension estimate for $DJ_{k,N}^{r,d}(\mu,X,l)$ is $$\dim DJ_{k,N}^{r,d}(\mu,X,l)\leq \dim DJ_{k,N_1}^{r,d-r}(\mu'_1,\mu'_2,C_1,l_1) + DJ_{k,N_2+1}^{r,2r}(\mu''_1,\mu''_2,C_2,l_2) -1 = N-d+r.$$
3. If $d-2r<\sum_{i=1}^k a_i d_{i,C_1}<d-r$, then $2r>\sum_{i=1}^k a_i d_{i,C_2}>r$ and we obtain de Jonquières divisors of length $N_j+1$ on the component $C_j$, for $j=1,2$. This yields $$\begin{aligned}
&\dim DJ_{k,N_1+1}^{r,d-r}(\mu'_1,\mu'_2,C_1,l_1)=N_1+1-d+2r=:x\geq 0 \\
&\dim DJ_{k,N_2+1}^{r,2r}(\mu''_1,\mu''_2,C_2,l_2) =N_2+1-r= (N-d+r) - x +2 \geq 0.\end{aligned}$$ Arguing as in the previous case (for both $C_1$ and $C_2$), we obtain the same upper bound for the dimension of $DJ_{k,N}^{r,d}(\mu_1,\mu_2,X,l)$: $$\begin{aligned}
\dim DJ_{k,N}^{r,d}(\mu,X,l)&\leq \dim DJ_{k,N_1}^{r,d-r}(\mu'_1,\mu'_2,C_1,l_1)-1 + DJ_{k,N_2+1}^{r,2r}(\mu''_1,\mu''_2,C_2,l_2) -1 \\
&= N-d+r.\end{aligned}$$
4. If $\sum_{i=1}^k a_i d_{i,C_1}=d-2r$, then $\sum_{i=1}^k a_i d_{i,C_2}=2r$ and we get de Jonquières divisors of length $N_1+1$ on $C_1$ and of length $N_2$ on $C_2$. This case is analogous to (1) and we again obtain the upper bound $N-d+r$ for $\dim DJ_{k,N}^{r,d}(\mu_1,\mu_2,X,l)$.
Step 2: proof for all $\rho(g,r,d)\geq 1$ {#sec:step2}
-----------------------------------------
Fix $r$, $s=g-d+r$, and $\rho(g,r,d)\geq 1$. We continue with the proof by induction on $\rho(g,r,d)$, where the base case is given by the dimension theorem for linear series with $\rho(g,r,d)=0$ (proved in Section \[sec:step1\]). The induction step constructs a curve $X$ of genus $g$ with a linear series $l$ of type ${g^r_d}$ from two components: $C_1$ equipped with a linear series $l_1$ and $C_2$ equipped with $l_2$ such that $\rho(l)=\rho(g,r,d)=\rho(l_1)+1$. As before, the induction hypothesis at each step is the dimension theorem for the components $C_j$ and their corresponding linear series $l_j$, with $j=1,2$.
We start with a general curve $C_1$ of genus $g-1$ equipped with a general linear series $l_1={g^{r}_{d-1}}$. We pick a general point $p\in C_1$ and attach to it an elliptic normal curve $C_2$ with its associated linear series $l_2={g^{r}_{r+1}}$. Note that the dimension theorem holds for the elliptic normal curve by virtue of the fact that $l_2$ is non-special (see the discussion in Section \[sec:nonspecial\]).
The resulting curve $X=C_1\cup_p C_2$ has genus $g$ and we construct on it a limit linear series $l$ of type ${g^r_d}$ aspect by aspect. On $C_1$ we take the aspect ${g^{r}_{d-1}}(p)$, hence $p$ is a base point of the ${g^r_d}$ on $X$ with vanishing sequence on $C_1$ given by $$(1,2,\ldots,r+1).$$ Since the limit ${g^r_d}$ must be refined in order to be smoothable, the aspect on $C_2$ must have the following vanishing sequence at $p$ $$(d-r-1,\ldots,d-1).$$ Thus the aspect on $C_2$ is given by the series ${g^{r}_{r+1}}((d-r-1)p)$.
We check that this limit ${g^r_d}$ also satisfies (\[eq:ramif\]): $$\begin{aligned}
&\text{on }C_1: (r+1)(1+g-1-d+r)=(r+1)s\leq g-1, \\
&\text{on }C_2: (r+1)(d-r-1+1-d+r)=0\leq 1, \end{aligned}$$ where in the first inequality we used the fact that $\rho(g,r,d)=g-(r+1)s\geq 1$. Hence $l$ is a smoothable limit linear series on $X$. Moreover, its Brill-Noether number is $$\rho(l)=\rho(g,r,d) = g-(r+1)s$$ while the linear series $l_1={g^{r}_{d-1}}$ on $C_1$ has Brill-Noether number $$\rho(l_1)=\rho(g-1,r,d-1) = \rho(g,r,d)-1.$$ Finally, we observe here that the induction step leaves the indices of speciality unchanged since $s_1=(g-1)-(d-1)+r=s$.
We now show that $\dim DJ_{k,N}^{r,d}(\mu_1,\mu_2,X,l)\leq N-d+r$. The argument is the same as in \[sec:step1\]. For $j=1,2$, denote by $N_j$ the length of the divisor $\sum_{i=1}^k a_i D_{i,C_j}$. As for the $\rho(g,r,d)=0$, there are a few possibilities:
1. If $\sum_{i=1}^k a_i d_{i,C_1}= d-1$, then $\sum_{i=1}^k a_i d_{i,C_2}=1$ and moreover $$\sum_{i=1}^k a_i D_{i,C_1} \in DJ_{k,N_1}^{r,d-1}(\mu'_1,\mu'_2,C_1,l_1)\text{ and }\sum_{i=1}^k a_i D_{i,C_2} + rp\in DJ_{k,N_2+1}^{r,r+1}(\mu''_1,\mu''_2,C_2,l_2),$$ where the vectors $\mu'_1,\mu'_2,\mu''_1,\mu''_2$ are defined as in \[sec:step1\]. By the induction hypothesis, $$\begin{aligned}
&\dim DJ_{k,N_1}^{r,d-1}(\mu'_1,\mu'_2,C_1,l_1) = N_1 - d + 1 +r =:x\geq 0,\\
&\dim DJ_{k,N_2+1}^{r,r+1}(\mu''_1,\mu''_2,C_2,l_{C_2}) = N_2 = (N-d+r)+1-x.
\end{aligned}$$ As discussed in \[sec:step1\], we have the bound $$\dim DJ_{k,N}^{r,d}(\mu_1,\mu_2,X,l) \leq DJ_{k,N_1}^{r,d-1}(\mu'_1,\mu'_2,C_1,l_1)+DJ_{k,N_2+1}^{r,r+1}(\mu''_1,\mu''_2,C_2,l_2)-1=N-d+r.$$
2. If $d-r-1<\sum_{i=1}^k a_i d_{i,C_1} < d-1$, then $r+1>\sum_{i=1}^k a_{i,C_2}>1$ and we get de Jonquières divisors of length $N_1+1$ on $C_1$ and length $N_2+1$ on $C_2$. Counting dimensions as before we obtain the upper bound $N-d+r$ for the dimension of $DJ_{k,N}^{r,d}(\mu_1,\mu_2,X,l)$.
3. If $\sum_{i=1}^k a_i d_{i,C_1} = d-r-1$, then $\sum_{i=1}^k a_i d_{i,C_2}=r+1$ and we have de Jonquières divisors of length $N_1+1$ on $C_1$ and length $N_2$ on $C_2$. We obtain the same upper bound $N-d+r$.
Smoothness of the space of de Jonquières divisors {#sec:smooth}
=================================================
We now prove Theorem \[thm:smooth\] which states that the space $DJ^{r,d}_{k,N}(\mu_1,\mu_2,C,l)$ is smooth for a complete linear series $l$ by showing that it arises as a transverse intersection of subvarieties of the symmetric product $C_d$. Recall that we already have the result for some cases (see Section \[sec:nonspecial\]) and it remains to show it for $r\geq 3$ and $s\geq 2$.
From the transversality condition (\[eq:transversefinal\]), we have to show that $H^0(C,K_C-D-D_1-\ldots-D_k)=0$. To do this, we prove that $$g-(d+N)+r'<0,$$ where $r'=h^0(D+D_1+\ldots+ D_k)-1=r+n'$, for some integer $n'\geq 0$.
Suppose towards a contradiction that $n'\geq N-g+d-r$.
Consider all flag curve degenerations $j:\overline{\mathcal{M}}_{0,g}\rightarrow\overline{\mathcal{M}}_g$ and let $\mathcal{Z}:=\overline{\mathcal{M}}_{0,g}\times_{\overline{\mathcal{M}}_g}\overline{\mathcal{C}}^{N}_g$, where $\overline{\mathcal{C}}_g=\overline{\mathcal{M}}_{g,1}$. Let $U\subset\mathcal{Z}$ be the closure of the divisors with $r'=r+n'$ and $n'\geq N-g+d-r$ on all curves from $\operatorname{\text{im}}(j)\subseteq\overline{\mathcal{M}}_g$. By assumption, the map $X\rightarrow\overline{\mathcal{M}}_{0,g}$ is dominant, hence $\dim X\geq g-3$. Applying Proposition 2.2 of [@Fa2], there exists a point $[\widetilde{R}:=R\cup E_1\cup\ldots\cup E_g,y_1,\ldots,y_{N}]\in U$, where $R$ is a rational spine (not necessarily smooth) and the $E_i$ are elliptic tails such that either:
1. the supports of the divisors $D_1,\ldots,D_k$ coalesce into one point, or else
2. the supports of the divisors $D_1,\ldots,D_k$ lie on a connected subcurve $Y$ of $\widetilde{R}$ of arithmetic genus $p_a(Y)=N$ and $|Y\cap\overline{(\widetilde{R}\setminus Y)}|=1$.
Denote by $q_1,\ldots,q_g$ the points of attachment of the elliptic tails to the rational spine.
A short computation using the Plücker formula allows us to immediately dismiss Case (i). We now deal with Case (ii). By assumption, there exists a proper flat morphism $\phi:\mathscr{X}\rightarrow B$ satisfying:
1. $\mathscr{X}$ is a smooth surface, $B$ is a smooth affine curve with $0\in B$ a point such that the fibre $\mathscr{X}_0$ is a curve stably equivalent to the curve $\tilde{R}$, and the fibre $\mathscr{X}_t$ is a smooth projective curve of genus $g$ for $t\neq 0$. Furthermore, we have the relative divisors $\mathscr{D}_i\in\mathscr{X}^{d_i}$ with $\mathscr{D}_i(0)=D_i$, for $i=1,\ldots,k$.
2. Let $\mathscr{X}^*=\mathscr{X}\setminus\mathscr{X}_0$. Then there exists a line bundle $\mathscr{L}^*$ on $\mathscr{X}^*$ of relative degree $d$ and with $\dim H^0(\mathscr{X}_t,\mathscr{L}_t)=r+1$ for $t\neq 0$. After (possibly) performing a base change and resolving the resulting singularities, the pair $(\mathscr{L}^*,\mathscr{V}^*:=H^0(\mathscr{X}^*,\mathscr{L}^*))$ yields a refined limit linear series $\mathfrak{m}:={g^r_d}$ on $\tilde{R}$. The limit linear series $\mathfrak{m}$ has moreover the property that it admits the de Jonquières divisor $\sum_{i=1}^k a_i \mathscr{D}_i(0)=\sum_{i=1}^k a_i D_i$.
3. The line bundle $\mathscr{L}^*$ also has the following property: $\mathscr{N}^*:=\mathscr{L}^*\bigl(\sum_{i=1}^k \mathscr{D}_i\bigr)$ is another line bundle on $\mathscr{X}^*$ of relative degree $d+N$ and with $h^0(\mathscr{X}^*_t,\mathscr{N}^*_t)=r+n'+1$. The pair $(\mathscr{N}^*,\widetilde{\mathscr{V}}^*:=H^0(\mathscr{X}^*,\mathscr{N}^*))$ also gives a limit linear series $\mathfrak{l}:={g^{r+n'}_{d+N}}$ on $\tilde{R}$. Furthermore, the limit linear series $\mathfrak{l}$ admits the de Jonquières divisor $\sum_{i=1}^k (a_i+1) D_i$.
The situation can be reformulated as follows: for $t\neq 0$, $$\dim H^0(\mathscr{X}^*_t,\mathscr{N}^*_t(-\sum_{i=1}^k \mathscr{D}_i(t))=r.$$ Then $\mathscr{N}^*\otimes\operatorname{\mathcal{O}}_{\mathscr{X}^*}(-\sum_{i=1}^k \mathscr{D}_i(B\setminus 0))$ induces the limit linear series ${g^r_d}$ that we started with.
Now, for a component $C\subset\mathscr{X}_0$, let $(\mathscr{L}_C,\mathscr{V}_C)\in G^r_d(Z)$ be the $C$-aspect $\mathfrak{m}_C$ of the limit $\mathfrak{m}={g^r_d}$. Then there exists a unique effective divisor $D_C\in C_N$ supported only at the points of $(C\cap \bigcup_{i=1}^{k}\mathscr{D}_i(B)\cap(C\cap\overline{\mathscr{X}_0\setminus C})$ such that the $C$-aspect of $\mathfrak{m}$ has the property that the restriction map $$\mathscr{V}_C\rightarrow \mathscr{V}_C|_{D_C}$$ has non-trivial kernel. For the $C$-aspect $\mathfrak{l}_C$ of the limit $\mathfrak{l}$ the situation is analogous, but now we have an effective divisor $D'_C\in C_{d+N}$ with $D'_C\geq D_C$. Moreover, the $C$-aspect of $\mathfrak{m}$ is of the form $$\mathfrak{m}_C = (\mathscr{M}_C:=\mathscr{N}_C\otimes\operatorname{\mathcal{O}}_C(-D'_C+D_C), \mathscr{W}_C\subset\widetilde{\mathscr{V}}_C\cap H^0(\mathscr{M}_C)).$$ Thus, the collection $\mathfrak{m}_Y:=\{\mathfrak{m}_C\}_{C\subset Y}$ forms a limit ${g^r_d}$ on $Y$, while the collection $\mathfrak{l}_Y:=\{\mathfrak{l}_C\}_{C\subset Y}$ forms a limit ${g^{r+n'}_{d+N}}$ on $Y$.
Let $p=Y\cap\overline{(\widetilde{R}\setminus Y)}$ and $Z:=\overline{\widetilde{R}\setminus Y}$. The vanishing sequence of the limit ${g^r_d}$ at $p$ is a subsequence of the vanishing sequence at $p$ of the limit ${g^{r+n'}_{d+N}}$. The complement of this subsequence yields another limit linear series ${g^{n'-1}_{d}}$ on $Y$ (see Lemma 2.1 of [@Fa2]). We distinguish two cases:
1. $N<g$.\
To begin with, we list two technical results that help us determine a lower bound for the ramification sequence at $p$ of the limit linear series ${g^{n'-1}_{d}}$ in this case.
\[lemma:EH\] Let $C\simeq \operatorname{\mathbb{P}}^1$ be an irreducible component of $Z$ such that $q_j \in C$ for some $j=1,\ldots,N$, where $q_j$ is the point of attachment of the elliptic tail $E_j$ to $C$. Let $l$ be a limit linear series on $Z$ and $C'$ be another component of $\widetilde{R}$ and $q=C\cap C'$. If $q'$ is another point on $C$, then for all but at most one value of $i$, $$a_i(l_C,q') < a_i(l_{C'},q).$$
\[lemma:1original\] Let $\{\sigma_C \mid C\subseteq Y \text{ irreducible component}\}$ be the set of compatible sections corresponding to the divisor $D+D_1+\ldots+D_k$. If $q\in C$, then $\operatorname{\text{ord}}_q(\sigma_C)=0$.
The proof works by induction on the components of $Y$. By construction, the tree curve $Y$ has at least two components, so the base case is $Y=C_1\cup_{q'} C_2$. Denote by $D_{C_1}$ and $D_{C_2}$ the specialisations of the divisor $D+D_1+\ldots+D_k$ on the two components $C_1$ and $C_2$. Then $\sigma_{C_1}$ vanishes on $D_{C_1}$ and $\operatorname{\text{ord}}_{q'}(\sigma_{C_1})=d+N-\deg D_{C_1}=\deg D_{C_2}$ and similarly $\sigma_{C_2}$ must vanish on $D_{C_2}$ and $\operatorname{\text{ord}}_{q'}(\sigma_{C_2})=\deg D_{C_1}$. Hence, if $q\in C_1$ is a smooth point, then $\sigma_{C_1}(q)=0$ and if $q\in C_2$ is a smooth point, then $\sigma_{C_2}(q)=0$ and we are done.
Suppose now that $Y$ has $m$ irreducible components denoted $C_1,\ldots,C_m$ and let $D_{C_1},\ldots,D_{C_m}$ be the specialisations of the divisor $D+D_1+\ldots+D_k$ to each component. Let $C_m \cap C_{m-1}=q'$. Then $$\operatorname{\text{ord}}_{q'}(\sigma_{C_{m-1}})=d+N-\deg D_{C_1}-\ldots-\deg D_{C_{m-1}}=\deg D_{C_m}.$$ Furthermore, $$\operatorname{\text{ord}}_{q'}(\sigma_{C_{m}})=d+N-\deg C_m.$$ Thus, if $q\in Y$ is a smooth point belonging to $C_m$, then $\sigma_{C_m}(q)=0$. If $q\in Y$ belongs to any of the components of the subcurve $C_1\cup\ldots\cup C_{m-1}$, then $\sigma_{C_j}(q)=0$ (with $j=1,\ldots,m-1$), where we used the induction hypothesis and the fact that $D_{C_1} + \ldots + D_{C_{m-1}} + (\deg C_{m})q'$ is a divisor of degree $d+N$ on the subcurve $C_1\cup\ldots\cup C_{m-1}$.
Let $C\subset Z$ be the irreducible component meeting $Y$ at $p$. Denote by $C'$ the component of $Y$ containing $p$. Suppose first that $C$ contains at least one of the points $q_j$ of attachment of the elliptic tails. Let $p'\in C$ be a general smooth point, which therefore has vanishing sequence $$a_i(({g^{r+n'}_{d+N}})_C,p') = (0,1,2,3,\ldots,r+n').$$ By Lemma \[lemma:EH\] with $q=p$ and $q'=p'$, the vanishing sequence at $p$ is $$a_i(({g^{r+n'}_{d+N}})_{C'},p) \geq (0,2,3,4,\ldots,r+n'+1).$$ By a similar argument, $$a_i(({g^{r}_{d}})_{C'},p) \geq (0,2,3,4,\ldots,r+1).$$ Combining this with Lemma \[lemma:1original\], we get the following ramification sequence for ${g^{n'-1}_{d}}$: $$\alpha_i(({g^{n'-1}_{d}})_{C'},p) \geq (1,1,\ldots,1).$$ In fact we obtain a limit linear series ${g^{n'-1}_{d}}$ on $Y$ with ramification $$\label{eq:ramifsequence}
\alpha_i(({g^{n'-1}_{d}})_{Y},p) \geq (1,1,\ldots,1).$$ We check a necessary condition for such a limit series to exist (cf. Theorem 1.1 of [@EH87]):
$$\label{eq:ramifcond3}
\sum_{i=0}^{n'-1}\widetilde{\alpha}_i + n'(N-d+n'-1)\leq N.$$
Since we assumed $n'\geq N-g+d-r$ and using moreover the inequality (\[eq:ramifsequence\]) we obtain that $$\sum_{i=0}^{n'-1}\widetilde{\alpha}_i + n'(N-d+n'-1) \geq (N-g+d-r)(2N-g-r).$$ Denoting by $s:=g-d+r$ and using $N > d-r$, we reformulate the necessary condition (\[eq:ramifcond3\]) as $$(N-s)(N-s-r)< N$$ which is equivalent to the quadratic inequality $$N^2 - (2s+r+1)N + s(s+1) < 0.$$ This implies that the solution $N$ must be contained in the interval $$\left(\frac{2s+r+1-\sqrt{(2s+r+1)^2-4s(s+r)}}{2},\frac{2s+r+1+\sqrt{(2s+r+1)^2-4s(s+r)}}{2}\right).$$ We now show that for $s\geq 2$ and $r\geq 3$ $$\label{eq:roots}
\frac{2s+r+1+\sqrt{(2s+r+1)^2-4s(s+r)}}{2}<d-r+1,$$ contradicting thus the hypothesis $N-d+r\geq 1$. To do this, first note that a simple calculation yields $$g\geq (r+1)s \geq 2s+r+1$$ for $s\geq 2$ and $r\geq 3$ which in turn yields $$2s+r+1\leq g-((r+1)s-2s+r+1)=g-s(r-1)+r+1.$$ Another simple calculation gives, for $r\geq 3$ and $s\geq 2$: $$\sqrt{(2s+r+1)^2-4s(s+r)} \leq (2s+r+1)-4.$$ Putting it all together, we get a sufficient condition for the inequality (\[eq:roots\]) to be satisfied, namely: $$\frac{2g-2s(r-1)+2(r+1)-4}{2} < d-r+1$$ which is equivalent to $$(2-r)(s-1) < 0.$$ This is clearly satisfied for $r\geq 3$ and $s\geq 2$ which means (\[eq:roots\]) is also satisfied for these value ranges of $r$ and $s$.
2. $N\geq g$.\
In this case $Y=\widetilde{R}$ and we check the necessary condition for the existence of a linear series ${g^{n'-1}_{d}}$ on the tree curve $Y$ without specified ramification at a point (also Theorem 1.1 of [@EH87]): $$\label{eq:ramifcond2}
n'(g-d+n'-1)\leq g.$$ By our assumptions, $n'\geq N-g+d-r\geq d-r$ and we therefore have $$\sum_{i=0}^{n'-1}\widetilde{\alpha}_i + n'(N-d+n'-1) \geq (d-r)(g-r-1).$$ Thus a necessary condition for (\[eq:ramifcond2\]) is that $$(d-r)(g-r-1)\leq g,$$ which is equivalent to $$g\leq \frac{(r+1)(d-r)}{d-r-1}.$$ However, we also know that $s=g-d+r\geq 2$ and $g\leq\frac{r+1}{r}(d-r)$, which immediately give $d\geq 3r$. This in turn yields $$g\leq \frac{(r+1)(d-r)}{d-r-1}\leq \frac{r}{2r-1}(d-r)\leq d-r,$$ which contradicts the assumption that $s=g-d+r\geq 2$.
Non-existence statement for non-complete linear series {#sec:noncomplete}
======================================================
In this section we prove Theorem \[cor:nonex\] which states that, for a smooth general curve of genus $g$, if $n-d+r<0$, the general linear series ${g^r_d}$ with $g-d+r<0$ does not admit de Jonquières divisors of length $n$ of the type $$a_1 p_1 + \ldots + a_n p_n,$$ where the points $p_i$ in the support are distinct. Recall that in this case, we need only one partition $\mu=(a_1,\ldots,a_n)$ of $d$ and we denote the space of de Jonquières divisors by $DJ^{r,d}_n(\mu,C,l)$. We proceed by induction. The base case is given by the non-existence statement in the case $n-d+r<0$ and $n<g$ shown in Lemma \[lemma:nonspecial\]. In the induction step we prove non-existence for $n\geq g$.
Consider the following quasi-stable curve $Y$ of genus $g\geq 4$ with $n\geq g$ marked points consisting of a smooth general curve $C$ of genus $g-1$ and a rational bridge with $n+1$ rational components $\gamma_j$, for $j=1,\ldots,n+1$. Since the curve is quasi-stable, at most one of the components of the rational chain is exceptional (i.e. it contains no marks). In our case, since we have $n$ marks, there must be one such component which we denote by $\gamma_{j'}$, while each of the other rational components $\gamma_j$ contains one of the marked points $p_i$. Let $C\cap \gamma_1=q_1$, $C\cap\gamma_{n+1}=q_{n+2}$, and $\gamma_j\cap \gamma_{j+1}=q_{j+1}$ for $j=1,\ldots,n+1$. The curve $Y$ is equipped with a linear series $l={g^r_d}=(L,V)$ with $g-d+r<0$ corresponding to a line bundle $L$ with $h^0(Y,L)>r+1$. The bundle $L$ has balanced multidegree $\underline{d}$, meaning that $\deg L_C = d-1$ and $\deg L_{\gamma_j}=0$ for all $j\neq j'$ and $\deg L_{\gamma_{j'}}=1$. An easy Mayer-Vietoris sequence calculation yields that $C$ is also equipped with a non-complete linear series $l_C={g^{r}_{d-1}}$.
This configuration gives a de Jonquières divisor on $Y$ corresponding to $l=(L,V)$ if there exists a twist $T$ satisfying the following system of linear equations: $$\begin{aligned}
&T(q_1,C) + T(q_{n+2},C) = d-1 \\
&T(q_j,\gamma_j) + T(q_{j+1},\gamma_j) + \sum_{p_i\in\gamma_{j}}a_i = 0 \text{ for all }j\neq j' \\
&T(q_{j'},\gamma_{j'}) + T(q_{j'+1},\gamma_{j'}) = 1.
\end{aligned}$$
Note that at least one of the terms $T(q_1,C)$ and $T(q_{n+2},C)$ must be non-zero. There are therefore two possibilities for solutions of this system:
1. Both $T(q_1,C)$ and $T(q_{n+1},C)$ are non-zero. In this case we have a de Jonquières divisor $T(q_1,C)q_1 + T(q_{n+2},C)q_{n+2}$ on $C$ of length $2$ corresponding to $l_{C}$. Note that since $2<g-1$ and $2-(d-1)+r=3-d+r<n-d+r<0$, the induction hypothesis yields that $l_C$ admits no such de Jonquières divisors.
2. Only one of the two terms is non-zero. We then have a de Jonquières divisor of length 1 corresponding to $l_C$. Since $1<g-1$ and $1-(d-1)+r<n-d+r<0$, the induction hypothesis yields that $l_C$ does not admit such de Jonquières divisors.
Hence $l$ does not admit any de Jonquières divisors on $Y$ of length $n\geq g$. We now explain how to conclude the non-existence statement for a general smooth curve with a general linear series of type ${g^r_d}$.
First note that $Y$ is embedded in $\operatorname{\mathbb{P}}^r$ by $l$ and using the methods of Hartshorne-Hirschowitz and Sernesi [@Se] (for the precise details, see for example Lemma 1.5 of [@FAO]) one shows that it is flatly smoothable to a general curve of genus $g$ and degree $d$ in $\operatorname{\mathbb{P}}^r$. Thus we have a family $\pi:\mathscr{X}\rightarrow\Delta$ of curves of genus $g$ with central fibre $\mathscr{X}_0=Y$. The family is equipped with a line bundle $\mathscr{L}$ of relative degree $d$ and such that $h^0(\mathscr{X}_t,\mathscr{L}_t)>r+1$ for all $t\in\Delta$. Thus the family $(\pi:\mathscr{X}\rightarrow\Delta,p_i:\Delta\rightarrow\mathscr{X},\mathscr{L})\notin\mathcal{DJ}^{r,d}_{g,n,\mu}(\Delta)$. Otherwise, if the smooth fibres of $\mathscr{X}\rightarrow\Delta$ admitted de Jonquières divisors, then by Proposition \[prop:closed\], the central fibre would as well. However, we have just proven this not to be the case, which concludes the induction step.
$Y$ is a quasi-stable curve obtained via semi-stable reduction from the stable curve $X$ of genus $g$ with no marked points and just one self-intersection node. Since $X$ is $d$-general (see Definition 4.13 of [@Ca2]), it follows that locally around $X$ the forgetful morphism $\Psi_{d,g,0}:\overline{P}_{d,g,0}\rightarrow\Delta$ (with $\Delta\subset\overline{{\mathcal{M}}}_{g}$) is proper. Moreover, if $\Psi_{d,g,0}$ is proper, then so are $\Psi_{d,g,n}:\overline{P}_{d,g,n}\rightarrow\Delta$ (with $\Delta\in\overline{\mathcal{M}}_{g,n}$ - see for example the discussion in Sections 7 and 8 of [@Me]) and $\mathcal{DJ}^{r,d}_{g,n,\mu}(\Delta)\rightarrow\Delta$.
Expected dimension for de Jonquières divisors with negative terms {#sec:lorentz}
=================================================================
It is also worthwhile to study de Jonquières divisors whose partition $\mu$ of $d$ contains negative terms. In fact, in Section \[sec:compactified\] we saw that negative coefficients occur naturally when considering de Jonquières divisors on nodal stable curves, as the twists $T$ may be negative. For simplicity of notation, we consider only de Jonquières divisors with distinct points in the support.
\[def:lorentz\] Fix a curve $C$ equipped with a linear series $l\in{G^r_d(C)}$ and let $$\mu=(a_1,\ldots,a_{n_1},-b_1,\ldots,-b_{n_2})$$ be a partition of $d$ of length $n$, where $a_i,b_i$ are positive integers satisfying $\sum_{i=1}^{n_1}a_i - \sum_{i=1}^{n_2}b_i=d$ and $n_1,n_2$ are fixed positive integers with $n_1+n_2=n$. We define the space $DJ_{n_1,n_2}^{r,d}(\mu,C,l)$ of de Jonquières divisors with $n_1$ positive and $n_2$ negative terms corresponding to the linear series $l$ on the curve $C$ by the rule $$\sum_{i=1}^{n_1}a_i p_i - \sum_{i=1}^{n_2}b_i q_i \in DJ_{n_1,n_2}^{r,d}(\mu,C,l)$$ if and only if $$\sum_{i=1}^{n_1}a_i p_i \in DJ_{n_1}^{r',d'}(\mu',C,l'),$$ where $p_i,q_i\in C$, $\mu'=(a_1,\ldots,a_{n_1})$ is a positive partition of $d'=\sum_{i=1}^{n_1}a_i=d+\sum_{i=1}^{n_2}b_i$, and $l'$ is the linear series of type ${g^{r'}_{d'}}$ given by $l'=l+\sum_{i=1}^{n_2}b_i q_i$.
\[thm:lorentz\] Fix a general curve $C$ of genus $g$ with a general linear series $l=(L,V)\in{G^r_d(C)}$, and let $\mu=(a_1,\ldots,a_{n_1},-b_1,\ldots,-b_{n_2})$ be a partition of $d$ of length $n$, where $a_i,b_i$ are positive integers and $n=n_1+n_2$. Assume that the points $q_i$ are general and $l'$ is complete. If $n_1-d'+r'\geq 0$, then the space $DJ^{r,d}_n(\mu,C,l)$ is of expected dimension $n-d'+r'$.
Set $L'=L(\sum_{i=1}^{n_2}b_i q_i)$. We first show that $\dim DJ^{r,d}_{n_1}(\mu',C,l')\geq n_1-d'+r'$. We distinguish a few cases.
- If $d'=2g-2$ and $L'=K_C$, then $h^0(L') = g$.
- If $d'=2g-2$, but $L'\neq K_C$, then $h^0(L') = g-1$.
- If $d' > 2g-2$, then $h^0(L') = d'-g+1$.
- If $d'<2g-2$, then $h^0(L') = r+\sum_{i=1}^{n_2}b_i+1$, by the generality of the points $q_i$.
Note that apart from the case $d'<2g-2$, the assumption that the points $q_i$ are general was not used. In all cases however, $h^0\bigl(L'|_{\sum_{i=1}^{n_1} a_i p_i}\bigr)=\sum_{i=1}^{n_1}a_i=d'$. With this in mind, we can describe the space $DJ^{r',d'}_{n_1}(\mu',C,l')$ as the locus in $C_{d'}$ where the vector bundle map $\Phi$ (constructed as in Section \[sec:deglocus\], but substituting $L'$ for $L$) has rank at most $h^0(L')-1=r'$. Hence the lower bound for the dimension of $DJ^{r',d'}_{n_1}(\mu',C,l')$ is given by
- $n_1-(h^0(L')-r')(d'-r')=n_1-d'+r'=n_1-g+1$ if $d'= 2g-2$ and $L' = K_C$,
- $n_1-(h^0(L')-r')(d'-r')=n_1-g$ if $d'\geq 2g-2$ and $L'\neq K_C$,
- $n_1-(h^0(L')-r')(d'-r')=n_1-d+r$ if $d'<2g-2$.
The fact that $$\dim DJ^{r',d'}_{n_1}(\mu',C,l') = n_1 - d' + r'$$ follows as in the case of effective de Jonquières divisors, by replacing the occurrences of $L$ by $L'$ in the proof of Theorem \[dimension\]. Finally, including the points $q_i$ in the dimension count, we get that the dimension of $DJ_{n_1,n_2}^{r,d}(\mu,C,l)$ is indeed $n-d'+r'$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'C. Ricci , S. Paltani , H. Awaki , P.-O. Petrucci , Y. Ueda , and M. Brightman'
bibliography:
- 'BaldwinI.bib'
date: 'Received; accepted'
title: 'Luminosity-dependent unification of Active Galactic Nuclei and the X-ray Baldwin effect'
---
Introduction
============
The observational signatures of reprocessed radiation in the X-ray spectra of Active Galactic Nuclei (AGN) are mainly two: a Compton hump peaking around 30keV and a fluorescent iron K$\alpha$ line. While the Compton hump is produced only if the reprocessing material is Compton-thick (CT, $N_{\rm\,H}\gtrsim 10^{24}\rm\,cm^{-2}$), the iron K$\alpha$ line is created also in Compton-thin material (e.g., [@Matt:2003fk]). The iron K$\alpha$ line is often observed as the superposition of two different components: a broad and a narrow one. The broad component has a full width at half-maximum (FWHM) of $\gtrsim 30,000\rm\,km\,s^{-1}$, and is thought to be created close to the black hole in the accretion disk (e.g., [@Fabian:2000uq]), or to be related to the presence of features created by partially covering warm absorbers in the line of sight (e.g., [@Turner:2009kx], [@Miyakawa:2012fk]). While the broad component is observed in only $\sim 35-45\%$ of bright nearby AGN (e.g., [@de-La-Calle-Perez:2010fk]), the narrow ($\mathrm{FWHM}\sim 2,000\rm\,km\,s^{-1}$, [@Shu:2010zr]) core of the iron line has been found to be almost ubiquitous (e.g., [@Nandra:2007ly], [@Singh:2011ly]). This component peaks at $6.4$keV (e.g.,[@Yaqoob:2004vn]), which points to the line being produced in cold neutral material. This material has often been identified as circumnuclear matter located at several thousand gravitational radii from the supermassive black hole, and is likely related to the putative molecular torus (e.g., [@Nandra:2006fk]), although a contribution of the outer part of the disk (e.g., [@Petrucci:2002fk]) or of the broad-line region (BLR, e.g., [@Bianchi:2008fk]) cannot be excluded. In a recent study, @Shu:2011fk found that the weighted mean of the ratio between the FWHM of the narrow Fe K$\alpha$ line and that of optical lines produced in the BLR is $\simeq 0.6$. This implies that the size of the iron K$\alpha$-emitting region is on average $\sim3$ times that of the BLR, and points towards most of the narrow iron K$\alpha$ emission being produced in the putative molecular torus.
[lcccc]{} & (2) & (3) & (4) & (5)\
Reference & $A$ & $B$ & Sample & Observatory/instrument\
\
[@Shu:2012fk] & $1.58\pm0.03^*$ & $-0.18\pm0.03$ & 32 (RQ) & [*Chandra*]{}/HEG\
[@Shu:2010zr] & $1.58\pm 0.03^*$ & $-0.22\pm0.03$ & 33 (RQ+RL) & [*Chandra*]{}/HEG\
[@Bianchi:2007vn] & $1.73\pm 0.03$ & $-0.17\pm0.03$ & 157 (RQ) & [*XMM-Newton*]{}/EPIC\
[@Jiang:2006vn] & – & $-0.20\pm0.04$ & 101 (RL+RQ)& [*XMM-Newton*]{}/EPIC + [*Chandra*]{}/HEG\
[@Jiang:2006vn] & – & $-0.10\pm0.05$ & 75 (RQ) & [*XMM-Newton*]{}/EPIC + [*Chandra*]{}/HEG\
[@Jimenez-Bailon:2005tg] & – & $-0.06\pm0.20$ & 38 (RQ) & [*XMM-Newton*]{}/EPIC\
[@Zhou:2005ys] & – &$-0.19\pm0.04$ & 66 (RQ+RL)& [*XMM-Newton*]{}/EPIC\
[@Page:2004kx] & – &$-0.17\pm0.08$ & 53 (RQ+RL)& [*XMM-Newton*]{}/EPIC\
[@Iwasawa:1993ys] & – &$-0.20\pm0.03$ & 37 (RQ+RL)& [*Ginga*]{}\
\
[@Shu:2012fk] & $1.64\pm0.03^*$ & $-0.11\pm0.03$ & 32 (RQ) & [*Chandra*]{}/HEG\
[@Shu:2010zr] & $1.63\pm 0.04^*$ & $-0.13\pm0.04$ & 33 (RQ+RL) & [*Chandra*]{}/HEG\
\
\
One of the most interesting characteristics of the narrow component of the iron K$\alpha$ line is the inverse correlation between its equivalent width ($EW$) and the X-ray luminosity (e.g., [@Iwasawa:1993ys]). The existence of an anti-correlation between the equivalent width of a line and the luminosity of the AGN continuum was found for the first time in the UV by @Baldwin:1977fk for the CIV$\,\lambda 1549$ line, and was then dubbed the [*Baldwin effect*]{}. A similar trend was later found for several other emission lines such as Ly$\alpha$, \[CIII\]$\,\lambda 1908$, SiIV$\,\lambda 1396$, MgII$\lambda 2798$ [@Dietrich:2002fk], UV iron emission lines [@Green:2001uq], mid-IR lines such as \[ARIII\]$\,\lambda8.99\mu m$, \[SIV\]$\,\lambda10.51\mu m$ and \[NeII\]$\,\lambda12.81\mu m$ [@Honig:2008kx], and forbidden lines as \[OII\]$\,\lambda 3727$ and \[NeV\]$\,\lambda 3426$ [@Croom:2002vn]. The slope of the Baldwin effect for most of these lines has been shown to be steeper for those originating from higher ionization species (e.g., [@Dietrich:2002fk]). The origin of the Baldwin effect is still unknown, although several possible explanations have been put forward, and it might be different for lines originating in different regions of the AGN. For the lines produced in the BLR it might be related to the lower ionization and photoelectric heating in the BLR gas of more luminous objects (e.g., [@Netzer:1992fk]).
In the X-rays, @Iwasawa:1993ys using [*Ginga*]{} observations of 37 AGN, found the first evidence of an anti-correlation between the equivalent width of the iron K$\alpha$ line and the $2-10\rm\,keV$ luminosity ($EW \propto L_{\,\mathrm{X}}^{-0.20\pm0.04}$). This trend is usually called the [*X-ray Baldwin effect*]{} or the [*Iwasawa-Taniguchi effect*]{}. Using [*ASCA*]{} observations, @Nandra:1997ve confirmed the existence of such an anti-correlation, and argued that most of the effect could be explained by variations of the broad component of the iron K$\alpha$ line with the luminosity. The advent of [*XMM-Newton*]{}, [*Chandra*]{} and [*Suzaku*]{} made however clear that most of the observed X-ray Baldwin effect is due to the narrow core of the iron K$\alpha$ line (e.g., [@Page:2004kx], [@Shu:2010zr], [@Fukazawa:2011ly]). The significance of the effect was questioned by @Jimenez-Bailon:2005tg, who discussed the possible importance of contamination from radio-loud (RL) AGN, which have on average larger X-ray luminosities and smaller signatures of reprocessed radiation (e.g., [@Reeves:2000uq]). However, @Grandi:2006zr, studying [*BeppoSAX*]{} observations of radio-loud AGN, also found evidence of an X-ray Baldwin effect. This, together with the study of a large sample of radio-quiet (RQ) AGN carried out by @Bianchi:2007vn using [*XMM-Newton*]{} data, showed that the X-ray Baldwin effect is not a mere artifact. @Jiang:2006vn suggested that the X-ray Baldwin effect might be due to the delay between the variability of the AGN primary energy source and that of the reprocessing material located farther away. Using [*Chandra*]{}/HEG observations, @Shu:2010zr [@Shu:2012fk] confirmed the importance of variability, showing that averaging the values of $EW$ obtained over multiple observations of individual sources would significantly attenuate the anti-correlation from $EW \propto L_{\,\mathrm{X}}^{-0.18\pm0.03}$ to $EW \propto L_{\,\mathrm{X}}^{-0.11\pm0.03}$. [*Chandra*]{}/HEG observations [@Shu:2010zr] have also shown that the normalization of the X-ray Baldwin effect is lower (i.e. the average $EW$ of the narrow Fe K$\alpha$ component is smaller) with respect to that obtained by previous works performed using [*XMM-Newton*]{}. In Table\[tab:XBEref\] we report the values of the slope and the intercept obtained from the most recent [*XMM-Newton*]{}/EPIC and [*Chandra*]{}/HEG works, along with the original [*Ginga*]{} work of @Iwasawa:1993ys.
{width="9cm"}
{width="9cm"}
An intriguing possibility is that the X-ray Baldwin effect is related to the decrease of the covering factor of the torus with luminosity (e.g., [@Page:2004kx], [@Zhou:2005ys]), in the frame of the [*luminosity-dependent unification*]{} schemes. The equivalent width of the iron K$\alpha$ line is in fact expected to be proportional to the covering factor of the torus (e.g., [@Krolik:1994fk], [@Ikeda:2009nx]), so that a covering factor of the torus decreasing with the luminosity might in principle be able to explain the X-ray Baldwin effect. Luminosity-dependent unification models have been proposed to explain the decrease of the fraction of obscured objects ($f_{\mathrm{obs}}$) with the increase of the AGN output power. The first suggestion of the existence of a relation between $f_{\mathrm{obs}}$ and the luminosity came about 30 years ago [@Lawrence:1982ys]. Since then the idea that the covering factor of the obscuring material decreases with luminosity has been gaining observational evidence from radio (e.g., [@Grimes:2004kx]), infrared (e.g., [@Treister:2008uq], [@Mor:2009fk], [@Gandhi:2009uq]), optical (e.g., [@Simpson:2005uq]) and X-ray (e.g., [@Ueda:2003qf], [@Beckmann:2009fk], [@Ueda:2011fk]) studies of AGN. Although luminosity-dependent unification models have been suspected for long to play a major role in the X-ray Baldwin effect, so far no quantitative estimation of this effect has been performed.
In this work we quantify for the first time the influence of the decrease of the covering factor of the torus with the luminosity on the equivalent width of the iron K$\alpha$ line. Using the Monte-Carlo spectral simulations of a torus with a variable half-opening angle ($\theta_{\mathrm{OA}}$) recently presented by @Ikeda:2009nx and @Brightman:2011oq, together with the most recent and comprehensive observations in different energy bands of the decrease of $f_{\mathrm{obs}}$ with the luminosity, we show that this mechanism is able to explain the X-ray Baldwin effect. The paper is organized as follows. In Sect.\[Sec:Simulations\] we present in detail our spectral simulations, and in Sect.\[sect:XBEslope\] we compare the slopes and intercepts obtained by our simulations with those measured by high sensitivity [*Chandra*]{}/HEG observations, as they provide the best energy resolution available to date. In Sect.\[Sec:discussion\] we discuss our results, and in Sect.\[Sec:summary\] we present our conclusions.
Simulations {#Sec:Simulations}
===========
The relation between $\theta_{\mathrm{OA}}$ and $L_{\mathrm{X}}$
----------------------------------------------------------------
As a proxy of the relationship between the half-opening angle of the molecular torus $\theta_{\mathrm{OA}}$ (see Fig.\[fig:tor\_CFL1\]) and the intrinsic X-ray luminosity of AGN we used the variation of the fraction of obscured sources $f_{\mathrm{obs}}$ with luminosity measured by the recent medium (2–10keV) and hard (15–55keV) X-ray surveys of @Hasinger:2008ve and @Burlon:2011cr. In the 2–10keV band, from a combination of surveys performed by [*HEAO-1*]{}, [*ASCA*]{}, [*BeppoSAX*]{}, [*XMM-Newton*]{} and [*Chandra*]{} in the luminosity range $42\leq \log L_{\mathrm{X}} \leq 46$, @Hasinger:2008ve found $$\label{Eq:Hasinger}
f_{\mathrm{obs}}\simeq-0.226\log L_{\mathrm{X}}+10.342,$$ where $L_{\mathrm{X}}$ is the luminosity in the 2–10keV energy range. In the 15–55keV band, using [*Swift*]{}/BAT to study AGN in the luminosity range $42\leq \log L_{\mathrm{X}} \leq 45$, and fitting the data with $$\label{Eq:Burlon}
f_{\mathrm{obs}}\simeq R_{\mathrm{low}}\,e^{-L_{\mathrm{HX}}/L_{\mathrm{C}}}+R_{\mathrm{high}}\,(1-e^{-L_{\mathrm{HX}}/L_{\mathrm{C}}}),$$ where $L_{\mathrm{HX}}$ is the luminosity in the 15–55keV band, @Burlon:2011cr obtained $R_{\mathrm{low}}=0.8$, $R_{\mathrm{high}}=0.2$ and $L_{\mathrm{C}}=10^{43.7}\rm\,erg\,s^{-1}$. We also used the results of the IR work of @Maiolino:2007bh, who found an anti-correlation between the ratio $\lambda L_{\lambda}(6.7\mu m)/\lambda L_{\lambda}(5100\AA)$ and the \[OIII\]$\lambda 5007\AA$ line luminosity. This trend was interpreted as an effect of the decrease of the covering factor of the circumnuclear dust as a function of luminosity, with $f_{\mathrm{obs}}$ varying as $$\label{Eq:Maiolino}
f_{\mathrm{obs}}\simeq \frac{1}{1+L_{\mathrm{opt}}^{0.414}},$$ where $L_{\mathrm{opt}}=(\lambda L_{\lambda}(5100\AA))/10^{45.63}$. Following @Maiolino:2007bh it is possible to rewrite $L_{\mathrm{opt}}$ as a function of the X-ray luminosity: $L_{\mathrm{opt}}=L_{\mathrm{X}}^{1.39}/10^{61.97}$.
The fraction of obscured sources at a given luminosity can be easily related to the half-opening angle of the torus using $$\label{Eq:fractioncos}
\theta_{\mathrm{OA}}=\cos ^{-1}( f_{\mathrm{obs}}).$$ For Eq.\[Eq:Burlon\] we converted the hard X-ray luminosity to the luminosity in the 2–10keV band assuming a photon index $\Gamma=1.9$ (e.g., [@Beckmann:2009fk]). The three $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationships used in this work are shown in Fig.\[fig:tor\_CFL2\].
![[*Top panel:*]{} extract in the 5.5-7.5keV region of a spectrum simulated using the model of @Ikeda:2009nx. The model has a torus with an half-opening angle of $\theta_{\mathrm{OA}}=46.2^{\circ}$ (equivalent to $\log L_{\mathrm{X}}=42.7$ according to the relationship of [@Hasinger:2008ve]), an equatorial column density of $\log N_{\rm\,H}^{\rm\,T}=23.8$, and an inclination angle of $\theta_{\,\mathrm{i}}=1^{\circ}$. The continuous line represents the fit to the simulated spectrum using for the continuum the same model we used for the simulations, and a Gaussian line for the iron K$\alpha$ fluorescent line. [*Bottom panel:*]{} contribution to the chi-squared for the best fit to the simulations. []{data-label="fig:simulations_line"}](sim_line1.ps){width="8.5cm"}
Spectral simulations and fitting
--------------------------------
@Ikeda:2009nx recently presented Monte-Carlo simulations of the reprocessed X-ray emission of an AGN surrounded by a three-dimensional spherical-toroidal structure. The simulations were performed using the ray-tracing method, taking into account Compton down-scattering and absorption, and are stored in tables, so that they can be used to perform spectral fitting. The free parameters of this model are the half-opening angle of the torus $\theta_{\mathrm{OA}}$, the line of sight inclination angle $\theta_{\,\mathrm{i}}$, the torus equatorial column density $N_{\rm\,H}^{\rm\,T}$ and the photon index $\Gamma$ of the primary continuum. [*Note that $N_{\rm\,H}^{\rm\,T}$ should not be confused with the observed hydrogen column density $N_{\rm\,H}$. In all objects, $N_{\rm\,H}\leq N_{\rm\,H}^{\rm\,T}$, with $N_{\rm\,H}=N_{\rm\,H}^{\rm\,T}$ only if $\theta_{\,\mathrm{i}}=90^{\circ}$*]{}. In Fig.\[fig:tor\_CFL1\] a schematic representation of the geometry considered is shown. The dependence of $N_{\rm\,H}$ on the inclination angle is given by Eq.3 of @Ikeda:2009nx. In the model of @Ikeda:2009nx the ratio $R_{\mathrm{in}}/R$ is fixed to 0.01, and the inclination angle of the observer ${\mathit \theta}_{\mathrm{i}}$ can vary between $1^{\circ}$ and $89^{\circ}$, $N_{\rm\,H}^{\rm\,T}$ between $10^{22}$ and $10^{25}\rm\,cm^{-2}$, while $\theta_{\mathrm{OA}}$ spans the range between $10^{\circ}$ and $70^{\circ}$. Because of the assumed dependence of the half-opening angle on the luminosity, the interval of values of $\theta_{\mathrm{OA}}$ allowed by the model limits the range of luminosities we can probe. In the following we will consider luminosities above $\log L_{\mathrm{X}} = 42$, as below this value few AGN are detected, and several works point towards a possible disappearance of the molecular torus (e.g., [@Elitzur:2006fk]). The upper-limit luminosity we can reach depends on the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship used, and considering $\theta_{\mathrm{OA}}^{\mathrm{\,max}}=70^{\circ}$, is $\log L_{\rm\,max}=44.2$ for @Hasinger:2008ve, $\log L_{\rm\,max}=43.8$ for @Burlon:2011cr and $\log L_{\rm\,max}=45.0$ for @Maiolino:2007bh. Due to its limitations in the values of $\theta_{\mathrm{OA}}$ permitted, the model of @Ikeda:2009nx restricts most of the simulations to the Seyfert regime ($\log L_{\mathrm{X}}\leq \log L_{\mathrm{X}}^{\mathrm{Q}}=44.2$).
To extend our study to the quasar regime ($L_{\mathrm{X}}> L_{\mathrm{X}}^{\mathrm{Q}}$) we simulated the X-ray Baldwin effect using the spectral model of @Brightman:2011oq. This model considers the same geometry as @Ikeda:2009nx, but has $R_{\mathrm{in}}=0$, a $\theta_{\,\mathrm{i}}$-independent $N_{\rm\,H}$, and it allows different values of $\theta_{\mathrm{OA}}$ ($26^{\circ}$ to $84^{\circ}$). The higher maximum half-opening angle permitted by this model allows to reach higher maximum luminosities in the simulations for the relationships of @Hasinger:2008ve ([-@Hasinger:2008ve]; $\log L_{\rm\,max}=45.3$) and @Maiolino:2007bh ([-@Maiolino:2007bh]; $\log L_{\rm\,max}=46.2$), while that of @Burlon:2011cr is flat ($f_{\rm\,obs}\simeq0.2$) in the quasar regime. However, the lower boundary of $\theta_{\mathrm{OA}}$ in the model of @Brightman:2011oq limits the lower luminosity we can reach in the simulations, in particular $\log L_{\rm\,min}\simeq 43$ for the relation of @Maiolino:2007bh. Thus, we used the model of @Ikeda:2009nx to study the X-ray Baldwin effect in the Seyfert regime ($42\leq\log L_{\mathrm{X}} \leq 44.2$), and the model of @Brightman:2011oq to probe the quasar regime ($\log L_{\mathrm{X}} > 44.2$).
![Equivalent width of the iron K$\alpha$ line versus the X-ray luminosity obtained simulating a torus with an equatorial column density of $\log N_{\rm\,H}^{\rm\,T}=23.1$ and a covering factor decreasing with the luminosity. The $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship used here is that of @Hasinger:2008ve (see Fig.\[fig:tor\_CFL2\]). The relations of @Bianchi:2007vn (black dashed line) and @Shu:2012fk ([-@Shu:2012fk]; blue line, obtained averaging different observations of the same source) are also shown for comparison (see also Table\[tab:XBEref\]). The normalization of the relation of @Bianchi:2007vn has been fixed to an arbitrary value for comparison.[]{data-label="fig:IT_hasinger"}](ITeff_hasinger.ps){width="9cm"}
{width="9cm"}
{width="9cm"}
To estimate the influence of the decreasing covering factor of the torus on the equivalent width of the iron line, we simulated, using the physical torus models, a large number of spectra using the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationships reported in Eq.\[Eq:Hasinger\]-\[Eq:Maiolino\]. Using the three different $\theta_{\mathrm{OA}} - L_{\mathrm{X}}$ relationships, we extrapolated the value of $\theta_{\mathrm{OA}}$ for each luminosity bin ($\Delta \log L_{\mathrm{X}}=0.01$). We fixed the equatorial column density of the torus $N_{\rm\,H}^{\rm\,T}$ to 26 different values, spanning between $10^{22.5}\rm\,cm^{-2}$ and $10^{25}\rm\,cm^{-2}$, with a step of $\Delta \log N_{\rm\,H}^{\rm\,T}=0.1$. To simulate an unabsorbed population, similarly to what is usually used to determine the X-ray Baldwin effect, for each value of $L_{\rm\,X}$ and $N_{\rm\,H}^{\rm\,T}$ we considered inclination angles $\theta_{\,\mathrm{i}}$ between $1^{\circ}$ and $\theta_{\mathrm{OA}}$, with a binning of $\Delta \theta_{\,\mathrm{i}}=3^{\circ}$. In XSPEC 12.7.1 [@Arnaud:1996kx] we simulated spectra for each of the three $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationships and for each bin of column density. For the continuum we used a power law with $\Gamma=1.9$ (e.g., [@Beckmann:2009fk]). Our choice of the photon index does not affect significantly the simulations, and adding a scatter of $\Delta \Gamma=0.3$ we found $EW-L_{\mathrm{X}}$ trends consistent with those obtained without scatter. The metallicity was set to the solar value in all our simulations, and the value of the normalization of the reflected component was fixed to that of the continuum.
We fitted the simulated spectra in the 0.3-10keV band using the same model adopted for the simulations, substituting the iron K$\alpha$ line component with a Gaussian line. We obtained a good fit for all the simulations, with a reduced chi-squared of $\chi^{2}_{\nu}\lesssim 1.1$. We report in Fig.\[fig:simulations\_line\] an example of a typical fit. The model used for the continuum fitting does not affect significantly the results, and using an alternative model like `pexrav` [@Magdziarz:1995pi] the equivalent width of the iron K$\alpha$ line differs on average of only $\sim4\%$. We evaluated the equivalent width of the iron K$\alpha$ line and studied its relationship with the luminosity. As an example, in Fig.\[fig:IT\_hasinger\] the simulated X-ray Baldwin effect obtained using the model of @Ikeda:2009nx and the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship of @Hasinger:2008ve for $\log N_{\rm\,H}^{\rm\,T}=23.1$ is shown, together with the fit to the X-ray Baldwin effect obtained by the recent works of @Bianchi:2007vn and @Shu:2012fk. The spread in $EW$ for a given luminosity is due to the range of values of $\theta_{\,\mathrm{i}}$ we considered: larger values of $EW$ usually correspond to lower values of $\theta_{\,\mathrm{i}}$.
Most of the studies of the X-ray Baldwin effect performed in the last years have used a relationship of the type $$\label{eq:fit_XBE}
\log EW=A+B \log L_{\mathrm{X,44}},$$ to fit the $EW-L_{\mathrm{X}}$ trend, where $L_{\mathrm{X,44}}$ is the luminosity in units of $10^{44}\rm\,erg\,s^{-1}$. In order to compare the simulated the $EW-L_{\mathrm{X}}$ trend with that observed in unabsorbed populations of AGN, we fitted, for each value of $N_{\rm\,H}^{\rm\,T}$, the simulated data with Eq.\[eq:fit\_XBE\] using the weighted least-square method, with weights of $w=\sin \theta_{\,\mathrm{i}}$. This allows to account for the non-uniform probability of randomly observing an AGN within a certain solid angle from the polar axis.
The X-ray Baldwin effect {#sect:XBEslope}
========================
Most of the studies on the X-ray Baldwin effect have found, using Eq.\[eq:fit\_XBE\], a slope of $B\sim -0.2$ (e.g., [@Iwasawa:1993ys], [@Page:2004kx], [@Bianchi:2007vn]). However, in all these works the values of $EW$ and $L_{\mathrm{X}}$ are obtained from individual observations of sources, and, as pointed out by @Jiang:2006vn and confirmed by @Shu:2010zr [@Shu:2012fk], flux variability might play an important role in the observed X-ray Baldwin effect. @Shu:2012fk recently found a clear anti-correlation between the equivalent width of the iron K$\alpha$ line and the luminosity for individual sources which had several [*Chandra*]{}/HEG observations. Using a sample of 32 radio-quiet AGN, they also found that the fit [*per source*]{} (i.e. averaging different observations of the same object) results in a significantly flatter slope ($B=-0.11\pm0.03$) than that done [*per observation*]{} (i.e. using all the available observations for every source of the sample; $B=-0.18\pm0.03$).
In order to compare the slope obtained by the simulations in the two different luminosity bands with real data, we fitted the data per source (33 AGN) of @Shu:2010zr (; obtained fixing $\sigma=1\rm\,eV$) with Eq.\[eq:fit\_XBE\] in the Seyfert and quasar regime. Consistently to what reported in @Shu:2010zr we did not use the 3 AGN for which only upper limits of $EW$ were obtained. Using the 28 AGN in the Seyfert regime we obtained that the best fit to the X-ray Baldwin effect is given by $A=1.64\pm0.05$ and $B=-0.12\pm0.04$. [*Chandra*]{}/HEG observations are available for only 5 objects in the quasar regime, and at these luminosities we obtained $A=1.5\pm0.3$ and $B=-0.16\pm0.22$.
{width="9cm"}
{width="9cm"}
Seyfert regime {#sec:Syregime}
--------------
Fitting the simulated data with Eq.\[eq:fit\_XBE\] in the Seyfert regime, we found that the slope obtained becomes flatter for increasing values of the equatorial column density of the torus (Fig.\[fig:slope\_NH\], left panel). This is related to the fact that the iron K$\alpha$ $EW$ is tightly connected to both $N_{\rm\,H}^{\rm\,T}$ and $\theta_{\mathrm{OA}}$, so that $EW$ is more strongly dependent on $\theta_{\mathrm{OA}}$ (and thus on its variation) for low values of $N_{\rm\,H}^{\rm\,T}$, which results in a steeper slope. On the other hand, when $N_{\rm\,H}^{\rm\,T}$ increases the dependence on $\theta_{\mathrm{OA}}$ becomes weaker, and the slope flatter. As shown in Fig.\[fig:slope\_NH\] (left panel), while the values of $B$ obtained by the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationships of @Burlon:2011cr and @Hasinger:2008ve are similar along the whole range of $N_{\rm\,H}^{\rm\,T}$ considered, flatter slopes are obtained for that of @Maiolino:2007bh. In particular for the latter relationship the correlation becomes positive for $\log N_{\rm\,H}^{\rm\,T}\gtrsim 24$.
By combining the observations of the X-ray Baldwin effect with our simulations we can extrapolate the average value of the equatorial column density of the torus of the unobscured AGN in the [*Chandra*]{}/HEG sample of @Shu:2010zr. A slope consistent within $1\sigma$ with our fit to the X-ray Baldwin effect in the Seyfert regime is obtained for $N_{\rm\,H}^{\rm\,T}\geq 10^{23.1}\rm\,cm^{-2}$ for the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship of @Hasinger:2008ve, for $N_{\rm\,H}^{\rm\,T}\gtrsim 10^{23.2}\rm\,cm^{-2}$ for that of @Burlon:2011cr, and for $N_{\rm\,H}^{\rm\,T}\lesssim10^{23}\rm\,cm^{-2}$ for the relationship of @Maiolino:2007bh. Comparing the value of the intercept obtained by the simulations with that resulting from our fit of the X-ray Baldwin effect in the Seyfert regime, we found that only a narrow range of average equatorial column densities of the torus can reproduce the observations (right panel of Fig.\[fig:slope\_NH\]). For the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship of @Maiolino:2007bh we found $22.9 \lesssim \log N_{\rm\,H}^{\rm\,T} \lesssim 23$, while for those of @Burlon:2011cr and @Hasinger:2008ve we obtained $\log N_{\rm\,H}^{\rm\,T} \simeq 23.2$. Using both the values of $A$ and $B$, it is possible to extract the average values of $N_{\rm\,H}^{\rm\,T}$ that can explain the X-ray Baldwin effect. As it can be seen from the two figures for both the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship of @Hasinger:2008ve and @Burlon:2011cr the only value of column density consistent with both the observed intercept and slope is $\log N_{\rm\,H}^{\rm\,T}\simeq 23.2$, while for the IR $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship of @Maiolino:2007bh the values allowed are in the range $22.9 \lesssim \log N_{\rm\,H}^{\rm\,T} \lesssim 23$. A similar result is obtained studying the $A/B$ chi-squared contour plot of our fit to the [*Chandra*]{}/HEG data. Using the model of @Brightman:2011oq to simulate the X-ray Baldwin effect in the Seyfert regime we obtained a range of $N_{\rm\,H}^{\rm\,T}$ consistent with that found using the model of @Ikeda:2009nx ($23.2 \lesssim \log N_{\rm\,H}^{\rm\,T} \lesssim 23.3$). The lower values of $N_{\rm\,H}^{\rm\,T}$ needed to explain the X-ray Baldwin effect using the relationship of @Maiolino:2007bh are due to the larger values of $f_{\mathrm{obs}}$ (and thus of $\theta_{\mathrm{OA}}$) predicted by Eq.\[Eq:Maiolino\]. This is again due to the fact that $EW$ depends on both $N_{\rm\,H}^{\rm\,T}$ and $\theta_{\mathrm{OA}}$, thus increasing the latter one would obtain lower values of the former.
Quasar regime {#sec:Qsoregime}
-------------
To study the behavior of $B$ in the quasar regime, we fitted the simulated data using Eq.\[eq:fit\_XBE\] for each value of $N_{\rm\,H}^{\rm\,T}$ for luminosities $L_{\mathrm{X}}>L_{\mathrm{X}}^{\mathrm{Q}}$. Our simulations show that $EW$ decreases more steeply in the quasar regime than at lower luminosities for the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationships of @Hasinger:2008ve and @Maiolino:2007bh, with values of the slope of $B\lesssim -0.3$ (left panel of Fig.\[fig:interc\_sy\_2\]). The slopes obtained by the simulations are consistent with those found by fitting [*Chandra*]{}/HEG data in the quasar regime for the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship of @Maiolino:2007bh for $N_{\rm\,H}^{\rm\,T}\geq 10^{23.7}\rm\,cm^{-2}$. The values of the slope expected using the relationship of @Hasinger:2008ve are steeper than the observed value for the whole range of column densities considered, while the flattening of the relationship of @Burlon:2011cr at high luminosities results in a slope of $B\sim0$ along the whole range of $N_{\rm\,H}^{\rm\,T}$, consistent within 1$\sigma$ with the observations. The intercepts obtained using the relationship of @Maiolino:2007bh are consistent with the observed value for $22.5 \leq \log N_{\rm\,H}^{\rm\,T} \leq 23.1$ (right panel of Fig.\[fig:interc\_sy\_2\]), with no overlap with the values of $N_{\rm\,H}^{\rm\,T}$ needed by the slope. We obtained intercepts that are consistent with the observations for $ 22.7 \lesssim \log N_{\rm\,H}^{\rm\,T}\lesssim 23.4$ and $23.2 \lesssim \log N_{\rm\,H}^{\rm\,T}\lesssim 24.2$ for the relation of @Hasinger:2008ve and @Burlon:2011cr, respectively. Thus only the hard X-ray $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship @Burlon:2011cr is able to explain, for average values of the equatorial column density of the torus in the range $23.2 \lesssim \log N_{\rm\,H}^{\rm\,T}\lesssim 24.2$, at the same time both the intercept and the slope of the X-ray Baldwin effect at high luminosities.
Discussion {#Sec:discussion}
==========
Since its discovery about 20 years ago, the existence of the X-ray Baldwin effect has been confirmed by several works performed with the highest spectral resolution available at X-rays. So far, several possible explanations have been proposed. @Jiang:2006vn argued that the observed anti-correlation could be related to the delay of the reprocessed radiation with respect to the primary continuum. The response of the circumnuclear material to the irradiated flux is not simultaneous, and one should always take this effect into account when performing studies of reprocessed features as the iron K$\alpha$ line or the Compton hump. @Shu:2010zr [@Shu:2012fk] have shown that averaging the values of $L_{\mathrm{X}}$ and $EW$ for all the observations of each source results in a significantly flattened anti-correlation. However, by itself variability fails to fully account for the observed correlation [@Shu:2012fk].
Explaining the X-ray Baldwin effect with a luminosity-dependent covering factor of the torus
--------------------------------------------------------------------------------------------
A mechanism often invoked to explain the X-ray Baldwin effect is the decrease of the covering factor of the molecular torus with the luminosity (e.g., [@Page:2004kx], [@Bianchi:2007vn]). The decrease of the fraction of obscured sources with the luminosity has been reported by several works performed at different wavelengths in the last decade, although some discordant results have been presented (e.g., [@Dwelly:2006fk], [@Lawrence:2010uq]). In this work we have showed that the covering factor-luminosity relationships obtained in the medium and hard X-ray band can explain well the X-ray Baldwin effect in the $10^{42}-10^{44.2}\rm\,erg\,s^{-1}$ luminosity range. In particular our simulations show that it is possible to reproduce the slope of the X-ray Baldwin effect with luminosity-dependent unification for average equatorial column densities of the torus of $\log N_{\rm\,H}^{\rm\,T}\gtrsim 23.1$, and both the slope and the intercept for $\log N_{\rm\,H}^{\rm\,T}\simeq 23.2$ (Fig.\[fig:slope\_NH\]). In the same luminosity range the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ IR relationship of @Maiolino:2007bh can explain the X-ray Baldwin effect for $22.9 \lesssim \log N_{\rm\,H}^{\rm\,T}\lesssim 23$. In the quasar regime we have shown that, while the medium X-ray $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship of @Hasinger:2008ve cannot explain the observations (Fig.\[fig:interc\_sy\_2\]), the slope obtained by the IR $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship is consistent with our fit to the [*Chandra*]{}/HEG data in the same luminosity band. However, as for the latter relationship the range of values of $N_{\rm\,H}^{\rm\,T}$ required to explain the slope and the intercept do not overlap, it cannot be considered as a likely explanation. The hard X-ray $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship of @Burlon:2011cr is flat above $L_{\mathrm{X}}^{\mathrm{Q}}$, and would thus produce a constant $EW$ of the iron K$\alpha$ line and a slope of $B\sim0$ for the whole range of $N_{\rm\,H}^{\rm\,T}$, consistent within $1\sigma$ with the value found using [*Chandra*]{}/HEG data. The intercept we found using this relation is also consistent with the observational value for a large range of average equatorial column densities of the torus ($23.2 \lesssim \log N_{\rm\,H}^{\rm\,T}\lesssim 24.2$). The fraction of obscured AGN is not well constrained at high-luminosity, thus any trend with the luminosity between that of @Burlon:2011cr and that of @Hasinger:2008ve would be able to reproduce the observed slope. The relation of @Hasinger:2008ve produces negative values of $f_{\rm\,obs}$ for $\log L_{\mathrm{X}}\gtrsim 45.8$, thus a flattening of the decline is expected below this luminosity. From Fig.7 (left panel) in @Hasinger:2008ve one can see that indeed above $\log L_{\mathrm{X}}\simeq 45$ the value of $f_{\rm\,obs}$ appears to be constant, similarly to what has been found at hard X-rays. A flattening of $f_{\mathrm{obs}}$ in the quasar regime would also be expected when considering the large amount of accreting material needed to power the AGN at these luminosities. It must however be stressed that at high luminosities the X-ray Baldwin effect is not well studied, and the [*Chandra*]{}/HEG sample we used in this luminosity range is small, not allowing us to reach a firm conclusion on the variation of the iron K$\alpha$ $EW$ with the luminosity. Possible evidence of a flattening of the X-ray Baldwin effect in the quasar regime has been recently found by @Krumpe:2010vn.
On the differences between X-ray and IR half-opening angle-luminosity relationships
-----------------------------------------------------------------------------------
The main difference between the different $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationships we used is that those obtained in the X-rays are based on direct observations of the absorbing material in the line of sight, while the one of @Maiolino:2007bh is extrapolated from the ratio of the thermal infrared emission to the primary AGN continuum. The $\theta_{\mathrm{OA}}-L_{\lambda}(5100\AA)$ relationship obtained by @Maiolino:2007bh was converted in $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ using the $L_{\,2\rm\,keV}-L{\lambda}(2500\AA)$ relation obtained by @Steffen:2006uq, and then converted in the $L_{\mathrm{X}}-L_{\lambda}(5100\AA)$ relation assuming the optical-UV spectral slope obtained by @Vanden-Berk:2001kx. All this is likely to introduce some error in the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ obtained by their IR work. It has been argued by @Maiolino:2007bh that the larger normalization of the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relation they found is related to the fact that medium X-ray surveys such as that of @Hasinger:2008ve are likely to miss a certain fraction of heavily obscured objects, which can instead be detected in the IR. This is due to the fact that for Compton-thick AGN at energies $\lesssim 10\rm\,keV$ most of the X-ray emission is depleted. However, hard X-ray surveys as that of @Burlon:2011cr, which are much less biased by absorption, have found a normalization of the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relation consistent with that obtained at lower X-ray energies (see Fig.\[fig:tor\_CFL1\]). Our results also show that the X-ray Baldwin effect can be explained by a Compton-thin torus, and that larger values of $f_{\mathrm{obs}}$ would imply even lower average values of $N_{\rm\,H}^{\rm\,T}$, so that missing heavily obscured objects would not significantly affect our results. X-rays are also probably better suited to probe the material responsible for the iron K$\alpha$ line emission. The line can in fact be emitted by both gas and dust, and while the IR can probe only the latter, X-rays are able to infer the amount of both gas and dust. We thus conclude that for the purpose of our study the X-ray $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relations are better suited than those extrapolated from IR observations.
It could be argued that the value of the hydrogen column density commonly used to determine the fraction of obscured objects (and thus as an indicator of the presence of the torus) is $N_{\rm\,H}\sim10^{22}\rm\,cm^{-2}$, while the torus is believed to have larger values of the equatorial column density, and that this obscuration might be related to the presence of dust lanes or molecular structures in the host galaxy (e.g., [@Matt:2000kx]). However, @Bianchi:2009fk, using the results of @Della-Ceca:2008fk, have showed that a similar decrease of $f_{\mathrm{obs}}$ with the luminosity is obtained when setting this threshold to a larger value of line-of-sight column density. They found that the fraction of CT objects decreases with the luminosity as $f_{CT}\propto L^{-0.22}$, similarly to what is found for $f_{\mathrm{obs}}$ in the X-rays (e.g., [@Hasinger:2008ve], see Eq.\[Eq:Hasinger\]). A similar result was obtained by @Fiore:2009fk from a [*Chandra*]{} and [*Spitzer*]{} study of AGN in the COSMOS field. This implies that the contribution of dust lanes or of galactic molecular structures to the observed $N_{\rm\,H}$ does not affect significantly the $f_{\mathrm{obs}}-L_{\mathrm{X}}$ relationships obtained in the X-rays.
The equatorial column density of the torus
------------------------------------------
![Iron K$\alpha$ $EW$ versus X-ray luminosities and predicted trends obtained for different values of the equatorial column density of the torus. The points are the values of the equivalent width of the iron K$\alpha$ line reported by @Shu:2010zr, obtained by averaging multiple [*Chandra*]{}/HEG observations of AGN. The two blue dash-dotted lines are the fits to our simulations of the X-ray Baldwin effect using the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship of @Hasinger:2008ve for the Seyfert regime ($\log EW=1.01-0.17\log L_{\mathrm{X,44}} $ for $\log N_{\rm\,H}^{\rm\,T}=22.5$, and $\log EW=2.05-0.08\log L_{\mathrm{X,44}}$ for $\log N_{\rm\,H}^{\rm\,T}=25$). The blue long-dashed lines represent the $EW-L_{\mathrm{X}}$ relations obtained in the quasar regime ($\log EW=1.17 -0.73\log L_{\mathrm{X,44}} $ for $\log N_{\rm\,H}^{\rm\,T}=22.5$ and $\log EW=2.14-0.39\log L_{\mathrm{X,44}}$ for $\log N_{\rm\,H}^{\rm\,T}=25$) using the relationship of @Hasinger:2008ve, while the green dashed lines represent those obtained using the relationship of @Burlon:2011cr. The intercepts obtained in the quasar regime have been modified in order to match those obtained at lower luminosities.[]{data-label="fig:Chandra_model"}](Shu_EW_L_qso_sy.ps){width="9cm"}
The distribution of values of the equatorial column density of the torus in AGN is still poorly constrained. X-ray observation can in fact infer solely the obscuration in the line of sight, and only studies of the reprocessed X-ray emission performed using physical torus models like those of @Ikeda:2009nx, @Brightman:2011oq and @Murphy:2009uq can help to deduce the value of the equatorial column density of the torus. However, this kind of studies are still very scarce (e.g., [@Rivers:2011fk; @Brightman:2012fk]), besides being largely geometry-dependent. Studies of AGN in the mid-IR band performed using the clumpy torus formalism of @Nenkova:2008kx have shown that the number of clouds along the equator is $N_0\sim 5-10$ [@Mor:2009fk]. If as reported by @Mor:2009fk each cloud of the torus has an optical depth of $\tau_V\sim 30-100$ (i.e., $\log N_{\rm\,H}^{\rm\,T}\sim 22-23$), the equatorial column density of the torus is expected to be $\log N_{\rm\,H}^{\rm\,T}\sim 22.5-24$, in agreement with our results. It must however be noticed that the value of the intercept of the X-ray Baldwin effect obtained by the simulations is strongly dependent on our assumptions on the metallicity of the torus. This implies that our constraints of the average value of $N_{\rm\,H}^{\rm\,T}$ are also tightly related to the choice of the metallicity: lower values of the metallicity would lead to larger values of $N_{\rm\,H}^{\rm\,T}$. To study this effect we repeated our study in the Seyfert regime using half-solar metallicities for the reflection model. We obtained that in this scenario, in order to explain the X-ray Baldwin effect, one needs values of the equatorial column density of the torus about two times larger than for the solar-metallicity case ($23.5\lesssim \log N_{\rm\,H}^{\rm\,T} \lesssim 23.7$). To have a Compton-thick torus one would then need values of the metallicity of $Z\lesssim 0.2\,Z_{\,\sun}$. A Compton-thick torus has often been invoked to explain the Compton hump observed in the spectrum of many unobscured AGN (e.g., [@Bianchi:2004fk]). It is still unclear however which fraction of the Compton hump is produced in the distant reflector and which in the accretion flow. From our study we have found that in unobscured objects the X-ray Baldwin effect can be explained by a luminosity-dependent covering factor of the torus for an average value of the equatorial column density of $\log N_{\rm\,H}^{\rm\,T}\simeq 23.2$. This value is lower than the line-of-sight $N_{\rm\,H}$ of many Seyfert2s (e.g., [@Ricci:2011zr; @Burlon:2011cr]), and it might be related either to the geometry we adopted or to presence of objects with sub-solar metallicities. In particular, due to the constant reflection angle relative to a local normal in any point of the reflecting surface, the spherical-toroidal geometry produces larger values of $EW$ (and thus larger values of the intercept) with respect to a toroidal structure [@Murphy:2009uq].
It is possible that there exists a wide spread of equatorial column densities of the torus, thus one could envisage that this, together with the different values of $\theta_{\,\mathrm{i}}$, would introduce the scatter observed in the anti-correlation. In Fig.\[fig:Chandra\_model\] we show the X-ray Baldwin effect obtained from our simulations using the relationship of @Hasinger:2008ve for $\log N_{\rm\,H}^{\rm\,T}=22.5$ and $\log N_{\rm\,H}^{\rm\,T}=25$, in both the Seyfert and the quasar regime, together with the time-averaged [*Chandra*]{}/HEG data of @Shu:2010zr. From the figure it is evident that all the data are well within the range expected from our simulations, both in the Seyfert and in the quasar regime. It is important to remark that we do not know whether there exists a relation between the equatorial column density of the torus and the AGN luminosity. However, we have shown that the variation of the covering factor of the torus with the luminosity alone can fully explain the observed trend, so that no additional luminosity-dependent physical parameter is needed.
Luminosity-dependent unification of AGN
---------------------------------------
The relation of $f_{\mathrm{obs}}$ with the luminosity might be connected to the increase of the inner radius of the torus with the luminosity due to dust sublimation. Both near-IR reverberation [@Suganuma:2006fk] and mid-IR interferometric [@Tristram:2011uq] studies have confirmed that the inner radius of the molecular torus increases with the luminosity as $R_{\mathrm{in}}\propto L^{0.5}$. Considering the geometry of Fig.\[fig:tor\_CFL1\], the fraction of obscured objects (see Eq.\[Eq:fractioncos\]) would be related to ratio of the height to the inner radius of the torus ($H/R$) by $$f_{\mathrm{obs}}\simeq\frac{H}{R}\sqrt{\frac{1}{1+(H/R)^2}}.$$ For values of $H/R\lesssim 1$, $f_{\mathrm{obs}}\propto H/R$. In the original formulation of the [*receding torus*]{} model [@Lawrence:1991vn], the height $H$ was considered to be constant. Assuming that $R$ has the same luminosity-dependence as $R_{\mathrm{in}}$, this would lead to $f_{\mathrm{obs}}\propto L^{-0.5}$. This has been shown to be inconsistent with the recent observations, which point towards a flatter slope ($f_{\mathrm{obs}}\propto L^{-0.25}$, e.g., [@Hasinger:2008ve], see Eq.\[Eq:Hasinger\]), and which would imply $H\propto L^{0.25}$. @Honig:2007kx have shown that in the frame of [*radiation-limited clumpy dust torus*]{} model, one would obtain $H\propto L^{0.25}$, in agreement with the observations. The decrease of the covering factor with luminosity might also have important implications on the AGN dichotomy. It has been shown that Seyfert 2s appear to have on average lower luminosities and lower Eddington ratios (e.g., [@Beckmann:2009fk], [@Ricci:2011zr]) than Seyfert1s and Seyfert1.5s, which suggests that they have on average a torus with a larger covering factor. This idea is also supported by the fact that the luminosity function of type-II AGN has been found to peak at lower luminosities than that of type-I AGN ([@Della-Ceca:2008fk], [@Burlon:2011cr]), and by the results obtained by the recent mid-IR work of @Ramos-Almeida:2011fk ([-@Ramos-Almeida:2011fk], see also [@Elitzur:2012vn]). A difference in the torus covering factor distribution between different types of AGN would also explain why the average hard X-ray spectrum of Compton-thin Seyfert2s shows a larger reflection component than that of Seyfert1s and Seyfert1.5s [@Ricci:2011zr]. A similar result was found by @Brightman:2012fk: studying high-redshift AGN in the [*Chandra*]{} Deep Field South they found that more obscured objects appear to have tori with larger covering factors, although they did not find a clear luminosity dependence.
Summary and conclusions {#Sec:summary}
=======================
In this work we have studied the hypothesis that the X-ray Baldwin effect is related to the decrease of the covering factor of the torus with the luminosity. We have used the physical torus models of @Ikeda:2009nx and @Brightman:2011oq to account for the reprocessed X-ray radiation, and the values of the fraction of obscured sources obtained by recent surveys in the X-rays and in the IR as a proxy of the covering factor of the torus. Our simulations show that the variation of the covering factor of the torus with the luminosity can explain the X-ray Baldwin effect. In the Seyfert regime ($L_{\mathrm{X}}\leq 10^{44.2}\rm\,erg\,s^{-1}$), the observed $EW-L_{\mathrm{X}}$ trend can be exactly (both in slope and intercept) reproduced by an average value of the equatorial column density of the torus of $\log N_{\rm\,H}^{\rm\,T}\simeq 23.2$, while a slope consistent with the observations is obtained for a larger range of column densities ($\log N_{\rm\,H}^{\rm\,T}\gtrsim 23.1$). At higher luminosities ($L_{\mathrm{X}}> 10^{44.2}\rm\,erg\,s^{-1}$) the situation is less clear due to the small number of high-quality observations available. Moreover, it is not clear whether in the quasar regime $f_{\mathrm{obs}}$ still decreases with the luminosity similarly to what is found in the Seyfert regime (as shown by [@Hasinger:2008ve]), or is constant (as found by [@Burlon:2011cr]). A flattening of the $\theta_{\mathrm{OA}}-L_{\mathrm{X}}$ relationship would be able to explain the observations at high luminosities (for $23.2 \lesssim \log N_{\rm\,H}^{\rm\,T}\lesssim 24.2$), and it might naturally arise from the large amount of accreting mass needed to power these luminous quasars.
In the next years [*ASTRO-H*]{} [@Takahashi:2010uq], with its high energy-resolution calorimeter SXS, will allow to constrain the origin of the narrow component of the iron K$\alpha$ line, being able to separate even better than [*Chandra*]{}/HEG the narrow core coming from the torus from the flux emitted closer to the central engine. [*ASTRO-H*]{} will also be able to probe the narrow iron K$\alpha$ line in the quasar regime, allowing to understand the behavior of the X-ray Baldwin effect at high luminosities.
We thank the anonymous referee for his/her comments that helped improving the paper. We thank XinWen Shu and Rivay Mor for providing us useful details about their work, Chin Shin Chang and Poshak Gandhi for their comments on the manuscript. CR thanks the Sherpa group and IPAG for hospitality during his stay in Grenoble. CR is a Fellow of the Japan Society for the Promotion of Science (JSPS).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Inference in general Ising models is difficult, due to high treewidth making tree-based algorithms intractable. Moreover, when interactions are strong, Gibbs sampling may take exponential time to converge to the stationary distribution. We present an algorithm to project Ising model parameters onto a parameter set that is guaranteed to be fast mixing, under several divergences. We find that Gibbs sampling using the projected parameters is more accurate than with the original parameters when interaction strengths are strong and when limited time is available for sampling.'
author:
- |
Justin Domke\
NICTA, The Australian National University\
`[email protected]`
- |
**Xianghang Liu**\
NICTA, The University of New South Wales\
`[email protected]`
bibliography:
- 'thebib.bib'
- 'bibliography\_pamipaper.bib'
- 'bibliography\_singleloop.bib'
title: Projecting Ising Model Parameters for Fast Mixing
---
Introduction
============
High-treewidth graphical models typically yield distributions where exact inference is intractable. To cope with this, one often makes an approximation based on a tractable model. For example, given some intractable distribution $q$, mean-field inference [@AMeanFieldTheoryLearningAlgorithm] attempts to minimize $KL(p||q)$ over $p\in\text{TRACT}$, where $\text{TRACT}$ is the set of fully-factorized distributions. Similarly, structured mean-field minimizes $ $the KL-divergence, but allows $\text{TRACT}$ to be the set of distributions that obey some tree [@ExploitingTractableSubstructures] or a non-overlapping clustered [@AGeneralizedMeanField] structure. In different ways, loopy belief propagation [@Yedidia05constructingfree] and tree-reweighted belief propagation [@Wainwright_anewclass] also make use of tree-based approximations, while Globerson and Jaakkola [@ApproximateInferenceUsingPlanarGraph] provide an approximate inference method based on exact inference in planar graphs with zero field.
In this paper, we explore an alternative notion of a “tractable” model. These are “fast mixing” models, or distributions that, while they may be high-treewidth, have parameter-space conditions guaranteeing that Gibbs sampling will quickly converge to the stationary distribution. While the precise form of the parameter space conditions is slightly technical (Sections \[sec:Background\]-\[sec:Mixing-Time-Bounds\]), informally, it is simply that interaction strengths between neighboring variables are not too strong.
In the context of the Ising model, we attempt to use these models in the most basic way possible– by taking an arbitrary (slow-mixing) set of parameters, projecting onto the fast-mixing set, using four different divergences. First, we show how to project in the Euclidean norm, by iteratively thresholding a singular value decomposition (Theorem \[thm:projection\_theorem-main\]). Secondly, we experiment with projecting using the “zero-avoiding” divergence $KL(q||p)$. Since this requires taking (intractable) expectations with respect to $q$, it is of only theoretical interest. Third, we suggest a novel “piecewise” approximation of the KL divergence, where one drops edges from both $q$ and $p$ until a low-treewidth graph remains where the exact KL divergence can be calculated. Experimentally, this does not perform as well as the true KL-divergence, but is easy to evaluate. Fourth, we consider the “zero forcing” divergence $KL(q||p)$. Since this requires expectations with respect to $p$, which is constrained to be fast-mixing, it can be approximated by Gibbs sampling, and the divergence can be minimized through stochastic approximation. This can be seen as a generalization of mean-field where the set of approximating distributions is expanded from fully-factorized to fast-mixing.
Background\[sec:Background\]
============================
The literature on mixing times in Markov chains is extensive, including a recent textbook [@MarkovChainsAndMixingTimes]. The presentation in the rest of this section is based on that of Dyer et al. [@MatrixNormsAndRapidMixingForSpinSystems].
Given a distribution $p(x)$, one will often wish to draw samples from it. While in certain cases (e.g. the Normal distribution) one can obtain exact samples, for Markov random fields (MRFs), one must generally resort to iterative Markov chain Monte Carlo (MCMC) methods that obtain a sample asymptotically. In this paper, we consider the classic Gibbs sampling method [@StochasticRelaxationGibbsDistributions], where one starts with some configuration $x$, and repeatedly picks a node $i$, and samples $x_{i}$ from $p(x_{i}|x_{-i})$. Under mild conditions, this can be shown to sample from a distribution that converges to $p$ as $t\rightarrow\infty$.
It is common to use more sophisticated methods such as block Gibbs sampling, the Swendsen-Wang algorithm [@SwendsenWang], or tree sampling [@FromFieldsToTrees]. In principle, each algorithm could have unique parameter-space conditions under which it is fast mixing. Here, we focus on the univariate case for simplicity and because fast mixing of univariate Gibbs is sufficient for fast mixing of some other methods [@ExtraUpdates].
Given two finite distributions $p$ and $q$, the **total variation distance** $||\cdot||_{TV}$ is $$||p(X)-q(X)||_{TV}=\frac{1}{2}\sum_{x}|p(X=x)-q(X=x)|.$$
We need a property of a distribution that can guarantee fast mixing. The **dependency** $R_{ij}$ of $x_{i}$ on $x_{j}$ is defined by considering two configurations $x$ and $x'$, and measuring how much the conditional distribution of $x_{i}$ can vary when $x_{k}=x'_{k}$ for all $k\not=j$.
Given a distribution $p$, the dependency matrix $R$ is defined by
$$R_{ij}=\max_{x,x':x_{-j}=x_{-j}'}||p(X_{i}|x_{-i})-p(X_{i}|x_{-i}')||_{TV}.$$
Given some threshold $\epsilon$, the **mixing time** is the number of iterations needed to guarantee that the total variation distance of the Gibbs chain to the stationary distribution is less than $\epsilon$.
Suppose that $\{X^{t}\}$ denotes the sequence of random variables corresponding to running Gibbs sampling on some distribution $p$. The mixing time $\tau(\epsilon)$ is the minimum time $t$ such that the total variation distance between $X^{t}$ and the stationary distribution is at most $\epsilon$. That is, $$\begin{aligned}
\tau(\epsilon)= & \min\{t:d(t)<\epsilon\},\\
d(t)= & \max_{x}||\mathbb{P}(X^{t}|X^{0}=x)-p(X)||_{TV}.\end{aligned}$$
Unfortunately, the mixing time can be extremely long, which makes the use of Gibbs sampling delicate in practice. For example, for the two-dimensional Ising model with zero field and uniform interactions, it is known that mixing time is polynomial (in the size of the grid) when the interaction strengths are below a threshold $\beta_{c}$, and exponential for stronger interactions [@CriticalIsingMixing]. For more general distributions, such tight bounds are not generally known, but one can still derive sufficient conditions for fast mixing. The main result we will use is the following [@ASimpleConditionImplyingRapidMixing].
\[thm:mixing-time-theorem\]Consider the dependency matrix $R$ corresponding to some distribution $p(X_{1},...,X_{n})$. For Gibbs sampling with random updates, if $||R||_{2}<1,$ the mixing time is bounded by
$$\tau(\epsilon)\leq\frac{n}{1-||R||_{2}}\ln\left(\frac{n}{\epsilon}\right).$$
Roughly speaking, if the spectral norm (maximum singular value) of $R$ is less than one, rapid mixing will occur. A similar result holds in the case of systematic scan updates [@MatrixNormsAndRapidMixingForSpinSystems; @ASimpleConditionImplyingRapidMixing].
Some of the classic ways of establishing fast mixing can be seen as special cases of this. For example, the Dobrushin criterion is that $||R||_{1}<1$, which can be easier to verify in many cases, since $||R||_{1}=\max_{j}\sum_{i}|R_{ij}|$ does not require the computation of singular values. However, for symmetric matrices, it can be shown that $||R||_{2}\leq||R||_{1}$, meaning the above result is tighter.
Mixing Time Bounds\[sec:Mixing-Time-Bounds\]
============================================
For variables $x_{i}\in\{-1,+1\},$ an Ising model is of the form $$p(x)=\exp\left(\sum_{i,j}\beta_{ij}x_{i}x_{j}+\sum_{i}\alpha_{i}x_{i}-A(\beta,\alpha)\right),$$ where $\beta_{ij}$ is the interaction strength between variables $i$ and $j$, $\alpha_{i}$ is the “field” for variable $i$, and $A$ ensures normalization. This can be seen as a member of the exponential family $p(x)=\exp\left(\theta\cdot f(x)-A(\theta)\right),$ where $f(x)=\{x_{i}x_{j}\forall(i,j)\}\cup\{x_{i}\forall i\}$ and $\theta$ contains both $\beta$ and $\alpha$.
\[lem:ising-dependency\]For an Ising model, the dependency matrix is bounded by $$R_{ij}\leq\tanh|\beta_{ij}|\leq|\beta_{ij}|$$
Hayes [@ASimpleConditionImplyingRapidMixing] proves this for the case of constant $\beta$ and zero-field, but simple modifications to the proof can give this result.
Thus, to summarize, an Ising model can be guaranteed to be fast mixing if the spectral norm of the absolute value of interactions terms is less than one.
Projection\[sec:Projection\]
============================
In this section, we imagine that we have some set of parameters $\theta$, not necessarily fast mixing, and would like to obtain another set of parameters $\psi$ which are as close as possible to $\theta$, but guaranteed to be fast mixing. This section derives a projection in the Euclidean norm, while Section \[sec:Divergences\] will build on this to consider other divergence measures.
We will use the following standard result that states that given a matrix $A$, the closest matrix with a maximum spectral norm can be obtained by thresholding the singular values.
\[thm:dense-projection\]If $A$ has a singular value decomposition $A=USV^{T}$, and $||\cdot||_{F}$ denotes the Frobenius norm, then $B=\underset{B:||B||_{2}\leq c}{\arg\text{}\min}||A-B||_{F}$ can be obtained as $B=US'V^{T},$ where $S_{ii}^{'}=\min(S_{ii},c^{2}).$
We denote this projection by $B=\Pi_{c}[A]$. This is close to providing an algorithm for obtaining the closest set of Ising model parameters that obey a given spectral norm constraint. However, there are two issues. First, in general, even if $A$ is sparse, the projected matrix $B$ will be dense, meaning that projecting will destroy a sparse graph structure. Second, this result constrains the spectral norm of $B$ itself, rather than $R=|B|$, which is what needs to be controlled. The theorem below provides a dual method that fixed these issues.
Here, we take some matrix $Z$ that corresponds to the graph structure, by setting $Z_{ij}=0$ if $(i,j)$ is an edge, and $Z_{ij}=1$ otherwise. Then, enforcing that $B$ obeys the graph structure is equivalent to enforcing that $Z_{ij}B_{ij}=0$ for all $(i,j)$. Thus, finding the closest set of parameters $B$ is equivalent to solving
$$\begin{aligned}
\min_{B,D} & & ||A-B||_{F}\text{ subject to }||D||_{2}\leq c,\,\, Z_{ij}D_{ij}=0,\,\, D=|B|.\label{eq:sparse-projection-minimization-singleline}\end{aligned}$$
We find it convenient to solve this minimization by performing some manipulations, and deriving a dual. The proof of this theorem is provided in the appendix. To accomplish the maximization of $g$ over $M$ and $ $$\Lambda$, we use LBFGS-B [@ALimitedMemoryAlgorithmFroBoundConstrainedOptimization], with bound constraints used to enforce that $M\geq0$.
The following theorem uses the “triple dot product” notation of $A\cdot B\cdot C=\sum_{ij}A_{ij}B_{ij}C_{ij}$.
\[thm:projection\_theorem-main\]Define $R=|A|$. The minimization in Eq. \[eq:sparse-projection-minimization-singleline\] is equivalent to the problem of $\max_{M\geq0,\Lambda}g(\Lambda,M)$, where the objective and gradient of $g$ are, for $D(\Lambda,M)=\Pi_{c}[R+M-\Lambda\odot Z],$
$$\begin{aligned}
g(\Lambda,M) & =\frac{1}{2}||D(\Lambda,M)-R||_{F}^{2}+\Lambda\cdot Z\cdot D(\Lambda,M) - M \cdot D(\Lambda,M) \label{eq:sparse-projection-dual-obj-1-1}\\
\frac{dg}{d\Lambda} & =Z\odot D(\Lambda,M)\label{eq:sparse-projection-dual-grad-1-1}\\
\frac{dg}{dM} & = -D(\Lambda,M).\label{eq:sparse-projection-dual-grad-2-1}\end{aligned}$$
Divergences\[sec:Divergences\]
==============================
Again, we would like to find a parameter vector $\psi$ that is close to a given vector $\theta$, but is guaranteed to be fast mixing, but with several notions of “closeness” that vary in terms of accuracy and computational convenience. Formally, if $\Psi$ is the set of parameters that we can guarantee to be fast mixing, and $D(\theta,\psi)$ is a divergence between $\theta$ and $\psi$, then we would like to solve $$\arg\min_{\psi\in\Psi}D(\theta,\psi).\label{eq:general_projection}$$
As we will see, in selecting $D$ there appears to be something of a trade-off between the quality of the approximation, and the ease of computing the projection in Eq. \[eq:general\_projection\].
In this section, we work with the generic exponential family representation
$$p(x;\theta)=\exp(\theta\cdot f(x)-A(\theta)).$$ We use $\mu$ to denote the mean value of $f$. By a standard result, this is equal to the gradient of $A$, i.e.
$$\mu(\theta)=\sum_{x}p(x;\theta)f(x)=\nabla A(\theta).$$
Euclidean Distance
------------------
The simplest divergence is simply the $l_{2}$ distance between the parameter vectors, $D(\theta,\psi)=||\theta-\psi||_{2}$. For the Ising model, Theorem \[thm:projection\_theorem-main\] provides a method to compute the projection $\arg\min_{\psi\in\Psi}||\theta-\psi||_{2}.$ While simple, this has no obvious probabilistic interpretation, and other divergences perform better in the experiments below.
However, it also forms the basis of our projected gradient descent strategy for computing the projection in Eq. \[eq:general\_projection\] under more general divergences $D$. Specifically, we will do this by iterating
1. $\psi'\leftarrow\psi-\lambda\frac{d}{d\psi}D(\theta,\psi)$
2. $\psi\leftarrow\arg\min_{\psi\in\Psi}||\psi'-\psi||_{2}$
for some step-size $\lambda$. In some cases, $dD/d\psi$ can be calculated exactly, and this is simply projected gradient descent. In other cases, one needs to estimate $dD/d\psi$ by sampling from $\psi$. As discussed below, we do this by maintaining a “pool” of samples. In each iteration, a few Markov chain steps are applied with the current parameters, and then the gradient is estimated using them. Since the gradients estimated at each time-step are dependent, this can be seen as an instance of Ergodic Mirror Descent [@ErgodicMirrorDescent]. This guarantees convergence if the number of Markov chain steps, and the step-size $\lambda$ are both functions of the total number of optimization iterations.
KL-Divergence
-------------
Perhaps the most natural divergence to use would be the “inclusive” KL-divergence $$D(\theta,\psi)=KL(\theta||\psi)=\sum_{x}p(x;\theta)\log\frac{p(x;\theta)}{p(x;\psi)}.\label{eq:zero-avoiding-KL}$$
This has the “zero-avoiding” property [@DivergenceMeasuresAndMessagePassing] that $\psi$ will tend to assign some probability to all configurations that $\theta$ assigns nonzero probability to. It is easy to show that the derivative is $$\frac{dD(\theta,\psi)}{d\psi}=\mu(\psi)-\mu(\theta),\label{eq:KL-divergence-gradient}$$ where $\mu_{\theta}=\mathbb{E}_{\theta}[f(X)].$ Unfortunately, this requires inference with respect to both the parameter vectors $\theta$ and $\psi$. Since $\psi$ will be enforced to be fast-mixing during optimization, one could approximate $\mu(\psi)$ by sampling. However, $\theta$ is presumed to be slow-mixing, making $\mu(\theta)$ difficult to compute. Thus, this divergence is only practical on low-treewidth “toy” graphs.
Piecewise KL-Divergences
------------------------
Inspired by the piecewise likelihood [@PiecewiseTraining] and likelihood approximations based on mixtures of trees [@SpanningTreeApproximations], we seek tractable approximations of the KL-divergence based on tractable subgraphs. Our motivation is the the following: if $\theta$ and $\psi$ define the same distribution, then if a certain set of edges are removed from both, they should continue to define the same distribution[^1]. Thus, given some graph $T$, we define the “projection” $\theta(T)$ onto the tree such by setting all edge parameters to zero if not part of $T$. Then, given a set of graphs $T$, the piecewise KL-divergence is $$D(\theta,\psi)=\max_{T}KL(\theta(T)||\psi(T)).$$ Computing the derivative of this divergence is not hard– one simply computes the KL-divergence for each graph, and uses the gradient as in Eq. \[eq:KL-divergence-gradient\] for the maximizing graph.
There is some flexibility of selecting the graphs $T$. In the simplest case, one could simply select a set of trees (assuring that each edge is covered by one tree), which makes it easy to compute the KL-divergence on each tree using the sum-product algorithm. We will also experiment with selecting low-treewidth graphs, where exact inference can take place using the junction tree algorithm.
Reversed KL-Divergence
----------------------
We also consider the “zero-forcing” KL-divergence $$D(\theta,\psi)=KL(\psi||\theta)=\sum_{x}p(x;\psi)\log\frac{p(x;\psi)}{p(x;\theta)}.$$
The divergence $D(\theta,\psi)=KL(\psi||\theta)$ has the gradient $$\frac{d}{d\psi}D(\theta,\psi)=\sum_{x}p(x;\psi)(\psi-\theta)\cdot f(x)\left(f(x)-\mu(\psi)\right).$$
Arguably, using this divergence is inferior to the “zero-avoiding” KL-divergence. For example, since the parameters $\psi$ may fail to put significant probability at configurations where $\theta$ does, using importance sampling to reweight samples from $\psi$ to estimate expectations with respect to $\theta$ could have high variance Further, it can be non-convex with respect to $\psi$. Nevertheless, it often work well in practice. Minimizing this divergence under the constraint that the dependency matrix $R$ corresponding to $\psi$ have a limited spectral norm is closely related to naive mean-field, which can be seen as a degenerate case where one constrains $R$ to have zero norm.
This is easier to work with than the “zero-avoiding” KL-divergence in Eq. \[eq:zero-avoiding-KL\] since it involves taking expectations with respect to $\psi$, rather than $\theta$: since $\psi$ is enforced to be fast-mixing, these expectations can be approximated by sampling. Specifically, suppose that one has generated a set of samples $x^{1},...,x^{K}$ using the current parameters $\psi$. Then, one can first approximate the marginals by $\hat{\mu}=\frac{1}{K}\sum_{k=1}^{K}f(x^{k}),$ and then approximate the gradient by $$\hat{g}=\frac{1}{K}\sum_{k=1}^{K}\left((\psi-\theta)\cdot f(x^{k})\right)\left(f(x^{k})-\hat{\mu}\right).\label{eq:SGD_gradient_estimate}$$
It is a standard result that if two estimators are unbiased and independent, the product of the two estimators will also be unbiased. Thus, if one used separate sets of perfect samples to estimate $\hat{\mu}$ and $\hat{g}$, then $\hat{g}$ would be an unbiased estimator of $dD/d\psi$. In practice, of course, we generate the samples by Gibbs sampling, so they are not quite perfect. We find in practice that using the same set of samples twice makes makes little difference, and do so in the experiments.
![The mean error of estimated univariate marginals on 8x8 grids (top row) and low-density random graphs (bottom row), comparing 30k iterations of Gibbs sampling after projection to variational methods. To approximate the computational effort of projection (Table \[tab:Running-Times\]), sampling on the original parameters with 250k iterations is also included as a lower curve. (Full results in appendix.)](esacc_plot/grid_mixed "fig:"){width="44.50000%"}![The mean error of estimated univariate marginals on 8x8 grids (top row) and low-density random graphs (bottom row), comparing 30k iterations of Gibbs sampling after projection to variational methods. To approximate the computational effort of projection (Table \[tab:Running-Times\]), sampling on the original parameters with 250k iterations is also included as a lower curve. (Full results in appendix.)](esacc_plot/grid_attractive "fig:"){width="44.50000%"}
![The mean error of estimated univariate marginals on 8x8 grids (top row) and low-density random graphs (bottom row), comparing 30k iterations of Gibbs sampling after projection to variational methods. To approximate the computational effort of projection (Table \[tab:Running-Times\]), sampling on the original parameters with 250k iterations is also included as a lower curve. (Full results in appendix.)](esacc_plot/rg_a1_0\lyxdot 3 "fig:"){width="44.50000%"}![The mean error of estimated univariate marginals on 8x8 grids (top row) and low-density random graphs (bottom row), comparing 30k iterations of Gibbs sampling after projection to variational methods. To approximate the computational effort of projection (Table \[tab:Running-Times\]), sampling on the original parameters with 250k iterations is also included as a lower curve. (Full results in appendix.)](esacc_plot/rg_1_0\lyxdot 3 "fig:"){width="44.50000%"}
Experiments
===========
Our experimental evaluation follows that of Hazan and Shashua [@ConvergentMessagePassingAlgorithms] in evaluating the accuracy of the methods using the Ising model in various configurations. In the experiments, we approximate randomly generated Ising models with rapid-mixing distributions using the projection algorithms described previously. Then, the marginals of rapid-mixing approximate distribution are compared against those of the target distributions by running a Gibbs chain on each. We calculate the mean absolute distance of the marginals as the accuracy measure, with the marginals computed via the exact junction-tree algorithm.
We evaluate projecting under the Euclidean distance (Section 5.1), the piecewise divergence (Section 5.3) (Section 5.4). On small graphs, it is possible to minimize the zero-avoiding KL-divergence $KL(\theta||\psi)$ by computing marginals using the junction-tree algorithm. However, as minimizing this KL-divergence leads to exact marginal estimates, it doesn’t provide a useful measure of marginal accuracy. Our methods are compared with four other inference algorithms, namely loopy belief-propagation (LBP), Tree-reweighted belief-propagation (TRW), Naive mean-field (MF), and Gibbs sampling on the original parameters.
LBP, MF and TRW are among the most widely applied variational methods for approximate inference. The MF algorithm uses a fully factorized distribution as the tractable family, and can be viewed as an extreme case of under the constraint of zero spectral norm. The tractable family that it uses guarantees “instant” mixing but is much more restrictive. Theoretically, Gibbs sampling on the original parameters will produce highly accurate marginals if run long enough. However, this can take exponentially long and convergence is generally hard to diagnose [@Cowles96markovchain] In contrast, Gibbs sampling on the rapid-mixing approximation is guaranteed to converge rapidly but will result in less accurate marginals asymptotically. Thus, we also include time-accuracy comparisons between these two strategies in the experiments.
Configurations
--------------
Two types of graph topologies are used: two-dimensional $8\times8$ grids and random graphs with $10$ nodes. Each edge is independently present with probability $p_{e}\in\{0.3,0.5,0.7\}$. Node parameters $\theta_{i}$ are uniformly drawn from unif$(-d_{n},d_{n})$ and we fix the field strength to $d_{n}=1.0$. Edge parameters $\theta_{ij}$ are uniformly drawn from unif$(-d_{e},d_{e})$ or unif$(0,d_{e})$ to obtain mixed or attractive interactions respectively. We generate graphs with different interaction strength $d_{e}=0,0.5,\dots,4$. All results are averaged over 50 random trials.
To calculate piecewise divergences, it remains to specify the set of subgraphs $T$. It can be any tractable subgraph of the original distribution. For the grids, one straightforward choice is to use the horizontal and the vertical chains as subgraphs. We also test with chains of treewidth 2. For random graphs, we use.
For each original or approximate distribution, a single chain of Gibbs sampling is run on the final parameters, and marginals are estimated from the samples drawn.
In Section 3.2, we show that for Ising models, a sufficient condition for rapid-mixing is the spectral norm of pairwise weight matrix is less than 1.0. However, we find in practice using a spectral norm bound of $2.5$ instead of $1.0$ can still preserve the rapid-mixing property and gives better approximation to the original distributions. (See Section \[sec:Discussion\] for a discussion.)
{width="44.50000%"}{width="44.50000%"}
{width="44.50000%"}{width="44.50000%"}
-- -- -- -- -- -- --
-- -- -- -- -- -- --
: Running Times on various attractive graphs, showing the number of Gibbs passes and Singular Value Decompositions, as well as the amount of computation time. The random graph is based on an edge density of 0.7. Mean-Field, Loopy BP, and TRW take less than 0.01s.\[tab:Running-Times\]
Discussion\[sec:Discussion\]
============================
Inference in high-treewidth graphical models is intractable, which has motivated several classes of approximations based on tractable families. In this paper, we have proposed a new notion of “tractability”, insisting not that a graph has a fast algorithm for exact inference, but only that it obeys parameter-space conditions ensuring that Gibbs sampling will converge rapidly to the stationary distribution. For the case of Ising models, we use a simple condition that can guarantee rapid mixing, namely that the spectral norm of the matrix of interaction strengths is less than one.
Given an intractable set of parameters, we consider using this approximate family by “projecting” the intractable distribution onto it under several divergences. First, we consider the Euclidean distance of parameters, and derive a dual algorithm to solve the projection, based on an iterative thresholding of the singular value decomposition. Next, we extend this to more probabilistic divergences. Firstly, we consider a novel “piecewise” divergence, based on computing the exact KL-divergnce on several low-treewidth subgraphs. Secondly, we consider projecting onto the KL-divergence. This requires a stochastic approximation approach where one repeatedly generates samples from the model, and projects in the Euclidean norm after taking a gradient step.
We compare experimentally to Gibbs sampling on the original parameters, along with several standard variational methods. The proposed methods are more accurate than variational approximations. Given enough time, Gibbs sampling using the original parameters will always be more accurate, but with finite time, projecting onto the fast-mixing set to generally gives better results.
Future work might extend this approach to general Markov random fields. This will require two technical challenges. First, one must find a bound on the dependency matrix for general MRFs, and secondly, an algorithm is needed to project onto the fast-mixing set defined by this bound. Fast-mixing distributions might also be used for learning. E.g., if one is doing maximum likelihood learning using MCMC to estimate the likelihood gradient, it would be natural to constrain the parameters to a fast mixing set.
One weakness of the proposed approach is the apparent looseness of the spectral norm bound. For the two dimensional Ising model with no univariate terms, and a constant interaction strength $\beta$, there is a well-known threshold $\beta_{c}=\frac{1}{2}\ln(1+\sqrt{2})\approx.4407$, obtained using more advanced techniques than the spectral norm [@CriticalIsingMixing]. Roughly, for $\beta<\beta_{c}$, mixing is known to occur quickly (polynomial in the grid size) while for $\beta>\beta_{c}$, mixing is exponential. On the other hand, the spectral bound norm will be equal to one for $\beta=.25$, meaning the bound is too conservative in this case by a factor of $\beta_{c}/.25\approx1.76$. A tighter bound on when rapid mixing will occur would be more informative.
Appendix {#appendix .unnumbered}
========
Recall that we are interested in the minimization
$$\begin{aligned}
\min_{B,D} & & ||A-B||_{F}\label{eq:sparse-projection-minimization-1}\\
s.t. & & ||D||_{2}\leq c\nonumber \\
& & Z_{ij}D_{ij}=0\nonumber \\
& & D=|B|.\nonumber \end{aligned}$$
\[lem:projection-minimization-2\]If we define $R=|A|$, this is equivalent to the minimization
$$\begin{aligned}
\min_{D} & & ||R-D||_{F}\label{eq:sparse-projection-minimization-1-1}\\
s.t. & & ||D||_{2}\leq c\nonumber \\
& & Z_{ij}D_{ij}=0\nonumber \\
& & D\geq0\nonumber \end{aligned}$$
For fixed $D$, the minimum $B$ will always be achieved by $B=D\odot\text{sign}(A)$, meaning $||A-B||_{F}=||A-D\odot\text{sign}(A)||_{F}=||R-D||_{F}$.
To actually project the parameters $A=(\beta_{ij})$ corresponding to an Ising model, one first takes the absolute value $R=|A|$, and passes it as input to this minimization. After finding the minimizing argument, the new parameters are $B=D\odot\text{sign}(A)$.
\[thm:projection\_theorem-1-1\]Define $R=|A|$. The minimization in Eq. \[eq:sparse-projection-minimization-singleline\] is equivalent to the problem of $\max_{M\geq0,\Lambda}g(\Lambda,M)$, where the objective and gradient of $g$ are, for $D(\Lambda,M)=\Pi_{c}[R+M-\Lambda\odot Z],$
$$\begin{aligned}
g(\Lambda,M) & =\frac{1}{2}||D(\Lambda,M)-R||_{F}^{2}+\Lambda\cdot Z\cdot D(\Lambda,M) - M \cdot D(\Lambda,M)\label{eq:sparse-projection-dual-obj-appendix}\\
\frac{dg}{d\Lambda} & =Z\odot D(\Lambda,M)\label{eq:sparse-projection-dual-grad-1-appendix}\\
\frac{dg}{dM} & = -D(\Lambda,M).\label{eq:sparse-projection-dual-grad-2-appendix}\end{aligned}$$
The minimization in Eq. \[eq:sparse-projection-minimization-1-1\] has the Lagrangian $$\mathcal{L}(D,\Lambda,M)=\frac{1}{2}||D-R||_{F}^{2}+I[||D||_{2}\leq c]+\Lambda\cdot Z\cdot D-M\cdot D,\label{eq:sparse-projection-lagrangian}$$ where $I$ is an indicator function returning $\infty$ if $||D||_{2}>c$ and zero otherwise, $\Lambda$ is a matrix of Lagrange multipliers enforcing that $Z_{ij}D_{ij}=0$, and $M$ is a matrix of Lagrange multipliers enforcing that $D\geq0$.
Standard duality theory states that Eq. \[eq:sparse-projection-minimization-1-1\] is equivalent to the saddle-point problem $\max_{M\geq0,\Lambda}\min_{D}\mathcal{L}(D,\Lambda,M)$. So, we are interested in evaluating $g(\Lambda,M)=\min_{D}\mathcal{L}(D,\Lambda,M)$ for fixed $\Lambda$ and $M$. Some algebra gives $$\begin{gathered}
\arg\min_{D}\mathcal{L}(D,\Lambda,M)\\
=\arg\min_{D}\frac{1}{2}||D-R||_{F}^{2}+\Lambda\cdot Z\cdot D+I\bigl[||D||_{2}\leq c\bigr]-M\cdot D\\
=\arg\min_{D}\frac{1}{2}||D-(R+M-\Lambda\odot Z)||_{F}^{2}+I\bigl[||D||_{2}\leq c\bigr],\end{gathered}$$ which shows that $g$ can be calculated as in Eq. \[eq:sparse-projection-dual-obj-appendix\].
Next, we are interested in the gradient of $g$. By applying Danskin’s theorem to Eq. \[eq:sparse-projection-lagrangian\], we have that $\frac{d}{dM}\arg\min_{D}\mathcal{L}(D,\Lambda,M)$ will be exactly $-D$ where $D$ is the minimizer of Eq. \[eq:sparse-projection-lagrangian\]. This establishes Eq. \[eq:sparse-projection-dual-grad-2-appendix\]. Similarly, it can be shown that $\frac{d}{d\Lambda}\arg\min_{D}\mathcal{L}(D,\Lambda,M)=Z\odot D,$ establishing Eq. \[eq:sparse-projection-dual-grad-1-appendix\].
The divergence $D(\theta,\psi)=KL(\psi||\theta)$ has the gradient $$\frac{d}{d\psi}D(\theta,\psi)=\sum_{x}p(x;\psi)(\psi-\theta)\cdot f(x)\left(f(x)-\mu(\psi)\right).$$
Firstly, it can be shown that $$D(\theta,\psi)=\sum_{x}p(x;\psi)(\psi-\theta)\cdot f(x)+A(\theta)-A(\psi).$$
From this, it follows that $$\begin{aligned}
{1}
\frac{d}{d\psi}D(\theta,\psi)= & \sum_{x}\frac{dp(x;\psi)}{d\psi}(\psi-\theta)\cdot f(x)\\
+ & \sum_{x}p(x;\psi)f(x)-\mu(\psi).\end{aligned}$$ This can be seen to be equivalent to the result by observing that the second two terms cancel, and that $dp(x;\psi)/d\psi=p(x;\psi)(f(x)-\mu(\psi))$.
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
{width="50.00000%"}{width="50.00000%"}
[^1]: Technically, here, we assume that the exponential family is minimal. However, in the case of an over-complete exponential family, enforcing this will simply ensure that $\theta$ and $\psi$ use the same reparameterization.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The alliance polynomial of a graph $G$ with order $n$ and maximum degree ${\Delta}$ is the polynomial $A(G; x) = \sum_{k=-{\Delta}}^{{\Delta}} A_{k}(G) \, x^{n+k}$, where $A_{k}(G)$ is the number of exact defensive $k$-alliances in $G$. We obtain some properties of $A(G; x)$ and its coefficients for regular graphs. In particular, we characterize the degree of regular graphs by the number of non-zero coefficients of their alliance polynomial. Besides, we prove that the family of alliance polynomials of ${\Delta}$-regular graphs with small degree is a very special one, since it does not contain alliance polynomials of graphs which are not ${\Delta}$-regular. By using this last result and direct computation we find that the alliance polynomial determines uniquely each cubic graph of order less than or equal to $10$.'
address:
- 'National council of science and technology (CONACYT) $\&$ Autonomous University of Zacatecas, Paseo la Bufa, int. Calzada Solidaridad, 98060 Zacatecas, ZAC, Mexico'
- 'Department of Mathematics, Universidad Carlos III de Madrid, Av. de la Universidad 30, 28911 Leganés, Madrid, Spain'
- 'Faculty of Mathematics, Autonomous University of Guerrero, Carlos E. Adame 5, Col. La Garita, Acapulco, Guerrero, Mexico'
- 'Humboldt International University, 4000 West Flagler Street, 33134, Miami, Fl., USA'
author:
- Walter Carballosa
- 'José M. Rodr[í]{}guez'
- 'José M. Sigarreta'
- 'Yadira Torres-Nuñez'
title: Alliance polynomial of regular graphs
---
Regular Graphs ,Cubic Graphs ,Defensive Alliances ,Alliance Polynomials
05C69 ,11B83
Introduction.
=============
Graph polynomials have been widely study since George D. Birkhoff introduced the chromatic polynomial (1912) in an attempt to prove the four color theorem (see [@Bi]). Although the original motivation for the study of this invariant (chromatic number) is still important, much of the current interest, for example, in the Tutte polynomial is not related to any of its applications. In particular, graph polynomials are considered interesting when they encode much information about the underlying graph. Some parameters of a graph $G$ allow one to define polynomials on the graph $G$, for instance, the parameters associated to matching sets [@F; @GG], independent sets [@BDN; @GH; @HL], domination sets [@AAP; @AP], chromatic numbers [@Bi; @R; @T] and many others. In recent years there has been an increase in the number of papers introducing new properties of graph polynomials [@KM; @MR; @MRB].
The study of defensive alliances in graphs, together with a variety of other kinds of alliances, was introduced in [@KHH]. In the cited paper authors initiated the study of the mathematical properties of alliances. In particular, several bounds on the defensive alliance number were given. The particular case of global defensive alliance (generalization of dominating sets) was investigated in [@FLH3; @HHH; @SRV2] where several bounds on the global defensive alliance number were obtained. Several tight bounds on different types of alliance numbers of a graph were obtained in [@KHH; @SRV2; @S; @SGBR], namely, (global) defensive alliance number, (global) offensive alliance number and (global) dual alliance number. Moreover, alliances, as a graph-theoretic concept, have recently attracted a great deal of attention due to some interesting applications in a variety of areas, including quantitative analysis of secondary RNA structures [@HKSZ] and national defense [@P]. Besides, defensive alliances could work as a mathematical model of web communities. Adopting the definition of a Web community proposed recently in [@FLG], “a Web community is a set of web pages having more hyperlinks (in either direction) to members of the set than to non-members”. In [@CHRT], the authors use the cardinality of the strong defensive alliances by define the strong alliance polynomial of a graph. The exact defensive alliance was defined in [@C]. The alliance polynomial was introduced in [@To] and have been studied in [@CRST; @CRSTx].
Throughout this paper, $G=(V,E)$ denotes a simple graph (not necessarily connected) of order $|V|=n$ and size $|E|=m$. We denote two adjacent vertices $u$ and $v$ by $u\sim v$. For a nonempty set $X\subseteq V$, and a vertex $v\in V$, $N_X(v)$ denotes the set of neighbors $v$ has in $X$: $N_X(v):=\{u\in X: u\sim v\},$ and the degree of $v$ in $ X$ will be denoted by $\delta_{X}(v)=|N_{X}(v)|$. We denote by ${\delta}$ and ${\Delta}$ the minimum and maximum degree of $G$, respectively. If $G$ is a regular graph with degree ${\Delta}$ we say that $G$ is ${\Delta}$-regular. The subgraph induced by $S\subset V$ will be denoted by $\langle S\rangle $ and the complement of the set $S\in V$ will be denoted by $\overline{S}$.
A nonempty set $S\subseteq V$ with $\langle S\rangle$ connected is a *defensive $k$-alliance* in $G=(V,E)$, $k\in [-\Delta,\Delta]\cap \mathbb{Z}$, if for every $ v\in S$, $$\label{cond-A-Defensiva} \delta _S(v)\ge \delta_{\overline{S}}(v)+k.$$ A vertex $v\in S$ is said to be $k$-*satisfied* by the set $S$, if holds.
We consider the value of $k$ in the set of integers $\mathcal{K}:= [-\Delta,\Delta]\cap \mathbb{Z}$. It may be that there are some values of $k \in\mathcal K$ such that there do not exist. For instance, for $k\ge 2$ there are no defensive $k$-alliances in the star graph $S_n$ with $n$ vertices. Besides, if $G$ is connected then $V(G)$ is a defensive ${\delta}$-alliance. Notice that for any $S$ with $\langle S\rangle$ connected there exists some $k\in\mathcal{K}$ such that it is a defensive $k$-alliance in $G$.
We define, for $S\subseteq V(G)$ inducing a connected subgraph $\langle S\rangle$, $$\label{eq:k_exact}
k_{S}:=\displaystyle\max_{} \{k\in\mathcal{K} \,:\, S \text{ is a defensive $k$-alliance}\}.$$ We say that $k_{S}$ is the *exact index of alliance of* $S$, or also, $S$ is an *exact defensive $k_{S}$-alliance* in $G$. The exact index of alliance of $S$ in $G$ is also $k_{S}= \displaystyle\min_{v\in S} \{{\delta}_{S}(v) - {\delta}_{\overline{S}}(v)\}$.
The *alliance polynomial* of a graph $G$ is defined as an ordinary generating function for the number of exact defensive $k$-alliances: $$\label{eq:Poly1}
A(G;x)= \displaystyle\sum_{k\in \mathcal{K}} A_{k}(G) x^{n+k}, \ \text{ with } A_{k}(G) \text{ the number of exact defensive } k \text{-alliances in } G.$$
There are several d.p.-equivalent definitions for an alliance polynomial. For instance, another natural definition for alliance polynomials of $G$ could be $A^\star(G;x)= \sum_{k\in \mathcal{K}} A_{k}(G) x^{k}$ (*i.e.*, $A(G;x)= x^n\, A^\star(G;x)$). We may also define it by $A^\dag(G;x)=\sum_{k\in \mathcal{K}} \big(A_k(G)+1\big) x^k$. Note that, $A^\star(G;x)$ and $A^\dag(G;x)$ are Laurent polynomials, and that the polynomial $A^\dag(G;x)$ does not satisfy any easy variant of Corollary \[c:CubUnim\]. Hence it is convenient to keep the previous choice, although each result in this paper has an analogous for $A^\star(G;x)$.
The following procedure allows to compute the alliance polynomial of a graph $G$ with order $n$, see [@CRST Algorithm 2.1]. This will be used in Section \[sect3\] in order to compute alliance polynomials. Let us consider $W=\{S_1,\dots,S_{2^n-1}\}$ the collection of nonempty subsets of $V$.
\[algorithm\] ${}$Input: adjacency matrix of $G$. Output: alliance polynomial of $G$. The algorithm starts with $A(G;x)=0$ and continues with the following steps, for $1 \le j \le 2^n-1$.
1. [If $\langle S_j\rangle$ is a connected subgraph, then go to step (2), else replace $j$ by $j+1$ and apply this step again.]{}
2. [Compute $k_{S_j}$, and add one term $x^{n+k_{S_j}}$ to $A(G;x)$.]{}
3. [If $j < 2^n-1$, then replace $j$ by $j+1$ and apply step (1) again. If $j = 2^n-1$, then the algorithm stops.]{}
The main aim of this work is to obtain further results about the alliance polynomial of regular graphs (graphs with all vertices with the same degree), since they are a very interesting class of graphs with many applications (see, e.g., [@B; @Cu; @DM; @RST]). In this paper we study the alliance polynomials of regular graphs and their coefficients, see Section \[sect:reg\]. In Section \[sect:con\_reg\] we focus on the alliance polynomials of connected regular graphs; besides, we prove that the family of alliance polynomials of connected ${\Delta}$-regular graphs with small degree is a very special one, since it does not contain alliance polynomials of graphs which are not connected ${\Delta}$-regular (see Theorems \[t:Reg0-3\] and \[t:Reg4-6\]). Finally, by using Theorem \[t:Reg0-3\] and direct computation we find that the alliance polynomial determines uniquely each cubic graph of order less than or equal to $10$.
Alliance polynomials for regular graphs {#sect:reg}
=======================================
In this section we deal with regular graphs, in particular, we obtain some properties of the alliance polynomial of regular graph and its coefficients. Below, a quick reminder of some previous results for general graphs (not necessarily regular) which will be useful, see [@CRST]. We denote by $\operatorname{Deg}(p)$ the degree of the polynomial $p$.
\[p:AlliPoly\] Let $G$ be any graph. Then $A(G;x)$ satisfies the following properties:
i) [$A(G;x)$ does not have zeros in the interval $(0,\infty)$.]{}
ii) [$A(G;1) < 2^{n}$, and it is the number of connected induced subgraphs in $G$.]{}
iii) [If $G$ has at least an edge and its degree sequence has exactly $r$ different values, then $A(G;x)$ has at least $r + 1$ terms.]{}
iv) [ $A(G;x)$ is a symmetric polynomial (i.e., an even or an odd function of $x$) if and only if the degree sequence of $G$ has either all values odd or all even.]{}
v) [$A_{-{\Delta}}(G)$ and $A_{-{\Delta}+ 1}(G)$ are the number of vertices in $G$ with degree ${\Delta}$ and ${\Delta}-1$, respectively.]{}
vi) [$A_{{\Delta}}(G)$ is equal to the number of connected components in $G$ which are ${\Delta}$-regular.]{}
vii) [$n + {\delta}\leq \operatorname{Deg}(A(G;x)) \leq n + {\Delta}$.]{}
As usual, by *cycle* we mean a simple closed curve, i.e., a path with different vertices, unless the last one, which is equal to the first vertex. The following lemma is a well known result of graph theory.
\[l:degreem\] If $r\ge 2$ is a natural number and $G$ is any graph with ${\delta}(v) \ge r$ for every $v \in V(G)$, then there exists a cycle $\eta$ in $G$ with $L(\eta) \ge r+1$.
We show now some results about the alliance polynomial of regular graphs and their coefficients. If $G$ is a graph and $v\in V(G)$, we denote by $G \setminus \{v\}$ the subgraph obtained by removing from $G$ the vertex $v$ and the edges incident to $v$. We say that $v$ is a *cut vertex* if $G \setminus \{v\}$ has more connected components than $G$. Besides, if $p$ is a polynomial we denote by $\operatorname{Deg}_{min}(p)$ the minimum degree of their non-zero coefficients.
\[t:PropRegular\] For any ${\Delta}$-regular graph $G$, its alliance polynomial $A(G;x)$ satisfies the following properties:
i) [$A_{-{\Delta}+2i}(G)$ is the number of connected induced subgraphs of $G$ with minimum degree $i$ $(0\le i\le{\Delta})$.]{}
ii) [$\operatorname{Deg}_{\min}\big(A(G;x)\big)= n - {\Delta}$ and $A_{-{\Delta}}(G) = n$.]{}
iii) $\operatorname{Deg}\big(A(G;x)\big) = n + {\Delta}$. Furthermore,
$$\label{eq:RegOrderV}
n = \frac{\operatorname{Deg}_{\min}\big(A(G;x)\big) + \operatorname{Deg}\big(A(G;x)\big)}2$$
and $$m = A_{-{\Delta}}(G) \frac{\operatorname{Deg}\big(A(G;x)\big) - \operatorname{Deg}_{\min}\big(A(G;x)\big)}4 = \frac{\operatorname{Deg}^2\big(A(G;x)\big) - \operatorname{Deg}_{\min}^2\big(A(G;x)\big)}8.$$
iv) [$1\le A_{{\Delta}}(G) \le n/({\Delta}+1)$. Furthermore, $G$ is a connected graph if and only if $A_{{\Delta}}(G) = 1$.]{}
v) [If ${\Delta}>0$, then $A_{-{\Delta}+2}(G) \ge m$ and $A_{{\Delta}-2}(G) \ge n+n_0$ with $n_0$ the number of cut vertices; in particular, $A_{{\Delta}-2}(G)\ge n$.]{}
vi) [$A(G;x)$ is either an even or an odd function of $x$. Furthermore, $A(G;x)$ is an even function of $x$ if and only if $n + {\Delta}$ is even.]{}
vii) [The unique real zero of $A(G;x)$ is $x=0$, and its multiplicity is $n-{\Delta}$.]{}
We prove each item separately.
i) [Let us consider $S\subset V$ with $S$ an exact defensive ($2i-{\Delta}$)-alliance in $G$. Then, we have for all $v\in S$ $$2{\delta}_S(v)\ge {\delta}(v)+2i-{\Delta}= {\Delta}+2i-{\Delta}\quad \Leftrightarrow \quad {\delta}_S(v)\ge i,$$ besides, the equality holds at some $w\in S$. We have the result since $A_{-{\Delta}+2i}(G)$ is the number of exact defensive ($2i-{\Delta}$)-alliance in $G$.]{}
ii) [One can check directly that if $S$ is a single vertex, then $S$ is an exact defensive ($-{\Delta}$)-alliance; furthermore, it is clear that any $S\subseteq V$ with $\langle S\rangle$ connected and more than one vertex is not an exact defensive ($-{\Delta}$)-alliance, since for any $v\in S$ we have $$\label{eq:i}
{\delta}_{S}(v) - {\delta}_{\overline{S}}(v) \geq 1 - ({\Delta}- 1) = -{\Delta}+ 2.$$ Consequently $A_{-{\Delta}}(G) = n$, since $G$ is a ${\Delta}$-regular graph.]{}
iii) [The maximum value in $\mathcal{K}$ is ${\Delta}$, so $Deg\big(A(G;x)\big)$ is at most $n+{\Delta}$. We have that each connected component of $G$ is an exact defensive ${\Delta}$-alliance since ${\delta}(v)={\Delta}$ for any vertex $v$. Then, $A_{{\Delta}}(G)>0$ and $Deg\big(A(G;x)\big)= n+{\Delta}$. Besides, the other results are consequences of the well known fact $2m=n{\Delta}$ and the previous results.]{}
iv) [By item i), $A_{{\Delta}}(G)$ is the number of connected induced subgraphs of $G$ with minimum degree ${\Delta}$; hence, $A_{{\Delta}}(G)$ is the number of connected components of $G$. Besides, since any connected component has cardinality greater than ${\Delta}$, we obtain the upper bound of $A_{{\Delta}}(G)$.]{}
v) [If $u,v\in V$ with $u\sim v$, then $S:=\{u,v\}$ is an exact defensive ($2-{\Delta}$)-alliance since $1={\delta}_S(u)={\delta}_{\overline{S}}(u)+2-{\Delta}$ and $1={\delta}_S(v)={\delta}_{\overline{S}}(v)+2-{\Delta}$. Thus, we obtain $A_{-{\Delta}+2}(G)\ge m$. Note that if ${\Delta}=1$, we have the second inequality. Assume that ${\Delta}\ge2$. Without loss of generality we can assume that $G$ is connected; otherwise, it suffices to analyze each connected component of $G$. Let us define $S_v:=V\setminus\{v\}$ for any $v\in V$. Since ${\delta}_{S_v}(u)\ge {\Delta}-1$, ${\delta}_{\overline{S_v}}(u)\le1$ for every $u\in S_v$ and both equalities hold for every $w\in N(v)$, we have that $S_v$ is an exact defensive (${\Delta}-2$)-alliance if $v$ is a non-cut vertex, or contains at least two exact defensive (${\Delta}-2$)-alliances if $v$ is a cut vertex.]{}
vi) [The first statement is a consequence of Theorem \[p:AlliPoly\] iv). Consider an exact defensive $k$-alliance $S$ in $G$. So, there exists $v\in S$ with $$2{\delta}_{S}(v) = {\delta}(v) + k = {\Delta}+k.$$ Then, ${\Delta}\equiv k \ (\text{mod } 2)$, $n + k \equiv n + {\Delta}\ (\text{mod } 2)$ and we have the result.]{}
vii) [Since $Deg_{\min}\big(A(G;x)\big)=n-{\Delta}$, we have that $x=0$ is a zero of $A(G;x)$ with multiplicity $n-{\Delta}$. The positivity of all coefficients of $A(G;x)$ gives $A(G;x)\neq 0$ for every $x>0$. Finally, by item vi), $A(G;x)=(-1)^{n+{\Delta}}A(G;-x)\neq0$ for every $x<0$.]{}
Note that Theorem \[t:PropRegular\] (items ii, iv, v and vi) has the following direct consequence.
\[c:Cubic\] Let $G$ be any cubic graph. Then, $$A(G;x)=A_{-3}(G)\, x^{n-3}+A_{-1}(G)\, x^{n-1}+A_1(G)\, x^{n+1}+A_{3}(G)\, x^{n+3},$$ with $A_{-3}(G)=n < m\le A_{-1}(G)$ and $A_{1}(G)\ge A_{3}(G)$.
A finite sequence of real numbers $(a_{0} , a_{1} , a_{2} , . . . , a_{n})$ is said to be *unimodal* if there is some $k \in \{0, 1, . . . , n\}$, called the *mode* of the sequence, such that $$a_{0} \leq . . . \leq a_{k-1} \leq a_{k} \geq a_{k+1} \geq . . . \geq a_{n}.$$ A polynomial is called unimodal if the sequence of its coefficients is unimodal. Therefore, we have the following result for alliance polynomial of a cubic graph, note that $A^\dag$ does not satisfy it.
\[c:CubUnim\] For any cubic graph $G$, $x^{(3-n)/2}\,A(G;\sqrt{x}\,)$ is an unimodal polynomial.
\[c:ConComp\] Let $G$ be any connected graph. Then $G$ is regular if and only if $A_{{\Delta}}(G) = 1$.
If $G$ is regular, then by Theorem \[p:AlliPoly\] vi) we obtain $A_{{\Delta}}(G)=1$. Besides, if $A_{{\Delta}}(G)=1$, then there is an exact defensive ${\Delta}$-alliance $S$ in $G$ with ${\delta}_S(v) \ge {\delta}_{\bar{S}}(v) + {\Delta}\ge {\Delta}$ (i.e., ${\delta}_S(v)={\Delta}$ and ${\delta}_{\overline{S}}(v)=0$) for every $v\in S$. So, the connectivity of $G$ gives that $G$ is a ${\Delta}$-regular graph.
\[t:Regulars\] Let $G_1,G_2$ be two regular graphs. If $A(G_1;x)=A(G_2;x)$, then $G_1$ and $G_2$ have the same order, size, degree and number of connected components.
Let $n_1,n_2$ be the orders of $G_1$ and $G_2$, respectively, and ${\Delta}_1,{\Delta}_2$ the degrees of $G_1$ and $G_2$, respectively. Then, by Theorem \[t:PropRegular\] ii) and iii) we have $$n_1 - {\Delta}_1 = n_2 - {\Delta}_2 \quad \text{ and } \quad n_1+{\Delta}_1 = n_2+{\Delta}_2$$ and we conclude $$n_1 = n_2 \quad \text{ and } \quad {\Delta}_1 = {\Delta}_2.$$ Hence, both graphs have the same size. Finally, since $A_{{\Delta}_1}(G_1) = A_{{\Delta}_2}(G_2)$, they have the same number of connected components by Theorem \[p:AlliPoly\] vi).
\[c:Reg1\] Let $G_1,G_2$ be two regular graphs with orders $n_1$ and $n_2$, and degrees ${\Delta}_1$ and ${\Delta}_2$, respectively. If $n_1 \neq n_2$ or ${\Delta}_1 \neq {\Delta}_2$, then $A(G_1;x) \neq A(G_2;x)$.
The next theorem characterizes the degree of any regular graph by the number of non-zero coefficients of its alliance polynomial.
\[t:PolyReg\] Let $G$ be any ${\Delta}$-regular graph with order $n$. Then $A(G;x)$ has ${\Delta}+1$ non-zero coefficients. Furthermore, $$A(G;x)=\displaystyle\sum_{i=0}^{{\Delta}} A_{{\Delta}-2i}(G)\ x^{n+{\Delta}-2i},$$ with $A_{-{\Delta}}(G)=n$, $A_{{\Delta}}(G)\ge1$, and $$A_{{\Delta}-2i}(G) \ge \frac{n {{\Delta}\choose i}}{\min\{{\Delta},n-i\}} \quad \text{for } \ 1\le i\le {\Delta}-1 \text{ if } {\Delta}>0.$$
Since $G$ is ${\Delta}$-regular, by Theorem \[t:PropRegular\] we have $A_{-{\Delta}}(G)=n$, $A_{{\Delta}}(G)\ge1$ and $A(G;x)$ is an even or an odd function of $x$. Assume now that ${\Delta}>0$ and fix $1\le i \le {\Delta}-1$. Let us consider $u\in V$ and $v_1,\ldots,v_i$ different vertices in $N(u)$. Denote by $S_u:=V\setminus\{v_1,\ldots,v_i\}$. Then, we have that ${\delta}_{S_u}(v)\ge {\Delta}-i$ and ${\delta}_{\overline{S_u}}(v)\le i$ for every $v\in S_u$; furthermore, the equalities hold at $u$. Let $S_u^*\subset S_u$ such that $\langle S_u^*\rangle$ is the connected component of $\langle S_u\rangle$ which contains $u$. So, $S_u^*$ is an exact defensive (${\Delta}-2i$)-alliance and $A_{{\Delta}-2i}(G)>0$. Since each set $S_u^*$ can appear at most $n-i$ times (once for each $S_w^*$ with $w\in V\setminus \{v_1,\ldots,v_i\}$), and at most ${\Delta}$ times (once for each $S_w^*$ with $w\sim v_1$), we obtain $A_{{\Delta}-2i}(G)\ge n {{\Delta}\choose i}/ \min\{{\Delta},n-i\}$.
A *Hamiltonian cycle* is a cycle in a graph that visits each vertex exactly once. A graph that contains a Hamiltonian cycle is called a *Hamiltonian graph*. The following theorem is a well known result in graph theory which will be useful.
\[Dirac\] A graph with order $n \ge 3$ is Hamiltonian if every vertex has degree $n / 2$ or greater.
In what follows we will use the following notation: for any $A,B\subset V$, we denote by $N(A,B)$ the number of edges with one endpoint in $A$ and the other endpoint in $B$.
\[t:An+D-2=n\] Let $G$ be any ${\Delta}$-regular graph with order $n < 2{\Delta}$. Then $A_{{\Delta}-2}(G)=n$.
Notice that ${\Delta}\ge 2$, since otherwise, such a graph $G$ does not exist; furthermore, $n\ge {\Delta}+1 \ge 3$. We have that $G$ is a Hamiltonian graph by Theorem \[Dirac\]. Besides, by Theorem \[t:PropRegular\] i), we have that $A_{{\Delta}-2}(G)$ is the number of connected induced subgraphs of $G$ with minimum degree ${\Delta}-1$. Let us consider $u\in V$ and define $S_u:=V\setminus\{u\}$. Since $G$ is a Hamiltonian graph, $\langle S_u\rangle$ is connected. Besides, we have ${\delta}_{S_u}(v)\ge {\Delta}-1 \ge {\delta}_{\overline{S_u}}(v) + {\Delta}-2$ for all $v\in S_u$ and the equality holds at $w\in N(u)$. So, $S_u$ is an exact defensive (${\Delta}-2$)-alliance in $G$ and $A_{{\Delta}-2}(G)\ge n$.
Seeking for a contradiction assume that there is an exact defensive (${\Delta}-2$)-alliance $S\subset V$ with $|S|\le n-2$. Notice that $|S|\ge{\Delta}>n/2$, by Theorem \[t:PropRegular\] i). Then, since any vertex in $S$ has degree ${\Delta}$ in $G$ with at most one edge among $S$ and $\overline{S}$, we have $$N(S,S) + N(S,\overline{S}) = \frac{|S| {\Delta}}2 + \frac{N(S,\overline{S})}2 \le \frac{|S| {\Delta}}2 + \frac{|S|}2 = \frac{|S|({\Delta}+1)}2.$$ Besides, since $|\overline{S}|=n-|S|$, we have $$N(\overline{S},\overline{S}) \le \frac{(n-|S|)(n-|S|-1)}2.$$ If $m$ denotes the size of $G$, then $$\begin{aligned}
0 & = 2\left(N(S,S) + N(S,\overline{S}) + N(\overline{S},\overline{S})\right) - 2m \\
&\le |S|({\Delta}+1) + (n-|S|)(n-|S|-1) - n{\Delta}\\
& = |S|^2 + |S|({\Delta}+2-2n) + n^2 - n - n{\Delta}.
\end{aligned}$$ Define $P(x):=x^2 + x({\Delta}+2-2n) + n^2 - n - n{\Delta}$; then $P(|S|)\ge0$. Since $$\begin{aligned}
P\left(\frac{n}2\right) &= \frac{n^2}4 + \frac{n}2 ({\Delta}+2-2n) + n^2 - n - n{\Delta}\\
&= \frac{n^2}4 + \frac{n{\Delta}}2 + n -n^2 + n^2 - n - n{\Delta}\\
&= \frac{n}4(n-2{\Delta})<0
\end{aligned}$$ and $$\begin{aligned}
P(n-2)&= (n-2)^2 + (n-2)({\Delta}+2-2n) + n^2 - n - n{\Delta}\\
&= (n-2)^2 + (n-2)({\Delta}-n) - (n-2)^2 + n^2 - n - n{\Delta}\\
&= n-2{\Delta}< 0,
\end{aligned}$$ we obtain that $P(|S|)<0$. This is the contradiction we were looking for, so, there not exists an exact defensive (${\Delta}-2$)-alliance $S$ with $|S|\le n-2$. This finishes the proof since $V$ is an exact defensive ${\Delta}$-alliance.
A *clique* in a graph $G = (V, E)$ is a subset $C$ of the vertex set $V$, such that $\langle C\rangle$ is a complete graph.
\[l:clique\] Let $G$ be any ${\Delta}$-regular graph with order $n$, ${\Delta}\ge3$ and $2{\Delta}\le n \le 2{\Delta}+1$. If $G$ contains two cliques of cardinality ${\Delta}$, then these cliques are disjoint. In particular, $G$ contains at most two cliques of cardinality ${\Delta}$.
Seeking for a contradiction, assume that there exist $S_1,S_2\subset V$ cliques of cardinality ${\Delta}$ with $S_1\cap S_2\neq\emptyset$. Denote by $r$ the number $r:= |S_1\cap S_2|$; then $1\le r \le {\Delta}-1$. Note that for any $v\in S_1\cap S_2$ we have ${\delta}_{S_1\cup S_2}(v)= |S_1|-1 + |S_2\setminus S_1|= {\Delta}-1+{\Delta}-r$, so, we obtain $r={\Delta}-1$. Then, we have $|S_1\cup S_2|={\Delta}+1$ and ${\Delta}-1\le|\overline{S_1\cup S_2}|\le {\Delta}$. Besides, we have $N(S_1\cup S_2,\overline{S_1\cup S_2}) = 2 = |(S_1\cup S_2)\setminus (S_1\cap S_2)|$ and, since $|\overline{S_1\cup S_2}|\le{\Delta}$, $N(\overline{S_1\cup S_2},S_1\cup S_2)\ge |\overline{S_1\cup S_2}|\ge {\Delta}-1$. Since $N(S_1\cup S_2,\overline{S_1\cup S_2}) = N(\overline{S_1\cup S_2},S_1\cup S_2)$, we obtain ${\Delta}=3$ and $n=6$; therefore, $G$ is a graph isomorphic to either $K_{3,3}$ or the Cartesian product $P_2\Box K_3$. Thus, we obtain that there are not two non-disjoint cliques in $G$ with cardinality ${\Delta}$. This finishes the proof since, by $n\le 2{\Delta}+1$, it is impossible to have three disjoint cliques of cardinality ${\Delta}$ contained in $G$.
\[r:exitClique\] If $G$ is a ${\Delta}$-regular graph with $n \le 2{\Delta}+1$, then $G$ does not contain a clique of cardinality greater than ${\Delta}$, since $2({\Delta}+1)>2{\Delta}+1\ge n$.
\[r:DisjointClique\] Let $G$ be any ${\Delta}$-regular graph with order $n$ and ${\Delta}\ge1$ such that $G$ has two disjoint cliques of cardinality ${\Delta}$. Then
1. If $n=2{\Delta}$, then $G$ is isomorphic to the Cartesian product graph $P_2 \Box K_{\Delta}$.
2. If $n=2{\Delta}+1$, then ${\Delta}$ is even (since $n{\Delta}=2m$) and $G$ can be obtained from $P_2 \Box K_{\Delta}$ by removing ${\Delta}/2$ copy edges of $P_2$ and connecting the ${\Delta}$ vertices with degree ${\Delta}-1$ with a new vertex. In particular, if $S$ is a clique of cardinality ${\Delta}$ in $G$, then $\overline{S}$ is not an exact defensive (${\Delta}-2$)-alliance.
\[t:An+D-2<=n+m\] Let $G$ be any ${\Delta}$-regular graph with order $n$, size $m$, ${\Delta}\ge 3$ and $2{\Delta}\le n \le 2{\Delta}+1$. Then $n\le A_{{\Delta}-2}(G)\le n+m+2$.
Note that if ${\Delta}=3$ then $n=6$, and $G$ is a graph isomorphic to either $K_{3,3}$ or $P_2\Box K_3$. Thus, a simple computation gives $6\le A_{1}(K_{3,3})=15\le 6 + 9 + 2$ and $6\le A_{1}(P_2\Box K_3) = 11 \le 6 + 9 + 2$.
Assume now that ${\Delta}\ge 4$. Clearly, $G$ is a connected graph and $\operatorname{diam}G = 2$, since $2{\Delta}>n-2$.
First we prove that $G$ does not have cut vertices. If $n=2{\Delta}$, then $G$ is a Hamiltonian graph by Theorem \[Dirac\]. If $n=2{\Delta}+1$, seeking for a contradiction assume that there is a cut vertex $w$ in $G$. Let $S_1,S_2 \subset V$ with $S_1\cup S_2\cup \{w\}=V$ such that $\langle S_1\rangle$ and $\langle S_2\rangle$ are disjoint. Without loss of generality we can assume that $|S_1|\le {\Delta}\le |S_2|$. Since ${\delta}_{S_1}(w),{\delta}_{S_2}(w)\ge1$, ${\delta}_{S_1}(w)+{\delta}_{S_2}(w)={\Delta}$ and ${\delta}_{S_1}(u)\le |S_1|-1 \le {\Delta}-1$ for all $u\in S_1$, we have ${\delta}_{S_1}(w)=|S_1|$ and ${\delta}_{S_1}(u)={\Delta}-1$ for all $u\in S_1$. Then, we obtain that $|S_1|={\Delta}$, but this is a contradiction since ${\delta}_{S_1}(w)={\Delta}-{\delta}_{S_2}(w)\le {\Delta}-1<{\Delta}=|S_1|={\delta}_{S_1}(w)$. Then, $G$ does not have cut vertices.
By Theorem \[t:PropRegular\] i), we have that $A_{{\Delta}-2}(G)$ is the number of connected induced subgraphs of $G$ with minimum degree ${\Delta}-1$; thus, any exact defensive (${\Delta}-2$)-alliance $S$ in $G$ verifies $|S|\ge {\Delta}$. Let us consider $u\in V$ and denote by $S_u:=V\setminus\{u\}$. Since $G$ does not have cut vertices, $\langle S_u\rangle$ is connected. Besides, we have ${\delta}_{S_u}(v)\ge {\Delta}-1 \ge {\delta}_{\overline{S_u}}(v) + {\Delta}-2$ for all $v\in S_u$ and the equality holds for every $v\in N(u)$; so, $S_u$ is an exact defensive (${\Delta}-2$)-alliance in $G$. Thus, $A_{{\Delta}-2}(G)\ge n$.
Let us consider $u_1,u_2\in V$ with $u_1\neq u_2$ and define $S_{u_1,u_2}:=V\setminus\{u_1,u_2\}$. If $u_1\nsim u_2$, then there is $w\in V$ with $u_1,u_2\in N(w)$ since ${\delta}(u_1)+{\delta}(u_2)= 2{\Delta}>|S_{u_1,u_2}|$; in fact, $S_{u_1,u_2}$ is not a defensive (${\Delta}-2$)-alliance in $G$. So, $S_{u_1,u_2}$ may be an exact defensive (${\Delta}-2$)-alliance in $G$, if $u_1\sim u_2$; then there are at most $m$ exact defensive (${\Delta}-2$)-alliances with $n-2$ vertices. Consider now $u_1,\ldots,u_r\in V$ with $3\le r\le {\Delta}-1$ and $u_i\neq u_j$ if $i\neq j$. Note that $S_r:=V\setminus\{u_1,\ldots,u_r\}$ is not a defensive (${\Delta}-2$)-alliance in $G$ if $r>3$, since $N(\overline{S_r},S_r)\ge r({\Delta}-r+1)=2{\Delta}-r+(r-2)({\Delta}-r)> 2{\Delta}+1-r \ge |S_r|$. Besides, if $r=3$ and ${\Delta}\ge5$ (thus ${\Delta}-r\ge2$) we have the same inequality and then $S_r$ is not a defensive (${\Delta}-2$)-alliance in $G$. Note that, if $r=3$ and $n=2{\Delta}$, then $N(\overline{S_r},S_r)\ge 2{\Delta}-r+(r-2)({\Delta}-r)>2{\Delta}-r=n-r\ge |S_r|$ and we also conclude that $S_r$ is not a defensive (${\Delta}-2$)-alliance in $G$. However, if $r=3$, ${\Delta}=4$ and $n=2{\Delta}+1$ (thus, $n=9$), then $S_r$ may be an exact defensive (${\Delta}-2$)-alliance in $G$. But a simple computation gives that these five graphs $G$ verify $A_{2}(G)< 9+18+2$.
We analyze separately the cases $n=2{\Delta}$ and $n=2{\Delta}+1$. Assume first that $n=2{\Delta}$. We only need to compute the possible exact defensive (${\Delta}-2$)-alliances in $G$ with cardinality ${\Delta}$, since every defensive (${\Delta}-2$)-alliance has at least ${\Delta}$ vertices and $n=2{\Delta}$. If $S$ is an exact defensive (${\Delta}-2$)-alliance in $G$, then $S$ is a clique of cardinality ${\Delta}$ and by Lemma \[l:clique\] there are at most $2$ exact defensive (${\Delta}-2$)-alliances with ${\Delta}$ vertices. Assume now that $n=2{\Delta}+1$. So, ${\Delta}$ is even. We only need to compute the possible exact defensive (${\Delta}-2$)-alliances in $G$ with cardinalities ${\Delta}$ and ${\Delta}+1$. If $S$ is an exact defensive (${\Delta}-2$)-alliance in $G$ with $|S|={\Delta}+1$, then ${\delta}_S(u)\ge {\Delta}-1$ for every $u\in S$ and ${\delta}_S(u_0)={\Delta}$ for some $u_0\in S$, since otherwise ${\delta}_S(u)={\Delta}-1$ for every $u\in S$ and we conclude $({\Delta}+1)({\Delta}-1)=|S|({\Delta}-1)=2m_S$, with $m_S$ the size of $\langle S\rangle$, which is not possible since ${\Delta}$ is even. Hence, $N(\overline{S},S)\le {\Delta}$; furthermore, since $|\overline{S}|={\Delta}$, ${\delta}_{S}(v)\ge 1$ for all $v\in \overline{S}$, and so, $\overline{S}$ is a clique. If $S$ is an exact defensive (${\Delta}-2$)-alliance in $G$ with $|S|={\Delta}$, then ${\delta}_S(u)\ge{\Delta}-1$ for every $u\in S$ and $S$ is a clique of cardinality ${\Delta}$. Lemma \[l:clique\] completes the proof since if $G$ has two cliques of cardinality ${\Delta}$, then they are disjoint and Remark \[r:DisjointClique\] gives that $\overline{S}$ is not an exact defensive (${\Delta}-2$)-alliance in $G$.
\[t:RegCharact\] Let $G$ be a ${\Delta}$-regular connected graph with order $n$ and let $G^*$ be a graph with order $n_1$ and, minimum and maximum degrees ${\delta}_1$ and ${\Delta}_1$, respectively. If $A(G^*;x)=A(G;x)$, then $G^*$ is a connected graph with exactly $n$ vertices of degree ${\Delta}_1={\Delta}+n_1-n$, $n_1\ge n$, ${\Delta}_1\ge{\Delta}$ and ${\delta}_1 \equiv {\Delta}_1 (\text{mod } 2)$.
Furthermore, if $n_1 > n$, then the following inequalities hold: $$\label{eq:Poly2}
\frac{{\Delta}_1+{\delta}_1+2}2\le {\Delta}.$$ $$\label{eq:Poly1}
{\delta}_1+2 < {\Delta}< {\Delta}_1,$$ $$\label{eq:Poly3}
{\Delta}+ 1 \le {\Delta}_1 \le 2{\Delta}-3,$$ $$\label{eq:Poly4}
{\delta}_1+4 \le {\Delta}_1.$$
Since $A(G^*;x)=A(G;x)$ is a symmetric polynomial by Theorem \[t:PropRegular\] vi), we conclude that ${\delta}_1 \equiv {\Delta}_1 (\text{mod } 2)$ by Theorem \[p:AlliPoly\] iv). By Theorems \[p:AlliPoly\] v) and \[t:PropRegular\] ii), $G^*$ has $n$ vertices of maximum degree ${\Delta}_1$, so, $n_1 \geq n$; besides, $n_1 - {\Delta}_1 = n-{\Delta}$. Note that if $n_1=n$ then $G^*$ is a ${\Delta}$-regular graph with $A_{{\Delta}}(G^*)=1$, so, Theorem \[c:ConComp\] gives that $G^*$ is a connected graph.
Assume that $n_1 > n$. Denote by $t:=n_1-n={\Delta}_1 -{\Delta}$. Let $v_1,\ldots,v_n \in V(G^*)$ be the vertices in $G^*$ with degree ${\Delta}_1$ and define $S:=\{v_1,\ldots,v_n\}$. Note that for any $v\in S$ we have ${\delta}_{S}(v) \ge {\Delta}_1 - t = t + ({\Delta}_1-2t) \ge {\delta}_{\overline{S}}(v) + {\Delta}_1-2t$; hence, $S$ contains a defensive (${\Delta}_1-2t$)-alliance $S_1$ and $k_{S_1} \ge {\Delta}_1-2t$. Therefore, there is at least one term of degree greater or equal than $n_1+{\Delta}_1-2t$ in $A(G^*;x)$. Since $x^{n_1+{\Delta}_1-2t}=x^{n+{\Delta}}$, $S_1$ is an exact defensive (${\Delta}_1-2t$)-alliance in $G^*$. Finally, note that if $\langle S\rangle$ is not a connected subgraph (i.e., $S_1\neq S$), then in $A(G^*;x)$ appear at least two terms $x^{n+{\Delta}}$, but this is a contradiction since $A(G;x)$ is a monic polynomial by Theorem \[p:AlliPoly\] vi). Hence, $\langle S\rangle$ is connected. Since the degree of $A(G^*;x)$ is $n+{\Delta}=n_1+{\Delta}_1-2t$, then $S$ is an exact defensive (${\Delta}_1-2t$)-alliance in $G^*$; therefore, there exists $1\le j\le n$ such that ${\Delta}_1={\delta}(v_j)=2{\delta}_{\overline{S}}(v_j)+{\Delta}_1-2t$, and we have ${\delta}_{\overline{S}}(v_j)=t$. Since $|S|=n=n_1-t$ and $|\overline{S}|=t$, $\overline{S}\subseteq N(v_j)$ and $G^*$ is a connected graph.
Also, since $G^*$ is connected, $A(G^*;x)=A(G;x)$, $k_S={\Delta}_1-2t$ and $k_{V(G^*)}={\delta}_1$, we have ${\delta}_1\le {\Delta}_1-2t$. We are going to prove ${\delta}_1 < {\Delta}_1 - 2t$; seeking for a contradiction assume that ${\delta}_1 = {\Delta}_1 - 2t$. Since $G^*$ is connected, $k_{V(G^*)}={\delta}_1= {\Delta}_1 - 2t = k_{S}$ and this contradicts that $A(G^*;x)$ is a monic polynomial. Therefore, ${\delta}_1 < {\Delta}_1 - 2t$. But, since ${\delta}_1\equiv {\Delta}_1 (\text{mod } 2)$ we obtain ${\delta}_1+2\le {\Delta}_1-2({\Delta}_1-{\Delta})=2{\Delta}-{\Delta}_1$, so holds.
Besides, since ${\Delta}_1>{\Delta}$, gives ${\delta}_1+2 <{\Delta}$, and so, holds. Furthermore, we have ${\Delta}+1\le {\Delta}_1$ and gives , since ${\delta}_1\ge1$. Finally, since ${\Delta}\le {\Delta}_1-1$, provides .
Alliance polynomials of regular graphs with small degree {#sect:con_reg}
========================================================
The theorems in this section can be seen as a natural continuation of the study in [@CRST] in the sense of showing the distinctive power of the alliance polynomial of a graph. In particular, we show that the family of alliance polynomials of ${\Delta}$-regular graphs with small degree ${\Delta}$ is a special family of alliance polynomials since there not exists a non ${\Delta}$-regular graph with alliance polynomial equal to one of their members, see Theorems \[t:Reg0-3\] and \[t:Reg4-6\].
\[t:Reg0-3\] Let $G$ be a ${\Delta}$-regular graph with $0\le{\Delta}\le3$ and $G^*$ another graph. If $A(G^*;x) = A(G;x)$, then $G^*$ is a ${\Delta}$-regular graph with the same order, size and number of connected components of $G$.
In [@CRST Theorem 2.7] the authors obtain the uniqueness of the alliance polynomials of $0$-regular graphs (the empty graphs).
Theorems \[p:AlliPoly\] iii) and \[t:PolyReg\] give that $1$-regular graphs are the unique graphs which have exactly two non-zero terms in their alliance polynomial; besides, Theorems \[p:AlliPoly\] vi) and \[t:PropRegular\] ii) give the uniqueness of these alliance polynomials.
In order to obtain the result for $2\le{\Delta}\le3$, denote by $n,n_1$, the orders of $G,G^*$, respectively, and let ${\delta}_1,{\Delta}_1$ be the minimum and maximum degree of $G^*$.
Assume first that ${\Delta}=2$. By Theorem \[t:PolyReg\] we have $A(G;x)= n x^{n-2} + A_0(G) x^n + A_2(G) x^{n+2}$, thus, by Theorem \[p:AlliPoly\] iii) the degree sequence of $G^*$ has at most two different values. If $G^*$ is regular then Theorem \[t:Regulars\] gives the result. Therefore, seeking for a contradiction assume that the degree sequence of $G^*$ has exactly two different values (i.e., $G^*$ is bi-regular). By Theorems \[p:AlliPoly\] iv) and \[t:PropRegular\] vi) we have ${\delta}_1 \equiv {\Delta}_1 (\text{mod } 2)$. By Theorems \[p:AlliPoly\] v) and \[t:PropRegular\] ii) we have $A_{-{\Delta}_1}(G^*)=A_2(G)=n<n_1$ and $n-2=n_1-{\Delta}_1$, so, we have ${\Delta}_1>2$. By Theorems \[p:AlliPoly\] vii) and \[t:PropRegular\] iii) we have $n_1+{\delta}_1 \le n+2$, so, we obtain $0\le {\delta}_1 \le 1$. If ${\delta}_1=0$, then there is a connected component $G'$ of $G^*$ which is ${\Delta}_1$-regular. So, $k_{V(G')}={\Delta}_1$ and $\operatorname{Deg}\big(A(G^*;x)\big)=n_1+{\Delta}_1 > n+2$, which is a contradiction. Thus, we can assume that ${\delta}_1=1$. Then, we have $n_1=n+1$; and so, ${\Delta}_1=3$. We prove now that $A_1(G^*)\ge n$. Let $u_0,v_0$ be the vertices of $G^*$ with ${\delta}(u_0)=1$ and $v_0\sim u_0$. If $G^*$ is not connected, then it has a $3$-regular connected component $G_0^*$; since $V(G_0^*)$ is an exact defensive $3$-alliance, then $\operatorname{Deg}\big( A(G^*;x)\big) \ge n_1+3 > n+2 = \operatorname{Deg}\big( A(G;x)\big)$, which is a contradiction and we conclude that $G^*$ is connected. Let us define $S_v:=V(G^*)\setminus\{v\}$ for any $v\in V(G^*)\setminus \{v_0\}$. Since ${\delta}_{S_v}(u)\ge 2$, ${\delta}_{\overline{S_v}}(u)\le1$ for every $u\in S_v\setminus\{u_0\}$ and both equalities hold for every $w\in N(v)$, and ${\delta}_{S_v}(u_0)=1$, ${\delta}_{\overline{S_v}}(u_0)=0$, we have that $S_v$ is an exact defensive $1$-alliance or contains an exact defensive $1$-alliance if $v$ is a cut vertex. Thus, $A_{1}(G^*)\ge n$. Besides, Theorem \[t:PropRegular\] iv) gives $A_2(G)\le n/3 < n\le A_{1}(G^*)$, so, $A(G;x)\neq A(G^*;x)$. This is the contradiction we were looking for, and so, we conclude $n_1=n$ and ${\Delta}_1=2$, and we obtain the result for ${\Delta}=2$.
Finally, assume that ${\Delta}=3$. By Corollary \[c:Cubic\] we have $A(G^*;x) = A(G;x)= n x^{n-3} + A_{-1}(G) x^{n-1} + A_{1}(G) x^{n+1} + A_{3}(G) x^{n+3}$, with $A_3(G)$ is the number of connected components of $G$. By Theorem \[p:AlliPoly\] i), we have $n_1 - {\Delta}_1 = n - 3$ and $n \leq n_1$. Hence, $n_1\geq n$ and ${\Delta}_1 \geq 3$. Also we have $n_1 + {\delta}_1 \le n+3$ by Theorem \[p:AlliPoly\] vii). Furthermore, if ${\Delta}_1 = 3$, then $n_1 = n$ and so, $G^*$ is $3$-regular since $A_{- 3}(G^*) = n$. By Theorem \[p:AlliPoly\] vi) they have the same number of connected components, and consequently $G,G^*$ have the same size, too. We will finish the proof by checking that ${\Delta}_1=3$.
Seeking a contradiction, assume that ${\Delta}_1 > 3$ (then $n_1 > n$) and let $k = n_1 - n = {\Delta}_1 - 3$.
Assume that ${\Delta}_1 \ge 6$ (i.e., $k \ge 3$). Then there exists a connected component $G_0$ of $G^*$ with ${\delta}_{G_0}(v)={\delta}(v)\ge1$ for every $v\in V(G_0)$; if $S=V(G_0)$, then ${\delta}_S(v)={\delta}(v)\ge1$, and so, $k_{S}^{(G^*)} \ge 1$. Hence, $A(G^*;x)$ has at least one term with exponent greater than $n_1 \ge n + 3 = \operatorname{Deg}\big(A(G;x)\big)$, and $A(G^*;x) \neq A(G;x)$, which is a contradiction. Thus, ${\Delta}_1=4$ or ${\Delta}_1=5$.
Assume that ${\Delta}_1 = 5$, then $n_1 = n + 2$. By Theorem \[p:AlliPoly\] v), we have that $G^*$ has exactly $n$ vertices with degree $5$; and so, by Theorem \[p:AlliPoly\] iv), we have that the other two vertices of $G$ have degree $1$ or $3$. Since $n_1+{\delta}_1\le n+3$ by Theorem \[p:AlliPoly\] vii), we obtain ${\delta}_1=1$.
Assume that $G^*$ has two vertices $v_1$ and $v_2$ with degree $1$. In this case, if $v_1 \sim v_2$, then $G^*$ is a disconnected graph with at least one connected component which is $5$-regular since $V(G^*)\setminus \{v_1,v_2\}$ induces a $5$-regular subgraph $G_1$ of $G^*$. Since $V(G_1)$ is an exact defensive $5$-alliance, $\operatorname{Deg}(A(G^*;x))\ge n_1+5$ and we have $\operatorname{Deg}(A(G^*;x)) \ge n_1 + 5 > n+3 = \operatorname{Deg}(A(G;x))$. If $v_1 \nsim v_2$ but there exists $w\in V(G^*)$ such that $w\sim v_1$ and $w\sim v_2$, then let us consider the connected component $G_2$ of $G^*$ containing $\{v_1,v_2,w\}$. The set $S=V(G_2)\setminus\{v_1,v_2,w\}$ is a defensive $3$-alliance in $G^*$, since for any $v\in S$ we have ${\delta}_{S}(v) \geq 4$ and ${\delta}_{\overline{S}}(v) \leq 1$. Then, $\operatorname{Deg}(A(G^*;x)) \geq n_1 + 3 > n + 3 = \operatorname{Deg}(A(G;x))$. If $v_1 \nsim v_2$ but there does not exist $w\in V(G^*)$ with $w\sim v_1$ and $w\sim v_2$, then let us consider the connected component $G_3$ of $G^*$ containing $v_1$ and $S=V(G_3)\setminus\{v_1,v_2\}$. The set $S$ is a defensive $3$-alliance in $G^*$, since for all $v\in S$ we have ${\delta}_{S}(v) \geq 4$ and ${\delta}_{\overline{S}}(v) \leq 1$. Then, $\operatorname{Deg}(A(G^*;x)) \geq n_1 + 3 > n + 3 = \operatorname{Deg}(A(G;x))$.
Consider now the case of $G^*$ containing two vertices $v_1$ and $v_2$ with degree $1$ and $3$, respectively. If $v_1 \sim v_2$, then let us consider the connected component $G_4$ of $G^*$ containing $\{v_1,v_2\}$ and $S=V(G_4)\setminus\{v_1,v_2\}$. Then, $S$ is a defensive $3$-alliance in $G^*$, since for all $v\in S$ we have ${\delta}_{S}(v) \geq 4$ and ${\delta}_{\overline{S}}(v) \leq 1$. Then, $\operatorname{Deg}(A(G^*;x)) \geq n_1 + 3 > n + 3 = \operatorname{Deg}(A(G;x))$. If $v_1 \nsim v_2$, let $G_5$ be the connected component of $G^*$ containing $v_1$ and $S=V(G_5)\setminus\{v_1\}$. Hence, $S$ is an exact defensive $3$-alliance in $G^*$, since ${\delta}_{S}(v_2) - {\delta}_{\overline{S}}(v_2) = 3-0$ if $v_2\in S$ and ${\delta}_{S}(v) - {\delta}_{\overline{S}}(v) \ge 4 - 1$ for any $v\in S\setminus\{v_2\}$. Then, $\operatorname{Deg}(A(G^*;x)) \geq n_1 + 3 > n + 3 = \operatorname{Deg}(A(G;x))$. So, it is not possible to have ${\Delta}_1 = 5$.
Assume that ${\Delta}_1 = 4$, then $n_1 = n + 1$. If $G^*$ is a disconnected graph, then there exists a connected component $\langle S^*\rangle$ of $G^*$ such that $\langle S^*\rangle$ is $4$-regular and so, $S^*$ is an exact defensive $4$-alliance in $G^*$. Therefore, $\operatorname{Deg}(A(G^*;x)) = n_1 + 4 > n + 3 = \operatorname{Deg}(A(G;x))$. Thus, $G^*$ is connected, and ${\delta}_1 = 2$ by Theorem \[p:AlliPoly\] iv). So, we have that $G^*$ has exactly $n$ vertices with degree $4$ and another vertex $w$ with degree $2$. Let $v_1,v_2 \in V(G^*)\setminus\{w\}$ with $v_1\neq v_2$, $v_1\sim w$ and $v_2\sim w$. Consider $\{u_1,\ldots,u_{n-2}\} := V(G^*)\setminus \{w,v_1,v_2\}$. Let $G_i$ be the connected component of $\langle V(G^*)\setminus\{u_i\}\rangle \subset G^*$ containing $w$, and $S_i=V(G_i)$, for each $1\le i\le n-2$. Note that $S_i$ is an exact defensive $2$-alliance since ${\delta}_{S_i}(w)-{\delta}_{\overline{S_i}}(w)=2$, for each $1\le i\le n-2$. Note that if $i\neq j$ and $u_j\notin S_i$ then $u_i \in S_j$, and so, $S_i\neq S_j$ since $u_i\notin S_i$; furthermore, if $u_j\in S_i$ then $S_i\neq S_j$ since $u_j\notin S_j$. Then, we obtain that $A_{2}(G^*)\ge n-1$, and thus $A_{3}(G)\ge n-1$. This contradicts Theorem \[p:AlliPoly\] vi) since $G$ is a cubic graph with order $n$. So, it is not possible to have ${\Delta}_1 = 4$.
Now we prove a similar result for ${\Delta}$-regular graphs with ${\Delta}>3$. First, we prove some technical results which will be useful.
\[l:d1=1\] Let $G_1$ be a graph with minimum and maximum degree ${\delta}_1$ and ${\Delta}_1$, respectively, and let $n\ge3$ be a fixed natural number. Assume that $G_1$ has order $n_1 > n$ with exactly $n$ vertices of degree ${\Delta}_1$, and such that its alliance polynomial $A(G_1;x)$ is symmetric. The following statements hold:
1. [If ${\delta}_1=1$, then $A(G_1;x)$ is not a monic polynomial of degree $2n-n_1+{\Delta}_1$.]{}
2. [If ${\delta}_1=2$, then we have $2n_1<2{\Delta}_1+n$ or $A(G_1;x)$ is not a monic polynomial of degree $2n-n_1+{\Delta}_1$.]{}
Seeking for a contradiction assume that $A(G_1;x)$ is a monic polynomial with degree $2n-n_1+{\Delta}_1$. By hypothesis, we have $n$ different vertices $v_1,\ldots,v_n$ in $G_1$ with degree ${\Delta}_1$. Denote by $S$ the set $S:=\{v_1,\ldots,v_n\}$. The argument in the proof of Theorem \[t:RegCharact\] gives that $G_1$ is a connected graph, $S$ is an exact defensive -alliance in $G_1$ and there is $w\in S$ with $\overline{S}\subseteq N(w)$. Let $u\in \overline{S}$ with ${\delta}(u)={\delta}_1$.
First assume that ${\delta}_1=1$. So, $S_w:=S\setminus\{w\}$ contains a defensive -alliance since ${\delta}_{S_w}(v)\ge {\Delta}_1 - \big(|\overline{S}\cup\{w\}|-|\{u\}|\big) = {\Delta}_1 - (n_1-n)$ and ${\delta}_{\overline{S}_w}(v)\le |\overline{S}\cup\{w\}|-1 = n_1-n$ for all $v\in S_w$; thus, in $A(G_1;x)$ appears at least one term of degree greater or equal than $2n-n_1+{\Delta}_1$ associated to $S_w$, but this is impossible since $A(G_1;x)$ is monic of degree $2n-n_1+{\Delta}_1$. This is the contradiction we were looking for.
Assume now that ${\delta}_1=2$. Let $w'\in V(G_1)\setminus\{w\}$ with $w'\sim u$. If $w'\in \overline{S}$ then $S_w$ is a defensive -alliance since $u\notin N(v)$ for every $v\in S_w$. This implies a contradiction as above. So, we can assume that $w'\in S_w$. Note that if $w'\nsim w$ then $S_w$ is a defensive -alliance since ${\delta}_{S_w}(w')-{\delta}_{\overline{S_w}}(w') \ge ({\Delta}_1-n_1+n)-(n_1-n)$ and ${\delta}_{S_w}(v)-{\delta}_{\overline{S_w}}(v) \ge ({\Delta}_1-n_1+n)-(n_1-n)$ for all $v\in S_w\setminus\{w'\}$, but this is impossible since $A(G_1;x)$ is a monic polynomial of degree $n_1+{\Delta}_1-2(n_1-n)$. Then, we can assume that $w'\sim w$. Note that if ${\delta}_{\overline{S}}(w')< n_1-n$ then $S_w$ is a defensive -alliance, but this is impossible, too. So, we can assume that $\overline{S} \subseteq N(w')$. Notice that if there is $u'\in S$ with $d(u',\{w,w'\})\ge 2$, then we can check that $S\setminus\{u'\}$ is a defensive -alliance, which is impossible. Thus, we can assume that $S\subseteq N(w)\cup N(w')$; in fact, $$n-2=|S\setminus\{w,w'\}| \le {\delta}_{S\setminus\{w'\}}(w) + {\delta}_{S\setminus\{w\}}(w') = 2[{\Delta}_1 -(n_1-n)-1].$$ Since $S\subseteq N(w)\cup N(w')$, if $n-2=2[{\Delta}_1 -(n_1-n)-1]$ then $S\cap N(w)\cap N(w')=\emptyset$, and $${\delta}_{S\setminus\{w,w'\}}(v) \ge {\Delta}_1-(n_1-n) \text{ and } {\delta}_{\overline{S\setminus\{w,w'\}}}(v) \le n_1-n,\quad \text{for every } v\in S\setminus\{w,w'\}.$$ Hence, $S\setminus\{w,w'\}$ is a defensive -alliance, which is impossible. Then $n-2 < 2[{\Delta}_1 -(n_1-n)-1]$ and this finishes the proof.
\[l:An+D-2>n\] Let $G_1$ be a graph with minimum and maximum degree $2$ and ${\Delta}_1$, respectively, and let $n\ge3$ be a fixed natural number. Assume that $G_1$ has order $n_1 > n$ with exactly $n$ vertices of degree ${\Delta}_1$, and such that its alliance polynomial $A(G_1;x)$ is symmetric. If $n< 2[{\Delta}_1 -(n_1-n)]$ and $A(G_1;x)$ is a monic polynomial of degree $2n-n_1+{\Delta}_1$, then $A_{2(n-n_1)+{\Delta}_1-2}(G_1)> n$.
By hypothesis, there exist different vertices $v_1,\ldots,v_n$ in $G_1$ with degree ${\Delta}_1$. The arguments in the proof of Lemma \[l:d1=1\] give that $G_1$ is a connected graph where $S:=\{v_1,\ldots,v_n\}$ is the unique exact defensive -alliance in $G_1$ and there are $w,w'\in S$ with $\overline{S}\subset N(w)\cap N(w')$. Note that $S_u:=S\setminus\{u\}$ is a defensive -alliance for any $u\in S$, since for all $v\in S_u$ we have $${\delta}_{S_u}(v)\ge {\Delta}_1 - \big|\overline{S_u}\big| \quad \text{ and } \quad {\delta}_{\overline{S_u}}(v)\le \big|\overline{S_u}\big|=n_1-n+1.$$ Note that ${\delta}_{S}(v)\ge {\Delta}_1-(n_1-n) > n/2$ for every $v\in S$. Since $\langle S\rangle$ is Hamiltonian by Theorem \[Dirac\], we have that $S_u$ induces a connected subgraph for any $u\in S$. Since $S$ is the unique exact defensive -alliance in $G_1$, $S_u$ is an exact defensive -alliance for any $u\in S$. Therefore, we have $A_{{\Delta}_1-2(n_1-n)-2}(G_1)\ge n$.
Denote by $u'$ a vertex of $G_1$ with ${\delta}(u')=2$. Since $v\nsim u'$ for any $v\in S\setminus\{w,w'\}$ we have $|S|-1\ge {\delta}_{S}(v)\ge{\delta}_{S}(w)+1$, and so, ${\delta}_S(w)\le |S|-2$ and there are $u_1,u_2\in S\setminus\{w,w'\}$ with $u_1,u_2\notin N(w)$; then $u_1,u_2\notin N(w)\cap N(w')$. Note that $S\setminus\{u_1,u_2\}$ is a defensive -alliance in $G_1$, since $${\delta}_{S\setminus\{u_1,u_2\}}(w) - {\delta}_{\overline{S\setminus\{u_1,u_2\}}}(w) = {\Delta}_1 - 2{\delta}_{\overline{S\setminus\{u_1,u_2\}}}(w) \ge {\Delta}_1 - 2(n_1-n+1),$$ $${\delta}_{S\setminus\{u_1,u_2\}}(w') - {\delta}_{\overline{S\setminus\{u_1,u_2\}}}(w') = {\Delta}_1 - 2{\delta}_{\overline{S\setminus\{u_1,u_2\}}}(w') \ge {\Delta}_1 - 2(n_1-n+1),$$ and $${\delta}_{S\setminus\{u_1,u_2\}}(v) - {\delta}_{\overline{S\setminus\{u_1,u_2\}}}(v) \ge {\Delta}_1 - 2( n_1-n+1) \quad \text{ for all } v\in S\setminus\{u_1,u_2,w,w'\}.$$ Then $S\setminus\{u_1,u_2\}$ is an exact defensive -alliance and this finishes the proof.
\[t:Reg4-6\] Let $G$ be a connected ${\Delta}$-regular graph with ${\Delta}\le5$ and $G^*$ another graph. If $A(G^*;x) = A(G;x)$, then $G^*$ is a ${\Delta}$-regular graph with the same order and size of $G$.
If $0\le{\Delta}\le3$, then the result follows from Theorem \[t:Reg0-3\]. Assume that $4\le{\Delta}\le5$. Let $n,n_1$ be the orders of $G,G^*$, respectively, and let ${\delta}_1,{\Delta}_1$ be the minimum and maximum degree of $G^*$, respectively. By Theorem \[t:RegCharact\], $G^*$ is a connected graph and $n_1\ge n$. Seeking for a contradiction assume that $n_1> n$.
Assume first ${\Delta}=4$. By Theorem \[t:RegCharact\] we have $n_1=n+{\Delta}_1-4$, ${\Delta}_1>4$ and ${\Delta}_1+{\delta}_1\le 6$. Thus, we have ${\Delta}_1=5$ and ${\delta}_1=1$, and then $n_1=n+1$. Then, Theorem \[t:RegCharact\] and Lemma \[l:d1=1\] give that $A(G;x)=A(G^*;x)$ is not a monic polynomial of degree $n_1+3=n+4$. This is the contradiction we were looking for, and we conclude $n_1=n$.
Assume now ${\Delta}=5$. By Theorem \[t:RegCharact\] we have $n_1=n+{\Delta}_1-5$, ${\Delta}_1>5$, ${\Delta}_1+{\delta}_1\le 8$, ${\delta}_1+4\le{\Delta}_1$ and ${\delta}_1\equiv {\Delta}_1 (\text{mod } 2)$. Thus, we have the following cases:
Case 1
: [${\delta}_1=1$ and ${\Delta}_1=7$,]{}
Case 2
: [${\delta}_1=2$ and ${\Delta}_1=6$.]{}
Lemma \[l:d1=1\] gives that $A(G;x)$ is not a monic polynomial of degree $n+5$ in Case 1; this is the contradiction we were looking for, and we conclude $n_1=n$. In Case 2 we have $n_1=n+1$. Since $A(G;x)$ is a monic polynomial of degree $n+5$, Lemma \[l:d1=1\] gives that $n<10$. Hence, Lemma \[l:An+D-2>n\] gives that $A_{2}(G^*)> n$; however, Theorem \[t:An+D-2=n\] gives $A_{3}(G)=n$. This is the contradiction we were looking for, and we conclude $n_1=n$.
Computing the alliance polynomials for cubic graphs with small order {#sect3}
--------------------------------------------------------------------
In this subsection we compute the alliance polynomial of cubic graphs of small order by using Algorithm \[algorithm\], and find that non-isomorphic cubic graphs of order at most $10$ have different alliance polynomials. By Theorem \[t:Reg0-3\] this implies these cubic graphs are uniquely determined by their alliance polynomial. A similar study on characterization of cubic graphs with small order by their domination polynomials is done in [@AP1], although it obtains a different result.
Computing the alliance polynomial of a graph $G$ on $n$ vertices and $m$ edges by calculating $k_S$ for each connected induced subgraph $\langle S\rangle$ takes time $O(m2^n)$. On ${\Delta}$-regular graphs the complexity is $O(n2^n)$. Note that in order to decreasing this time for small size of $G$ could be used its topology by traveling each connected induced subgraph. Testing whether $\langle S\rangle$ is connected can be done using Depth-First Search (DFS), and this has time complexity $O(m)$. Finding $k_S$ requires $O(n)$ time.
Let $G$ be a cubic graph with order $n$. If $n=4$ then $G$ is isomorphic to $K_4$ and Theorem \[t:Reg0-3\] gives uniqueness. If $n=6$ then $G$ is isomorphic either to $K_{3,3}$ or to the Cartesian product $P_2\Box C_3$; hence, Theorem \[t:Reg0-3\] implies that they are uniquely determined by their alliance polynomial since $A(K_{3,3};x)=6 x^3 + 33 x^5 + 15 x^7 + x^9$ and $A(P_2\Box C_3;x)=6 x^3 + 33 x^5 + 11 x^7 + x^9$. Notice that these alliance polynomials are equal except for the coefficient of $x^7$; it is an interesting fact since many parameters of these graphs are different.
Figure \[fig:ord8\] shows the cubic graphs with order $8$ and Table \[tab:ord8\] their alliance polynomials; since they are different, Theorem \[t:Reg0-3\] gives their uniqueness.
Figure \[fig:ord10\] shows the cubic graphs with order $10$ and Table \[tab:ord10\] their alliance polynomials. Since they are different, Theorem \[t:Reg0-3\] gives their uniqueness.
We say that a graph $G$ is characterized by a graph polynomial $f$ if for every graph $G'$ such that $f(G') = f(G)$ we have that $G'$ is isomorphic to $G$. A set of graphs $K$ is characterized by a graph polynomial $f$ if every graph $G \in K$ is characterized by $f$.
\[p:uniqCub10\] Every cubic graph of order at most $10$ is characterized by its alliance polynomial.
Particularly, by Theorem \[p:AlliPoly\] ii) we have that the cubic graphs of order at most $10$ are characterized by the evaluation at $x=1$ of their alliance polynomials.
\[p:Eval1Cub10\] Two non-isomorphic cubic graphs of order at most $10$ have a different number of connected induced subgraphs.
Here we have proved that a polynomial cannot be the alliance polynomial of both a cubic and a non-cubic graph.
In [@CRST] the authors prove that paths, cycles, complete graphs, complete graphs minus an edge, and stars are all characterized by their alliance polynomials. Here we have proved that a polynomial cannot be the alliance polynomial of both a cubic and a non-cubic graph. In fact, cubic graphs of order at most $10$ are characterized by their alliance polynomials (Proposition \[p:uniqCub10\]).
In [@CRST Proposition 4.1] the authors compare the distinctive power of the alliance polynomial with other well-known graph polynomials, such as the domination polynomial [@AP], the independence polynomial [@GH], the matching polynomial [@F], the characteristic polynomial, the Tutte polynomial [@T], the bivariate chromatic polynomial [@DPT] and the subgraph component polynomial [@TAM]. In fact, this result exhibits for each of these polynomials $p(G;x)$ two graphs $G_1,G_2$ with $p(G_1;x) = p(G_2;x)$ and $A(G_1;x) \neq A(G_2;x)$.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was partly supported by a grant for Mobility of own research program at the University Carlos III de Madrid, a grant from Ministerio de Econom[í]{}a y Competitividad (MTM 2013-46374-P) Spain, and a grant from CONACYT (CONACYT-UAG I0110/62/10), Mexico.
[99]{}
S. Akbari, S. Alikhani and Y-H. Peng, Characterization of graphs using domination polynomials, [*European J. Combin.*]{} [**31**]{}(7) (2010), 1714–1724.
S. Alikhani and Y-H. Peng, Domination polynomials of cubic graphs of order 10, [*Turkish J. Math.*]{} [**35**]{}(3) (2011), 355–366.
S. Alikhani and Y-H. Peng, Introduction to Domination Polynomial of a Graph, [*Ars Comb.*]{} [**CXIV**]{} (2014), 257-266.
G. D. Birkhoff, A determinant formula for the number of ways of coloring a map, [*Ann. of Math.*]{} [**2**]{}(14) (1912), 42-46.
P. L. Bowers, Negatively curved graph and planar metrics with applications to type, [*Michigan Math. J.*]{} [**45**]{} (1998), 31-53.
J.I. Brown, K. Dilcher, R.J. Nowakowski, Roots of Independence Polynomials of Well Covered Graphs, [*J. Alg. Comb.*]{} [**11**]{}(3) (2000), 197-210.
W. Carballosa, Exact Defensive Alliances in Graphs, [*Appl. Math. Sci.*]{} [**7**]{}(74) (2013), 3673-3679.
W. Carballosa, J.C. Hernández-Gómez, O. Rosario and Y. Torres-Nuñez, Computing the strong alliance polynomial of a graph. To appear in [*Invest. Oper.*]{}
W. Carballosa, J. M. Rodríguez, J. M. Sigarreta and Y. Torres-Nuñez, Computing the alliance polynomial of a graph, to appear in [*Ars Comb.*]{} Preprint in http://arxiv.org/abs/1410.2940
W. Carballosa, J. M. Rodríguez, J. M. Sigarreta and Y. Torres-Nuñez, Distinctive power of the alliance polynomial for regular graphs, [*Electron. Notes Discrete Math.*]{} [**46**]{} (2014), 313-320.
Curtin, B., Algebraic characterizations of graph regularity conditions, [*Designs, Codes and Cryptography*]{} [**34**]{}(2-3) (2005), 241-248.
M. DeVos and B. Mohar, An analogue of the Descartes-Euler formula for infinite graph and Higuchi’s conjecture, [*Trans. Amer. Math. Soc.*]{} [**359**]{} (2007), 3275-3286.
K. Dohmen, A. Pönitz, and P. Tittmann, A new two-variable generalization of the chromatic polynomial, [*Discrete Math. Theor. Comput. Sci.*]{} [**6**]{} (2003), 69–90.
E. J. Farrell, An introduction to matching polynomials, [*J. Comb. Theory Ser. B*]{} [**27**]{} (1979), 75-86.
G. W. Flake, S. Lawrence, and C. L. Giles, Efficient Identification of Web Communities. In [*Proceedings of the 6th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining*]{} (KDD-2000), (2000), 150-160.
G. H. Fricke, L. M. Lawson, T. W. Haynes, S. M. Hedetniemi and S. T. Hedetniemi, A Note on Defensive Alliances in Graphs, [*Bull. Inst. Combin. Appl.*]{} [**38**]{} (2003), 37-41.
C. D. Godsil and I. Gutman, On the theory of the matching polynomial, [*J. Graph Theory*]{} [**5**]{} (1981), 137-144.
I. Gutman and F. Harary, Generalizations of the matching polynomial, [*Utilitas Math.*]{} [**24**]{} (1983) 97–106.
T. W. Haynes, S. T. Hedetniemi and M. A. Henning, Global defensive alliances in graphs, [*Electron. J. Combin.*]{} [**10**]{} (2003), 139-146.
T. Haynes, D. Knisley, E. Seier, and Y. Zou, A quantitative analysis of secondary RNA structure using domination based parameters on trees, [*BMC bioinformatics*]{}, [**7**]{}:108 (2006).
C. Hoede and X. Li, Clique polynomials and independent set polynomials of graphs, [*Discrete Math.*]{} [**125**]{} (1994), 219-228.
P. Kristiansen, S. M. Hedetniemi and S. T. Hedetniemi, Alliances in graphs, [*J. Combin. Math. Combin. Comput.*]{} [**48**]{} (2004), 157-177.
T. Kotek and J. A. Makowsky, Recurrence relations for graph polynomials on bi-iterative families of graphs, [*Eur. J. Combin.*]{} [**41**]{} (2014), 47-67. J. A. Makowsky and E. V. Ravve, On the location of roots of graph polynomials, [*Electron. Notes Discrete Math.*]{} [**43**]{} (2013), 201-206.
J. A. Makowsky, E. V. Ravve and N. K. Blanchard, On the location of roots of graph polynomials, [*Eur. J. Combin.*]{} [**41**]{} (2014), 1-19.
M. Powel, Alliance in graph, [*Proceeding presented in th 255 of the United state Military Academy*]{} (2004), 1350-1415.
R. C. Read, An introduction to chromatic polynomials, [*J. Comb. Theory*]{} [**4**]{}(1) (1968), 52–71.
J. M. Rodríguez, J. M. Sigarreta and Y. Torres-Nuñez, Computing the hyperbolicity constant of a cubic graph, [*Int. J. Comp. Math.*]{} [**91**]{}(9) (2014), 1897-1910.
J. A. Rodríguez-Velázquez and J. M. Sigarreta, Global defensive k-aliances in graphs, [*Discr. Appl. Math.*]{} [**157**]{}(2) (2009), 211-218.
J. M. Sigarreta, Upper k-alliances in graphs, [*Int. J. Contemp. Math. Sciences*]{} [**6**]{}(43) (2011), 2121- 2128.
J. M. Sigarreta, I. González-Yero, S. Bermudo and J. A. Rodríguez-Velázquez, Partitioning a graph into offensive k-alliances, [*Discr. Appl. Math.*]{} [**159**]{} (2011), 224-231.
P. Tittmann, I. Averbouch and J.A. Makowsky, The enumeration of vertex induced subgraphs with respect to the number of components, [*European J. Combin.*]{} [**32**]{} (7) (2011), 954–974.
Y. Torres-Nuñez, Alliance polynomial and hyperbolicity in regular graphs, Ph. D. Thesis, Universidad Carlos III de Madrid, Madrid, Spain. 2014. http://orff.uc3m.es/handle/10016/19744
W. T. Tutte, A contribution to the theory of chromatic polynomials, [*Canad. J. Math.*]{} [**6**]{}(80-91) (1954), 3-4.
| {
"pile_set_name": "ArXiv"
} |
---
author:
-
bibliography:
- 'collet.bib'
title: '**The Barium Isotopic Abundance in the Metal-Poor Star HD140283**'
---
Introduction
============
Elements beyond the Fe peak are predominantly produced through successive neutron-capture reactions in two processes known as the slow ($s$-) and the rapid ($r$-) process. The distinction between the $s$- and $r$- process depends on whether the time-scale for neutron captures is longer or shorter, respectively, than the time-scale of radioactive decay of freshly synthesized unstable nuclei. Observational evidence and theoretical studies have identified the $s$-process site in low- to intermediate-mass (${\sim}1.3$–$8$ M$_\odot$) stars in the asymptotic giant branch (AGB). In particular, the He-shell of thermally pulsating low-mass AGB stars is believed to be the site of the so-called *main* $s$-process, which synthesizes nuclides heavier than Sr [see, e.g., @busso_gall_wasserb_1999]. The $r$-process instead is usually associated with the explosive environment of Type II supernovae (SNeII), although this astrophysical site hasn’t been fully confirmed yet [see, e.g., @sneden_cowan_gallino_2008].
Massive stars that end their evolution as SNeII are shorter-lived compared with low-mass stars that eventually evolve into the AGB phase. One therefore expects the $r$-process to dominate the enrichment of the interstellar medium (ISM) in terms of heavy elements during the early stages of Galactic chemical evolution[^1]. According to this picture, the oldest most metal-poor stars should then only contain heavy elements in the relative proportions determined by the $r$-process [@truran_1981].
[@magain_1995] used a novel approach to infer the relative contributions of the $r$- and $s$-process spectroscopically by looking at the detailed shape of the Ba [ii]{} $4554$ [Å]{} resonance line. The odd Ba isotopes contribute to broaden the line and alter the symmetry of its profile via hyperfine splitting. As the $r$- and $s$-process produce rather different mixtures of odd and even Ba isotopes, the actual width and shape of the Ba [ii]{} $4554$ [Å]{} line are therefore dependent on the relative contribution of the two neutron-capture processes. Based on the analysis of solar abundances [@anders89] by [@arlandini_1999], one expects a value of $f_\mathrm{odd}^\mathrm{s}=0.11$ for the fractional abundance of the odd isotopes of barium[^2], in case of a pure $s$-process mixture, and $f_\mathrm{odd}^\mathrm{r}=0.46$ in case of a pure $r$-process one. Magain challenged the Galactic chemical evolution scenario depicted above by deriving a fractional abundance of odd Ba isotopes $f_{\mathrm{odd}}=0.06{\pm}0.06$ for the metal-poor subgiant HD140283 (\[Fe/H\]$=-2.5$): this value is fully compatible with a pure $s$-process production and excludes any significant contribution from the $r$-process. Magain’s result cannot indeed be reconciled with the predictions of standard Galactic chemical evolution models [e.g., @travaglio_1999] which indicate that the $s$-process signature in stars should become manifest only at higher metallicities (\[Fe/H\]${\gtrsim}-1.5$). More recently, [@lap_2002] re-observed and re-analysed the Ba [ii]{} $4554$ [Å]{} line in the spectrum of HD140283 and derived, in contrast to Magain, a fractional abundance of odd Ba isotopes $f_{\mathrm{odd}}=0.30{\pm}0.21$, consistent with a pure $r$-process isotopic mixture. [@lap_2002] used a spectrum of superior quality than Magain’s both in terms of resolution (${\lambda}/{\Delta}{\lambda}{\approx}200{\,}000$) and signal-to-noise ratio (S/N${\approx}550$); their analysis, however, was still based on 1D model stellar atmospheres. Classical spectroscopic analyses based on 1D model atmospheres, however, rely on a number of tunable fudge parameters. In particular, they cannot account for Doppler broadening due to photospheric convective motions without introducing the micro- and macro-turbulence parameters. Moreover, convective flows also induce asymmetries and overall wavelength shifts in spectral line profiles, which cannot be reproduced by ordinary 1D analyses.
Three-dimensional hydrodynamical simulations of stellar surface convection, on the other hand, can self-consistently predict photospheric velocity fields and correlated temperature and density inhomogeneities [@stein98; @asplund99]; line shapes, asymmetries, and wavelength shifts can be accurately reproduced by using such simulations as 3D hydrodynamical model atmospheres, without relying on ad hoc free parameters [@asplund00fe2]. In the present contribution, we re-derive the fractional abundance of the odd Ba isotopes in HD140283 by means of both a 1D and a 3D analysis of the Ba [ii]{} $4554$ [Å]{} line in the spectrum obtained by [@lap_2002] at the W. J. McDonald Observatory (Mt Locke, Texas). In particular, we use a 3D model atmosphere of the metal-poor subgiant star to adequately disentangle the contributions of photospheric convective flows and hyperfine splitting to the broadening and asymmetry of the Ba [ii]{} $4554$ [Å]{} line.
[lcc]{}\
Isotope & Wavelength ([Å]{}) & Relative strength\
$^{134}$Ba: & 4554.0314 & 1.0000\
\
$^{135}$Ba: & 4554.0003 & 0.1563\
& 4554.0015 & 0.1562\
& 4554.0019 & 0.0625\
& 4554.0473 & 0.4375\
& 4554.0500 & 0.1563\
& 4554.0512 & 0.0313\
\
$^{136}$Ba: & 4554.0317 & 1.0000\
\
$^{137}$Ba: & 4553.9975 & 0.1563\
& 4553.9986 & 0.1562\
& 4553.9988 & 0.0625\
& 4554.0498 & 0.4375\
& 4554.0531 & 0.1563\
& 4554.0542 & 0.0313\
\
$^{138}$Ba: & 4554.0330 & 1.0000\
Methods {#sec:methods}
=======
We synthesize the Ba [ii]{} $4554$ [Å]{} line profile with the help of both a 1D and a 3D model atmosphere of HD140283, varying the Ba isotopic mixture to produce the best fit to the observed feature. We account for the isotopic and hyperfine splitting components of the Ba [ii]{} $4554$ [Å]{} line listed in table \[tab:hfscomp\]. Isotopic wavelength shifts are actually too small to be resolved even with the highest-resolution spectrographs currently available: in the present analysis, therefore, we effectively do not make any distinction between the even isotopes as well as between the three odd isotopes.
For the 1D analysis, we employ a plane-parallel hydrostatic LTE [marcs]{} model atmosphere [@gustafsson75; @asplund97] of the subgiant star with the following stellar parameters: $T_\mathrm{eff}=5690$ K, ${\log}{g}=3.67$ (cgs), and \[Fe/H\]$=-2.50$. Spectral line profiles are computed under the assumption of LTE. We stress that, contrary to the overall barium abundance, isotopic abundances derived from the Ba [ii]{} $4554$ [Å]{} line are expected to be insensitive to departures from LTE and to the actual choice of stellar parameters. In the 1D calculations, we adopt a micro-turbulence $\xi=1.49$ , based on the LTE analysis of the $51$ Fe [i]{} and $13$ Fe [ii]{} lines given in Tab. 2 of [@lap_2002]. With the above value, the iron abundance becomes independent from the equivalent width and chemical equilibrium is fulfilled (${\log}{\epsilon}(\mathrm{Fe~{I}}) =4.98{\pm}0.13$ and ${\log}{\epsilon}(\mathrm{Fe~{II}})=4.98{\pm}0.11$).
Before proceeding with the synthesis of the Ba [ii]{} $4554$ [Å]{} line we need to estimate the broadening from mechanisms other than hyperfine splitting. In order to do this, we fit the Fe lines from the same sample used by [@lap_2002] (Tab. 3 in their article) by assuming a rotational velocity $v_\mathrm{rot}{\sin}{i}=0.5$ and varying the iron abundance, the central wavelength, and the FWHM of a Gaussian which we use toconvolve the line profiles. The convolution with a Gaussian accounts for macro-turbulent and instrumental broadening (and, possibly, residual rotational broadening). We account for natural and linear Stark broadening and we model collisional broadening with neutral hydrogen atoms using the quantum mechanical calculations by [@barklem00]. From the analysis of the Fe lines, we derive an average value of $4.87{\pm}0.11$ for the FWHM of the Gaussian broadening. Using this estimate of the Gaussian broadening, we compute synthetic flux profiles of the Ba [ii]{} $4554$ [Å]{} line for different barium isotopic mixtures. We quantify the comparison between the theoretical and observed profiles by means of a $\chi^2$-analysis similar to the one carried out by [@asplund99] to investigate $^6$Li/$^7$Li ratios in metal-poor halo stars. We compute the $\chi^2$ according to the expression $\chi^2={\sum}(O_\mathrm{i}-S_\mathrm{i})^2/\sigma^2$, with $O_\mathrm{i}$ and $S_\mathrm{i}$ denoting the observed and synthetic flux at wavelength point $i$, respectively, and $\sigma$ is the inverse signal-to-noise ratio. The most likely isotopic mix is the one that minimizes $\chi^2$.
![*Upper panel*: Synthetic (*continuous black line*) versus observed (*grey symbols*) profiles of the Ba [ii]{} line at $4554.0$ [Å]{} for the best fitting Ba isotopic mix in the 1D analysis. *Lower panel*: Relative difference between observed and synthetic profiles.[]{data-label="fig:bestfit_1d"}](HD140283_Ba.eps)
![*Thick continuous line*: Resulting $\Delta\chi^2=\chi^2-\chi_{\mathrm{min}}^2$ from the fit of the Ba [ii]{} $4554$ [Å]{} line as a function of the fractional abundance of odd Ba isotopes for the combined Gaussian (FWHM$=4.87$ ) of the mean macro-turbulence and instrumental broadening in the 1D case; also shown are the cases for FWHM${\pm}{\Delta}{\mathrm{FWHM}}$ (*thick dashed* and *dot-dashed lines*). A value $v_{\mathrm{rot}}{\sin}{i}=0.5$ is adopted for the rotational velocity. *Thick grey lines*: fractional abundances of odd Ba isotopes in case of pure $s$-process and $r$-process isotopic mixes (S. Bisterzo 2008, private communication). []{data-label="fig:chi2_1d"}](HD140283_fodd_chi2.eps)
For the 3D analysis, we use a simulation of stellar surface convection by [@asplund99], constructed for the same stellar parameters as for the 1D case, as a time-dependent 3D hydrodynamical model atmosphere. We select a $35$-minute long sequence of $30$ simulation snapshots sampled at regular time intervals and downgrade the resolution of the original simulation to $50{\times}50{\times}82$ grid-points prior to the spectral line formation calculations. We compute LTE flux profiles in 3D by solving the radiative transfer equation along $33$ rays (four $\mu$-angles, eight $\phi$-angles, and the vertical) for all grid-points at the surface, performing then a disk integration and a time average over all snapshots. As we already account for the effect of Doppler shifts induced by the velocity fields in the 3D simulation, we ignore altogether the micro- and macro-turbulence parameters. We then determine the amount of broadening contributed by processes other than hyperfine splitting: we fit the profiles of the Fe lines similarly as in the 1D analysis above but varying the rotational velocity instead of the Gaussian broadening, which, in 3D, we keep fixed and equal to the instrumental broadening (corresponding roughly to a FWHM$=1.5$ , as estimated from the resolution of the observed spectrum). Natural and collisional broadening are included and treated in the same way as in the 1D calculations. From the 3D analysis of the Fe lines, we estimate an average rotational velocity of $2.58{\pm}0.30$ . We then follow the same $\chi^2$-minimization procedure as in the 1D analysis to determine the most probable barium isotopic mix from the fitting of the observed Ba [ii]{} $4554$ [Å]{} line profile.
Results
=======
1D analysis
-----------
Figure \[fig:bestfit\_1d\] shows the best-fitting synthetic profile to the observed Ba [ii]{} $4554.0$ [Å]{} line in the 1D case. The derived fractional abundance of odd Ba isotopes in 1D is $f_\mathrm{odd}=0.33$, which suggests a contribution predominantly from the $r$-process (namely $64$%) to the barium isotopic mix. Figure \[fig:chi2\_1d\] shows the $\Delta\chi^2=\chi^2-\chi_{\mathrm{min}}^2$ curves considered in the $\chi^2$-minimization procedure. The $1\sigma$, $2\sigma$, and $3\sigma$ confidence limits correspond to $\Delta\chi^2=1$, $4$, and $9$, respectively (indicated by the dotted lines in the figure). The half-width of the $f_\mathrm{odd}$ interval for which $\Delta\chi^2<1$ is a measure of the uncertainty in fractional abundance of odd isotopes due to the finite signal-to-noise; for S/N$=550$ the uncertainty on $f\mathrm{odd}$ is ${\pm}0.023$. Figure \[fig:chi2\_1d\] also shows how the position of the $\chi^2$ minimum depends on the Gaussian broadening: a change of ${\pm}0.11$ translates into a change of ${\mp}0.06$ in $f_\mathrm{odd}$. The fractional abundance of odd Ba isotopes is also sensitive to the choice of micro-turbulence parameter (which affects the Gaussian broadening determination) and to the adopted surface gravity (which affects pressure broadening of the lines). For the present analysis, we adopt the estimates of ${\mp}0.11$ and ${\mp}0.01$ by [@lap_2002] of the errors on $f_\mathrm{odd}$ due to the uncertainties on micro-turbulence (${\pm}0.2$ ) and ${\log}{g}$ (${\pm}0.2$ [dex]{}), respectively. Our estimated total uncertainty on $f_\mathrm{odd}$ is then ${\Delta}f_\mathrm{odd}=$$\sqrt{0.023^2+0.06^2+0.11^2+0.01^2}$ ${\approx}0.13$ which directly corresponds to an uncertainty in the contribution of $r$-process fraction of Ba isotopes of about ${\pm}36$%.
3D analysis
-----------
![Same as Fig. \[fig:bestfit\_3d\] but for the best fitting Ba isotopic mix in the 3D analysis.[]{data-label="fig:bestfit_3d"}](f140283_Ba.eps)
![*Thick continuous line*: Resulting $\Delta\chi^2$ from the fit of the Ba [ii]{} $4554$ [Å]{} line as a function of the fractional abundance of odd Ba isotopes for a rotational velocity ${v_\mathrm{rot}\sin{i}}=2.58$ in the 3D case; also shown are the cases for ${v_\mathrm{rot}\sin{i}}{\pm}{\Delta}{v_\mathrm{rot}\sin{i}}$ (*thick dashed* and *dot-dashed lines*). A value FWHM$=1.5$ is assumed for the instrumental broadening.[]{data-label="fig:chi2_3d"}](f140283_fodd_chi2.eps)
The best-fitting synthetic Ba [ii]{} $4554.0$ [Å]{} line profile in the 3D analysis is shown in Fig. \[fig:bestfit\_3d\]. It is clear from the comparison of the residuals, that the synthetic profile in 3D gives a significantly better fit to the observed spectral line than the synthetic 1D profile. The estimated fractional abundance of odd Ba isotopes in the 3D analysis is $f_\mathrm{odd}=0.15$, which —contrary to the 1D case— points toward a predominantly $s$-process isotopic mix (only $12$% contribution from the $r$-process). Figure \[fig:chi2\_3d\] shows the $\Delta\chi^2=\chi^2-\chi_{\mathrm{min}}^2$ curves considered for the $\chi^2$-minimization procedure in the 3D analysis. Similarly as for the 1D case, we use the information in the figure to quantify the uncertainties on $f_\mathrm{odd}$ due to the finite signal-to-noise and the error on the estimated rotational velocity: these are $0.024$ and $0.12$, respectively. For the error on $f_\mathrm{odd}$ due to the uncertainty on surface gravity, we adopt the same value as in the 1D case. We caution that a change in surface gravity in practice also affects the velocities predicted by the simulation. In this preliminary 3D analysis of the Ba [ii]{} $4554.0$ [Å]{} line, however, we neglect this aspect; we defer the study of the effects of changes in stellar parameters on the estimate of the uncertainty on $f_\mathrm{odd}$ to a future and more comprehensive work. Finally, we do not consider any contribution to the error on $f_\mathrm{odd}$ from micro-turbulence, since this parameter is absent in the 3D analysis. The estimated total uncertainty on $f_\mathrm{odd}$ in the 3D analysis is then ${\Delta}f_\mathrm{odd}=$ $\sqrt{0.024^2+0.12^2+0.01^2}$ ${\approx}0.12$ which directly corresponds to an uncertainty in the contribution of $r$-process fraction of about ${\pm}34$%.
Discussion
==========
The fractional abundance $f_\mathrm{odd}=0.33{\pm}0.13$ of odd Ba isotopes derived in the present 1D analysis is in excellent agreement with the result obtained by [@lap_2002]. Contrary to the finding of [@magain_1995], our 1D estimate of $f_\mathrm{odd}$ indicates a large contribution from the $r$-process ($64$%) to the barium isotopic mix in HD140283 and seems to exclude at a 3$\sigma$ level a pure $s$-process contribution.
Interestingly, the 3D analysis of the same feature leads to a radically different result, namely that the Ba [ii]{} $4554.0$ [Å]{} line profile is best-fitted assuming instead a solar-like barium isotopic mix with only a $15$% (${\pm}34$%) contribution from the $r$-process. Moreover, in 3D, a pure $r$-process isotopic mixture can be excluded at a 2$\sigma$ level from the $\chi^2$ analysis. The reason for the difference between the results of the 1D and 3D analyses is essentially due to the differences between the 1D and 3D modelling of line broadening and asymmetries as sketched in Sec. \[sec:methods\]. For instance, in the 3D calculations, the simulation’s velocity fields naturally induce a “C-shaped” asymmetry in the flux profile of the line. This implies that a different fraction of odd barium isotopes is necessary in 3D comparing to the 1D case to properly model the shape of the Ba [ii]{} $4554.0$ [Å]{} line profile. We would like to caution at this point the reader that the our result for HD140283 cannot be applied straightforwardly to the 1D analysis of the Ba [ii]{} $4554.0$ [Å]{} feature in the spectra of other stars. The sign and magnitude of the 3D$-$1D correction to the derived fractional abundance of odd isotopes may depend in general on the strength and detailed shape of the observed Ba [ii]{} $4554.0$ [Å]{} profile.
Our 3D analysis of the Ba [ii]{} $4554.0$ [Å]{} feature in the spectrum of HD140283 seems unable to settle down the controversy raised by [@magain_1995] in favour of standard Galactic chemical evolution scenarios. On the contrary, our result seems to corroborate the possibility that the metal-poor star HD140283 possesses a strong $s$-process signature in terms of isotopic Ba abundance against the expectations from theoretical models of Galactic chemical evolution. We would like to stress however that, although our and Magain’s result are apparently in agreement with each other, our finding is based on rather different premises. First, we rely on a higher-resolution and higher signal-to-noise spectrum, Second, we use a 3D hydrodynamical model atmosphere of the star to synthesize the profiles of the Ba [ii]{} $4554.0$ [Å]{} feature and other Fe lines, which implies that our modelling of non-thermal broadening is more robust and, contrary to the 1D analysis, independent on the tunable micro- and macro-turbulence parameters.
As a final note, we cannot completely rule out the possibility that HD140283 has been polluted by $s$-process material from an AGB star and may not be representative of halo stars at the same metallicity in terms of isotopic Ba abundance. Further investigation is therefore necessary to draw any significant comparison with Galactic chemical evolution models. We intend to extend in an upcoming paper the 3D$-$1D analysis of the Ba [ii]{} $4554.0$ [Å]{} feature to a larger sample of halo stars and study in particular the barium isotopic abundance as a function of metallicity.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank C. Allende Prieto for kindly making parts of the McDonald spectrum of HD140283 available to us for the present analysis.
[^1]: Assuming, of course, that SNeII indeed are the site of the $r$-process
[^2]: $f_\mathrm{odd}
{\equiv} [ N(^{135}\mathrm{Ba}) + N(^{137}\mathrm{Ba}) ] / N(\mathrm{Ba})$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Muon spin rotation/relaxation spectroscopy has been employed to study electron localization around a donor center — the positive muon — in the $3d$ magnetic spinel semiconductor CdCr$_2$Se$_4$ at temperatures from 2 to 300 K in magnetic fields up to 7 T. A bound state of an electron around a positive muon — a magnetic polaron — is detected far above the ferromagnetic transition up to 300 K. Electron localization into a magnetic polaron occurs due to its strong exchange interaction with the magnetic $3d$ electrons of local Cr$^{3+}$ ions, which confines its wave function within $R \approx 0.3$ nm, allowing significant overlap with both the nearest and next nearest shells of Cr ions.'
author:
- 'Vyacheslav G. Storchak'
- 'Jess H. Brewer'
- 'Peter L. Russo'
- 'Scott L. Stubbs'
- 'Oleg E. Parfenov'
- 'Roger L. Lichti'
- 'Tel’man G. Aminov'
date: 10 August 2009
title: |
**Bound Magnetic Polarons in the $3d$-electron\
Ferromagnetic Spinel Semiconductor CdCr$_2$Se$_4$**
---
The semiconductors currently in use as working media in electronics and information technology (Si, Ge, GaAs [*etc.*]{}) are nonmagnetic; therefore the spin of the carriers has so far played a minor role in semiconductor devices. One way to enhance spin-related phenomena for semiconductor spintronics applications [@Prinz1998; @Wolf2001] is to incorporate magnetic ions (typically Mn) into nonmagnetic semiconductors to realize dilute magnetic semiconductors (DMS) [@Furdyna1988; @Ohno1998]. The interplay between electric and magnetic properties in ferromagnetic (FM) Mn-doped III-V DMS has recently been demonstrated [@Ohno1999; @Tang2003; @Chiba2003]. Combined with nonmagnetic semiconductors, these DMS may also serve as polarized spin injectors in spintronics devices [@Fiederling1999; @Ohno1999a]. The FM in these $p$-type materials results from a long-range coupling between the Mn atoms mediated by holes generated by Mn substitution at the trivalent cation site [@Jungwirth2006].
Unfortunately, the ferromagnetism in III-V Mn-doped DMS is limited by low concentration of magnetic ions. Molecular beam epitaxy results in a non-equilibrium enhancement of the otherwise low solubility of transition metals in III-V hosts, but still allows incorporation of no more than about 7-8% of Mn atoms; above this critical concentration Mn tends to cluster and phase separate [@Matsukura2002]. Even at lower concentrations, spatial homogeneity may be affected by adding Mn [@Jungwirth2006] and nanoscale-range magnetic inhomogeneities can occur [@Storchak2008]. By contrast, intrinsic magnetic semiconductors (MS) such as the $4f$ Eu chalcogenides or $3d$ Cr spinels exhibit spontaneous homogeneous ferromagnetic order without any doping. When doped, these MS show semiconducting behavior (both $n$- and $p$-type), which indicates strong mutual influence between electrical and magnetic properties [@Coey1999; @Nagaev2002]. Successful demonstration of the epitaxial growth of EuO and CdCr$_2$Se$_4$ on technologically important semiconductors Si, GaN, GaAs and GaP [@Schmehl2007; @Park2002; @Kioseoglou2004] makes them very attractive working media for spintronics applications. These materials offer several important advantages over DMS, such as higher magnetization, spatial homogeneity and wider ranges of conductivity tunable by doping. The longer spin lifetimes and spin-scattering lengths of electrons in Si, GaAs and GaN [@Kioseoglou2004; @Kikkawa1999; @Oestreich1999], as well as much higher electron mobilities compared with those of holes, makes the ability to support $n$-type conductivity in MS especially attractive for semiconductor spintronics.
Since charge and spin transport for electrons in Si, GaAs, GaP and GaN is excellent, the effectiveness of prospective all-semiconductor spintronics devices is determined by the electron transport in doped magnetic semiconductors. These materials, however, can support states that lead to severe electron localization. In fact, MS provide optimal conditions for the formation of a new type of quasiparticle — the [*magnetic polaron*]{} (MP) — in which conduction electron “autolocalization” stabilizes an atomic-scale region of the ferromagnetic phase well above $T_c$ [@deGennes1960; @Kasuya1968; @Nagaev2002]. This electron localization in MS profoundly modifies their magnetic, electrical and optical properties. In particular, such MP determine most of the transport properties of magnetic semiconductors, leading to metal-insulator transitions with a remarkable resistivity change of up to 13 orders of magnitude (in doped EuO) and colossal magnetoresistance which suppresses resistivity by 3-4 orders of magnitude in magnetic fields of $\sim 10$ T [@Nagaev2002]. Measurements of both resistivity and Hall effect in the ferromagnetic spinel CdCr$_2$Se$_4$ clearly show that these remarkable properties reflect changes in the density, not the mobility, of charge carriers [@Nagaev2002]. These effects can be explained in terms of electron localization into entities roughly the size of a unit cell: MP. In fact, the MP concept now forms the basis for numerous studies of MS and related materials [@vonMolnar2007]. A magnetic polaron is formed by an electron localized due to its strong exchange interaction $J$ with magnetic ions in its immediate environment, whose direct coupling is rather weak. Of relevance to the current study is the [*bound*]{} MP, in which the increase in the kinetic energy of the electron due to localization is compensated by [*both*]{} the $s$-$d(f)$ exchange interaction $J$ [*and*]{} the Coulomb interaction with the corresponding donor, so that the net change in the energy $$\Delta F = \frac{\hbar^2}{2m^* R^2}
- J \frac{a^3}{R^3}
- \frac{e^2}{\varepsilon R}
\label{eq:FreeEnergy}$$ has a minimum as a function of $R$ — thus determining the radius of the electron confinement [@Nagaev2002; @vonMolnar2007]. The dominant exchange term is optimized by maximum overlap of the MP electron with nearby magnetic ions.
Magnetic polarons have recently been detected in $4f$ magnetic semiconductors EuS, EuO, EuSe [@Storchak2009] and SmS [@Storchak2009a], following muon spin rotation/relaxation ($\mu^+$SR) [@Brewer1994] experiments in nonmagnetic semiconductors [@Storchak1997; @Storchak2003], that revealed the details of electron capture to form a muonium (Mu $\equiv \mu^+ e^-$) atom. Formation and dissociation of Mu more generally models electron capture by and release from donor centers, since a positive muon acts in this respect just like any other Coulomb attractive impurity [@Storchak2004]. In non-magnetic semiconductors, formation of such Mu atoms is driven solely by the Coulomb interaction; in magnetic semiconductors, it is the combined effect of the Coulomb and exchange interactions which drives Mu formation [@Storchak2009; @Storchak2009a] In treating the $s-d(f)$ exchange interaction, two limiting cases are important: when the $s$-electron bandwidth $W$ is large compared with $J$, and when $J \gg W$. The former case is typical for the $s-f$ exchange in rare-earth compounds where the extremely localized $f$-electrons are screened by electrons of other shells. In particular, in Eu compounds the $s$-electron delocalized in the rather wide (a few eV) hybridized $5d$-$6s$ band exchanges with the partially filled inner $4f$ shell, which is separated from the band states by completely filled $5s$ and $6p$ shells, thus reducing $J$ to a few tenths of an eV. The opposite case $W \ll J$ [@Anderson1955] provides the basis for the well-known double exchange in transition metal compounds. Perhaps surprisingly, this inequality is likely to be quite realistic in such compounds where the charge carriers are often of the same $d$-type as the localized spins of the magnetic ions. In particular, in CdCr$_2$Se$_4$ the Fermi level falls in the middle of the narrow $d$-band which lies just below a much broader unoccupied $s$-$p$ conduction band [@Continenza1994].
In this Letter, we present experimental evidence of severe electron localization into a magnetic polaron bound to a positive muon in the paramagnetic phase of the $3d$-electron magnetic semiconductor CdCr$_2$Se$_4$.
CdCr$_2$Se$_4$ is a chalcogenide spinel with cubic symmetry (56 atoms per unit cell), a lattice constant of 10.72 Å and a direct gap of about 1.5 eV. Its magnetic moment per chemical formula is close to 6 $\mu_B$, which corresponds to the sum of two Cr$^{3+}$ moments, each having a moment of 3 $\mu_B$. Ordering of the Cr$^{3+}$ moments into the FM phase occurs at $T_c = 130$ K. Above $T_c$ its paramagnetic susceptibility exhibits Curie-Weiss behavior.
Single crystals of CdCr$_2$Se$_4$ for the current study were grown by the closed-tube vapor transport technique. They all have perfect octahedral shape with typical sizes of 3-5 mm. They are slightly $n$-type with carrier concentration of about $10^{18}$ cm$^{-3}$ at room temperature. Magnetization (SQUID) measurements in $H=50$ Oe were used to determine $T_c = 130$ K for these crystals, in close agreement with literature data. Time-differential $\mu^+$SR experiments using 100% spin-polarized positive muons were carried out on the M15 surface muon channel at TRIUMF using the [*HiTime*]{} $\mu^+$SR spectrometer.
{width="1.35\columnwidth"}
\[fig:1-3lines\]
In a high magnetic field applied transverse to the initial muon polarization, the TF-$\mu^+$SR spectra exhibit two Mu-like signals shifted to higher frequencies relative to the narrow line positioned exactly at the diamagnetic frequency $\nu_\mu = \gamma_\mu B/2\pi$ (where $\gamma_\mu = 2\pi \times 135.5$ MHz/T is the muon gyromagnetic ratio and $B$ is the magnetic field) as shown in Fig. 1. This narrow line is a background signal from muons stopped outside the sample and provides a good reference since it does not depend on temperature. The two broader signals (blue and red online) at higher frequencies that depend on both temperature and magnetic field present a characteristic signature of the muon-electron bound state. For a $\mu^+ e^-$ (Mu) spin system \[as for $p^+ e^-$ (H)\] governed by the Breit-Rabi Hamiltonian, these signals correspond to two muon spin-flip transitions between states with fixed electron spin orientation [@Brewer1994].
Accordingly, the rotating reference frame fits of the $\mu^+$SR spectra in the time domain at various temperatures show 3-frequency precession (1 background and 2 Mu-like signals, Fig. 2). The evolution of these signals is presented in the inset to Fig. 3. We claim that the two Mu-like lines are the spectroscopic signature of the magnetic polaron — one electron localized around the positive muon by the combined effects of Coulomb and exchange interactions. In CdCr$_2$Se$_4$ this muon-bound MP forms at temperatures well above $T_c$, up to at least 300 K. The shift of the centroid of the two-line MP spectrum, the splitting and linewidths all scale with the bulk magnetic susceptibility over the 160-300 K temperature range, implying a common origin — magnetic polarons and their orientational dynamics.
{width="0.75\columnwidth"}
\[fig:2\]
Similar 2-frequency signals originating from magnetic polarons are detected in the $4f$ magnetic semiconductors EuS, EuO, EuSe [@Storchak2009] and SmS [@Storchak2009a]. In all MS studied, the MP lines are shifted with respect to the reference signal, reflecting the local ferromagnetic environment around the muon. In Eu chalcogenides, the MP lines exhibit negative shifts with respect to the reference signal, indicating that the MP electron spin is opposite to the net polaron spin. Addition of the MP electron effectively reduces the spin of the neighboring Eu ion from 7/2 (according to Hund’s rule, the maximum allowed spin is for a half-filled $f$-shell) to 3 due to Pauli exclusion. By contrast, in Cr spinels the MP electron is bound to have its spin parallel to the net MP spin (Hund’s rule for a less than half-filled $d$-shell), which effectively increases the spin of one of the neighboring Cr ions from its original value of 3/2 to 2. Accordingly, in CdCr$_2$Se$_4$ the FM shift is positive (see Fig. 1). The absolute value of this shift at room temperature and 1 T is about 0.007 T.
The splitting $\Delta\nu$ between the two MP lines provides information on the muon-electron hyperfine coupling $A$, which is determined by the probability density of the electron wavefunction at the muon [@Brewer1994]. Figure 3 shows the temperature dependence of this splitting. A distinctive feature of magnetic semiconductors in general, and CdCr$_2$Se$_4$ in particular, is a strong dependence of the conduction electron energy on the magnetization due to the exchange interaction between the mobile electron and localized $d(f)$ spins, the minimum electron energy being achieved at the ferromagnetic ordering [@Nagaev2002]. In the paramagnetic state, an “extra” electron tends to establish and support this ordering, thus forming a FM “droplet” (MP) over the extent of its wave function. The exchange contribution to the localization \[see Eq. (\[eq:FreeEnergy\])\] amounts to the difference between the paramagnetic disorder of the CdCr$_2$Se$_4$ and the enhanced (FM) order in the MP. By contrast, in the FM state the exchange contribution to the localization is negligible, as the lattice spins are already aligned. In fact, as the magnetization develops towards low temperature, the exchange contribution to electron localization diminishes and therefore can no longer compensate the increase of its kinetic energy. The electron thus avoids strong localization as the temperature approaches $T_c$ from above. Accordingly, we do not detect the MP lines below about 150 K (see Fig. 3).
\[t\]
\[fig:3\]
Temperature and magnetic field dependences of the MP signal frequency splitting $\Delta\nu$ provide information on the characteristic size (the localization radius for electron confinement) of the MP in CdCr$_2$Se$_4$. Since the Mu electron spin is “locked” to the MP moment and the experiment samples the component along the applied field, the observed splitting scales with the mean orientation of the MP moment with respect to the external field, reaching the intrinsic Mu hyperfine constant only for complete alignment (see Appendix of Ref.[@Storchak2009]). Amplitudes for the two-frequency spectrum qualitatively scale as expected for a bound electron whose spin is locked to that of the polaron core, providing strong support for this picture. Within a mean field approximation, $\Delta\nu$ is thus proportional to a Brillouin function [@Smart1966; @Storchak2009a] . For $g \, \mu_{_{\rm B}} B \ll k_{_{\rm B}} T$, this is linear in $B/(T-\Theta)$ (see Fig. 4): $$\Delta\nu = A \left[
\frac{g \, \mu_{_{\rm B}} B}{3k_{_{\rm B}} (T-\Theta)}
\right] ({\cal S}+1) \; ,
\label{eq:Splitting}$$ where $\Theta=140$ K is the paramagnetic Curie temperature of CdCr$_2$Se$_4$ and ${\cal S}$ is the net polaron spin. At low $T - \Theta$ and high $B$, $\Delta\nu$ saturates at a value of $A$ [@Brewer1994; @Storchak2009; @Storchak2009a].
The vacuum state of a Mu atom is characterized by its hyperfine coupling $A_{\rm vac} = 4463$ MHz which corresponds to electron confinement within $R_{\rm Bohr} = 0.0529$ nm. In a solid nonmagnetic medium, Mu usually has $A < A_{\rm vac}$ so that, in a magnetic field high enough to satisfy $\gamma_\mu B/2\pi \gg A$, the splitting is $\Delta \nu = A$, independent of temperature and magnetic field [@Brewer1994; @Storchak2009; @Storchak2009a]. Figure 4 shows the MP frequency splitting in CdCr$_2$Se$_4$ as a function of both $1/(T-\Theta$) and $H$. Saturation is clearly seen in both the temperature dependence and the magnetic field dependence at $T=200$ K: on both plots $\Delta\nu$ levels off to give $A = 24{\pm}2$ MHz.
\[fig:4\]
This value of $A$ gives a measure of the electron confinement around the muon at T=200 K: assuming an expanded hydrogen-like MP wavefunction, the hyperfine coupling $A$ scales as $1/R^3$, where $R$ is the characteristic radius of the corresponding $1s$ electron wave function. We find $R \approx 0.3$ nm, which falls between the nearest neighbor (NN) and next nearest neighbor (NNN) ion coordination spheres. This value of $R$ is about an order of magnitude less than that calculated in [@Yanase1972] for temperature just above T$_c$. It is known, however, that $R$ grows very fast as one approaches T$_c$ from above [@Nagaev2002; @Kasuya1968]. The value of the MP spin extracted from the slopes of both linear dependences of $\Delta\nu$ on $T$ and $H$ at higher temperature and lower magnetic field (see Fig. 4 and the inset for $T=200$ K) using Eq. (\[eq:Splitting\]) amounts to ${\cal S} = 30 \pm 4$. This is reasonably consistent with a fully polarized core of 4 NN and 12 NNN Cr$^{3+}$ ions, yielding a net spin of 24, plus a partially ordered halo.
Here we find an important difference between MP in $4f$ and $3d$ MS: while in $4f$ MS the radius of the electron confinement is strongly “glued” to the corresponding NN coordination sphere [@Storchak2009a; @Storchak2009] because of the requirement of maximum electron overlap with extremely localized ($\sim 0.03$ nm) $4f$ ion electrons, in $3d$ MS this requirement is much less stringent because of the significantly larger “spread” of the $3d$ wavefunction. In CdCr$_2$Se$_4$, therefore, $R$ falls in between the NN and NNN Cr$^{3+}$ ions. More generally, exchange-driven electron localization in $3d$ magnets might enhance FM coupling between host magnetic ions by formation of MP around impurity atoms.
This work was supported by the Natural Sciences and Engineering Research Council of Canada and the U.S. Department of Energy (Grant DE-SC0001769).
[99]{}
G.A. Prinz, [*Science*]{} [**282**]{}, 1660 (1998).
S.A. Wolf [*et al.*]{}, [*Science*]{} [**294**]{}, 1488 (2001).
J.K. Furdyna, [*J. Appl. Phys.*]{} [**264**]{}, R29 (1988).
H. Ohno, [*Science*]{} [**281**]{}, 951 (1998).
H. Ohno [*et al.*]{}, [*Nature*]{} [**408**]{}, 944 (1999).
H.X. Tang [*et al.*]{}, [*Phys.Rev.Lett.*]{} [**90**]{}, 107201 (2003).
D. Chiba [*et al.*]{}, [*Science*]{} [**301**]{}, 943 (2003).
R. Fiederling [*et al.,*]{} [Nature]{} [**402**]{}, 787 (1999).
Y. Ohno [*et al.,*]{} [Nature]{} [**402**]{}, 790 (1999).
T. Jungwirth [*et al.,*]{} [Rev. Mod. Phys.]{} [**78**]{}, 809 (2006).
F. Matsukura, H. Ohno and T. Dietl, in [*Handbook of Magnetic Materials*]{}, (Ed. K.H.J. Buschow), 2002.
V.G. Storchak [*et al.,*]{} [Phys. Rev. Lett.]{} [**101**]{}, 027202 (2008).
J.M.D. Coey [*et al.,*]{} [Adv. Phys.]{} [**48**]{}, 167 (1999).
E.L. Nagaev, in [Magnetic Semiconductors]{} (London: Imperial College Press, 2002).
A. Schmehl [*et al.,*]{} [Nat. Mater.]{} [**6**]{}, 882 (2007).
Y.D. Park [*et al.,*]{} [Appl. Phys. Lett.]{} [**81**]{}, 1471 (2002).
G. Kioseoglou [*et al.,*]{} [Nat. Mater.]{} [**3**]{}, 799 (2004).
J. Kikkawa and D. Awschalom, [Nature]{} [**397**]{}, 139 (1999).
M. Oestreich, [Nature]{} [**402**]{}, 735 (1999).
P.G. de Gennes, [Phys. Rev.]{} [**118**]{}, 141 (1960).
T.Kasuya and A.Yanase, [Rev. Mod. Phys.]{} [**40**]{}, 684 (1968).
S. von Moln[á]{}r and P.A. Stampe, in [Handbook of Magnetism and Advanced Magnetic Materials]{}, eds. H. Kronmueller and S. Parkin (John Wiley & Sons, Ltd., 2007).
V.G. Storchak [*et al.,*]{} [Phys. Rev. B]{} [**80**]{}, 235203 (2009); [Phys. B]{} [**404**]{}, 899 (2009).
V.G. Storchak [*et al.,*]{} [*Phys. Rev.*]{} [**B79**]{}, 193205 (2009).
J.H. Brewer, [Encyclopedia of Applied Physics]{} [**11**]{}, 23-53 (VCH Publishers, New York, 1994).
V.G. Storchak, [*et al.*]{}, [Phys. Rev. Lett.]{} [**78**]{}, 2835 (1997). V.G. Storchak [*et al.*]{}, [Phys. Rev. B]{} [**67**]{}, 121201 (2003). V.G. Storchak [*et al.*]{}, [J. Phys.]{}, [**16**]{}, S4761 (2004). P.W. Anderson and H. Hasegawa, [Phys. Rev.]{} [**100**]{}, 675 (1955).
A. Continenza [*et al.*]{}, [Phys. Rev. B]{} [**49**]{}, 2503 (1994).
J.S. Smart, [*Effective Field Theories of Magnetism*]{}, (Saunders, Philadelphia, 1966).
A. Yanase, [Int. J. Magn.]{} [**2**]{}, 99 (1972).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we propose a textual clue approach to help metaphor detection, in order to improve the semantic processing of this figure. The previous works in the domain studied the semantic regularities only, overlooking an obvious set of regularities. A corpus-based analysis shows the existence of surface regularities related to metaphors. These clues can be characterized by syntactic structures and lexical markers. We present an object oriented model for representing the textual clues that were found. This representation is designed to help the choice of a semantic processing, in terms of possible non-literal meanings. A prototype implementing this model is currently under development, within an incremental approach allowing step-by-step evaluations. [^1]'
author:
- |
Stéphane Ferrari\
LIMSI-CNRS\
PO Box 133\
F-91403 Orsay cédex, [France]{}\
[email protected]
title: Using textual clues to improve metaphor processing
---
Introduction {#sec:intro}
============
Metaphor is a frequently used figure of speech, reflecting common cognitive processes. Most of the previous works in Natural Language Understanding (NLU) looked for regularities only on the semantic side of this figure, as shown in a brief overview in section \[sec:overview\]. This resulted in complex semantic processings, not based on any previous robust detection, or requiring large and exhaustive knowledge bases. Our aim is to provide NLU systems with a set of heuristics for choosing the most adequate semantic processing, as well as to give some probabilistic clues for disambiguating the possibly multiple meaning representations.
A corpus-based analysis we made showed the existence of textual clues in relation with the metaphors. These clues, mostly lexical markers combined with syntactic structures, are easy to spot, and can provide a first set of detection heuristics. We propose, in section \[sec:clues\], an object oriented model for representing these clues and their properties, in order to integrate them in a NLU system. For each class, attributes give information for spoting the clues, and, when possible, the source and the target of the metaphor, using the results of a syntactic parsing. A prototype, STK, partially implementing the model, is currently under development, within an incremental approach. It is already used to evaluate the clues relevance.
In conclusion, we will discuss how the model can help chosing the adequate semantic analysis to process at the sentence level or disambiguating multiple meaning representations, providing probabilities for non-literal meanings.
Classical methods: a brief overview {#sec:overview}
===================================
The classical NLU points of view of metaphor have pointed out the multiple kinds of relations between what is called the source and the target of the metaphor, but rarely discuss the problem of detecting the figure that bears the metaphor. For our purpose, we choose to present these approaches in two main groups, depending on how they initiate the semantic processing.
The previous works led to a classification introduced by Dan Fass [@Fas91]. In the [*comparison view*]{}, the metaphor corresponds to an analogy between the structures representing the source and the target of the figure, as in Gentner’s works [@Gen88a] and their implementation [@Fal89]. The [*interaction view*]{}, as in Hobbs [@Hob91], points at the novelty brought by the metaphor. Fass also distinguishes a [*selection restrictions violations view*]{} presenting the metaphor as a kind of anomaly. We would argue that the two previous views already considered metaphor as a kind of anomaly. Indeed, the semantic analysis proposed for dealing with metaphors were processed depending on the results of another, say a “classical” one[^2]. Thereby, detecting a metaphor meant detecting an anomaly in the meaning representation issued from such a classical analysis.
Fass proposed a method for discriminating literal meanings, metaphors, metonymies and “anomalies”, merging different points of view [@Fas91]. In this approach, multiple semantic analysis can be processed, resulting in possibly multiple meaning representations. In [@Pri92b], a method to overcome similar kinds of ambiguities reveal the difficulties encountered if no previous detection is made. James Martin’s approach [@Mar92], called the [*conventional view*]{} by Fass, is based on Lakoff’s theory on cognitive metaphors [@Lak80]. It requires a specific knowledge representation base and also results in multiple representation meanings. Detecting a metaphor is meaningless here, and conventional metaphoric meanings can be viewed as polysemies. Martin revealed at least that the heuristic of the ill-formness of meaning representations issued from classical analysis is not sufficient at all to deal with all the possible metaphors.
In our point of view, all the previous approaches were founded. The main remaining problem, however, is to choose an adequate processing when confronted with a metaphor, and thus, to detect the metaphors before trying to build their meaning representation. This can be partially solved using textual clues.
Textual clues: object oriented description {#sec:clues}
==========================================
If the classical views of the metaphor overlook the textual clues, in other domains, especially those concerning explanation, they have been wisely re-introduced. In [@Per90], Pery-Woodley shows the existence of such clues related to the explanatory discourse. They can help in generating explanations in natural language as well as in modelling the student in a intelligent tutoring system [@Dan92]. A corpus of 26 explanatory texts in French, of about 200 words each, has been collected under a shared research project between psychologists and computer scientists, in order to study metaphors and analogies in teaching. The analysis we made showed the existence of textual clues in relation with metaphoric contexts and analogies (e.g. “like”, “such as”, “illustrated by”). They can be characterized by syntactic regularities (e.g. the comparative is used in structures such as “less than”, “more than”; the identification is made through attributes or appositions, ...). They also involve lexical markers (e.g. “literaly”, “illustrating”, “metaphorically”,). These properties, already found in the previous works, can help detecting the clues themselves. Studying the relation between the syntactic regularities and the lexical markers, one can observe that the first build the ground where to find the second. We thus propose an object-oriented model for representing these clues. A generic textual clue can thereby be described by the two following attributes:
- the [*Surface Syntactic Pattern*]{} representing the syntactic regularity, with a label on the item where to find the lexical marker
- the [*Lexical Marker*]{} itself
Typically, the word “metaphor” itself can be used as a lexical marker in expressions such as [*“to extend the conventional metaphor, pruning such a tree means to generalize”*]{}. On the other hand, “metaphor” will not be a marker if used as the subject of the sentence, like in this one. Thus, describing the syntactic regularities surrounding a lexical marker improves its relevance as a marker. We propose to represent this relevance for probabilistic purposes. Each clue that was found is currently evaluated on a large corpus (about 450,000 words). The frequencies of use of the lexical markers in metaphoric contexts are represented in the relevance attribute (see example below).
The syntactic structures may also give information about the source and the target of the metaphor. For instance, in the sentence [ *“Yesterday, at home, Peter threw himself on the dessert like a lion.”*]{}, the subject inherits the properties of speed and voracity of a lion attacking its victim. It is here possible to spot the source and the target of the metaphor using the syntactic properties of the comparison. Two attributes are added to textual clues related to metaphors, corresponding to the elements of the sentence bearing the source and the target.
[**Example of textual clue representations**]{}
type
: metaphor-analogy
name
: B.2.2.2
comment
: [*comparison involving the meaning of a marker, adjective, attribute of the object, object before the verb*]{}
SSP
: $GN_0 \quad GN_1 \quad V_1 \quad
Adj_0 \quad [prep] \quad GN_2$
LM
: $Adj_0$: pareil [*(meaning “similar”)*]{}
target
: $GN_1$
source
: $GN_2$
LM relevance
: (15/28)
number of occurrences
: 28
conventional metaphors
: 3
new metaphors
: 2
metaphoric contexts
: 12
total
: 15
[: $GN$ and $GV$ stand for nominal or verbal groups, $Adj$ and $Adv$ for adjectives and adverbs, and $prep$ for prepositions.]{}
The model has been partially implemented in a tool, STK, for detecting the textual clues related to metaphors and adding specific marks when found. In its current version, STK allows us to tokenize, tag, and search for lexical markers on large corpora. The tagger we use is the one developped by Eric Brill [@Bri92] with a set of tags indicating the grammatical categories as well as other information such as the number and the gender for nouns and adjectives. It is evaluated under GRACE[^3] protocol for corpus-oriented tools assigning grammatical categories. It is currently used for the evaluation of the textual clues that were found. The latter can be easily retrieved using STK, avoiding lexical ambiguities. They are then analyzed by hand, in order to determine their relevance attribute. In the previous example of textual clue, the relevance values are issued from this corpus-based analysis.
Conclusion, perspectives
========================
Classical approaches to the metaphor in NLU revealed multiple underlying processes. We therefore focussed our study on how to help detecting metaphors in order to chose the most adequate semantic processing. Textual clues can give information about the figures that bear the metaphor, which are easy to spot. Indeed, they can be found using the results of syntactic parsing. We proposed an object-oriented model to represent these clues and their multiple properties.
If textual clues give information about possible non-literal meanings, metaphors and analogies, one may argue they do not allow for a robust detection. Indeed, a textual clue is not sufficient to prove the presence of such figures of speech. The relevance of each clue can be used to help disambiguating multiple meaning representation when it occurs. This must not be the only disambiguation tool, but when no other is avalaible, it provides NLU systems with a probabilistic method.
Our future works will focuss on the study of the relation between the metaphors introduced by a clue and others that are not conventional. The guideline is that novel metaphors not introduced by a clue at the sentence level may have been introduced previously in the text.
Brill, E. (1992). In [*[Proceedings of the Third Conference on Applied Natural Language Processing]{}*]{}, [Trento]{}. [ACL]{}. Daniel, M., Nicaud, L., Prince, V., and Pery-Woodley, M. (1992). . , 608:252–260. . Falkenhainer, B., Forbus, K., and Gentner, D. (1989). . , 41:1–63. Fass, D. (1991). . , 17(1):49–90. Fass, D., Hinkelman, E., and Martin, J., editors. , [Sydney, Australia]{}. [1991]{}. Gentner, D. (1988). , , chapter 3, pages 63–88. Edited by Prieditis A., [Pitman Publishing, London, Morgan Kaufmann Publishers, Inc., Los Altos, California]{}. Hobbs, J. (1991). . In [@IJCAI91], pages 52–61. Lakoff, G. and Johnson, M. (1980). . Martin, J. (1992). . , 16:233–270. Pery-Woodley, M. (1990). Textual clues for user modeling in an intelligent tutoring system. Master’s thesis, [University of Manchester]{}, [England, Great-Britain]{}. Prince, V. and Sabah, G. (1992). . In [*[Acts of PRICAI’92]{}*]{}, [Seoul, Corea]{}. September, 1992.
[^1]: This work takes part in a research project sponsored by the AUPELF-UREF (Francophone Agency For Education and Research)
[^2]: We prefer to call it a [*classical*]{} rather than [ *literal meanings*]{} processing because it can deal with some conventional metaphors, even if not explicitly mentioned.
[^3]: GRACE stands for “Grammars and Resources for Corpora Analysis and their Evaluation”. It is a national research project for the development of tools for French language processing.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Paris Varytis
- 'Dan[-]{}Nha Huynh'
- Wladislaw Hartmann
- Wolfram Pernice
- Kurt Busch
bibliography:
- 'bibfile.bib'
title: Design study of random spectrometers for applications at optical frequencies
---
Random spectrometers are suitable for portable sensing and efficient lab-on-a-chip functionality. While previous efforts aimed at the telecom wavelength regime with silicon-based structures [@redding2013compact], random spectrometers operating in the near-infrared (NIR) spectral region where water, cells and tissue exhibit low absorption (and molecules feature low fluorescence) could offer important advantages for biological and medical applications such as Raman spectroscopy [@dhakal2016single] and bioimaging [@yodh1995spectroscopy; @gayen1996emerging]. Similar arguments can be made with respect to operations at visible frequencies.
In the present work, we analyze the performance of planar random spectrometers for frequencies in the visible, NIR and telecom ranges for both polarizations (TE and TM) by means of Mie and multiple-scattering theory in conjunction with the numerical solution of the Maxwell equations. To provide the required broadband optical transparency, we conduct our design study for silicon-nitride based waveguides. In fact, silicon nitride is a robust material [@zhao2015visible; @moss2013new] with low thermal sensitivity that is compatible with CMOS processes for low-cost mass fabrication. Further, silicon nitride is transparent in the wavelength regime between and where it features a relatively high refractive index ($n\sim2$) so that it is suitable for many applications in the visible and NIR regime [@goykhman2010ultrathin; @subramanian2012near; @romero2013silicon; @romero2013visible]. As scattering area, we consider a set of identical pores that have been etched into the silicon nitride waveguide.
In Fig. \[fig1\] we depict the considered random spectrometer geometry: We use a semicircular scattering area with a radius of $L=\SI{25} {\micro \meter}$, which consists of a random array of identical pores with radius that cover 9 % of the scattering area. A single input waveguide (with a width of ) launches light into the center of the scattering area and after multiple scattering light reaches the 13 readout waveguides (with widths of ), which are placed with equal angular distance around the outer circumference. The backscattering into the launch waveguide provides the 14th readout port. Into this two-dimensional system, we launch a fundamental mode pulse, and at the locations of the detectors D$_1$, ..., D$_{13}$ we determine the radiation that exits through the readout port. After a Fourier transform of the (time-dependent) output signals, we recover the frequency-resolved transmission matrix $T$ with a single time-domain computation. For the actual numerical computations we utilize a Discontinuous Galerkin Time-Domain (DGTD) finite-element approach [@busch2011discontinuous; @niegemann2012efficient].
To quantify the spectral resolution of the spectrometer we compute the spectral correlation function by [@redding2013compact] $$\label{SP}
C(\Delta \lambda,D_i)
=
\dfrac{\langle I(\lambda,D_i) I(\lambda+ \Delta \lambda,D_i)
\rangle}{\langle I(\lambda,D_i) \rangle
\langle I(\lambda+ \Delta \lambda,D_i) \rangle}
- 1~,$$ where $I( \lambda , D_i)$ denotes the light intensity at wavelength $\lambda$ and detector $D_i$ ($i = 1,2, ...,13$) and the average is taken over the wavelength. $C$ is averaged over all detectors and normalized at $\Delta \lambda=0$. As a result, the half-width at half-maximum (HWHM) of the correlation function provides an estimate of the spectral resolution.
![ (Color online) Schematic of the random spectrometer layout considered in this work. Radiation is launched into the center of the semicircular scattering area through an input waveguide and after multiple scattering reaches the 13 output waveguides where it is detected at the detectors D$_1$, ... D$_{13}$ (green lines), or leaks into free space. Grey-shaded regions correspond to silicon nitride, white-shaded regions correspond to free space, and red-shaded regions represent perfectly matched layers that absorb any outgoing radiation. See text for further details.[]{data-label="fig1"}](Schematic.jpg){width="\linewidth"}
Since the pore filling fraction is $f=0.09$, multiple scattering can be treated via the independent scattering approximation [@van1991speed; @van1992speed], where the multiple scattering process can be described by approriate combinations of single-scattering quantities (we will touch upon the case of higher filling ratios below). The scattering efficiency for an infinite homogeneous cylindrical scatterer at normal incidence reads $$\label{Q}
Q^\mathrm{s,p}_\mathrm{sc}
=
\dfrac{2}{kR}\Big[\vert a_{0}^\mathrm{s,p}\vert^{2}
+
2\sum_{n=1}^{\infty}\vert a_{n}^\mathrm{s,p}\vert^{2}\Big]~,$$ where $k$ is the wavenumber in the background material, $R$ denotes the radius of the cylinder, $a_{n}$ ($n = 0,1,2, ...$) represent the Mie scattering coefficients. Finally, the superscripts $\mathrm s$ and $\mathrm p$ correspond, respectively, to $\mathrm s$-wave (TM-polarization, electric field perpendicular to the $xy$-plane in Fig. \[fig1\]) and $\mathrm p$-wave (TE-polarization, electric field parallel to the $xy$-plane in Fig. \[fig1\]). The associated differential scaterring efficiency is $$\label{SA}
\dfrac{\partial Q^\mathrm{s,p}_\mathrm{sc}(\phi)}{\partial \phi}
=
\dfrac{2}{\pi kR} \bigg\vert a_{0}^\mathrm{s,p}
+
2\sum_{n=1}^{\infty} a_{n}^\mathrm{s,p}cos(n \theta)\bigg\vert^{2}~,$$ where $\theta=\pi -\phi$ is the scattering angle. The average cosine of the scattering angle defines [@hulst1957light; @bohren2008absorption] the so-called asymmetry parameter $g$: $$\label{g}
\begin{aligned}
g^\mathrm{s,p}=\langle \cos(\theta) \rangle
&=
\dfrac{ \int_{0}^{\pi}
\dfrac{\partial Q^\mathrm{s,p}_\mathrm{sc}(\phi)}{\partial \phi}
\cos(\theta) d\theta} { Q^\mathrm{s,p}_\mathrm{sc}} \\
&=
\dfrac{4}{kRQ^\mathrm{s,p}_\mathrm{sc}}\sum_{n=0}^{\infty}
a^\mathrm{s,p}_{n}(a^\mathrm{s,p}_{n+1})^{*}~.
\end{aligned}$$ For isotropic scattering, the asymmetry parameter vanishes and for predominantly forward (backward) scattering the asymmetry parameter takes on positive (negative) values. From multiple-scattering theory [@ishimaru1978wave; @arruda2016electromagnetic] it follows that for lossless scatterers the transport mean free path is given by $$\label{lt}
l^\mathrm{s,p}_\mathrm{t} = \dfrac{\pi R}{2fQ^\mathrm{s,p}_\mathrm{sc}(1-g^\mathrm{s,p})}~$$ within the independent scattering approximation [@van1991speed; @van1992speed]. In the diffusive regime, the spectral resolution of random spectrometers scales as $l^\mathrm{s,p}_\mathrm{t} / L^{2}$ [@pine1988diffusing], so that reduced transport mean free paths $l^\mathrm{s,p}_\mathrm{t}$ are generally desirable. According to Eq. \[lt\], short transport mean free paths can be expected when the (wavelength-dependent) scattering efficiency is high (strong scattering) and when the asymmetry parameter is significantly below 1 (near isotropic or even predominant backward scattering).
![ (Color online) (a) Scattering efficiency $Q_\mathrm{sc}$ and (b) asymmetry parameter $g$ of an air cylinder of radius 125 nm embedded in a silicon nitride matrix. The system is irradiated with $\mathrm s$- and $\mathrm p$-waves under normal incidence (blue and red lines, respectively). (c) Calculated transport mean free path $l_\mathrm{t}$ normalized to the radius $L$ of the scattering area of a 2D disordered medium composed of air pores in a silicon nitride matrix where the pores occupy a fraction $f=0.09$ of the available area.[]{data-label="fig2"}](Scat.jpg){width="\linewidth"}
For an air pore of radius embedded in a silicon-nitride matrix, we display in Fig. \[fig2\] both, the scattering efficiency and the asymmetry parameter, for TE and TM polarization. We observe that when moving from telecom frequencies to visible frequencies, the scattering efficiencies increase monotonously while simultaneously the asymmetry parameter concurrently moves to even stronger forward scattering. Despite these opposing trends, the overall effect is that the transport mean free path reduces when moving from telecom to visible wavelengths (with a minimum around for TE polarization) so that high spectral resolution of the random spectrometer is expected for the red end of the visible spectrum and NIR frequencies. Further reduction of the operation wavelength ($\sim \SI{400} {\nano
\meter}$) would lead to an elongated transport mean free path for TE polarization due to a stronger increase in the forward scattering characteristics relative to the corresponding increase in the scattering efficiency.
In Figs. \[fig3\] and \[fig4\], we display the wavelength- and detector-resolved transmission matrices for TE and TM polarization, respectively. As described above, these results have been obtained for the random spectrometer sketched in Fig. \[fig1\] by solving the Maxwell equations numerically using the DGTD upwind scheme approach [@busch2011discontinuous]. Within this approach we use an adaptive tetrahedron mesh for the spatial discretization with a minimal element size of and a polynomial order of 3. Meanwhile, the time-stepping is handled by a 14-stage fourth order low-storage-Runge scheme [@niegemann2012efficient].
![(Color online) Panels (a), (b) and (c): Wavelength- and detector-resolved transmission matrix for the random spectrometer depicted in Fig. \[fig1\] for TE polarization for the visible, NIR and telecom wavelength region, respectively. Light is launched from the input port and is detected by the detectors D$_{1}$ to D$_{13}$. The ordinate labels correspond to the detector index. The color coding corresponds to the relative intensity transmitted into the different waveguides and expressed in decibels. Panels (d), (e) and (f): The normalized spectral correlation function corresponding to the wavelength regimes of panels (a), (b) and (c).[]{data-label="fig3"}](TE.jpg){width="\linewidth"}
At telecom wavelengths (-), we find that the transport mean free path becomes comparable to the radius of the scattering area so that ballistic transport is dominant. As a result, the transmission matrix (Fig. \[fig3\](a) for TE polarization and Fig. \[fig4\](a) and (b) for TM polarization), is characterized by concentrations of the intensity in the central output ports (6-8).
Upon moving to NIR wavelengths (-) stronger multiple scattering occurs and we observe the onset of diffusion where the transmission matrix exhibits a more uniform distribution over the different output ports (cf. Fig. \[fig3\](b)), while at the same time the random spectrometer exhibits higher resolutions: Here, the HWHM is for TE polarization and for TM polarization, respectively, which is essentially less than half the values obtained at telecom wavelengths. Finally, we would like to note that the differences between TE and TM polarization in the angular distribution of light over the output ports can explained by the angular characteristics of light scattering by a single scatterer (cf. Fig. \[fig5\]). For TE polarization, scattering by $\mathrm p$-waves is dominant and light scatters more into the forward direction so that it is primarily detected in the central output ports as long as the diffusive regime is not fully developed. For TM polarization, scattering by $\mathrm s$-waves is dominant, which leads to more isotropic single-particle scattering and subsequently the light is more evenly distributed over all output ports even if the diffusive regime is not fully developed (cf. Figs. \[fig3\](b) and \[fig4\](b)).
![(Color online) Panels (a), (b) and (c): Wavelength- and detector-resolved transmission matrix for the random spectrometer depicted in Fig. \[fig1\] for TM polarization for the visible, NIR and telecom wavelength region, respectively. Light is launched from the input port and is detected by the detectors D$_{1}$ to D$_{13}$. The ordinate labels correspond to the detector index. The color coding corresponds to the relative intensity transmitted into the different waveguides and expressed in decibels. Panels (d), (e) and (f): The normalized spectral correlation function corresponding to the wavelength regimes of panels (a), (b) and (c).[]{data-label="fig4"}](TM.jpg){width="\linewidth"}
At the red end of the visible spectrum (-), we find that the transport mean free path is further reduced and the transmission matrix (Figs. \[fig3\](c) and \[fig4\](c)) is characterized by sharp peaks with a uniform spatial distribution, i.e., we find diffusive speckle patterns. For this wavelength range, the random spectrometer exhibits the highest resolution, for both polarizations (see Figs. \[fig3\](f) and \[fig4\](f)).
Clearly, the above numerical results are in full agreement with the predictions of the combined Mie and multiple-scattering theory and demonstrate that the highest spectral resolution of the spectrometer is obtained when the diffusive regime is fully developed. However, acceptable resolution can already be obtained when the size of the scattering region just barely allows for the onset of diffusion. If ballistic transport is dominant, the spectral resolution will be lower. Thus, which transport regime to select, should, thus, be based on a compromise between desired resolution and available footprint of the device. At this point it should be noted that scattering from low-index inclusions in high-index matrices, quite generally, does not exhibit pronounced Mie resonances notably at low and moderate index contrasts. This suggests that our findings are of a rather general nature and, for instance, apply to other material systems such as ZnO, chalcogenides and LiNbO$_3$ and broad wavelength ranges.
In summary, through a combination of electrodynamic simulations, Mie theory and multiple-scattering theory, we have studied the response of planar random spectrometers in low-footprint integrated-optical layout for TE and TM polarization at telecom, NIR and visible wavelengths. We have shown that the attainable spectral resolution depends on the interplay of two single scattering properties, scattering efficiency and the asymmetry parameter. The highest resolution is obtained for systems where the diffusive regime is fully developed. However, the transition regime between ballistic and diffusive transport might be sufficient for certain applications. These results are based on low filling fractions of the scatterers. For higher filling fractions, strong multiple-scattering corrections blur the connection to the single-scattering quantities. In this high-$f$ regime, an effective medium theory capable of reliably determining transport mean free paths [@kirchner1998transport] can be employed instead of the independent scattering approximation. Our design study provides a basis for the realization of random spectrometers in the visible and NIR wavelength range.
![(Color online) Panel (a): Angular scattering characteristics, Eq. (\[SA\]), for an air pore of 125 nm radius embedded in a silicon nitride matrix, for TE (blue line) and TM (red line) polarizations and fixed wavelength $\lambda=1500$ nm. Light scatters to larger angles for TM polarization due to predominance of $\mathrm s$-wave scattering. Panels (b),(c): Wavelength- and detector-resolved transmission matrix for TE and TM polarizations, respectively.[]{data-label="fig5"}](Polar.jpg){width="\linewidth"}
Funding {#funding .unnumbered}
=======
We acknowledge support by the Deutsche Forschungsgemeinschaft (DFG) under project Bu 1107/10-1 within the framework of the DFG priority program SPP 1839 ([*Tailored Disorder*]{}).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cet essai offre une introduction accessible aux raisons motivant la recherche d’une théorie quantique de la gravité. Il se concentre sur une des façons d’approcher la problème de la gravité quantique, la gravitation quantique à boucles, et sur les questions philosophiques que cette approche pose, questions qui sont paradigmatiques de toute tentative d’association de la physique de l’espace-temps avec la physique quantique.'
author:
- |
Christian Wüthrich\
(Traduit de l’anglais par Soazig Le Bihan)
date: 'A paraître [*in*]{} Soazig Le Bihan (dir.), [*La Philosophie de la Physique: D’aujourd’hui à demain*]{}, Paris : Vuibert.'
title: 'A la recherche de l’espace-temps perdu : questions philosophiques concernant la gravité quantique'
---
La physique théorique se voit de nos jours confrontée à un défi qui n’est pas sans rappeler celui auquel faisaient face les prédécesseurs de Newton au 17 siècle : deux théories incompatibles décrivent avec succès deux domaines séparés parmi les phénomènes observables. Les lois de la mécanique quantique gouvernent les phénomènes de petite échelle de la physique des particules, tandis que les lois de la relativité générale (RG) régissent la structure de l’univers à grande échelle. Le défi que la gravité quantique se doit de relever est d’achever la révolution qui a eu lieu en physique au siècle dernier ; la tâche ne consiste en rien moins que de fusionner les deux cadres théoriques incommensurables que sont la physique quantique et la RG. Nombreux sont ceux parmi les protagonistes de cette initiative qui espèrent qu’en relevant ce défi, un pas important, sinon final, sera fait vers l’unification théorique de la physique fondamentale. Attirés plus que jamais aujourd’hui par cette tâche herculéenne, les physiciens se ruent, creusant le sol, à la recherche d’une mine d’or menant au Saint Graal. De leurs efforts est née une variété d’approches, de techniques et de théories, parmi lesquelles deux sont proéminentes : la théorie des cordes et la gravitation quantique à boucles (GQB). Malgré des développements pour le moins palpitants du côté de la physique, les philosophes se sont montrés particulièrement lents à confronter les ressources conceptuelle et philosophique des richesses qui ont été déterrées à cette occasion.
Cet article se veut un appel aux armes et a pour but de mettre l’eau à la bouche aux lecteurs, en exposant certains des casse-têtes philosophiques les plus captivants qui se présentent en gravité quantique. L’analyse sera préfacée, en première section, par des considérations générales quant à la nécessité de trouver une théorie quantique de la gravité et aux méthodes employées dans la poursuite de ce but. Après avoir cartographié le domaine en Section \[sec:mapping\], je présenterai la GQB comme une candidate compétitive et particulièrement riche en problèmes philosophiques en Section \[sec:lqg\]. Ce qu’on appelle le problème du temps, i.e. le casse-tête consistant en ce que les quantités physiques ne peuvent jamais véritablement changer, est traité en Section \[sec:time\]. Enfin, en Section \[sec:emerge\] est analysé comment l’espace-temps continu qui nous est familier pourrait ré-émerger de la structure fondamentale non spatiotemporelle de l’univers.
Pourquoi la gravité quantique ? {#sec:why}
===============================
Avant de nous embarquer dans l’étude des fondements de la gravité quantique, il nous faut nous convaincre qu’une théorie de la gravité quantique est, de fait, nécessaire. Appelons [*théorie quantique de la gravité*]{} toute théorie cohérente qui combine la gravité avec une description quantique de la matière. Il est important de noter qu’il n’est pas nécessaire que la théorie en question considère la gravité elle-même comme quantifiée. Il est tout à fait acceptable qu’une telle théorie marrie une conception classique de la gravité avec une conception quantique de la matière. Ajoutons cependant une importante exigence méthodologique : une théorie quantique de la gravité se doit d’avoir des limites convenables , c’est-à-dire qu’aux échelles pour lesquelles il n’est plus pertinent de prendre en compte la nature quantique de la matière, la théorie se doit de fusionner avec la RG, et pour les régimes où la gravité est faible, elle se doit de se transformer en une théorie quantique adéquate.[^1]
Mais pourquoi la physique aurait-elle besoin d’une théorie quantique de la gravité? La réponse commune à cette question peut être analysée comme une combinaison de trois (groupes d’) arguments. Premièrement, il est souvent avancé qu’une telle théorie est nécessaire pour satisfaire une exigence d’[*unification*]{}. L’exigence d’unification est devenue pour beaucoup une exigence méthodologiquement importante du fait du succès que son application a permis à la physique de rencontrer dans le passé. James Clerk Maxwell forgea l’unification des forces magnétiques et électriques avec sa théorie dynamique de l’électromagnétisme dans les années 1870. Dans les années soixante, Abdus Salam, Sheldom Glashow, et Steven Weinberg formulèrent la théorie électrofaible, unifiant ainsi l’électrodynamique et l’interaction nucléaire faible associée à la radioactivité. Par la suite, la chromodynamique quantique, décrivant l’interaction nucléaire forte, qui est responsable de la cohésion des noyaux des atomes et de leurs constituants, et la théorie électrofaible furent unifiées à leur tour au sein du modèle standard de la physique des particules, qui parvient à rendre compte de trois des quatre forces fondamentales. Bien naturelle est ainsi l’ambition d’unifier les théories quantiques du modèle standard avec la RG – la meilleure théorie qu’on ait actuellement de la dernière des forces, la gravité. Notons cependant que si cette vénérable histoire justifie que l’unification soit prise comme desideratum méthodologique – et, dans une certaine mesure, comme programme de recherche, elle ne justifie en rien qu’on en fasse un dogme métaphysique. Le succès du processus d’unification dans le passé n’implique pas que la nature soit elle-même suffisamment unifiée pour pouvoir faire l’objet d’une théorie fondamentale unique qui sous-tendrait toute la physique. Il est en effet parfaitement concevable que la nature soit multiple au sens où la gravité résiste à toute subsomption sous l’ombrelle quantique de la physique des particules.
Le deuxième argument se nourrit des théorèmes sur les singularités, démontrés dans les années soixante et soixante dix par Stephen Hawkins, Roger Penrose, et Robert Geroch, et qui établissent de façon solide que les singularités sont génériques en RG classique. De nombreux auteurs ont soutenu que la RG cesse là ou alors d’être valide et que, par conséquent, elle contient les graines de sa propre destruction. Du coup, si on poursuit cet argument, il nous faut une théorie de remplacement pour la gravité et on a toutes les raisons de s’attendre à ce qu’une telle théorie soit quantique. En particulier, certains pensent que procéder à la quantification de la gravité, i.e. décrire la gravité comme possédant elle-même une nature quantique, permet de dissoudre certaines singularités comme le big bang. Cet argument n’a cependant, du moins en lui-même, pas beaucoup de force. En RG, les singularités ne font pas partie du tissu spatio-temporel, c’est-à-dire qu’elles ne se trouvent pas en un lieu particulier, et donc il n’est aucun besoin d’avoir une théorie valide là ou alors . La RG est une théorie parfaitement cohérente dans son domaine d’application, et par conséquent il est faux de dire qu’elle contient les graines de sa propre destruction. Ou du moins, si graines il y a, elles ne sauraient porter aucun fruit dialectique sans qu’on y ajoute une quantité considérable d’engrais argumentatif.
La troisième ligne d’argument, qui est de loin la plus convaincante, consiste à dire qu’il est des phénomènes dont nous avons de bonnes raisons de croire qu’ils présentent des aspects à la fois quantiques et gravitationnels, et que par conséquent, toute théorie qui rend compte de ces phénomènes de façon satisfaisante se doit d’intégrer ces deux éléments. Le plus important étant ici que parmi ces phénomènes comptent la dynamique des trous noirs et le commencement de l’univers. Insistons ici sur le fait que, bien que ces deux phénomènes soient liés à des singularités – dans leur description classique – ce n’est pas de là que vient la nécessité d’une théorie quantique de la gravité. c’est bien plutôt la présence simultanée de hautes densités de matière et d’un champ gravitationnel fort, tout ceci sur des échelles petites, qui impose une telle théorie. Au bout du compte, c’est bien l’existence de phénomènes plutôt extrêmes, et non un quelconque critère méthodologique ou esthétique, qui explique qu’on ait besoin d’une théorie de la gravité quantique.
Notons avant de conclure ici que, bien que quantifier la gravité – si tant est qu’on puisse le faire – permettrait clairement d’obtenir une théorie de la gravité quantique, ceci n’est en rien nécessaire. l’existence de régimes où les effets quantiques de la matière et les champs gravitationnels forts jouent conjointement un rôle important n’implique en rien que la gravité se doive d’être elle-même quantique. Tout ce dont on a besoin est une théorie qui gouverne l’ interaction entre la matière quantique et la gravité qui pourrait rester classique. En d’autres termes, rien de ce qui a été dit ici jusqu’à présent ne permet d’exclure les théories de la gravité dites semi-classiques de l’ensemble des théories adéquates possibles, même si ces théories violent les principes fondamentaux de la RG.
Maintenant que nous avons bien distingué la question de savoir si nous avons besoin d’une théorie quantique de la gravité de celle de savoir si la gravité doit être quantifiée, je me hâte d’ajouter qu’il existe de nombreux arguments qui ont pour but de montrer que la gravité doit être quantifiée dans toute théorie de la gravité quantique, et que donc les approches semi-classiques ne peuvent pas aboutir. Un des arguments typiques qui vont en ce sens consiste à tenter de faire entrer en contradiction d’un côté l’hypothèse d’une gravité de type classique interagissant avec la matière quantique et de l’autre un des principes physiques bien établis comme le principe de correspondance ou celui de la conservation de l’énergie. A ma connaissance cependant, il n’existe aucun argument de ce type qui ne fasse pas appel en plus à des prémisses que les défenseurs des approches semi-classiques ne sont nullement obligés d’accepter.[^2]
Maintenant que nous sommes assurés de la nécessité de développer une théorie quantique de la gravité, venons-en à dessiner la carte des principales approches de la gravité quantique en compétition.
Cartographie du domaine : les différentes approches de la gravité quantique {#sec:mapping}
============================================================================
Sur la base d’un schéma taxonomique bien utile, Chris Isham (1994) a proposé de diviser les différentes façons de formuler une théorie pleinement quantique, donc pas semi-classique, de la gravité en quatre types d’approche : premièrement, celles qui quantisent la RG ; deuxièmement, celles qui relativisent la physique quantique ; troisièmement, celles qui construisent une théorie quantique conventionnelle incluant la gravité et qui considèrent la RG comme sa limite aux basses énergies ; et quatrièmement, celles qui voient à la fois la RG et les théories quantiques conventionnelles comme les limites de basse énergie d’une théorie radicalement nouvelle. Considérons chacun de ces groupes l’un après l’autre.
La première famille d’approche prend pour point de départ la RG, and cherche à appliquer un procédé de quantification – autrement dit un recette générale pour concocter une théorie quantique conventionnelle à partir d’une théorie classique comme l’est la RG, ceci d’une façon qui soit à la fois en accord avec la rigueur mathématique et les principes de la physique. Evidemment, procéder à une quantification, c’est procéder pour ainsi dire à l’envers , du point de vue métaphysique, puisque cela consiste à partir d’une théorie classique douteuse – en ce sens qu’elle est reconnue comme défectueuse et comme nécessitant, pour cette raison, remplacement – puis à tenter de construire l’édifice solide de la nouvelle théorie (quantique) de la gravité sur les ruines de l’ancienne. On doit cependant comprendre cette stratégie comme motivée par l’usage d’un moyen méthodologiquement prometteur de parvenir à certaines fins, un peu à la façon de l’échelle de Wittgenstein. Les procédés de quantification ont prouvé leur efficacité dans le passé, et ont permis de produire, dans d’autres domaines de la physique, d’importantes théories telles que l’électrodynamique quantique. Les partisans de ce type de stratégie espèrent pouvoir rencontrer un succès similaire en physique gravitationnelle.
Le première famille d’approche se divise en deux genres: l’Ansatz covariant, dont on peut dire qu’il git désormais sur son lit de mort,[^3] et l’approche, bien vigoureuse, elle, en termes de quantification canonique. Pour pouvoir procéder à une quantification canonique d’une théorie, il faut que celle-ci soit formulée dans un formalisme particulier, appelé le formalisme hamiltonien contraint. Nous discuterons plus bas de comment l’adaptation de la RG au moule d’un tel formalisme se trouve au cœur des problèmes conceptuels les plus déroutants que rencontre cette théorie. La gravitation quantique à boucle (GQB) est la représentante la plus importante de cette approche, mais il y en a d’autres.
Concernant la seconde stratégie, il ne se dessine à ce jour aucune voie qui promette d’obtenir une théorie de la gravité quantique [*complète*]{} par relativisation des théories quantiques des champs, c’est-à-dire par l’emploi de techniques qui permettrait l’incorporation totale des leçons de la RG dans la théorie quantique. La seule représentante actuelle de cette approche consiste à tenter de formuler une théorie quantique des champs dans le cadre d’un espace-temps courbe, et non plat comme l’est l’espace-temps usuel. l’idée générale qui sous-tend cette approche est d’incorporer, en un sens local, le principe de covariance générale de la RG. Il est important de noter que, en revanche, l’espace-temps, tout courbé qu’il soit, n’est en rien dynamique. Autrement dit, il ne peut aucunement être interprété à la façon dont on interprète l’espace-temps de la RG, c’est-à-dire comme interagissant avec les champs de matière.
Le troisième groupe se place également dans la perspective de la physique quantique, mais au lieu d’essayer d’y incorporer directement les leçons de RG, tente de développer la physique quantique par des moyens aussi conventionnels que possible pour y intégrer la gravité. l’espoir est alors qu’on retrouvera la RG comme limite aux basses énergies de la théorie nouvelle. Le membre de loin le plus prometteur de ce groupe est la théorie des cordes, qui, cependant, va beaucoup plus loin que la théorie quantique des champs conventionnelle, tant du point de vue de ses méthodes que de celui de ses ambitions. Ceci dit, bien qu’elle fasse exploser les frontières naturelles du groupe, la théorie des cordes se place malgré tout dans la perspective de la théorie quantique des champs conventionnelle, ce autant du point de vue de son histoire que de celui de son système de pensée, contrairement aux tentatives de construction d’une nouvelle théorie de la gravité quantique qui serait séparée de la vieille physique . Encore une fois, il existe d’autres approches dans ce groupe, telles que la théorie quantique des champs topologique, mais aucune ne rassemble de soutien substantiel de la part des physiciens.
La meilleure façon de caractériser le quatrième et dernier groupe de la taxonomie de Isham est par son attitude iconoclaste. Les approches hétérodoxes de ce type ne prenne en effet aucun élément de théorie physique connue pour point de départ ; au lieu de cela, des points de vue radicalement nouveaux sont envisagés dans l’espoir de formuler une théorie de la gravité quantique [*ab initio*]{}. Pour autant que je sache, ces approches ne proposent à l’heure actuelle que des [*schémas*]{} programmatiques, et non des théories à part entière. Leur caractère attractif vient principalement de la redoutable apparence d’incompatibilité profonde entre les principes gouvernant la physique quantique des phénomènes à petite échelle et ceux gouvernant la RG et les phénomènes à grandes échelle. Une telle incompatibilité, nous dit-on, ne saurait être résolue que si l’on prend un tout nouveau départ.
Toutes ces approches présentent des aspects attractifs, et pour cette raison possèdent chacune des disciples. Mais toutes ont aussi des lacunes. Il est impossible d’en faire une liste exhaustive dans le cadre de cet article. En dehors des deux défis principaux que doit relever la GQB, que je discuterai par la suite, je ne m’étendrai donc pas plus sur le sujet, sauf pour souligner le fait que toutes ces approches ont un problème majeur commun, à savoir leur manque total de lien avec les observations et l’expérience. Quelques propositions sont faites ici et là quant à la façon dont telle ou telle approche pourrait entrer en relation avec l’empirie, mais, jusqu’ici, ces propositions en restent à des suggestions le plus souvent pour le moins spéculatives concernant la manière dont, peut-être, une telle relation pourrait être établie. Ou bien la théorie proposée est trop malléable, de sorte qu’elle est capable de s’adapter à presque toute donnée empirique, comme c’est le cas de la théorie des cordes, dont d’une part les prédictions concernant les particules supersymétriques ont été révisées de façon constante au gré des échecs répétés des détecteurs de particules à les trouver aux valeurs d’énergie prédites, et qui, d’autre part, se trouve dans un [*embarras de richesses*]{}[^4], avec le fameux problème du paysage , qui consiste à devoir choisir parmi $10^{500}$ modèles différents. Ou bien encore le lien entre les données relativement bien comprises et les théories reste à la fois ténu et controversé, comme quand on se demande si, et si oui dans quelle mesure, les données qui saisissent de façon étroite des violations possibles de la symétrie de Lorentz sont en relation avec les théories de la gravité quantique qui prédisent, ou font l’hypothèse que, la structure de l’espace-temps est discrète, ce qui est supposé impliquer une violation, ou au moins une modification, de la symétrie de Lorentz par ailleurs très bien confirmée aux échelles plus grandes. Ou bien les prédictions faites par les théories ne peuvent être testées que dans des cadres expérimentaux qui échappent de loin aux possibilités techniques actuelles, comme c’est le cas des prédictions de la GQB selon lesquelles l’espace-temps est discret à l’échelle de Planck, soit un quintillion de fois ($10^{18}$) les énergies que l’on cherche à obtenir avec le grand collisionneur de hadrons du CERN. Ou bien enfin, personne n’a tout simplement la moindre idée de comment la théorie pourrait être mise en relation avec l’empirie, comme dans le cas des approches embryonnaires du quatrième groupe, telles que la théorie à ensembles causaux.[^5]
Introduction à la gravitation quantique à boucles {#sec:lqg}
=================================================
La GQB est, avec la théorie des cordes, une des théories de la gravité quantique à la fois les plus prometteuses et les plus importantes du point de vue du nombre de ses disciples. Il s’agit d’une approche canonique, qui prend pour point de départ la RG – la meilleure théorie classique de la gravitation que l’on possède – à laquelle est appliquée un procédé bien connu qui permet de concocter une théorie quantique à partir d’une théorie classique, ceci dans l’espoir d’obtenir une théorie quantique de la gravité viable. C’est là une approche fondamentalement conservatrice en ce sens qu’elle a pour ambition de rester aussi fidèle que possible aux théories physiques qui sont bien connues et déjà couronnées de succès.
Le procédé de quantification choisi ici – appelée [*quantification canonique*]{} – ne peut être appliqué que si la RG est reformulé en termes de système hamiltonien. Un système hamiltonien est un système physique qui obéit aux équations de Hamilton, qui sont des équations différentielles mettant en relation les positions (généralisées) et les quantités de mouvement (généralisées) de tous les degrés physiques de liberté avec l’énergie du système, donnant ainsi l’évolution temporelle de tous les degrés de liberté du système.[^6] Il se trouve que les systèmes physiques obéissant les équations de Hamilton et pouvant donc être mis sous la forme de systèmes hamiltoniens sont à la fois nombreux et importants.
La RG, dans sa formulation habituelle, n’est pas formulée en termes de systèmes hamiltoniens. Au cœur de la RG standard se trouvent ce qu’on appelle les [*équations du champ d’Einstein*]{}, qui mettent en relation la géométrie de l’espace-temps, encodée dans le champ métrique, avec la distribution de matière et d’énergie dans cet espace-temps. Elles sont souvent interprétées comme décrivant une interaction dynamique entre la gravité, représentée par le champ métrique, et la distribution de matière-énergie.[^7] Les fameuses paroles de John Wheeler, disant que, en RG, la masse s’accroche à l’espace-temps, et lui dit comment se courber, tandis que l’espace-temps s’accroche à la masse, et lui dit comment se mouvoir,[^8] incarne cette interprétation des équations du champ d’Einstein comme gouvernant la coévolution dynamique de la métrique de l’espace et des champs de matière. Il nous faudra revenir sur cette interprétation de la RG dans la section suivante au moment où nous discuterons du problème du temps.
Une solution aux équations du champ d’Einstein se présente sous la forme d’un triplet $\langle \mathcal{M}, g, T\rangle$ – où $\mathcal{M}$ est une variété différentiable à quatre dimensions, $g$ un champ métrique, et $T$ ce qu’on appelle un [*tenseur énergie-impulsion*]{}, expression mathématique de la distribution de la matière et de l’énergie sur la variété – triplet donc tel que la relation entre $g$ et $T$ soit en accord avec les équations du champ d’Einstein en tout point de $\mathcal{M}$. Si on regarde les choses d’un point de vue neutre, donc, les équations d’Einstein ne sont pas des équations [*dynamiques*]{} ; au lieu de cela, elles ne font qu’imposer des conditions locales sur les valeurs du couple champ métrique $g$ – distribution de matière-énergie de l’univers telle que donnée par $T$. Et pourtant, il est nécessaire de comprendre la RG dynamiquement si l’on veut lui donner une formulation hamiltonienne. On attend d’une théorie dynamique qu’elle soit formulable clairement en termes de valeurs initiales – c’est-à-dire qu’il existe une formulation de la théorie qui permette d’obtenir l’évolution dynamique complète d’un système physique à tout instant sur la base des équations du mouvement et étant donnés un ensemble de conditions initiales à un instant donné. Le problème posé par la formulation hamiltonienne de la RG est donc naturellement lié au problème de Cauchy. Ce dernier, cependant, est difficile à bien poser dans le cadre de la formulation standard de RG, du fait que la division nécessaire de l’espace-temps à quatre dimensions en un espace qui évolue dans le temps semble bien violer l’une des leçons les plus centrales de la RG qui est qu’une telle division n’a aucune justification physique satisfaisante. Cette division forcée de l’espace-temps fait naître une inquiétude : tenter de comprendre la RG de façon dynamique pourrait bien mener nécessairement à mésinterpréter cette théorie. En réalité, bien que la formulation hamiltonienne de la RG requiert que l’espace-temps soit feuilleté en espaces à trois dimensions ordonnés selon un paramètre temps unidimensionnel, la leçon relativiste concernant le caractère quadridimensionnel de l’espace-temps est intégrée mathématiquement dans le cadre hamiltonien, à savoir dans les équations de contrainte de ce dernier, point sur lequel nous reviendrons plus bas.
On peut dynamiser la RG, et obtenir une version hamiltonienne de la théorie, de plusieurs façons. En général, le jeu consiste à trouver des coordonnées canoniques qui saisissent en quelque sorte une géométrie spatiale évoluant dans le temps. A cette fin, les formulations hamiltoniennes de la RG découpe l’espace-temps en un feuilletage d’hypersurfaces tridimensionnelles (qui sont des sous-variétés de genre espace de $\mathcal{M}$). La façon standard de procéder à ce feuilletage est traditionnellement appelée le [*formalisme ADM*]{}, du nom de ses fondateurs Richard Arnowitt, Stanley Deser, et Charles Misner. Le formalisme ADM prend les métriques tridimensionnelles (spatiales) induites par $g$ sur les hypersurfaces spatiales comme variables de position , et la courbure extérieure de ces hypersurfaces (plus précisément une combinaison linéaire des composantes de cette courbure), représentant la façon dont elles sont contenues dans l’espace-temps quadridimensionnel, comme variables de quantité de mouvement , qui se conjuguent alors de façon canonique avec les métriques spatiales. Les équations de Hamilton peuvent être écrites sur cette base.
Il s’avère cependant que ces équations ne sont pas, à elles seules, équivalentes aux équations du champ d’Einstein. Pour obtenir l’équivalence, il faut leur joindre des équations supplémentaires contraignant les relations entre les variables canoniques. Ces équations de contraintes témoignent du fait que les données initiales ne peuvent pas être choisies arbitrairement, mais doivent bien satisfaire certaines conditions.[^9] On peut montrer que ces équations de contrainte sont l’expression mathématique de la présence de ce qu’on appelle une liberté de gauge , soit une redondance représentationnelle dans la description mathématique de la situation physique.[^10] En particulier, elles apparaissent en conséquence du fait que le groupe des difféomorphismes quadridimensionnels est le groupe de symétrie dynamique de la RG, en accord avec l’exigence de la covariance générale.[^11] L’exigence de covariance générale est que la physique reste la même si les champs – y compris le champ métrique – sont tous déplacés de la même manière et de façon continue sur la variété. L’idée qui se trouve derrière cette exigence est donc que, bien que la représentation mathématique diffère selon que les champs sont déplacés ainsi ou non, la situation physique est la même dans les deux cas.
En réalité, deux équations (ou familles d’équations) de contraintes apparaissent. La première, représentant la liberté de choisir le feuilletage, est appelée la [*contrainte hamiltonienne*]{}. Il s’avère que l’hamiltonien des équations de Hamilton habituelles est lui-même une contrainte.[^12] On peut voir ainsi que l’absence d’une variable temps externe qui serve de repère fait de l’équation dynamique elle-même une contrainte, liée à la liberté de choisir une gauge sans conséquence observable. La seconde famille – qui comprend trois membres –, liée à la liberté de choisir les coordonnées spatiales dans l’espace à trois dimension, sont appelés [*contraintes vectorielles*]{}. Ceci nous fait un total de quatre équations de contrainte.
Une fois qu’on a obtenu une formulation hamiltonienne de RG, i.e. une fois qu’on a identifié les variables canoniques et que l’on a écrit toutes les équations de contraintes que ces dernières doivent satisfaire, on peut se lancer dans la quantisation de la théorie classique grâce au procédé de quantisation canonique, décrit par Paul Dirac (1964). Ce procédé a pour principe essentiel de prendre les variables canoniques et de les transformer en opérateurs quantiques agissant sur un espace d’états quantiques. Le jeu de relations entre ces variables, décrit par les crochets de Poisson au niveau classique, se mue en des relations de commutation canoniques entre des opérateurs de base, tandis que les équations de contraintes deviennent des équations d’onde d’opérateurs de contraintes identiques, du point de vue fonctionnel, aux fonctions de contraintes classiques, et agissant sur les états quantiques. Seuls les états quantiques qui satisfont ces équations de contraintes quantiques comptent comme états physiques acceptables.
Les tentatives d’obtention d’une théorie quantique de la gravité en usant du formalisme ADM et par quantification canonique se sont heurtées à des difficultés techniques insurmontables, comme par exemple le fait que les équations de contraintes ne sont pas polynomiales. A un moment donné, il semblait que tout espoir d’obtenir une théorie quantique de la gravité par quantification canonique était tout simplement perdu. Mais c’est alors que, dans les années 80, Abhay Ashtekar trouva, en s’appuyant sur le travail de Amitabha Sen, de nouvelles variables. Ces variables d’Ashtekar permettent de simplifier les équations de contrainte de façon significative, même si on perd la signification directement géométrique des variables ADM.[^13] Je vous épargne ici les détails – qu’on peut trouver dans n’importe quelle présentation un peu sérieuse de la GQB.[^14] Mentionnons seulement ici que le principe de base consiste en ce que la géométrie de l’espace-temps est donnée par un champ triade représentant les référentiels inertiels locaux définis sur les hypersurfaces spatiales, au lieu de l’être par la métrique spatiale. Ces approches permettent toutes deux de représenter la géométrie de l’espace-temps, et la traduction de l’une à l’autre est possible, bien qu’il y ait une famille de contrainte supplémentaire dans la GQB, qui est liée aux symétries internes. Pour ne retenir que l’essentiel, le passage des variables ADM aux variables d’Ashtekar consiste en une réinterprétation des équations du champ d’Einstein comme des énoncés concernant, non pas la métrique comme c’était avant le cas, mais une connexion – un outil mathématique qui sert à décrire ce qui arrive à des vecteurs tangents à une variété quand ils sont transportés d’un point de la variété à un autre le long d’une courbe. La théorie de la relativité générale, réinterprétée ainsi, est ensuite soumise au procédé de quantification canonique décrit ci-dessus.
Il se trouve que les équations de contraintes ne peuvent pas toutes être résolues facilement. Pour tout dire, seule deux des trois familles d’équations de contraintes ont été à ce jour résolues. Si on définit l’[*espace physique de Hilbert*]{} comme l’espace de tous les états quantiques de la théorie qui sont solutions de toutes les équations de contrainte, et qui donc doivent être considérés comme les états [*physiques*]{}, alors il faut admettre que l’espace physique de Hilbert de la GQB n’est pas encore connu. l’espace, plus grand, des états qui sont solutions des deux premières familles d’équations de contrainte est souvent appelé l’[*espace cinématique de Hilbert*]{}. La contrainte qui a jusqu’ici opposé résistance à toute résolution est l’équation de contrainte hamiltonienne, à la forme apparemment simple $\hat{H} |\psi\rangle = 0$, appelée l’[*équation de Wheeler-DeWitt*]{}, où $\hat{H}$ est l’opérateur hamiltonien interprété habituellement comme la source de l’évolution dynamique et $|\psi\rangle$ est un état quantique de l’espace cinématique de Hilbert. Bien entendu, l’opérateur hamiltonien $\hat{H}$ est une fonction(elle) complexe des opérateurs de bases correspondant aux variables canoniques. En réalité, la forme fonctionnelle elle-même de $\hat{H}$ est l’objet de controverses, du fait qu’il existe plusieurs candidates possibles non-équivalentes les unes aux autres. Dans la mesure où l’espace physique de Hilbert n’a pas encore été totalement construit, la GQB reste donc incomplète.
Etant donné que l’espace physique de Hilbert est un sous-espace de l’espace cinématique de Hilbert, tous les états physiques sont aussi des éléments de l’espace cinématique de Hilbert. Heureusement, nous en savons bien plus sur ce dernier. Ses éléments sont les [*états des réseaux de spin*]{}, c’est-à-dire les états quantiques du champ gravitationnel, ou au moins de la distribution spatiale de ce dernier. Les états des réseaux de spin peuvent être représentés par un graphe annoté où ils sont inscrits dans un cadre spatio-temporel (cf. Figure \[fig:spinnetwork\] ). L’espace physique est supposé être, au niveau fondamental, un état de réseaux de spin ou une superposition quantique de tels états.[^15]
L’essentiel de la structure de l’espace physique est ainsi saisi par des graphes annotés du type de la Figure \[fig:spinnetwork\]. Comme il apparaît sur la figure, les représentation en spin se situent sur les sommets du graphe (les nœuds) ainsi que sur les arrêtes (les lignes qui joignent les nœuds). Ce que les sommets, les $i_k$, représentent, ce sont les nombres quantiques indiquant la taille des atomes d’espace , tandis que ce que représentent les arrêtes, les $j_l$, ce sont les tailles de la surface de connexion entre des morceaux d’espace adjacents. Les états de réseaux de spin sont des structures discrètes. On voit donc que, selon la GQB, l’espace physique est granulaire à l’échelle minuscule qu’est celle de Planck. Ainsi, l’espace continu de la théorie classique est supplanté par une structure quantique discrète. Par conséquent, l’espace tel qu’il figure dans notre conception du monde est un phénomène émergent, et non un élément de réalité fondamentale. c’est du moins là ce qu’affirme la GQB.
Les deux problèmes les plus pressants que face la GQB sont notre manque de compréhension de la dynamique, ou, de façon équivalente, notre incapacité à résoudre l’équation de contrainte hamiltonienne, ainsi que notre échec à rendre compte de l’émergence de l’espace-temps continu classique, ou, de façon équivalente, des raisons du succès des théories gravitationnelles classiques telles que la RG. Ces problèmes présentent tous deux des aspects techniques aussi bien que philosophiques, et tous deux surgissent, d’une façon ou d’une autre, dans le cadre de nombre des principales approches de la gravité quantique. Par exemple, le problème technique de la résolution de l’équation de contrainte hamiltonienne en GQB est intimement lié au problème du temps, qui a de nombreuses dimensions philosophiques. De plus, dans la mesure où la théorie des cordes contient la RG, elle doit aussi traiter le problème du temps, au moins dans sa formulation très générale qui exige que soit résolue la tension conceptuelle existant entre, d’un côté, la notion pré-relativiste d’un temps extérieur au, et indépendant du système physique considéré, dont il est fait usage par les théories quantiques et la théorie des cordes, et, d’un autre côté, la conceptualisation nouvelle du temps comme un acteur physique inséparable de l’espace et interagissant avec les champs de matière ainsi qu’avec les autres formes d’énergie. Naturellement, la forme précise du problème varie cependant, et parfois de façon radicale, d’une approche à l’autre. Ce sont ces deux problèmes majeurs qui seront traités, chacun leur tour, dans les deux dernières sections ci-dessous, avec une attention particulière portée aux aspects conceptuels et philosophiques plutôt que techniques.
Le problème du temps {#sec:time}
====================
Il est bien connu que le philosophe présocratique Parménide d’Élée soutenait que le monde, fondamentalement, est un tout indestructible, immuable, et existant de toute éternité. Le changement, selon lui, n’est que pure apparence et ce qui existe en réalité est gelé temporellement. Très peu de philosophes, en particulier ces derniers siècles, ont suivi Parménide sur cette voie métaphysique radicale. De façon surprenante, son hypothèse pour le moins courageuse reçoit un certain soutien de la part de la RG (dans sa formulation hamiltonienne) et des théories quantiques qui sont fondées sur cette dernière.
Déjà en RG standard, il est loin d’être facile d’isoler le temps physique, qui, de façon générale, et quelle que soit sa nature, ne fait qu’induire un pré-ordre de priorité temporelle sur l’ensemble des événements sur la variété. Un [*pré-ordre*]{} est une relation binaire $Rxy$ définie sur un ensemble $X$, qui est réflexive et transitive, mais en général ni faiblement antisymétrique, ni comparable. On parle d’antisymétrie faible quand est satisfaite la condition que, pour deux événements $a$ et $b$, si $a$ précède temporellement $b$ (ce qui inclut le cas où les deux événements sont simultanés), et $b$ précède temporellement $a$, alors ils sont simultanés. La condition d’antisymétrie faible est satisfaite à moins que le temps soit circulaire ou ait une topologie étrange, et de ce fait, elle est généralement considérée comme une condition nécessaire pour avoir un ordre qui puisse être légitimement considéré comme [*temporel*]{}. Mais le temps en RG n’est pas faiblement antisymétrique, puisqu’il peut y avoir deux événements qui soient chacun en relation de précédence temporelle avec l’autre sans pour autant être simultanés. Il n’est pas non plus comparable puisque toute paire d’événements séparés par un intervalle de genre espace n’entretient absolument aucune relation temporelle, ce qui serait requis pour que l’on ait comparabilité. On peut récupérer l’antisymétrie faible dans les espaces-temps dont la topologie est du type $\Sigma\times \mathbb{R}$, où $\Sigma$ est un espace à trois dimensions, $\mathbb{R}$ est l’ensemble des réels, et $\times$ désigne le produit cartésien. Les espaces-temps de ce type peuvent être divisés en un espace et un temps – même si en général il existe un nombre infini de façons également valides de procéder à une telle division. En revanche, ce n’est que si un unique feuilletage, c’est-à-dire une seule façon de diviser, peut être privilégié selon un certain principe physique que l’on peut retrouver la comparabilité, et par là, le caractère total de la relation d’ordre.
Puisque la formulation hamiltonienne de la RG requiert que l’espace-temps soit feuilleté de telle sorte que des systèmes spatiaux évoluent dans le temps, elle ne peut traiter que des espaces-temps à topologie du type $\Sigma\times\mathbb{R}$. Il pourrait sembler que les difficultés liées à la notion de temps soient atténuées au sein de la formulation hamiltonienne de la RG (et par conséquent au sein des approches de la gravité quantique qui en dérive) par rapport au cas de la RG standard. Mais une telle apparence est trompeuse ; en réalité, la situation est encore pire. En un sens, la notion de temps s’évanouit et toutes les grandeurs physiques se retrouve forcées à rester parfaitement constantes au cours du temps.
Pour ce qui est de la disparation complète du temps au sein de la gravité quantique canonique, Wheeler et Bryce DeWitt remarquèrent dans les années 60 que l’équation dynamique de base, celle-là même qui porte leurs noms et que nous avons écrit plus haut ici, ne contient pas de paramètre temps. Par rapport à l’équation de Schrödinger, $$\hat{H}|\psi\rangle = i\hbar \frac{\partial}{\partial t} |\psi\rangle,$$ qui donne la dynamique des systèmes quantiques non-relativistes, toute la partie à droite du signe égal disparaît dans l’équation de Wheeler-DeWitt, et, avec elle le paramètre temps $t$. Pour ce qui concerne donc les théories canoniques de la gravité quantique, la notion de temps disparaît complètement du tableau. La disparition du temps dans les équations dynamiques de l’approche canonique peut être considérée comme un témoignage de la tension conceptuelle à laquelle se confronte quiconque tente de réconcilier ensemble la version édulcorée de la notion de temps des espaces-temps relativistes avec la notion de temps externe, se déployant de façon constante et indépendamment des systèmes physique dont elle permet l’évolution, que l’on trouve en mécanique quantique. C’est peut-être là une conséquence du fait que le temps y faisait partie du système physique, à savoir l’espace-temps, que l’on a quantifié. En physique relativiste, il est bien clair qu’il n’y a pas de paramètre temps externe qui puisse servir de repère pour l’évolution dynamique. Ceci dit, conclure directement de la forme de l’équation de Wheeler-DeWitt que le temps a disparu, serait aller un peu trop vite. Le temps pourrait évidemment encore être présent, mais sans que la dynamique ne le porte de façon ostentatoire, comme il ne convient de le faire qu’en physique relativiste.
Certains physiciens comme Carlo Rovelli et Julian Barbour, cependant, ont fait leur la conclusion plus radicale, et ont par conséquent tenté de formuler la mécanique quantique de façon à ce que cette dernière ne requiert pas l’existence d’un temps externe rythmant l’évolution dynamique, et que, au lieu de cela, le temps soit remplacé par une relation directe entre les événements.[^16] Ceci dit, on peut accepter l’idée que le temps n’existe pas au niveau quantique sans cependant se soumettre à celle que la mécanique quantique est relationnelle, du moment que l’on puisse montrer que la RG est capable de décrire le [*changement*]{} et comprendre comment l’espace-temps classique émerge de la structure quantique sous-jacente, ou, pour le dire autrement, comment la RG classique est valide comme approximation limite de la GQB aux basses énergies. La RG peut aisément satisfaire la première de ces deux exigences: même si le temps en RG ne permet pas, en général, d’ordonner même partiellement les événements de façon valide, objective et universelle, on peut y rendre compte du changement de façon relationnelle comme la modification des propriétés des systèmes physiques le long de leur lignes d’univers. Il n’existe, en revanche, pas de solution facile pour la deuxième exigence, comme nous le verrons en Section \[sec:emerge\] ci-dessous.
Même si le changement peut être décrit dans la version standard de la RG, il est un autre aspect du problème du temps qui semble indiquer que l’apparence qu’il n’y a pas de temps en gravité quantique n’est en réalité pas trompeuse dans la mesure où il n’y a pas non plus de changement au niveau le plus fondamental de la réalité physique.[^17] En effet, dans sa formulation hamiltonienne, la RG ne peut même pas rendre compte du changement au niveau des propriétés des systèmes physiques. On voit donc que le problème du temps, ou du moins du [*changement*]{}, surgit déjà au niveau classique, même si ce n’est que dans une formulation de la RG. Parmenides, semble-t-il, avait raison : le changement n’existe pas, et tout reste gelé, si j’ose dire, dans le temps.
Formellement, ceci vient du fait que la re-paramétrisation de l’(espace-)temps est une symétrie de gauge de la théorie. Plus précisément, le groupe de symétrie dynamique des équations d’Einstein est Diff$(\mathcal{M})$, le groupe des difféormorphismes à quatre dimensions sur $\mathcal{M}$, qui devient dans la formulation hamiltonienne de la RG un ensemble de contraintes qui engendrent ces difféomorphismes spatiotemporels. Autrement dit, le changement n’est rien d’autre qu’une redondance au sein de la représentation mathématique. Au cœur du problème se trouve l’exigence qu’aucune grandeur physique ne saurait dépendre de la simple représentation mathématique d’une situation réelle – plus précisément du système de coordonnées dans lequel la situation est décrite. Cette exigence est tout à fait raisonnable, puisque les changements de représentation n’ont aucune conséquence observable. La physique reste la même, quel que soit le système de coordonnées utilisé par les humains pour la décrire.
En dépit de sa conclusion pour le moins contre-intuitive, il nous faut prendre l’argument du paragraphe ci-dessus sérieusement, comme nous presse de le faire John Earman (2002).[^18] La conclusion n’est pas le résultat d’un raisonnement futile, et chacune des étapes peut être défendue de façon solide. Il n’est donc pas possible de s’en défaire d’un revers de main. Une façon directe et efficace de se débarrasser de cette conclusion gênante serait bien évidemment de refuser toute pertinence physique à la formulation hamiltonienne de la RG, comme semble vouloir le faire Tim Maudlin (2002). En effet, les conséquences qui s’imposent à nous dans ce cadre peuvent être considérées comme une réduction à l’absurde pour l’approche toute entière. Mais ce serait là aller trop vite : la quantification canonique a rencontré de nombreux succès dans d’autres domaines tels que l’électrodynamique et offre une route vers le Saint Graal de la gravité quantique qui est fondée physiquement et relativement bien maîtrisée mathématiquement.
Il existe de multiple façons de traiter du problème du temps, comme en témoigne le déluge de réponses de la part tout aussi bien de physiciens que de philosophes à l’appel à contribution du Foundational Questions Institute (FQXi) pour le prix du meilleur essai concernant la nature du temps.[^19] Je ne m’étendrai pas ici sur ces propositions, mais il suffira de dire que les réponses au problème du temps sont sans doute tout aussi bigarrées qu’est consensuelle l’idée qu’on ne verra probablement pas de progrès substantiel en physique fondamentale sans passer par une réflexion soutenue sur la nature du temps et son rôle en gravité quantique. Quelle que soit la position où l’on se place dans le débat, une autre chose est claire : comprendre comment l’espace-temps classique émerge de la structure quantique fondamentale non spatiotemporelle éclairera le problème du temps de façon importante.
Espace-temps : disparition et ré-émergence {#sec:emerge}
==========================================
En théorie des cordes comme en GQB, ainsi que dans les autres approches de la gravité quantique, tout semble indiquer que non seulement le temps, mais aussi l’espace ne sont non pas des entités fondamentales, mais plutôt des phénomènes émergents qui surgissent depuis la physique fondamentale. Dans la langue de la physique, les théories de l’espace-temps telles que la RG sont des théories effectives et l’espace-temps lui-même est émergent , tout comme la thermodynamique est une théorie effective et la température une propriété émergente au niveau effectif, du fait qu’elle provient du comportement collectif des molécules de gaz. Cependant, à la différence de l’idée que la température n’est qu’un phénomène émergent, la pensée que l’univers n’est pas dans l’espace et le temps choque sans doute notre notion même d’existence physique de façon plus profonde qu’aucune autre révolution scientifique. On se demande même si on peut formuler une théorie physique de façon cohérente en dehors de l’espace et du temps.[^20]
L’espace disparaît en GQB dans la mesure où les structures physiques qui y sont décrites ne ressemblent que peu, peut être pas du tout, aux géométries spatiales que l’on peut trouver en RG. Comme nous l’avons vu dans la Section \[sec:lqg\], ces structures sont discrètes, et non continues, comme le sont les espaces-temps classiques. Elles représentent les éléments constituant notre univers au niveau fondamental, qui correspondent, en quelque sorte, à des morceaux d’espace physique, et ainsi donne lieu, d’une façon qu’il convient encore d’élucider, à l’apparition des géométries spatiales que l’on trouve en RG classique. Il faut souligner que le fait que l’espace-temps est remplacé par une structure [*discrète*]{} au niveau quantique est une conséquence bien établie et tout à fait attendue de certains postulats de base communs à une vaste classe de théories quantiques de la gravité, y compris la GQB.[^21] Le problème conceptuel qui consiste à comprendre comment on peut faire de la physique en absence d’un cadre spatio-temporel sous-jacent, dans lequel la physique peut avoir lieu , est intimement lié à la difficulté technique qui consiste mettre en relation LA GQB et la RG. Les physiciens n’ont pas encore fini de travailler à comprendre tout à fait comment les espace-temps classiques émergent depuis la structure fondamentale non spatio-temporelle de LA GQB, et les philosophes commencent à peine à étudier les fondements conceptuels de cette théorie ainsi que les conséquences de la gravité quantique en général, et de la disparition de l’espace-temps en particulier.[^22] Même si le gros œuvre reviendra aux physiciens, les philosophes ont ici un rôle à jouer dans l’exploration et la cartographie du paysage des possibilités conceptuelles, en faisant entrer en jeu l’immense littérature philosophique disponible sur les questions d’émergence et de réduction, qui offre une boite à outil conceptuelle bien fournie. Je finirai ici par quelques mots qui peuvent servir de préalable à la conception de lignes de recherche en vue d’une solution à ces problèmes.
Concevoir comment l’espace-temps classique émerge depuis la structure quantique fondamentale implique que l’on prenne la limite classique , comme le disent les physiciens. D’une certaine façon, faire en retour le lien entre les états de réseaux de spin de la GQB et les espaces-temps de la RG revient à engager la procédure inverse de la procédure de quantification utilisée au départ pour formuler la théorie quantique. Ainsi, tandis qu’on peut considérer la quantification comme le contexte de la découverte , trouver la limite classique qui permet de faire le lien entre la théorie quantique et la RG doit être conçu comme le contexte de justification (partielle) . Il faut souligner ici que comprendre comment l’espace-temps (classique) ré-émerge en montrant comment la RG peut être récupérée comme la limite aux basses énergies d’une théorie plus fondamentale est non seulement important pour sauver les apparences et pour satisfaire le sens commun – bien que bien entendu ces aspects comptent également –, mais doit surtout être considérée comme un exigence méthodologique centrale du projet de la gravité quantique. S’il n’est pas possible de montrer qu’il existe entre la RG et la GQB une certaine relation mathématique bien comprise où la RG apparaît comme la théorie approximativement correcte quand les niveaux d’énergie sont suffisamment bas, ou, de façon équivalente, à des échelles suffisamment grandes, alors la GQB ne peut pas expliquer pourquoi la RG a pu être couronnée de succès comme elle l’a été.[^23] Or une théorie couronnée de succès ne saurait être supplantée par une autre théorie que si cette dernière non seulement fait des prédictions nouvelles ou donne des explications plus profondes, mais aussi est capable de reproduire le succès empirique de la théorie qu’elle prétend pouvoir remplacer.
En fin de compte, bien entendu, l’analyse du problème dépendra de la façon dont la théorie complète sera articulée. Considérons malgré tout l’analyse conceptuelle de Jeremy Butterfield et Chris Isham (1999, 2001), en nous concentrant cependant sur le seul niveau cinématique, afin d’éviter d’avoir à traiter du problème du temps en son entier comme eux sont contraints le faire, et appliquons cette analyse non seulement à la question de l’émergence du temps, comme ils le font, mais aussi à celle de l’émergence de l’espace-temps dans sa totalité. Butterfield et Isham distinguent entre trois types de relation de réduction entre des théories : [*l’extension définitionnelle*]{}, [*la supervénience*]{}, et [*l’émergence*]{}, parmi lesquelles seule la troisième a une chance de pouvoir s’appliquer pour ce qui nous concerne. Selon Butterfield et Isham, une théorie $T_1$ est dite [*émerger*]{} d’une autre théorie $T_2$ si et seulement s’il existe un procédé de dérivation comme limite, ou d’approximation (ou quelque combinaison des deux) d’une théorie à l’autre. Un [*procédé de dérivation comme limite*]{} consiste à prendre la limite mathématique, en général dans un ordre particulier, de certains paramètres de la théorie sous-jacente, paramètres qui sont pertinents du point de vue physique, afin d’en arriver à la théorie émergente. Un tel procédé ne peut pas marcher dans notre cas, du moins pas s’il est appliqué seul, du fait de problèmes techniques concernant d’un côté la densité de boucle maximale, et de l’autre le problème de la mesure bien familier en mécanique quantique non-relativiste.
Un [*procédé d’approximation*]{} consiste ou bien à négliger, justification à l’appui, certaines grandeurs physiques, ou bien à sélectionner, ici encore justification à l’appui, un sous-ensemble propre de l’espace des états de la théorie approximans – celle qui sera l’approximation de l’autre, ou bien à faire les deux, le tout afin de se trouver avec une théorie dans laquelle les valeurs des grandeurs physiques restent suffisamment proches de celles de la théorie à approximer, la théorie approximanda . Notons que, dans notre cas, la théorie approximanda ne sera pas la RG, mais seulement au secteur du vide des espaces de topologie $\Sigma \times \mathbb{R}$. l’une des questions centrales sera celle de savoir comment justifier la sélection faite sur les états. Une telle justification peut être obtenue si on parvient à identifier un mécanisme par lequel le système est dirigé vers les états voulus. Toute tentative de ce genre nous imposera de traiter d’une foule de difficultés liées au problème bien connu des liens entre mécaniques quantique et classique. Il est possible qu’un mécanisme de ce genre puisse être trouvé dans une forme de décohérence , même si les formes standard de décohérence comprennent une notion d’ environnement avec lequel le système considéré est en interaction. Or le système qui nous intéresse ici est, bien entendu, l’univers lui-même, ce qui rend difficile de voir comment il pourrait y avoir un environnement extérieur avec lequel le système pourrait interagir. Le défi qui se présente à nous consiste donc à conceptualiser la décohérence d’une façon qui contourne ce problème.
Même s’il reste encore beaucoup de travail à accomplir, du point de vue technique comme du point de vue philosophique, je m’aventurerai ici à avancer la thèse – ou devrais-je dire la reconnaissance de dette – selon laquelle, dans la mesure au moins où la GQB est une théorie cohérente, la RG (ou une proche cousine) peut être considérée comme émergeant de la GQB si une subtile combinaison de procédés de dérivation comme limite et d’approximation est mise en place. Cette affirmation est illustrée en Figure \[fig:buttisham\],
où on peut voir que l’idée consiste à appliquer d’abord un procédé d’approximation au niveau de la théorie quantique – théorie consistant en un espace de Hilbert $\mathcal{H}$ et un ensemble d’operateurs $\{\hat{O}\}$ définis sur $\mathcal{H}$ –, procédé qui permette de diriger le système physique vers un sous-espace semi-classique qui peut être à son tour mis en relation avec l’espace classique des états $\Gamma$ au moyen d’une dérivation comme limite. Il est bien certain que ce n’est là qu’une esquisse grossière, et il y faut ajouter beaucoup de détails, mais on peut se reporter à Wüthrich (2006, Ch. 9) pour un début d’analyse dans cette direction.
Une fois qu’on aura compris comment l’espace et le temps classiques disparaissent en gravité quantique canonique, et comment ils pourraient bien ré-émerger depuis la structure fondamentale non spatio-temporelle, la façon dont la classicalité émerge depuis la théorie quantique de la gravité ne diffère pas vraiment de la façon dont elle émerge de la mécanique quantique ordinaire. Chercher à comprendre cela est un projet pertinent et intéressant pour au moins deux raisons. Premièrement, le projet touche certaines questions fondamentales importantes concernant l’interprétation des théories, et les relations que ces théories entretiennent entre elles, questions dont la résolution peuvent mener à des clarifications conceptuelles pour les fondements de la physique. Un progrès conceptuel de ce type pourrait bien se trouver être une étape cruciale sur la route menant vers la formulation d’une théorie de la gravité quantique complète. Deuxièmement, la gravité quantique est un sol fertile pour tout métaphysicien puisqu’elle aura inévitablement des implications concernant les questions philosophiques, et en particulier les questions métaphysiques de la nature de l’espace et du temps.
Remerciements {#remerciements .unnumbered}
=============
Je veux remercier ici Soazig Le Bihan pour son invitation à contribuer à ce volume, pour la traduction de cet article ainsi que pour sa grande patience. Ce projet a été subventionné en partie par une Collaborative Research Fellowship offerte par le American Council of Learned Societies, en partie par une UC President’s Fellowship in the Humanities offerte par l’University of California, et en partie par une Arts and Humanities Award offerte par l’University of California, San Diego.
[99]{}
Barbour, Julian (2008), The nature of time , disponible sur [www.fqxi.org/community/forum/topic/360](http://www.fqxi.org/community/forum/topic/360).
Butterfield, Jeremy et Chris Isham (1999), On the emergence of time in quantum gravity , [*in*]{} J. Butterfield (dir.), [*The Arguments of Time*]{} (Oxford University Press), 111-168.
Butterfield, Jeremy et Chris Isham (2001), Spacetime and the philosophical challenge of quantum gravity , [*in*]{} Callender et Huggett, 33-89.
Callender, Craig (2010), Is time an illusion? , [*Scientific American*]{}, June: 58-65.
Callender, Craig et Nick Huggett (dir.) (2001), [*Philosophy Meets Physics at the Planck Scale*]{} (Cambridge University Press).
Dirac, Paul A M (1964), [*Lectures on Quantum Mechanics*]{} (Belfer Graduate School of Science Monograph Series, New York). Ré-imprimé par Dover Publications, Mineola, NY (2001).
Earman, John (2002), Thoroughly modern McTaggart, or what McTaggart would have said if he had read the general theory of relativity , [*Philosophers’ Imprint*]{} 2/3.
Huggett, Nick et Craig Callender (2001), Why quantize gravity (or any other field for that matter)? , [*Philosophy of Science*]{} 68: S382-S394.
Isham, Chris (1994), Prima facie questions in quantum gravity , [*in*]{} J. Ehlers et H. Friedrich (dir.), [*Canonical Gravity: From Classical to Quantum*]{} (Springer), 1-21.
Kiefer, Claus (2008), Does time exist in quantum gravity? , disponible sur [www.fqxi.org/community/forum/topic/265](http://www.fqxi.org/community/forum/topic/265).
Mattingly, James (2006), Why Eppley and Hannah’s thought experiment fails , [*Physical Review*]{} D73: 064025.
Maudlin, Tim (2002), Thoroughly muddled McTaggart or how to abuse gauge freedom to create metaphysical monstrosities , [*Philosophers’ Imprint*]{}, 2/4. Avec une réponse de John Earman.
Maudlin, Tim (2007), Completeness, supervenience and ontology , [*Journal of Physics A: Mathematical and Theoretical*]{} 40: 3151-3171.
Rovelli, Carlo (1996), Relational quantum mechanics , [*International Journal of Theoretical Physics*]{} 35: 1637-1678.
Rovelli, Carlo (2004), [*Quantum Gravity*]{} (Cambridge University Press).
Rovelli, Carlo (2008), Forget time , disponible sur [www.fqxi.org/community/forum/topic/237](http://www.fqxi.org/community/forum/topic/237).
Smolin, Lee (2009), Generic predictions of quantum theories of gravity , [*in*]{} D. Oriti (dir.), [*Approaches to Quantum Gravity*]{} (Cambridge University Press), 548-570.
Wald, Robert M (1984), [*General Relativity*]{} (University of Chicago Press).
Wheeler, John A (1990), [*A Journey Into Gravity and Spacetime*]{} (Scientific American Library).
Will, Clifford M (2006), The confrontation between general relativity and experiment , [*Living Rev. Relativity*]{} 9/3. [www.livingreviews.org/lrr-2006-3](http://www.livingreviews.org/lrr-2006-3).
Wüthrich, Christian (2005), To quantize or not to quantize: fact and folklore in quantum gravity , [*Philosophy of Science*]{} 72: 777-788.
Wüthrich, Christian (2006), [*Approaching the Planck Scale from a Generally Relativistic Point of View: A Philosophical Appraisal of Loop Quantum Gravity*]{}, PhD dissertation, University of Pittsburgh.
[^1]: Plus précisément, elle se doit de se transformer en théorie quantique des champs définie sur l’espace-temps de Minkowski caractéristique de la relativité restreinte.
[^2]: Cf. Callender et Huggett (2001), Huggett et Callender (2001), Mattingly (2006), et Wüthrich (2005).
[^3]: Ceci parce que les quantifications covariantes de la RG ne sont pas renormalisables, un défaut qui est le plus souvent considéré comme fatal. Cela ne veut pas dire, cependant, que les techniques covariantes ne jouent aucun rôle dans la théorie de gravité quantique contemporaine.
[^4]: (NdT) en français dans le texte.
[^5]: (NdT) causal set theory en anglais. Certains traduisent par théorie causale des ensembles mais il s’agit sans doute là d’une traduction malheureuse, la causal set theory n’étant aucunement une théorie des ensembles mais plutôt une théorie physique fondée sur les ensembles causalement reliés.
[^6]: Plus précisément, les équations de Hamilton sont un système d’équations du premier ordre imposant des contraintes dynamiques sur l’espace des phases du système, espace de dimension $2n$ si $n$ est le nombre de degrés de liberté.
[^7]: Mathématiquement parlant, les équations d’Einstein sont un système de dix équations différentielles partielles, non-linéaires, et indépendantes du second degré, qui se réduisent à six équations indépendantes quand on prend en compte la liberté de choix des coordonnées d’espace-temps. Quatre de ces six équations imposent des contraintes liées à l’invariance par difféomorphisme à quatre dimensions, sur laquelle nous reviendrons plus bas.
[^8]: Le bon mot de Wheeler apparaît dans bon nombre de ses écrits, comme par exemple dans Wheeler (1990, xi).
[^9]: On trouvera des détails concernant le formalisme ADM et la façon dont les équations de contraintes y apparaissent dans Wald (1984, Chapitres 10 et Appendice E.2).
[^10]: Sur ce sujet, voir Wüthrich (2006, Section 4.1) \[NdT ainsi que l’article de A. Guay dans ce volume\].
[^11]: Un difféomorphisme est une application bijective et continue entre deux variétés différentiables dont l’inverse est aussi continue.
[^12]: Je passe ici sur certains details: à strictement parler, c’est une combinaison linéaire de contraintes des deux familles. Mais cela ne change pas le fait que ce soit une contrainte liée à un choix de gauge.
[^13]: Et même si ces simplifications dependent de la résolution encore aujourd’hui aussi problématique que controversée d’un certain nombre de difficultés techniques.
[^14]: Rovelli (2004) est le manuel standard.
[^15]: Plus précisément, du fait que les états de réseaux de spin ne sont pas invariant par difféormorphisme, des [*classes d’équivalence*]{} d’états de réseaux de spin par difféomorphisme à trois dimensions doivent être prises pour représenter la structure fondamentale de l’espace physique. Le fait que les états de réseaux de spin ne sont pas invariants par difféomorphisme se voit facilement quand on considère que, à strictement parler, les faire se ballader – tous ou une partie d’entre eux – sur l’espace les contenant sans changer leur structure nodale telle qu’elle est indiquée par les flèches de la Figure \[fig:spinnetwork\] résultera en un réseaux de spin différents à chaque fois. Mais comme c’est la structure nodale qui représente la situation physique, et non la façon particulière dont cette structure est inscrite dans l’espace, on doit considerer ce que les mathématiciens appelent des [*graphes abstraits*]{}, i.e. des classes d’équivalence de graphes dont la structure nodale est la même, mais inscrite de façons différentes dans l’espace. Ce point, important bien que subtil, sera ignoré dans la suite.
[^16]: Pour une reference canonique sur la mécanique quantique relationnelle , voir Rovelli (1996); pour une présentation tout public, voir Callender (2010).
[^17]: Cf. Wüthrich (2006, §4.3).
[^18]: Mais il y a aussi des opposants: cf. Maudlin (2002).
[^19]: Cf. par exemple Barbour (2008), Kiefer (2008), et Rovelli (2008).
[^20]: Maudlin (2007) ne pense pas que ce soit le cas.
[^21]: Cf. Smolin (2009, 549).
[^22]: A ma connaissance, la littérature concernant l’émergence en gravité quantique canonique se réduit aux deux articles de Butterfield et Isham (1999, 2001) cités en bibliographie, et à Wüthrich (2006).
[^23]: Et de fait, elle a bien été couronnée de success; cf. Will (2006).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
[Detlev Buchholz$^a$ and Stephen J. Summers$^b$ ]{}\
${}^a$ Institut für Theoretische Physik, Universität Göttingen,\
37077 Göttingen, Germany\
${}^b$ Department of Mathematics, University of Florida,\
Gainesville FL 32611, USA
date: Dedicated to Jacques Bros on the occasion of his seventieth birthday
title: 'Stable Quantum Systems in Anti–de Sitter Space: Causality, Independence and Spectral Properties'
---
Introduction and basic assumptions {#intro}
==================================
Quantum field theory in anti-de Sitter space–time (AdS) has been studied for almost 40 years (see [*e.g.*]{} [@F1; @AIS]), primarily because it was found that AdS occurs as the ground state geometry in certain supergravity theories with gauged internal symmetry [@BrFr; @dWNi2]. But it has become the object of an extraordinary amount of attention since the AdS-CFT correspondence has emerged.[^1] There is therefore motivation to clarify in a model–independent setting and in a mathematically rigorous manner the universal properties of such theories, as implied by generally accepted and physically meaningful assumptions. This investigation has lead us to results which apparently have not been remarked in any form in the literature before.
AdS is a maximally symmetric and globally static solution of the vacuum Einstein equations. We consider here AdS of any dimension $n \geq 2$, except when explicitly stated otherwise. It can conveniently be described in terms of Cartesian coordinates in the ambient space $\RR^{n+1}$ as the quadric surface $${\mbox{{AdS}$^n\,$}}= \{x\in\RR^{n+1} \mid
x^2 \doteq x_0^2 - x_1^2 - \dots - x_{n-1}^2 + x_n^2 = R^2 \}$$ with metric $g = \mbox{diag}(1,-1,\dots,-1,1)$ in diagonal form. As the value of the radius $R$ is not relevant for the results of this paper, we shall set it equal to $1$ for convenience. The isometry group is O$(2,n-1)$ whose identity component will be denoted by . is a homogeneous space of the group . It is not globally hyperbolic; indeed, it has closed timelike curves and has a timelike boundary at spatial infinity through which physical data can propagate. Although the covering space of eliminates the closed timelike curves, it still has a timelike boundary at spatial infinity. We shall find some notable differences between the properties of quantum field theories on and those of theories on the covering space.
As some of our basic assumptions are motivated by physical considerations concerning certain families of observables, we must collect some basic facts about observers in AdS. Let $x_O \in {\mbox{{AdS}$^n\,$}}$ be any point and let $\lambda (t)$, $t \in \RR$, be any one–parameter subgroup of ${\mbox{SO$_0(2,n\!-\!1)$}}$ such that $t \mapsto \lambda(t) x_O$ is an orthochronous curve. (Note that AdS is time orientable.) We interpret this curve as the worldline of some observer. Among these observers will be those moving along a geodesic (henceforth, geodesic observers) and those experiencing a constant acceleration (uniformly accelerated observers). Points in a neighborhood of $x_O$ will, in general, also give rise to orthochronous curves under the action of the chosen subgroup of ${\mbox{SO$_0(2,n\!-\!1)$}}$, and we denote by $W$ the connected neighborhood of $x_O$ in consisting of all such curves. Typically, $W$ is the causal completion of the originally specified worldline. We view the region $W$ as the maximal possible localization for any laboratory within the purview of the given observer. The associated dynamics are given by $e^{\, itM} \doteq U(\lambda(t))$ with suitable generator $M$. Since we are choosing a fixed parametrization of the pertinent subgroups of , the proper time of the observer is obtained by rescaling $t$ with $((\dot{\lambda}(0) \, x_O)^2){}^{1/2}$.
To become more precise, the geodesics of are conic sections by two–planes containing the origin of the ambient space $\RR^{n+1}$. So, one-parameter subgroups[^2] $\lambda(t)$, $t \in \RR$, of ${\mbox{SO$_0(2,n\!-\!1)$}}$ of the form $\lambda \lambda_{0n}(t) \lambda^{-1}$, $t \in \RR$, for some $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$ generate admissible geodesic worldlines in the sense just indicated. These worldlines are closed, timelike curves, whose causal completion is the entire space ${\mbox{{AdS}$^n\,$}}$. Hence, the maximal laboratory localization region $W$ for such geodesic observers must be the entire space-time, ${\mbox{{AdS}$^n\,$}}$. For uniformly accelerated observers, the corresponding one-parameter subgroups are of the form $\lambda \lambda_{01}(t) \lambda^{-1}$, $t\in\RR$, for some $\lambda\in{\mbox{SO$_0(2,n\!-\!1)$}}$. Their laboratory regions, called AdS wedges, are described immediately below. The algebras $\As(W)$ corresponding to any such wedge region as well as to $W = {\mbox{{AdS}$^n\,$}}$ are taken to be weakly closed.
We define a “wedge” in to be the causal completion of the worldline of a uniformly accelerated observer in . To be concrete and in order to simplify the necessary computations, we consider the particular choice of region $$W_R = \{ x \in {\mbox{{AdS}$^n\,$}}\mid x_1 > |x_0|\, , x_n > 0 \} \, ,$$ on which the one-parameter subgroup of boosts $\lambda_{01}(t)$, $t\in\RR$, in the $0$–$1$–plane acts in an orthochronous manner. For any $x_O \in W_R$, the curve $t \mapsto \lambda_{01}(t) x_O$ is the worldline of a uniformly accelerated observer for which the causal completion is precisely $W_R$. By the assumed covariance, all results concerning this wedge have natural extensions to all images of $W_R$ under . We therefore define the set of wedges to be $$\Ws \doteq \{ \lambda W_R \mid \lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}\} \, .$$ These are maximal laboratory localizations for the uniformly accelerated observers.
We can now describe the four standing assumptions of this paper. A discussion of their physical motivation is given in [@AdS], so we shall only expand upon the less familiar ones.
1. There exists a strongly continuous, unitary, nontrivial representation $U$ of the symmetry group acting on a separable Hilbert space ${{\cal H}}$.[^3]
2. On $\Hs$ act the global von Neumann algebra of observables ${{\cal A}}= \As({\mbox{{AdS}$^n\,$}})$, which contains any observable measurable in ${\mbox{{AdS}$^n\,$}}$, and an isotonous family of von Neumann algebras $\wnet$ associated with the wedges $\Ws$. Furthermore, one has $$\label{weakadditivity}
\bigvee_{W \in \Ws} \As(W) = \As \, .$$
3. For each wedge $W \in \Ws$ and $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$, one has the equality $$U(\lambda) {{\cal A}}(W) U(\lambda)^{-1} = {{\cal A}}(\lambda W) \, .$$
The weak additivity condition (\[weakadditivity\]) is a generalization of the natural idea that all observables are constructed out of local ones. But in contrast to [@AdS], we do not assume that all observables can be constructed out of observables with arbitrarily small localization region. This is because there are nets of physical interest on curved space–times for which the condition (v) specified below holds but the algebras $\As(\Os)$ associated with all bounded open regions $\Os$ are trivial [@BMS; @Rehren]. Also, there exist examples in which the algebra $\As(\Os)$ is nontrivial only for sufficiently large bounded regions $\Os$ [@BMS]. In both of these cases the assumption made in [@AdS] is violated. We have therefore eliminated all assumptions referring to bounded regions.
We emphasize that we do not postulate from the outset any local commutation relations of the observables. For, in contrast to the case of globally hyperbolic space–times, the principle of Einstein causality does not provide any clues as to which observables in AdS should commute with each other. Instead, we shall [*derive*]{} such commutation relations from stability properties of the vacuum, which we now specify.
We shall assume that the state $\omega$ determined by $\Omega$ is passive (cf. [@PuWo] and Section 5.4.4 in [@BratRob]) for the dynamical system $(\As(W), \ad U(\lambda(t)))$, for all geodesic and all uniformly accelerated observers described above. We recall that passivity is an expression of the Second Law of Thermodynamics. Since the vacuum is the most elementary system, all order parameters should have sharp values in this state. This is expressed by the weak mixing property: $$\label{weakmixing}
\lim_{T \rightarrow \infty}
\frac{\mbox{\footnotesize $1$}}{\mbox{\footnotesize $T$}}
\int_0^T \, \left( \omega( A(t) B) -
\omega(A(t))\, \omega(B) \right) \, dt = 0 \, ,$$ for all $A,B \in {{\cal A}}$, where $A(t) \doteq e^{\, itM } A e^{\, -itM }$. The restriction of the state $\omega$ to $\As(W)$ is said to be central if $\omega(AB) = \omega(BA)$, for all $A,B \in \As(W)$. If this holds, then either $\Omega$ is annihilated by most of the observables in $\As(W)$ or $\As(W)$ is a finite algebra (cf. Section 8.1 in [@KadRing]). In quantum field theory this is a physically pathological circumstance, which we shall exclude from consideration.
These basic features of the vacuum can be summarized as follows.
\(iv) The vacuum vector $\Omega$ is cyclic for $\As$ and determines a passive, weakly mixing and noncentral state $\omega$ for all geodesic and all uniformly accelerated observers.
The Standing Assumptions (i)–(iv) are model–independent and physically natural. In this paper we shall show that these assumptions entail that for geodesic observers the vacuum $\omega$ is a ground state; uniformly accelerated observers in AdS will register a universal value of the Unruh temperature; they will discover a PCT symmetry; and they will find that observables localized in complementary wedge–shaped regions must commute in the vacuum state. Not only do such observables commute in this sense, but the corresponding algebras manifest strong properties of statistical independence, the nature of which will be studied in detail. We shall also see that these assumptions imply that quantum theories on AdS obey a geodesic causal structure.
Related results appeared in [@AdS], and we revisit some of those arguments here in more detail than in that announcement. But our research in the intervening time has led not only to further results and a weakening of the assumptions, but also to a shift in our point of view, which now places emphasis on the locality and independence properties which can be derived from our assumptions. We establish independence properties going far beyond those announced in [@AdS], and we prove an additional locality property of such theories on proper which was not observed in [@AdS]. Moreover, we show that in two dimensions these suffice to construct a nontrivial, local, covariant net indexed by bounded spacetime regions. We also explain how known examples of quantum fields on AdS fit into our scheme.
The primary lesson to be drawn from this paper is the observation that covariance and passivity properties of states induce strong algebraic relations between the observables, which may be interpreted as manifestations of Einstein causality. In our research program, the theories on AdS treated here serve as a theoretical laboratory to test this striking feature. But the insight gained in this analysis goes beyond this class of field theoretical models to quantum fields on other space–times. Further, it seems to be of relevance in the discussion of causality problems appearing in nonlocal theories, such as string theory and quantum field theory on noncommutative space–times.
Unruh effect, PCT symmetry and weak locality {#main}
============================================
We now enter into the analysis of the implications of our standing assumptions (i)–(iv) by appealing to a deep result of Pusz and Woronowicz for general quantum dynamical systems [@PuWo]. In the present context this result says that the vacuum vector $\Omega$ is, as a consequence of its passivity and mixing properties, invariant under the dynamics of any of the observers discussed above [@PuWo Theorem 1.1]. In particular, this entails that $M_{01} \Omega = 0$ (and hence, by Lemma \[invariance\], $\Omega$ is invariant under the entire group $U({\mbox{SO$_0(2,n\!-\!1)$}})$), and $\omega$ is either [@PuWo Theorem 1.3] a ground state for $M_{01}$ (which is excluded by Lemma \[nogroundstate\]), or satisfies, for some [*a priori*]{} unknown $\beta \geq 0$, the Kubo–Martin–Schwinger (KMS) condition. In fact, our assumptions exclude the possibility of $\beta = 0$. In the proof of Lemma 4.1 in [@PuWo] it is shown that if $\beta = 0$, then either $\omega$ is a trace state on $\As(W_R)$ or $M_{01} = 0$. In the second case, one would have the triviality of the representation of the boost group and thus the triviality of $U({\mbox{SO$_0(2,n\!-\!1)$}})$, which is excluded by (i). The first case is excluded by assumption (iv). Therefore, for any pair of operators $A,B \in \As(W_R)$ there exists an analytic function $F$ in the strip $\{ z \in \CC \mid 0 < \Im(z) < \beta \}$ with continuous boundary values at $\Im(z) = 0$ and $\Im(z) = \beta$, which are given by $$\label{kms}
F(t) = \omega( AB(t)) \, , \quad
F(t+i\beta) = \omega( B(t)A) \, ,$$ respectively, for all $t \in \RR$ and with $B(t) \doteq e^{itM_{01}}Be^{-itM_{01}}$. By the –covariance the same assertions are valid for the action of the groups $e^{itM_{0j}}$, $j = 2,\ldots,n-1$, on the suitable wedge algebras.
In Appendix \[sectionReehSchlieder\] it is proven that this analyticity entails that the theories we are considering here satisfy the Reeh–Schlieder property (cf. Lemma \[ReehSchlieder\]). So the vacuum vector $\Omega$ is cyclic for the algebra $\As(W)$, given any $W \in \Ws$, and, by the KMS-property, it is also separating for $\As(W)$ [@BratRob Corollary 5.3.9]. Hence, the Tomita–Takesaki modular theory is applicable to $(\As(W),\Omega)$, for every $W \in \Ws$ (cf. [@BR; @KadRing]). Let $J_{W_R}$ denote the modular conjugation and $\Delta_{W_R}^{it}$ the modular unitaries associated to the pair $(\As(W_R),\Omega)$. Since the adjoint action of the strongly continuous unitary group $e^{itM_{01}}$, $t \in \RR$, leaves the algebra $\As(W_R)$ invariant and satisfies the KMS condition, we must have $\Delta_{W_R}^{it} = e^{-i\beta tM_{01}}$, for all $t \in \RR$ [@KadRing Theorem 9.2.16]. Hence, $J_{W_R}$ is determined by the equation $$\label{jwr}
J_{W_R} A \Omega = e^{\, -(\beta/2) M_{01}} A^* \Omega
\, , \ A \in {{\cal A}}(W_R) \, .$$
Unruh temperature
-----------------
The main task of this subsection is to determine the Unruh temperature $\beta^{-1}$ and specific properties of the operator $J_{W_R}$. To this end we shall adapt methods employed in [@BB].
We show in Lemma \[neighborhood\] that there exists a wedge $W_0 \in \Ws$ such that $\lambda W_0 \subset W_R$ for all $\lambda$ in a neighborhood of the identity in ${\mbox{SO$_0(2,n\!-\!1)$}}$. Therefore, for any $j = 2,\ldots,n-1$ one has $\lambda_{0 j} (s) \, W_0 \subset W_R$ for the boosts $\lambda_{0 j} (s)$ in the $0$–$j$–plane for all sufficiently small parameters $s$. From equation (\[010j\]) in Appendix \[grouptheory\] we have $$e^{\, itM_{0 1}} e^{\, isM_{0 j}} =
e^{\, is(\cosh(t)M_{0 j} + \sinh(t) M_{1 j})} e^{\, itM_{0 1}} \, .$$ Thus we get for any vector $\Phi \in {{\cal H}}$ and operator $A^* \in {{\cal A}}(W_0)$ $$\label{equation11}
\langle \Phi , \, e^{\, itM_{0 1}}
\, e^{\, isM_{0 j}} A^* e^{\, -isM_{0 j}} {{\Omega \rangle}}= \langle \Phi , \,
e^{\, is(\cosh(t)M_{0 j} + \sinh(t) M_{1 j})}\, e^{\, itM_{0 1}}
A^* {{\Omega \rangle}}\, .$$ We are now in the position of employing the argument given in [@AdS] to yield the equalities $$\label{equation13}
J_{W_R} e^{\, isM_{0 j}} =
e^{\, is(\cos(\beta/2) M_{0 j} + i \sin(\beta/2) M_{1 j})} J_{W_R} \, .$$ As pointed out in [@AdS], the operator on the left–hand side of this equation is anti–unitary, which entails that $\beta$ is an [*integer*]{} multiple of $2 \pi$, for otherwise the operator appearing in the exponential function on the right–hand side would not be skew-adjoint. By using the proof of Theorem 6.2 in [@BB] with $\As(\Os)$ replaced by $\As(W_0)$, one sees that its only possible value is $\beta = 2 \pi$. Proceeding to the proper time scale of the observer, we conclude that he is exposed to the Unruh temperature $(1/2\pi)((\dot{\lambda}_{0 1}(0)\, x_O)^2){}^{-1/2}$, in accordance with the value found in computations for some particular models [@DeLe; @Jac] and also by more general considerations [@BEM].
For geodesic observers, Lemma \[nogroundstate\] is not applicable. In fact, $\omega$ cannot be a KMS-state for $e^{itM_{0n}}$ on $\As$, the laboratory observable algebra for geodesic observers. Indeed, since the covariance assumption (iii) implies $U(\lambda)\As U(\lambda)^{-1} = \As$, for all $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$, if $e^{itM_{0n}}$ were the modular group for $\Omega$ on $\As$, then modular theory would necessitate $U(\lambda)e^{itM_{0n}} = e^{itM_{0n}}U(\lambda)$, for all $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$ (cf. Thm. 3.2.18 in [@BR]). But this would only be possible if the representation $U({\mbox{SO$_0(2,n\!-\!1)$}})$ were trivial, which is excluded by assumption (i). So $\omega$ must be a ground state for $e^{itM_{0n}}$.
We have therefore established the following general facts:
\[temperature\] Let Standing Assumptions (i)–(iv) hold. Then $\Omega$ is invariant under the action of $U({\mbox{SO$_0(2,n\!-\!1)$}})$ and each uniformly accelerated observer testing $\Omega$ in finds a universal value $(1/2\pi)((\dot{\lambda}_{0 1}(0)\, x_O)^2){}^{-1/2}$ of the Unruh temperature which depends only on his particular orbit. For geodesic observers $\omega$ is a ground state; in particular, $M_{0n}$ is a positive operator.
This result is a consequence of the passivity of $\omega$, and this vacuum state is the only normal state on $\As$ which is passive for all observers. In light of Theorem \[temperature\], it is physically justified to identify the operator $M_{0n}$ with the global energy operator.
The result $\beta = 2\pi$ and [@KadRing Theorem 9.2.16] permit us to completely determine the modular unitaries corresponding to the pair $(\As(W),\Omega)$, for all $W \in \Ws$.
\[modularunitary\] Given the Standing Assumptions (i)–(iv), the modular unitaries for the pair $(\As(W_R),\Omega)$ are given by $$\Delta_{W_R}^{it} = e^{-i2\pi tM_{01}} \, , \quad t \in \RR \, .$$
Covariance and the uniqueness of the modular objects yield similar results for $\Delta_W^{it}$, for all $W \in \Ws$.
PCT symmetry
------------
Having computed the modular unitaries and the value of the inverse temperature $\beta$ seen by the uniformly accelerated observers, let us return now to the analysis of $J_{W_R}$ and clarify its relation to spacetime reflections. Plugging $\beta = 2 \pi$ into equation (\[equation13\]), we see that for small $s$ $$J_{W_R} e^{\, isM_{0 j}} =
e^{\, - is M_{0 j}} J_{W_R} \, , \quad j = 2,\ldots,n-1 \, .$$ This relation can be extended to arbitrary $s$ by iteration, if one decomposes $e^{\, isM_{0 j}}$ into an $m$–fold product $(e^{\, i(s/m)M_{0 j}})^m$ for sufficiently large $m$. A similar argument with $\lambda_{0j}$ replaced by $\lambda_{0n}$ and equation (\[010j\]) replaced by (\[010n\]) yields $$\label{WR0n}
J_{W_R} e^{\, isM_{0 n}} =
e^{\, - is M_{0 n}} J_{W_R} \, .$$ On the other hand, modular theory yields $$J_{W_R} e^{\, isM_{0 1}} =
e^{\, is M_{0 1}} J_{W_R} \, .$$ Since these one-parameter subgroups generate ${\mbox{SO$_0(2,n\!-\!1)$}}$, the intertwining properties of $J_{W_R}$ with all unitaries $U(\lambda) \in U({\mbox{SO$_0(2,n\!-\!1)$}})$ are determined.
\[TCPgroup\] Given Standing Assumptions (i)–(iv), one has $$J_{W_R} U(\lambda) = U(\theta_{01}\lambda\theta_{01}) J_{W_R} \, ,$$ for all $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$, where $\theta_{01} = \diag(-1,-1,1,\ldots,1)$ is the reflection which changes the sign of the $0$–$1$–coordinates of the points in (the reflection about the edge of the wedge $W_R$).
Hence, if we define $U(\theta_{01}\lambda) \doteq J_{W_R}U(\lambda)$, for any $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$, the following partial analogue of the PCT–theorem can be proven.
\[proper\] If Standing Assumptions (i)–(iv) hold, then the unitary representation $U$ of extends to a representation of in which the reflection $\theta_{01}$ is implemented by the anti–unitary involution $J_{W_R}$.
Note that ${\mbox{SO$(2,n\!-\!1)$}}$ is the disjoint union of $\theta_{01}{\mbox{SO$_0(2,n\!-\!1)$}}$ and ${\mbox{SO$_0(2,n\!-\!1)$}}$. Since $$U(\theta_{01} \lambda_1 \cdot \lambda_2) = J_{W_R} U(\lambda_1 \lambda_2)
= J_{W_R} U(\lambda_1)U(\lambda_2) = U(\theta_{01}\lambda_1)U(\lambda_2)$$ and $$\begin{aligned}
U(\theta_{01} \lambda_1 \cdot \theta_{01}\lambda_2) & = &
U(\theta_{01} \lambda_1 \theta_{01})U(\lambda_2) \\ \nonumber
& = & J_{W_R} U(\lambda_1) J_{W_R} U(\lambda_2)
= U(\theta_{01}\lambda_1)U(\theta_{01}\lambda_2) \, ,\end{aligned}$$ for all $\lambda_1,\lambda_2 \in {\mbox{SO$_0(2,n\!-\!1)$}}$, the assertion follows.
Theorem \[proper\] is a purely group–theoretic statement which does not yet say anything about the adjoint action of $J_{W_R}$ on the observables. Results of that type require an additional assumption and will appear in a later publication.
Weak locality {#weaklocalitysec}
-------------
In order to gain insight into the locality properties of the net, we consider the observables which are localized in the region $$W_R{}^{\prime} \doteq \{x\in{\mbox{{AdS}$^n\,$}}\mid - x_1 > | x_0 |, \, x_n > 0\} \, .$$ Throughout this subsection we shall only consider of dimension $n \geq 3$. Since the regions $W_R$ and $W_R{}^{\prime}$ are each one–half (on the same “side” of ${\mbox{{AdS}$^n\,$}}$ — see Figure 1 below) of the regions obtained by intersecting opposite wedge–shaped regions in the ambient space $\RR^{n+1}$ with , we call them opposite wedges. In general, if $W = \lambda W_R$, for some $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$, then $W' = \lambda W_R{}'$.
In [@AdS] it was shown that for $n \geq 3$ the above results entail that observables which are localized in complementary wedges are weakly local in the vacuum state. So we can state:
\[weaklocality1\] Under the assumptions (i)–(iv) and for $n \geq 3$, observables which are localized in opposite wedges of are weakly local with respect to each other. Explicitly, for any $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$ and any $A' \in \As(\lambda W_R{}')$, $B \in \As(\lambda W_R)$, one has $\langle\Omega,A'B\Omega\rangle = \langle\Omega,BA'\Omega\rangle$.
This result is also valid for theories in the covering space of which satisfy (i)–(iv). A similar result was proven in [@BEM] for two–point functions of quantum fields on the covering space of satisfying a different set of assumptions in the Wightman framework.
The uniqueness of the modular objects and the covariance assumption (iii) give us a relation we shall use repeatedly in the following: $$\label{unique}
U(\lambda) J_W U(\lambda)^{-1} = J_{\lambda W} \, ,$$ for all $W \in \Ws$ and all $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$. Hence, since $\As(W_R{}') = e^{i\pi M_{12}}\As(W_R)e^{-i\pi M_{12}}$, equation (\[unique\]) and the observations made above imply $$J_{W_R{}'} = e^{i\pi M_{12}} J_{W_R} e^{-i\pi M_{12}} =
J_{W_R} e^{-i2\pi M_{12}} = J_{W_R} \, .$$ Thus, the ${\mbox{SO$_0(2,n\!-\!1)$}}$-covariance of the net entails $$\label{equation15}
J_{W'} = J_{W} \, , \quad \textnormal{for every $W \in \Ws$} \, .$$ If the algebras $\As(W)$ and $\As(W')$ commuted strongly with each other, it can be shown that (\[equation15\]) follows from modular theory. It is therefore of interest that (\[equation15\]) obtains when the algebras only weakly commute.
We continue with some further locality results, which distinguish theories on proper AdS from those on the covering space. We have chosen our wedge regions $W \in \Ws$ to be connected for physical reasons, as previously explained. But the intersection of the (connected) wedge [**W**]{} in the ambient space $${\bf W} = \{ x \in \RR^{n+1} \mid x_1 > |x_0| \}$$ with has two connected components, one of which is $W_R$ and the other is the conjugate wedge $$\widetilde{W}_R{}' = \{ x \in {\mbox{{AdS}$^n\,$}}\mid x_1 > |x_0|\, , -x_n > 0 \} \, .$$ One has the geometric relations
\[F1\]
$$\widetilde{W}_R = \theta_{01}\widetilde{W}_R{}^\prime = -W_R \quad
\textrm{and} \quad W_R{}' = -\widetilde{W}_R{}' \, .$$
Moreover, it is easy to see that $$\label{equation16}
e^{i\pi M_{0n}} \As(W_R) e^{-i\pi M_{0n}} = \As(\widetilde{W}_R{}')
\quad \rm{and} \quad
e^{i\pi M_{0n}} \As(W_R{}') e^{-i\pi M_{0n}} = \As(\widetilde{W}_R) \, .$$ The rotation in the $0$–$n$–plane by $\pi$ also reverses the orientation of the world lines $\lambda_{01}(t)x_O$ (cf. equation (\[0n01\])), so that the world lines in $\widetilde{W}_R{}'$ run in the same direction as those of $W_R{}'$, while those in $\widetilde{W}_R$ have the same orientation as those in $W_R$. Explicitly, Corollary \[modularunitary\] implies that the modular unitaries for the pair $(\As(W_R{}'),\Omega)$ are given by
$$\Delta_{W_R{}'}^{it} = e^{i2\pi tM_{01}} \, , \quad t \in \RR \, .$$
Hence, for all $t \in \RR$, one has $$\Delta_{\widetilde{W}_R}^{it} =
e^{i\pi M_{0n}} \Delta_{W_R{}'}^{it} e^{-i\pi M_{0n}} =
e^{i\pi M_{0n}} e^{i2\pi t M_{01}} e^{-i\pi M_{0n}}
= e^{-i2\pi t M_{01}} = \Delta_{W_R}^{it} \, ,$$ by (\[0n01\]). So, $\{ e^{-i2\pi tM_{01}}\}_{t \in \RR}$ is the group of modular unitaries for $(\As(\widetilde{W}_R),\Omega)$, and we have the relation $$J_{\widetilde{W}_R} A \Omega = e^{\, -\pi M_{01}} A^* \Omega \ , \quad
\mbox{for} \ \ A \in {{\cal A}}(\widetilde{W}_R) \, .$$ Moreover, we may appeal to (\[equation15\]) and the modular theory to find $$J_{\widetilde{W}_R} A' \Omega = e^{\, \pi M_{01}} A'{}^* \Omega \ , \quad
\mbox{for} \ \ A' \in {{\cal A}}(\widetilde{W}_R{}') \, .$$ From equations (\[unique\]) and (\[WR0n\]) follow the equalities $$\label{0nR}
J_{\widetilde{W}_R{}'} = e^{i\pi M_{0n}} J_{W_R} e^{-i\pi M_{0n}} =
e^{i2\pi M_{0n}}J_{W_R} = J_{W_R}
\, ,$$ where we have used $e^{i2\pi M_{0n}} = 1$, valid in proper AdS but not in its covering space. Thus, by (\[equation15\]) we have for $n \geq 3$: $$\label{Js}
J_{W_R{}'} = J_{W_R} = J_{\widetilde{W}_R} = J_{\widetilde{W}_R{}'} \, .$$ We can now prove that $\As(\widetilde{W}_R{}')$ and $\As(W_R)$ are weakly local with respect to each other.
\[weaklocality2\] Under the assumptions (i)–(iv) and for $n \geq 3$, observables which are localized in $W_R$ are weakly local with respect to observables localized in $\widetilde{W}_R{}'$. Explicitly, for any $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$ and any $A' \in \As(\lambda \widetilde{W}_R{}')$, $B \in \As(\lambda W_R )$, one has $\langle\Omega,A'B\Omega\rangle = \langle\Omega,BA'\Omega\rangle$.
With the above preparations, one sees that for any $A' \in \As(\widetilde{W}_R{}')$ and $B \in \As(W_R )$, one has $$\begin{split}
{{\langle\Omega }}, A^{\prime *} B {{\Omega \rangle}}& = \overline{{{\langle\Omega }}, B^{*} A^{\prime}{{\Omega \rangle}}}
= \overline{{{\langle\Omega }}, B^{*} J_{W_R} J_{W_R} A^{\prime}{{\Omega \rangle}}} \\ & =
{{\langle\Omega }}, B e^{\, -\pi M_{0 1}} e^{\, \pi M_{0 1}} A^{\prime *} {{\Omega \rangle}}= {{\langle\Omega }}, B A^{\prime *} {{\Omega \rangle}}\, .
\end{split}$$
Theorem \[weaklocality2\] does not hold in the covering space of , since, in general, $e^{i\pi M_{0n}}$ and $J_{W_R}$ will not commute in such theories. Indeed, in [@F2; @BEM] can be found examples of a free field theory on the covering space of for which it can be shown that assumptions (i)–(iv) are satisfied, but the elements of $\As(W_R)$ and $\As(\widetilde{W}_R{}')$ do not commute in the vacuum state — see Appendix \[examples\] for further discussion. However, for theories on proper AdS, Theorem \[weaklocality2\] is consistent with a property of two-point functions observed in [@BEM] and arrived at there by very different means and assumptions.
In Theorems \[weaklocality1\] and \[weaklocality2\] we see that the passivity of the vacuum state and the group relations in ${\mbox{SO$_0(2,n\!-\!1)$}}$ have determined which regions in are to have (weakly) commensurable observables. In theories on AdS where the basic fields satisfy standard c–number commutation relations, it follows from this result that such observables actually commute in the usual (operator) sense [@StWi]. Indeed, it follows then that $$\label{locality}
\As(W_R ') \subset \As(W_R)' \quad \textnormal{and} \quad
\As(W_R) \subset \As(\widetilde{W}_R{}')' \, .$$ But assumptions (i)–(iv) do not imply the strong locality relations (\[locality\]) in general. Indeed, consider a tempered hermitian Fermi–type field $\phi$ with anticommutator $$\label{counterexample}
\phi(x)\phi(y) + \phi(y)\phi(x) = (W(x,y) + W(y,x)) \cdot 1 \, ,$$ where $W(x,y)$ is taken to be any ${\mbox{SO$_0(2,n\!-\!1)$}}$–invariant two-point function satisfying the spectrum condition. (Specific examples are provided by two-point functions given in [@F2 eq. (4.9)]). This anticommutator does not vanish when $x$ and $y$ are in complementary wedges. The field $\phi$ generates a CAR-algebra with a quasifree state fixed by the two-point function $$\omega( \phi(x) \phi(y)) = W(x,y) \, .$$ Proceeding to the GNS representation with cyclic vector $\Omega$, we conclude from [@F2; @BEM] that this example satisfies all of our standing assumptions. Since $W(x,y)$ is symmetric when $x,y$ are in complementary wedges, the weak locality is explicit.
As explained in [@AdS], if (\[locality\]) holds there are grounds to expect that the theory generically cannot have interaction. Theories which are weakly local but not strongly local escape that argument. In the following sections we shall therefore continue to explore consequences of Standing Assumptions (i)–(iv) when (\[locality\]) does not hold, beginning with the nature of the independence of the algebras $\As(W_1)$, $\As(W_2)$ associated with suitable spacelike separated wedges. But, first, we make some further observations.
Assumptions (i)–(iv) entail that each wedge algebra $\As(W)$ is a factor (cf. the proof of Theorem \[split\]). If wedge locality (\[locality\]) held, then it would follow that $\As(W) \cap \As(W') = \CC 1$, [*i.e.*]{} no nontrivial observable can be localized in both $W$ and $W'$. It is therefore of physical interest that this fact also follows directly from (i)–(iv) in the absence of (\[locality\]).
\[extendedlocality\] Let Standing Assumptions (i)–(iv) hold and let $n \geq 3$. Then for any $W \in \Ws$, one has $\As(W) \cap \As(W') = \CC 1$ and $\As(W) \bigvee \As(W') = \Bs(\Hs)$.
By covariance, it suffices to prove the assertion for $W = W_R$. Let $A \in \As(W_R) \cap \As(W_R{}')$ and $\Hs_0$ be the closure in $\Hs$ of $\big( \As(W_R) \cap \As(W_R{}') \big)\Omega$. Since $e^{i2t\pi M_{01}}$, respectively $e^{-i2t\pi M_{01}}$, are the modular unitaries corresponding to $(\As(W_R),\Omega)$, respectively $(\As(W_R{}'),\Omega)$, then with $\Delta = e^{2\pi M_{01}}$ one has $$J_{W_R} \Delta^{1/2} A \Omega = A^* \Omega \, , \quad
J_{W_R{}'} \Delta^{-1/2} A \Omega = A^* \Omega \, .$$ Hence, equation (\[Js\]) entails that $\Delta^{1/2} = \Delta^{-1/2}$ on $\Hs_0$. Therefore, $\Delta A \Omega = A \Omega$, which yields $A \Omega = \Delta^{it} A \Omega = e^{i2t\pi M_{01}} A \Omega$, for all $t \in \RR$. Hence, the Mean Ergodic Theorem (see, [*e.g.*]{}, [@Kato]) entails that $A \Omega = F_0 A \Omega$, where $F_0$ is the projection onto the subspace of vectors in $\Hs$ each left invariant under $U({\mbox{SO$_0(2,n\!-\!1)$}})$, using Lemma \[invariance\]. Since the mixing property in condition (iii) entails that the rank of $F_0$ is 1 and since $\Omega$ is separating for $\As(W_R)$, it follows that $A$ is a multiple of the identity. But this entails $$\CC 1 = J_W (\As(W) \cap \As(W')) J_W =
J_W \As(W) J_W \cap J_W \As(W') J_W =
\As(W)' \cap \As(W')' \, ,$$ using (\[Js\]), so that $\As(W) \bigvee \As(W') = \Bs(\Hs)$, for every $W \in \Ws$.
Before we close this section, we have a final proposition to prove.
\[unequal\] Let Standing Assumptions (i)–(iv) hold, $n \geq 3$ and $W_1,W_2 \in \Ws$. If $W_2 \neq \pm W_1$, then $\As(W_1) \neq \As(W_2)$.
Since two unequal wedges $W_1,W_2 \in \Ws$ have unequal reflections about their edges, unless $W_2$ coincides with $W_1{}', \widetilde{W}_1$ or $\widetilde{W}_1{}'$, Lemma \[TCPgroup\] entails $J_{W_1} \neq J_{W_2}$ and, thus, $\As(W_1) \neq \As(W_2)$.
If $\As(W_1) \subset \As(W_1{}')$, then Theorem \[weaklocality1\] entails that the restriction of $\omega$ to $\Rs(W_1)$ is a trace, which is excluded by assumptions (i) and (iii). Similarly Theorem \[weaklocality2\] yields $\As(W_1) \neq
\As(\widetilde{W}_1{}')$.
The Schlieder property {#Schliedersec}
======================
Many versions of the notion of independence of algebras of observables in spacelike separated regions have emerged in algebraic quantum theory (see [@Sum] for a review), and most are logically independent of the usual notion of commensurability, which is that the algebras commute with each other elementwise. In this section we shall prove that algebras associated with properly spacelike separated wedges $W_1,W_2$ satisfy an extended form of the algebraic independence condition known as the Schlieder property, namely that $A_1 \in \As(W_1), A_2 \in \As(W_2)$ and $A_1 A_2 = 0$ imply either $A_1 = 0$ or $A_2 = 0$. We shall say that two wedges $W_1,W_2$ are properly spacelike separated if $\lambda W_1 \subset W_2{}'$ for all $\lambda$ in some neighborhood of the identity in . Note that $W$ and $W'$ are not properly spacelike separated. Although in de Sitter and Minkowski spaces of dimension $n \geq 3$ such properly spacelike separated wedges do not exist, we show in Appendix \[sectioninclusion\] that they are plentiful in AdS.
The proof of an extended Schlieder property in , $n \geq 3$, will be carried out in a series of steps.
\[indprepare\] Let $W_1,W_2 \in \Ws$ be properly spacelike separated and let $A_{1,k} \in \As(W_1)$, $A_{2,k} \in \As(W_2)$, $k = 1,\ldots,n$. If $B \in \Bs(\Hs)$ is such that $$\sum_{k=1}^n A_{1,k} U(\lambda) B U(\lambda)^{-1} A_{2,k} = 0 \, ,$$ for all $\lambda$ in some neighborhood $\Ns \subset {\mbox{SO$_0(2,n\!-\!1)$}}$ of the identity, then this equality holds for all $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$.
Choosing a smaller neighborhood $\Ns$, if necessary, it may be assumed that there exists a $W \in \Ws$ such that $W_1 \subset \lambda_0 \lambda_1 W$ and $W_2 \subset \lambda_0 \lambda_1 W^\prime$ for $\lambda_0, \lambda_1 \in \Ns$. Let $s \mapsto \lambda_W(s)$ be the group of boosts inducing a positive timelike flow on the wedge $W$ (and hence a negative timelike flow on $W^\prime$). Setting $\lambda_1(s) \doteq \lambda_1 \lambda_W(s) {\lambda_1}^{-1} $ for $\lambda_1 \in \Ns$, it will first be shown that $B_s \doteq U(\lambda_1(s)) B U(\lambda_1(s))^{-1}$, $s \in \RR$, satisfies the hypothesis of the lemma as well. Putting $\lambda(s) \doteq \lambda_0 \lambda_1 \lambda_W(s)
{\lambda_1}^{-1} \lambda_0^{-1}$, $\lambda_0 \in N$, and picking arbitrary elements $X_1$, $X_2$ in $\As(W_1)$, $\As(W_2)$, respectively, one has $$\begin{split}
\sum_{k=1}^n {{\langle\Omega }}, X_1 A_{1,k} \, U(\lambda_0) & B_s U(\lambda_0)^{-1}
A_{2,k} X_2 {{\Omega \rangle}}\\
= \sum_{k=1}^n {{\langle\Omega }}, X_1 A_{1,k} \,
U(\lambda(s)) U(\lambda_0) & B U(\lambda_0)^{-1} U(\lambda(s))^{-1}
A_{2,k} X_2 {{\Omega \rangle}}= 0
\end{split}$$ for sufficiently small $|s|$. Now $s \mapsto U(\lambda(s))$ is, after rescaling $s$, the modular group corresponding to $(\As(\lambda_0 \lambda_1 W),\Omega)$ and, similarly, $s \mapsto U(\lambda(s))^{-1}$ is the modular group corresponding to $(\As(\lambda_0 \lambda_1 W^\prime),\Omega)$. Since $\As(W_1) \subset \As(\lambda_0 \lambda_1 W)$ and $\As(W_2) \subset \As(\lambda_0 \lambda_1 W^\prime)$, it follows that $$s \mapsto {{\langle\Omega }}, X_1 A_{1,k} \,
U(\lambda(s)) U(\lambda_0) B U(\lambda_0)^{-1} U(\lambda(s))^{-1}
A_{2,k} X_2 {{\Omega \rangle}}$$ extends to an analytic function on a strip of the upper complex half plane for each $k = 1,\ldots,n$. By the preceding result, the corresponding sum of functions thus has to vanish for all $s \in \RR$. As $X_1$, $X_2$ were arbitrary within the above limitations and $\Omega$ is cyclic for $ \As(W_1)$ and $ \As(W_2)$, respectively, one concludes that $\sum_{k=1}^n A_{1,k} U(\lambda_0) B_s U(\lambda_0)^{-1} A_{2,k} = 0$, $s \in \RR$ and $\lambda_0 \in \Ns$.
Next, let $\lambda_1,\ldots,\lambda_m \in \Ns$. Setting $\lambda_i(s) \doteq \lambda_i \lambda_W(s) {\lambda_i}^{-1}$, $i = 1,\ldots,m$, one deduces by induction on $m$ that also $$B_{s_1,\ldots,s_m} \doteq U(\lambda_m(s_m)) \cdots U(\lambda_1(s_1))
\, B \, U(\lambda_1(s_1))^{-1} \cdots U(\lambda_m(s_m))^{-1}$$ satisfies the hypothesis of the lemma for $s_1,\ldots,s_m \in \RR$. Indeed, the case $m = 1$ has just been proven. By the induction hypothesis and the group property of $U$, the assertion follows for $s_1,\ldots,s_{m-1} \in \RR$ and small $\vert s_m \vert$. The argument presented in the preceding paragraph then entails that the assertion holds for all $s_m \in \RR$.
But, according to Lemma \[generate\], the closure of the group generated by $\lambda_0 \lambda_W(s) {\lambda_0}^{-1}$, $\lambda_0 \in \Ns$, $s \in \RR$, is . Hence, taking into account that $U$ is a continuous representation, it follows by (weak operator) continuity of $\lambda \mapsto U(\lambda) B U(\lambda)^{-1}$ that these operators satisfy the hypothesis of the lemma as well, thereby completing its proof.
In the following, we shall say that the wedge $W_1$ is properly contained in the wedge $W_2$ and shall write $W_1 \Subset W_2$ if there exists a neighborhood $\Ns$ of the origin in ${\mbox{SO$_0(2,n\!-\!1)$}}$ such that $\lambda W_1 \subset W_2$, for all $\lambda \in \Ns$. Again, note that in de Sitter space and Minkowski space of dimension $n \geq 3$ such pairs of wedges do not exist, but in AdS they are abundant (see Appendix \[sectioninclusion\]).
Let $W_1$, $W_2$ be properly spacelike separated, let $W$ be any wedge such that $W_1$, $W_2$ are properly contained in $W$ and $W^\prime$, respectively. If $A_{1,k}$, $A_{2,k}$, $k = 1,\ldots,n$, are elements of $\As(W_1)$ and $\As(W_2)$, respectively, such that $\sum_{k=1}^n A_{1,k} A_{2,k} = 0$, one has $$\sum_{k=1}^n A_{1,k} U(\lambda_1) B_1 U(\lambda_1)^{-1} \cdots
U(\lambda_m) B_m U(\lambda_m)^{-1} A_{2,k} = 0 \, ,$$ for $m \in \IN$ and $B_i \in \As(W)'$, $\lambda_i \in {\mbox{SO$_0(2,n\!-\!1)$}}$, $i=1,\ldots,m$.
The proof of the lemma proceeds by induction on $m$. Let $\Ns$ be a neighbourhood of the identity in such that ${\lambda_0}^{-1} W_1 \subset W$ for $\lambda_0 \in \Ns$. Then $$U(\lambda_0)^{-1} A_{1,k} U(\lambda_0) \in \As({\lambda_0}^{-1} W_1)
\subset \As(W) = (J_W \As(W) J_W)^\prime \, .$$ So $A_{1,k}$ and $U(\lambda_0) B_0 U(\lambda_0)^{-1}$ commute for any $B_0 \in J_W \As(W) J_W$ and consequently $$0 = U(\lambda_0) B_0 U(\lambda_0)^{-1} \sum_{k=1}^n A_{1,k} A_{2,k} =
\sum_{k=1}^n A_{1,k} U(\lambda_0) B_0 U(\lambda_0)^{-1} A_{2,k} \, .$$ By the preceding lemma, this equality extends to all $\lambda_0 \in {\mbox{SO$_0(2,n\!-\!1)$}}$. Assuming now that the statement holds for $m$, one has with the same choices of $B_0$ and $\lambda_0$ as in the preceding step $$\begin{split}
0 & = \sum_{k=1}^n U(\lambda_0) B_0 U(\lambda_0)^{-1}
A_{1,k} U(\lambda_1) B_1 U(\lambda_1)^{-1} \cdots
U(\lambda_m) B_m U(\lambda_m)^{-1} A_{2,k} \\
& = \sum_{k=1}^n A_{1,k} U(\lambda_0) B_0 U(\lambda_0)^{-1}
U(\lambda_1) B_1 U(\lambda_1)^{-1} \cdots
U(\lambda_m) B_m U(\lambda_m)^{-1} A_{2,k} \, ,
\end{split}$$ for $B_i \in J_W \As(W) J_W$, $\lambda_i \in {\mbox{SO$_0(2,n\!-\!1)$}}$, $i=1,\ldots,m$. Taking into account that $U$ is a representation of , this implies (after an obvious redefinition of $\lambda_1,\ldots,\lambda_m$) $$\sum_{k=1}^n A_{1,k} U(\lambda_0) \big( B_0 U(\lambda_1) B_1 U(\lambda_1)^{-1}
\cdots U(\lambda_m) B_m U(\lambda_m)^{-1} \big) U(\lambda_0)^{-1} A_{2,k}
= 0 \, .$$ Applying once more the preceding lemma, one concludes that $$\sum_{k=1}^n A_{1,k} U(\lambda_0) B_0 U(\lambda_0)^{-1}
U(\lambda_1) B_1 U(\lambda_1)^{-1} \cdots
U(\lambda_m) B_m U(\lambda_m)^{-1} A_{2,k} = 0 \, ,$$ for $B_i \in J \As(W) J$, $\lambda_i \in {\mbox{SO$_0(2,n\!-\!1)$}}$, $i=0,1,\ldots,m$, completing the proof.
In the next step of our argument we make use of the relation $J_W U(\lambda) J_W = U(\theta \lambda \theta)$, where $\theta$ is the reflection about the edge of $W$ (cf. Lemma \[TCPgroup\]). Because of weak additivity, Proposition \[extendedlocality\] and the preceding relation we have $$\begin{split}
\bigvee_{\lambda} U(\lambda) J_W \As(W) J_W U(\lambda)^{-1}
& = J_W \Big( \bigvee_{\lambda} U(\lambda) \As(W)
U(\lambda)^{-1} \Big) J_W \\ & = J_W \Bs(\Hs) J_W = \Bs(\Hs) \, ,
\end{split}$$ showing that $\Bs(\Hs)$ is the weak operator closure of the algebra generated by the operators $U(\lambda) J_W B J_W U(\lambda)^{-1}$, $B \in \As(W)$, $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$. It therefore follows that for any collection of operators $A_{1,k}$, $A_{2,k}$ as in the preceding lemma one has $$\sum_{k=1}^n A_{1,k} B A_{2,k} = 0 \, , \quad B \in \Bs(\Hs) \, .$$ Taking into account that $\Bs(\Hs)$ contains in particular all operators of rank $1$, we conclude that for any normal state $\tilde{\omega}$ on $\Bs(\Hs)$ we have $$\label{Schlieder}
\sum_{k=1}^n \tilde{\omega}(A_{1,k})A_{2,k} = 0 =
\sum_{k=1}^n A_{1,k} \tilde{\omega}(A_{2,k}) \, .$$ So we have established that properly spacelike separated wedge algebras manifest a strong form of algebraic independence which implies the Schlieder property.
Let Standing Assumptions (i)–(iv) hold and let $W_1$ and $W_2$ be properly spacelike separated wedges. For any $A_{1,k} \in \As(W_1)$ and $A_{2,k} \in \As(W_2)$, $k = 1,\ldots,n$, such that $\sum_{k=1}^n A_{1,k} A_{2,k} = 0$, relation (\[Schlieder\]) must hold for all normal states $\tilde{\omega}$ on $\Bs(\Hs)$. In particular, if $A_1 A_2 = 0$, then either $A_{1} = 0$ or $A_{2} = 0$.
If the algebras $\As(W_1)$ and $\As(W_2)$ were mutually commuting, then the Schlieder property is equivalent to $C^*$-independence [@Roos]. However, in the noncommuting case, the Schlieder condition is strictly weaker than $C^*$-independence [@Ham0], and it is an open question in our setting whether $C^*$-independence holds if $\As(W_1)$ and $\As(W_2)$ do not commute.
The split property {#splitsec}
==================
We shall next show that if $W_1$ is properly contained in $W_2$, then there exists a type I factor $\Ms$ such that $\As(W_1) \subset
\Ms \subset \As(W_2)$, as long as the multiplicity of the eigenvalues of $M_{0n}$ does not grow too fast. Hence, with this additional assumption, the algebras $\As(W_1)$ and $\As(W_2)'$ manifest a particularly strong form of statistical independence.
Since ${\mbox{{AdS}$^n\,$}}$ is periodic in the time variable and $M_{0n}$ is a positive operator, the spectrum of $M_{0n}$ is a subset of $\IN_0 = \{ 0 \} \cup \IN$. If the multiplicities of the eigenvalues of $M_{0n}$ do not increase too rapidly, then $e^{-\gamma M_{0n}}$ is a trace class operator for any $\gamma > 0$. In Appendix \[multiplicity\] we exhibit simple examples, constructed from irreducible unitary positive energy representations of ${\mbox{SO$_0(2,n\!-\!1)$}}$, where this situation obtains. We formulate this assumption explicitly as condition (NC).
(NC) There exist constants $c_0 > 0$ and $0 < k_0 < 1$ such that the spectral multiplicities $\boldsymbol{\mu}_m$ of the eigenvalues $m$ of $M_{0n}$ are bounded by $\boldsymbol{\mu}_m \leq e^{c_0 \, m^{k_0}}$, $m \in \IN_0$.
It is particularly straightforward to establish the “split property” in the presence of condition (NC). In fact, in [@DLR Thm. 3.2] it was shown that in conformally invariant theories a trace-class condition on the exponentiated conformal Hamiltonian entails that suitable inclusions are split. We indicate here a somewhat different and more explicit argument. To this end, we recall the following lemma [@BDF Lemma 2.3].
Let $U(t) = e^{itH}$, $t \in \RR$, determine a strongly continuous one-parameter group of unitary operators with positive generator $H$ and invariant unit vector $\Omega \in \Hs$. Moreover, let $\As$ and $\Bs$ be von Neumann algebras satisfying $$\label{comm}
U(t) \As U(t)^{-1} \subset \Bs \, ,$$ for all $\vert t \vert < \delta$ and some $\delta > 0$. Then there exists a continuous function $f : \RR \rightarrow \RR$ which decreases almost exponentially, [*i.e.*]{} $\sup_\omega \vert f(\omega) \vert e^{\vert \omega \vert^k} < \infty$, for any $0 < k < 1$, such that $${{\langle\Omega }},AB'{{\Omega \rangle}}= {{\langle\Omega }},Af(H)B'{{\Omega \rangle}}+ {{\langle\Omega }},B'f(H){{\Omega \rangle}}\, ,$$ for all $A \in \As$ and $B' \in \Bs '$. \[bdf\]
We use this to establish the following general result.
\[splitabstract\] Let $U(t) = e^{itH}$, $t \in \RR$, be a strongly continuous one-parameter group of unitary operators with an (up to a phase) unique invariant unit vector $\Omega \in \Hs$ and with a generator $H$ having spectrum in $\IN_0$ and spectral multiplicities bounded by $\boldsymbol{\mu}_m \leq e^{c_0 \, m^{k_0}}$, $m \in \IN_0$, for fixed $c_0 > 0$ and $0 < k_0 < 1$. Moreover, let $\As$ and $\Bs$ be type III factors with $\Omega$ cyclic and separating for both and satisfying $$\label{comm1}
U(t) \As U(t)^{-1} \subset \Bs \, ,$$ for all $\vert t \vert < \delta$ and some $\delta > 0$. Then there exists a type $I$ factor $\Ms$ such that $\As \subset \Ms \subset \Bs$.
Consider the algebraic tensor product $\As \odot \Bs'$ and two of its representations
\(a) $(\pi,\Hs,\Omega)$ with $\pi(A \odot B') \doteq A \, B'$
\(b) $(\pi_p,\Hs \otimes \Hs,\Omega\otimes\Omega)$ with $\pi_p(A \odot B') \doteq A \otimes B'$
in an obvious notation. By the preceding lemma, one has $${{\langle\Omega }},\pi(A \odot B'){{\Omega \rangle}}= {{\langle\Omega }},AB' {{\Omega \rangle}}=
{{\langle\Omega }},Af(H)B'{{\Omega \rangle}}+ {{\langle\Omega }},B'f(H){{\Omega \rangle}}\, . \label{zip}$$ Let $P_\Omega$ be the projection onto $\CC\Omega \subset \Hs$ and let $V$ be the unitary flip on $\Hs\otimes\Hs$: $$V \Phi \otimes \Psi \doteq \Psi \otimes \Phi \, , \quad
\Phi,\Psi \in \Hs \, .$$ Taking into account that $f(H)$ is a trace class operator on $\Hs$, because of the assumptions on the spectrum of $H$ and the rapid decay of $f$, one may conclude from (\[zip\]) that $${{\langle\Omega }},\pi(A \odot B'){{\Omega \rangle}}= tr_{\Hs\otimes\Hs} \left(
\left[ V(f(H) \otimes P_\Omega) + (P_\Omega \otimes f(H))V \right]
A \otimes B' \right) \, ,$$ where the operator in square brackets is of trace class on $\Hs\otimes\Hs$. Thus, $${{\langle\Omega }},\pi(A \odot B'){{\Omega \rangle}}= \omega_p(\pi_p (A \odot B')) \, ,$$ where $\omega_p$ is some normal functional with respect to the representation $\pi_p$. Since the left hand side defines a state on $\As\odot\Bs'$, so does the right hand side; hence, $\omega_p$ is, in fact, a state on $\pi_p(\As\odot\Bs')^- = \As \overline{\otimes} \Bs'$. Since $\As,\Bs'$ are type III factors, so is their tensor product. Moreover, $\Omega \otimes \Omega$ is a cyclic and separating vector for $\As \overline{\otimes} \Bs'$, since $\Omega$ is cyclic and separating for $\As$ and $\Bs'$. Hence $\omega_p$ is represented by a vector $\Omega_p \in \Hs \otimes \Hs$, and one has $${{\langle\Omega }},\pi(A \odot B'){{\Omega \rangle}}= {{\langle\Omega_p }},\pi_p (A \odot B'){{\Omega_p \rangle}}\, ,
\quad A \in \As, B' \in \Bs' \, .$$ Since $\Omega$ is cyclic for $\pi(\As\otimes\Bs')$, one concludes that $\pi$ is unitarily equivalent to some subrepresentation of $\pi_p$. But, due to the fact that $\pi_p(\As\otimes\Bs')$ is a factor of type III, any subrepresentation of $\pi_p$ is equivalent to $\pi_p$; hence, $\pi$ and $\pi_p$ are unitarily equivalent.
So let $W : \Hs\otimes\Hs \rightarrow \Hs$ be a unitary such that $$AB' = \pi(A \odot B') = W\pi_p(A \odot B')W^{-1} = W A \otimes B'\, W^{-1} \,
.$$ Since $\As \otimes 1 \subset \Bs(\Hs) \otimes 1 \subset (1 \otimes \Bs')'$, one concludes, after applying the adjoint action of $W$ to this chain of inclusions, $$\As \subset W (\Bs(\Hs) \otimes 1) W^{-1} \subset \Bs'' = \Bs \, ,$$ where $\Ms \doteq W (\Bs(\Hs) \otimes 1) W^{-1}$ is a type I factor.
We can now prove the following theorem for our immediate purposes.
\[split\] Let the assumptions (i)–(iv) and (NC) hold. Then for any wedges $W_1$ and $W_2$ such that $$\label{propercontain}
e^{itM_{0n}} \As(W_1) e^{-itM_{0n}} \subset \As(W_2) \, ,$$ for all sufficiently small $t \in \RR$, there exists a type $\rm{I}_{\infty}$ factor $\Ms$ such that $\As(W_1) \subset \Ms \subset \As(W_2)$.
[**Remark**]{}: Of course, if $W_1 \Subset W_2$, then (\[propercontain\]) holds.
Our passivity and mixing assumptions entail that $\As(W)$ is a type $\rm{III}_1$ factor, for all $W \in \Ws$ [@PuWo Theorem 4.3]. The theorem then follows at once from Proposition \[splitabstract\].
Note that in de Sitter space and Minkowski space of dimension $n \geq 3$ no inclusions of wedge algebras can be split. For Minkowski space this was observed in [@Bu]; in de Sitter space $W_1 \subset W_2$ entails $W_1 = W_2$.
The above results hold only for theories on proper AdS. If the covering space of AdS is considered, then condition (NC) must be replaced by the condition that the map $$\As(W) \ni A \mapsto e^{-\gamma M_{0n}}A\Omega \, ,$$ is nuclear, for all $\gamma > 0$ [@BuWi]. The proof presented in [@BuWi] is formulated in terms of double cone algebras in Minkowski space but carries over to the present situation without difficulty.
Local nets on two-dimensional AdS {#2d}
=================================
In this section we prove that Standing Assumptions (i)–(iv) and the spectral condition (NC) entail the existence of a nontrivial, covariant and local subnet in two-dimensional AdS. It is noteworthy that locality properties can be derived in these circumstances, and it would be of interest to see whether the same is also true in higher-dimensional AdS.
In two-dimensional AdS, the set of all wedge-shaped regions consists of two disconnected pieces $\{ \lambda W_R \mid \lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}\}$ and $\{ \lambda W_L \mid \lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}\}$, where $W_L = W_R{}'$. The former (respectively, latter) will be denoted by $\Ws_R$ ($\Ws_L$) and called the class of right wedges (left wedges). Note that $W \in \Ws_R$ if and only if $W' \in \Ws_L$. We assume as before that $$\label{leftweakadditivity}
\As = \bigvee_{W \in \Ws_R} \As(W) \, ,$$ and that $\Omega$ is cyclic for $\As$. Hence, once again, Proposition \[ReehSchlieder\] yields the cyclicity of $\Omega$ for every $\As(W)$, $W \in \Ws_R$.
In ${\mbox{{AdS}$^2\,$}}$ the edge of a wedge is a single point. For each point $a \in {\mbox{{AdS}$^2\,$}}$ we shall denote by $W_a$ the unique element of $\Ws_R$ whose edge is $a$. Then $W_a{}'$ is the unique element of $\Ws_L$ whose edge is $a$. For $a,b \in {\mbox{{AdS}$^2\,$}}$ such that $W_b \Subset W_a$ we define the open sets (double cones) $\Os_{a,b} = W_a \cap W_b{}'$. Corresponding to $\Os_{a,b}$ we define the von Neumann algebra $$\Bs(\Os_{a,b}) = \As(W_a) \cap \As(W_b)' \, .$$ Note that $\Os_{a,b} \subset \Os_{c,d}$ if and only if $W_a \subset W_c$ and $W_d \subset W_b$. The isotony of the original net of wedge algebras then implies $$\Bs(\Os_{a,b}) \subset \Bs(\Os_{c,d}) \, ,$$ [[*i.e.* ]{}]{}the isotony of the net $\{ \Bs(\Os_{a,b}) \mid a,b \in {\mbox{{AdS}$^2\,$}}\}$. The covariance of the original net of wedge algebras entails $$\begin{split}
U(\lambda)\Bs(\Os_{a,b})U(\lambda)^{-1} & =
U(\lambda)(\As(W_a) \cap \As(W_b)')U(\lambda)^{-1} \\ & =
\As(\lambda W_a) \cap \As(\lambda W_b)'
= \Bs(\lambda \Os_{a,b}) \,
\end{split}$$ for all $\lambda \in {\mbox{SO$(2,n\!-\!1)$}}$.
In general, there is no reason for such relative commutants of wedge algebras to be nontrivial. But in our setting they turn out to be very large algebras, having a dense $G_\delta$ set of cyclic vectors.
If conditions (i)–(iv) and (NC) hold in a theory on two-dimensional AdS and $W_b \Subset W_a$, then $\Bs(\Os_{a,b})$ is a type $\rm{III}_1$ factor.
If $W_b$ is properly contained in $W_a$, then from Theorem \[split\] one has the existence of a type I factor $\Ms$ such that $\As(W_b) \subset \Ms \subset \As(W_a)$. From the proof of Theorem 2.1 in [@DopLon] there exists a unitary mapping $\Hs \rightarrow \Hs \otimes \Hs$ such that $U \As(W_b) U^{-1} = \As(W_b) \otimes 1$ and $U \As(W_a) U^{-1} = \Bs(\Hs) \otimes \As(W_a)$. Using Takesaki’s commutant theorem for tensor products (cf. [@KadRing Theorem 11.2.16]), one therefore sees that $$\begin{split}
\Bs(\Os_{a,b}) & = \As(W_a) \cap \As(W_b)' =
U^{-1} \big( (\Bs(\Hs) \otimes \As(W_a)) \cap (\As(W_b)'
\otimes \Bs(\Hs))\big)U
\\
& = U^{-1} \As(W_b)' \otimes \As(W_a) U \, .
\end{split}$$ Since $\As(W_a)$ and $\As(W_b)$ are type $\rm{III}_1$ factors (cf. proof of Theorem \[split\]), the algebra $\As(W_b)' \otimes \As(W_a)$ and, thus, also $\Bs(\Os_{a,b})$ is a type $\rm{III}_1$ factor (cf. [@KadRing Corollary 11.2.17] and [@Con Théorèmes 1.3.4 and 3.4.1]).
We have seen previously that assumptions (i)–(iv) prescribe the sense in which “spacelike separated” is to be understood in . Hence, the region $\Os_{a,b}$ is spacelike separated from $\Os_{c,d}$ if there exists a wedge $W$ such that $\Os_{a,b} \subset W$ and $\Os_{c,d} \subset W'$. Without loss of generality, we may assume for concreteness that $W \in \Ws_R$, so that $W_a \subset W \subset W_d$. Then one finds that the local algebras satisfy locality: $$\begin{split}
\Bs(\Os_{a,b}) & = \As(W_a) \cap \As(W_b)' \subset \As(W)
\subset \As(W_c)' \vee \As(W_d) = \Bs(\Os_{c,d})' \, .
\end{split}$$ We summarize these findings in the following theorem.
\[localnet\] If conditions (i)–(iv), (NC) and (\[leftweakadditivity\]) hold in a theory on two-dimensional AdS, then the above construction yields a nontrivial, covariant and local net $\Os \mapsto \Bs(\Os )$ based on double cones $\Os \subset {\mbox{{AdS}$^2\,$}}$, in which each algebra $\Bs(\Os)$ is a factor of type $\rm{III}_1$.
One can similarly define a second nontrivial, covariant and local net from a given net of wedge algebras based upon the wedges in $\Ws_L$ . These two local nets coincide if the initial nets of wedge algebras are local with respect to each other, $\As(W^\prime) \subset \As(W)^\prime$.
We wish to sketch some consequences of this construction. The weakly local, but quite nonlocal net fixed by the field $\phi$ given in equation (\[counterexample\]) is also well-defined in two-dimensional AdS, as long as the two-point function $W(x,y)$ is suitably chosen. The above construction yields many nontrivial observables localized in precompact subsets of ${\mbox{{AdS}$^2\,$}}$ and associated with this field. In terms of the original, simply expressed field, these local observables are quite complicated objects. This suggests the possibility of constructing complex local objects from relatively simple nonlocal fields.
In [@BL] the existence of split inclusions of wedge algebras was replaced by a modular nuclearity condition to employ the above construction in theories on two-dimensional Minkowski space. This modular nuclearity condition has been verified [@L] in a factorizing S-matrix model with S-matrix not equal to the identity. The basic field in that model has algebraic relations similar to those of our field $\phi$. Since in some suitable sense quantum theories on AdS become close to theories on Minkowski space as the AdS radius becomes sufficiently large, one may expect that the model on ${\mbox{{AdS}$^2\,$}}$ determined by $\phi$ also describes physics which goes beyond that of generalized free fields. These matters shall be further investigated elsewhere.
Conclusions and further remarks
===============================
We have thus shown that stability properties of a state carrying the interpretation of a “vacuum” imply a PCT theorem, the uniqueness of the Unruh temperature, as well as commensurability and independence properties of the observables in any quantum field theory on AdS. Such implications exist also in other space–times, but they are of particular interest in the case of AdS, where the causal structure is such that it is not clear [*a priori*]{} how to define “mutually spacelike regions,” and therefore it is not clear which locality relations are physically meaningful. Nor is it sufficient to sidestep the issue by appealing to theories on the covering space of . In fact, we have seen that observables in opposite wedges $W$ and $W^{\prime}$ or in conjugate wedges $W$ and $\widetilde{W}'$ necessarily commute with each other, either weakly or strongly. The former fact can seem natural from the point of view of the covering space of , but the latter can certainly not be observed in theories on the covering space. It is remarkable that locality properties can be determined by stability assumptions. Indeed, we showed that in two-dimensional AdS these stability assumptions and a weak growth condition on the eigenvalues of $M_{02}$ entail the existence of a nontrivial, covariant, local net indexed by bounded regions in AdS. It would be of interest to determine if the same is true of theories on higher dimensional AdS.
It was conjectured by Bros, Epstein and Moschella in [@BEM] that the assumptions made in that paper should follow from our standing assumptions. This would be an interesting matter to settle, since it would allow one to derive more detailed information about the analyticity properties of the two-point function of quantum fields on AdS.
[**Acknowledgements**]{}: DB and SJS wish to thank the Department of Mathematics and the Institute for Fundamental Theory of the University of Florida and the Institute for Theoretical Physics at the University of Göttingen, respectively, for hospitality and financial support. DB is also grateful for financial support by the Deutsche Forschungsgemeinschaft (DFG). We have benefitted from discussions with J. Bros. We also acknowledge conversations with M. Florig at an early stage of these investigations.\
Unitary representations of the AdS group {#grouptheory}
========================================
The algebraic properties of the strongly continuous unitary representation of assumed to exist in (i) are of crucial importance to us. As a convenience to the reader, we collect the relevant properties here. Many of the arguments in this section are adapted from [@BB]. Given any such representation $U$ and a coordinate system on , we denote by $M_{\mu\nu}$, $\mu,\nu = 0,1,\ldots,n$, the corresponding self–adjoint generators. On a dense, invariant domain of analytic vectors in $\Hs$, they fulfill the following Lie-algebraic relations. $$\label{liealgebra}
[M_{\mu\nu},M_{\rho\sigma}] = -ig_{\mu\rho}M_{\nu\sigma} +
ig_{\mu\sigma}M_{\nu\rho} - ig_{\nu\sigma}M_{\mu\rho} + ig_{\nu\rho}M_{\mu\sigma} \, ,$$ where $g = \diag(1,-1,\cdots,-1,1)$, $\mu,\nu = 0,1,\ldots,n$. In particular, the operator $M_{01}$ generates the action of the boosts $\lambda_{01}$, $$U(\lambda_{01}(t)) = e^{itM_{01}} \, , \quad t \in \RR \, ,$$ and similarly for the operators $M_{0j}$, $j = 2,\ldots,n-1$. $M_{0n}$ is the generator of the time translations. Any of the operators $M_{0j}$, $j = 1,\ldots,n$, may be taken to be the operator $M$ discussed in the Introduction (see Section \[intro\]). If $n \geq 3$, the operators $M_{jk}$, $j,k = 1,\ldots,n-1$, are the generators of spatial rotations, whereas $M_{jn}$, $j = 2,\ldots,n-1$ generate other subgroups of boosts obtained from the first-mentioned by temporal rotation. If $n = 2$, then $M_{01}$ and $M_{12}$ are generators of boosts and there are no spatial rotations.
The Lie-algebraic relations (\[liealgebra\]) yield the equality $$e^{isM_{0n}}M_{01}e^{-isM_{0n}} = \cos(s) M_{01} - \sin(s) M_{1n}
\, ,$$ which implies $$e^{i\pi M_{0n}}M_{01}e^{-i\pi M_{0n}} = - M_{01} \, .$$ Hence, $M_{01}$ cannot be a positive operator on $\Hs$. Since the representation $U({\mbox{SO$_0(2,n\!-\!1)$}})$ is nontrivial, we may conclude the following lemma from the assumed ${\mbox{SO$_0(2,n\!-\!1)$}}$–covariance.
\[nogroundstate\] The operators $M_{0j}$, $j = 1,\ldots,n-1$, cannot be positive on $\Hs$ and thus $\Omega$ cannot be a ground state for the group $e^{itM_{0j}}$.
The Lie-algebraic relations (\[liealgebra\]) also imply the following group relations: $$\begin{aligned}
\label{0n01}
& e^{isM_{0n}}e^{itM_{01}}e^{-isM_{0n}} =
e^{it(\cos(s) M_{01} - \sin(s) M_{1n})} \, ; & \\
\label{equation6}
& e^{isM_{0j}}e^{itM_{01}}e^{-isM_{0j}} =
e^{it(\cosh(s) M_{01} - \sinh(s) M_{1j})} \, ,
\quad j = 2,\ldots,n-1 \, ; & \\
\label{equation7}
& e^{isM_{j1}}e^{itM_{01}}e^{-isM_{j1}} =
e^{it(\cos(s) M_{01} - \sin(s) M_{j0})}
\, , \quad j = 2,\ldots,n-1 \, ; & \\
\label{010n}
& e^{isM_{01}}e^{itM_{0n}}e^{-isM_{01}} =
e^{it(\cosh(s) M_{0n} - \sinh(s) M_{1n})} \, . &\end{aligned}$$ Of course, equations (\[equation6\]) and (\[equation7\]) are vacuous if $n = 2$. We shall establish a few necessary tools.
\[generate\] Let $\Ns \subset {\mbox{SO$_0(2,n\!-\!1)$}}$ be any neighborhood of the identity in . Then the closure $\Us_{\Ns}$ in the strong operator topology of the group generated by the unitaries $U(\lambda \lambda_{01}(t)\lambda^{-1})$, $t \in \RR$, $\lambda \in \Ns$, contains $U({\mbox{SO$_0(2,n\!-\!1)$}})$.
Consider first $n \geq 3$. From (\[equation6\]) it follows that for all sufficiently small $\vert s \vert$ and for all $t \in \RR$, $\Us_{\Ns}$ contains the operator $e^{it(\cosh(s) M_{01} - \sinh(s) M_{1j})}$ for $j = 2,\ldots,n-1$. With fixed $s$, the Trotter product formula [@RS] applied to the product of the one–parameter groups $e^{it(\cosh(s) M_{01} - \sinh(s) M_{1j})}$ and $e^{it\cosh(s) M_{01}} \in \Us_{\Ns}$ implies that the rotations $e^{-it\sinh(s) M_{1j}}$, $j = 2,\ldots,n-1$, are also contained in $\Us_{\Ns}$. Similarly, equations (\[0n01\]) and (\[equation7\]) entail that the operators $e^{itM_{1n}}$ and $e^{itM_{j0}}$, $j = 2,\ldots,n-1$, $t \in \RR$, are contained in $\Us_{\Ns}$. Since the group $U({\mbox{SO$_0(2,n\!-\!1)$}})$ is generated by these subgroups, the proof is complete for $n \geq 3$.
If $n = 2$, then the same argument applied to (\[0n01\]) entails that the group $e^{itM_{12}}$, $t \in \RR$, is contained in $\Us_{\Ns}$. One may then apply the same reasoning to the equality $$e^{isM_{12}}e^{itM_{01}}e^{-isM_{12}} = e^{it(\cosh(s) M_{01} - \sinh(s)
M_{02})}$$ to conclude that also the group $e^{itM_{02}}$, $t \in \RR$, is contained in $\Us_{\Ns}$. The assertion now follows for $n = 2$.
\[invariance\] Let $\Psi \in \Hs$ satisfy $U(\lambda_{01}(t))\Psi = \Psi$, for all $t \in \RR$. Then $U(\lambda)\Psi = \Psi$, for all $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$.
Consider first $n \geq 3$. Using the equation $$\label{010j}
e^{isM_{01}}e^{itM_{0j}}e^{-isM_{01}} =
e^{it(\cosh(s) M_{0j} + \sinh(s) M_{1j})}
\, , \quad j = 2,\ldots,n-1 \, ,$$ and setting $t = 2re^{-\vert s \vert}$, the continuity of the representation $U({\mbox{SO$_0(2,n\!-\!1)$}})$ implies that $$\label{equation10}
\lim_{s \rightarrow \pm\infty} e^{isM_{01}}e^{i2re^{-\vert s \vert}M_{0j}}e^{-isM_{01}} = e^{ir(M_{0j} \pm M_{1j})} \, ,$$ for $j = 2,\ldots,n-1$, where the limit is the strong operator limit on $\Hs$. But then, by hypothesis, one has $$\lim_{s \rightarrow \pm\infty} \Vert e^{isM_{01}}e^{i2re^{-\vert s \vert}M_{0j}}e^{-isM_{01}}\Psi - \Psi\Vert =
\lim_{s \rightarrow \pm\infty} \Vert e^{i2re^{-\vert s \vert}M_{0j}} \Psi - \Psi \Vert = 0 \, ,$$ since $e^{i2re^{-\vert s \vert}M_{0j}}$ converges strongly to the identity $1$ on $\Hs$ as $s \rightarrow \pm\infty$. These two equations then imply $$e^{ir(M_{0j} \pm M_{1j})}\Psi = \Psi \, ,$$ for all $r \in \RR$ and $j = 2,\ldots,n-1$. Similarly, the equation $$\label{011n}
e^{isM_{01}}e^{itM_{1n}}e^{-isM_{01}} = e^{it(\cosh(s) M_{1n} + \sinh(s) M_{0n})}$$ yields $$e^{ir(M_{1n} \pm M_{0n})}\Psi = \Psi \, ,$$ for all $r \in \RR$. By using the Trotter product formula again and taking suitable limits, it is clear that $M_{0j}\Psi = 0 = M_{1j}\Psi$, for all $j = 0,\ldots,n$. Equation (\[liealgebra\]) then implies that $\Psi$ is annihilated by all of the generators $M_{\mu\nu}$, which yields the assertion for $n \geq 3$.
If $n = 2$, equation (\[011n\]) yields $$e^{ir(M_{12} \pm M_{02})}\Psi = \Psi \, ,$$ for all $r \in \RR$, and thus $M_{12}\Psi = 0 = M_{02}\Psi$. The assertion then follows for $n = 2$.
We mention that Lemma \[invariance\] is also stated (without explicit proof) as Lemma 5.4 in [@BEM].
Wedge inclusions in AdS {#sectioninclusion}
=======================
Here we give the proof of some useful geometric properties of the subregions of which we have identified as the correct choice of wedges in . Indeed, we wish to show that for any wedge $W \in \Ws$ there exist wedges $W_0 \in \Ws$ which are properly contained in $W$, $W_0 \Subset W$, respectively, such that ${W_0}^\prime$ is properly spacelike separated from $W$, $W \Subset W_0$. In fact, such wedges are quite abundant. These results are to be contrasted with the situation in de Sitter space, where de Sitter wedges satisfy $W_1 \subset W_2$ if and only if $W_1 = W_2$ [@BDFS]. In Minkowski space of dimension $n \geq 3$, two wedges form an inclusion $W_1 \subset W_2$ only if $W_1$ is a suitable translation of $W_2$. Hence, also in the latter case there do not exist properly spacelike separated wedges.
It is convenient to use the following description of $W_R$: $$W_R = \{ x \in {\mbox{{AdS}$^n\,$}}\mid e_{\pm} \cdot x < 0 \, , \, x \cdot e_4 > 0 \} \, ,$$ where $e_\pm = (\pm 1,1,0, \dots ,0)$ and $e_4 = (0, \dots ,0,1)$. Consider the lightlike vectors $f_\pm = (\pm 1, c, 0, \dots ,0, s)$, where $s > 0$ and $c^2 = 1 + s^2$, and the wedge they determine: $$W_0 \doteq \{ x \in {\mbox{{AdS}$^n\,$}}\mid x \cdot f_\pm < 0\, , x \cdot e_4 > 0 \} \, .$$ The edge of this wedge is the spacelike submanifold $$\Es_1 \doteq \left\{
(0,s(1+\sigmavec^2)^{1/2}, \sigmavec ,c(1+\sigmavec^2)^{1/2})
\mid \sigmavec \in \RR^{n-2} \right\} \, ,$$ which is contained in $W_R$.
\[neighborhood\] With the above definitions, for any $t \in \RR$ there exists a neighborhood $\Ns$ of the identity in ${\mbox{SO$_0(2,n\!-\!1)$}}$ such that $\lambda \lambda_{01}(t)W_0 \subset W_R$ for all $\lambda \in \Ns$. Hence, for any wedge $W_2 \in \Ws$ there exist wedges $W_1,W_3 \in \Ws$ such that $W_1 \Subset W_2$ and $W_2$ and $W_3$ are properly spacelike separated.
In order to show that $\lambda \lambda_{01}(t)\overline{W_0} \subset W_R$ for $\lambda$ in a neighborhood of the identity, it suffices to show that the characteristic boundary of $\lambda \lambda_{01}(t)\overline{W_0}$ is contained in $W_R$: $\lambda \lambda_{01}(t)(\Es_1 + \RR_+ f_\pm) \subset W_R$, [*i.e.*]{} $\lambda\lambda_{01}(t)(x + l f_\pm) \in W_R$, for all $l \geq 0$, $x \in \Es_1$. Since $\lambda$ is to be in a neighborhood of the identity $1$, consider $\lambda = 1 + M$, where $\Vert M \Vert < \varepsilon$ and $\Vert\cdot\Vert$ is the norm on the $(n+1) \times (n+1)$ matrices with real entries. Then one has $$\begin{split}
& \lambda\lambda_{01}(t)(x + l f_+) \cdot e_\pm =
(\pm \sinh(t) - \cosh(t)) s(1 +\sigmavec^2)^{1/2} \\
& + l(\pm \cosh(t) \pm c \sinh(t) - \sinh(t) -c \cosh(t)) +
M \lambda_{01}(t) x \cdot e_\pm +
lM \lambda_{01}(t) f_+ \cdot e_\pm \label{hier} \\
& \leq (\pm \sinh(t) - \cosh(t))
s(1+\sigmavec^2)^{1/2} +
\Vert M \Vert \, \Vert e_\pm \Vert \{ \Vert \lambda_{01}(t) x \Vert + l
\Vert \lambda_{01}(t) f_+ \Vert\} \, ,
\end{split}$$ where use was made of the fact that $$\pm \cosh(t) \pm c \sinh(t) - \sinh(t) -c \cosh(t) =
(\mp 1 + c)(\pm \sinh(t) - \cosh(t)) < 0 \, .$$ But $\Vert \lambda_{01}(t) f_\pm \Vert \leq 2\cosh(t)$, $\Vert \lambda_{01}(t) x \Vert \leq \cosh(t)(1 + c^2 + 2s^2)^{1/2}
(1+\sigmavec^2)^{1/2}$ and $\Vert e_\pm \Vert = \sqrt{2}$. In addition, $\pm \sinh(t) - \cosh(t) < 0$, for all $t \in \RR_+$. Hence, if $\varepsilon$ is sufficiently small, there exists a $\delta < 0$ (depending on $t$) such that $$\lambda \lambda_{01}(t)(x + l f_+) \cdot e_\pm \leq \delta < 0 \, , \quad
x \in \Es_1 \, , l \geq 0 \, .$$ Similarly, one shows that $$\lambda \lambda_{01}(t)(x + l f_-) \cdot e_\pm \leq \delta < 0 \, , \quad
x \in \Es_1 , l \geq 0 \, ,$$ for suitably small $\varepsilon$.
Since $(\lambda W)^\prime = \lambda W^\prime$, for all $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$ and $W \in \Ws$, it is clear that $W_2$ and $W_3^\prime$ are properly spacelike separated if and only if $W_2 \Subset W_3$. Thus, since $\Ws = \{ \lambda W_R \mid \lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}\}$, the remaining assertions follow at once.
It is of interest to note that in ${\mbox{{AdS}$^n\,$}}$, $n \geq 3$, there exists a wedge $W_2 \in \Ws$ such that $\overline{W_2} \subset W_R$, but in any neighborhood of the identity of ${\mbox{SO$_0(2,n\!-\!1)$}}$ there exists some $\lambda$ such that $\lambda W_2 \not\subset W_R$.
The Reeh–Schlieder property {#sectionReehSchlieder}
===========================
We prove that the theories we are considering here must satisfy the Reeh–Schlieder property for the wedge algebras. Let $W \in \Ws$ be a wedge, and let $\Bs(W)$ denote the \*–algebra consisting of all $B \in \As(W)$ for which there exists a neighborhood $\Ns(B)$ of the identity in ${\mbox{SO$_0(2,n\!-\!1)$}}$ such that $B(\lambda) \doteq U(\lambda)BU(\lambda)^{-1} \in \As(W)$, for all $\lambda \in \Ns(B)$. Note that Lemma \[neighborhood\] entails that there exists a wedge $W_0 \in \Ws$ such that $\As(W_0) \subset \Bs(W)$.
\[ReehSchlieder\] Let Assumptions (i)–(iv) obtain. Then $\Omega$ is cyclic for $\As(W)$, given any $W \in \Ws$.
[^4] Let $W,W_0 \in \Ws$ and $\Bs(W)$ be as described above and let $\Ns$ be a neighborhood of the identity in ${\mbox{SO$_0(2,n\!-\!1)$}}$ such that $\lambda^{-1} W_0 \subset W$, for all $\lambda \in \Ns$. By the covariance assumption (iii), it suffices to consider $W = W_R$. Further, let $\Psi\in\Hs$ be orthogonal to the set of vectors $\As(W)\Omega$.
Since $\Bs(W) \subset \As(W)$, $\Psi$ is also orthogonal to $\Bs(W)\Omega$. From the definition of $\Bs(W)$ and the continuity of the representation $U$, it is clear that for any $B \in \Bs(W)$ and $\lambda \in \Ns$ as described above there exists an $\varepsilon > 0$ such that $B(\lambda\lambda_{01}(t)\lambda^{-1}) \in \Bs(W)$, for all $\vert t \vert < \varepsilon$. Therefore, one has $$\langle\Psi,B(\lambda\lambda_{01}(t)\lambda^{-1})\Omega\rangle = 0 \, ,$$ for $\vert t \vert < \varepsilon$. Since $\lambda\lambda_{01}(t)\lambda^{-1}W_0 \subset \lambda W$, one also has $B(\lambda\lambda_{01}(t)\lambda^{-1}) \in \As(\lambda W)$, for all $t \in \RR$. Hence, by the KMS property of the restriction of $\omega$ to $\As(\lambda W)$, the function $$t \mapsto B(\lambda\lambda_{01}(t)\lambda^{-1})\Omega \, , \quad t
\in \RR \, ,$$ extends analytically to a vector–valued function in the strip $\{ z \in \CC \mid 0 < \Im(z) < \beta/2 \}$ with continuous boundary values. Therefore, one must have $$\langle U(\lambda\lambda_{01}(t)\lambda^{-1})^{-1} \Psi, B\Omega\rangle
= \langle\Psi,B(\lambda\lambda_{01}(t)\lambda^{-1})\Omega\rangle = 0 \, ,$$ for all $t \in \RR$ and $B \in \Bs(W)$. By iterating this argument, it follows that for $\lambda_1,\ldots,\lambda_k \in \Ns$ and $t_1,\ldots,t_k \in \RR$, $$\langle U(\lambda_1\lambda_{01}(t_1)\lambda_1^{-1})^{-1}\cdots
U(\lambda_k\lambda_{01}(t_k)\lambda_k^{-1})^{-1} \Psi, B\Omega\rangle
= 0 \, .$$ Lemma \[generate\] then implies that $U(\lambda)\Psi$ is orthogonal to $\Bs(W)\Omega$, for any $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$, and hence $\Psi$ is orthogonal to $U(\lambda)^{-1}\Bs(W)\Omega = \Bs(\lambda^{-1}W)\Omega$.
Moreover, since $\Bs(W)$ is a \*–algebra, $B^* U(\lambda)\Psi$ is orthogonal to $\Bs(W)\Omega$, for any $B \in \Bs(W)$ and $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$. Hence, by induction, for any $\lambda_1,\ldots,\lambda_k \in {\mbox{SO$_0(2,n\!-\!1)$}}$ and $B_1,\ldots,B_k \in \Bs(W)$ one has $$\langle\Psi,B_1(\lambda_1)\cdots B_k(\lambda_k)\Omega\rangle = 0 \, .$$ Putting these results together, it now follows that $\Psi$ is orthogonal to $\big( \bigcup_{\lambda\in\text{SO}_0(2,n-1)} U(\lambda)\Bs(W)U(\lambda)^{-1} \big)
\Omega$. And since $\As(W_0) \subset \Bs(W)$, one observes $$\big( \bigvee_{\lambda\in\text{SO}_0(2,n-1)}
\Bs(\lambda W)\big)\Omega \quad \supset
\big( \bigvee_{\lambda\in\text{SO}_0(2,n-1)}
\As(\lambda W_0)\big)\Omega = \As\Omega
\, ,$$ by the assumption of weak additivity (\[weakadditivity\]). Thus, $\Psi$ is orthogonal to $\As\Omega$. As $\Omega$ is cyclic for $\As$, the assertion is proven.
Multiplicity of energy levels and nuclearity {#multiplicity}
============================================
In this appendix we wish to prove that given any irreducible unitary positive energy representation $U_1$ of the anti–de Sitter group ${\mbox{SO$_0(2,n\!-\!1)$}}$ on a Hilbert space $\Hs_1$, then the corresponding unitary representation $U$ obtained by “second quantization” on the bosonic or fermionic Fock space $\Hs$ based upon $\Hs_1$ satisfies condition (NC). Hence the free field examples discussed in the main text and Appendix \[examples\] satisfy (NC). To verify this, we shall have need of some basic results about unitary representations of ${\mbox{SO$_0(2,n\!-\!1)$}}$. We merely recall these and refer the reader to [@F1; @F2; @Ev] for details and proofs. We shall present the case $n = 4$; the situation is similar for the other cases.
Let $SO(2)\otimes SO(3) \subset {\mbox{SO$_0(2,3)$}}$ be the maximal compact subgroup consisting of the “time rotations” around AdS$^4$ and the spatial rotations. Restricting $U_1$ to this compact group, the Hilbert space $\Hs_1$ decomposes into a direct sum of corresponding irreducible subspaces which are labelled by the energy $n \in \IN$ and angular momentum $l \in \IN_0$. It has been shown in [@Ev] that in this decomposition there appear only representations where $l < n$. Moreover, for given $(n,l)$, the number of these representations is bounded by $l+1$. Taking these facts into account, the multiplicity $\mu_n$ of the eigenvalue $n$ of the generator of the time rotations on $\Hs_1$ can be estimated by $\mu_n \leq \sum_{l=0}^{n-1} \, (2l +1)(l+1) \leq 2 n^3$, $n \in \IN$.
Next, let $M_{0 4}$ be the generator of the time rotations on the Fock space $\Hs$. The bound given above entails by standard arguments in statistical mechanics that in the bosonic case the corresponding partition function satisfies, for any $\gamma > 0$, $$\label{nucl1}
\begin{split}
\ln {\rm tr} e^{-\gamma M_{04}}
& = - \sum_n \mu_n \,\ln(1 - e^{-\gamma n})
\leq 2 \sum_n n^3 \, \frac{e^{-\gamma n}}{1 - e^{-\gamma}} \\
& = \frac{2}{1 - e^{-\gamma}} (- \, \partial_\gamma)^3
\frac{1}{1 - e^{-\gamma}} \leq \frac{12 e^{-\gamma}}{(1 - e^{-\gamma})^5}
\leq \frac{12 \cdot 5^5}{\gamma^5} \, .
\end{split}$$ Denoting by $\boldsymbol{\mu}_n$ the multiplicity of the eigenvalue $n$ of $M_{0 4}$, $n\in \IN_0$, we obtain the estimate $${\rm tr} \, e^{-\gamma M_{04}} = \sum_{n} \boldsymbol{\mu}_n
e^{- \gamma n} \leq e^{12 \cdot 5^5 / \gamma^5} \, ,$$ which implies $$\boldsymbol{\mu}_n \leq e^{\gamma n + 12 \cdot 5^5 / \gamma^5} \, ,$$ for all $n \in \IN_0$ and $\gamma > 0$. With the choice $\gamma = 5 n^{-1/6}$, we conclude that $$\boldsymbol{\mu}_n \leq e^{17 \, n^{5/6}} \, , \quad n \in \IN_0 \, .$$ A similar argument applies also to the fermionic case and in any number of spacetime dimensions. Hence we have the following result.
In any free boson or fermion model based upon an irreducible positive energy representation of the anti–de Sitter group, condition (NC) holds.
Examples
========
In this appendix we shall discuss some examples of nets and states which fulfill Standing Assumptions (i)–(iv), as well as the condition (NC). Because AdS is not globally hyperbolic, the standard means of obtaining examples do not suffice. Free field models on AdS have been discussed in a series of papers by Fronsdal [@F1; @F2; @F3; @F4] and by Avis, Isham and Storey [@AIS]. More recently, the Wightman functions of quantum field models on AdS satisfying certain general conditions have been treated rigorously in [@BEM]. In addition, models of quantum field theories on AdS can also be obtained [*via*]{} holography [@Rehren] (see also [@BBMS]). We begin our discussion with the latter.
In an elegant paper [@Rehren] Rehren has given rigorous mathematical meaning to the notion of holography, namely the correspondence between theories on ${\mbox{{AdS}$^n\,$}}$ and conformally invariant theories on the boundary $CM_{n-1}$ of , compactified $(n-1)$–dimensional Minkowski space. He shows that between the set of wedges $W$ in AdS and the set of (conformal images of) double cones $C$ in the boundary there exists a canonical bijection $\alpha$ which preserves inclusions and causal complements, and intertwines the actions of the anti-de Sitter group and of the conformal group (which are isomorphic groups): $$\alpha(\lambda(W)) = \tilde{\lambda}(\alpha(W)) \, , \quad
\alpha^{-1}(\tilde{\lambda}(C)) = \lambda(\alpha^{-1}(C)) \, ,$$ where $\tilde{\lambda}$ is the restriction of the action of $\lambda \in {\mbox{SO$_0(2,n\!-\!1)$}}$ to the boundary. The double cone $C = \alpha(W)$ is defined to be the intersection of $W$ with the boundary. Hence, given a net $\{ \widetilde{\As}(\alpha(W)) \}$ associated with a (for example) free quantum field on *CM*$_{n-1}$ and a vacuum state $\tilde{\omega}$ on this net, one can define a net $\{ \As(W)\}$ and state $\omega$ on ${\mbox{{AdS}$^n\,$}}$ by $$\As(W) \doteq \widetilde{\As}(\alpha(W)) \, , \quad
\omega(A) \doteq \tilde{\omega}(A) \, ,$$ for every $A \in \As(W) = \widetilde{\As}(\alpha(W))$. From the results in [@Rehren] it is easy to show that the resulting theory on fulfills our assumptions (i)–(iv), provided the underlying net on $CM_{n-1}$ complies with the standard assumptions of conformal Minkowski space theories. In such theories one also has the equality $\As(W) = \As(-W)$, for every $W \in \Ws$. The CGMA and the , formulated in [@BDFS], both obtain in these models.
Using an irreducible representation of ${\mbox{SO$_0(2,3)$}}$, Fronsdal [@F2] defines Hermitian free fields on the covering space of AdS$^4$; only if the energy spectrum of the theory is contained in $\IN_0$ does his field restrict to AdS proper. In this latter case, Fronsdal’s model satisfies assumptions (i)-(iv), as well as (NC) – see below and Appendix \[multiplicity\]. Moreover, the elements of $\As(W)$ commute weakly with those of $\As(\widetilde{W}')$. Yet if the energy spectrum is not a subset of $\IN_0$, this feature is no longer present, confirming our expectation that Theorem \[weaklocality2\] can not hold in general for fields not manifesting the periodicity in time required to enable them to be defined on AdS. We note that since in his examples the fields are invariant under the map $x \mapsto -x$, one has $\As(\widetilde{W}') = \As(W') = \As(W)'$.
In [@AIS] Avis, Isham and Storey use an embedding of (the covering space of) AdS into the static Einstein universe to construct a free quantum field on AdS. Since the static Einstein universe is globally hyperbolic, one can rigorously construct free fields and the associated nets of local algebras in (subsets of) that spacetime. And since the covering space of AdS can be conformally embedded into the Einstein universe, one can define free massless, [[*i.e.* ]{}]{}conformally invariant, fields on AdS. However, the matter is complicated by the fact that one must find suitable boundary conditions at spacelike infinity in AdS – we must refer the reader to [@AIS] for details. Since the resultant fields manifest the necessary periodicity in the time variable, they may be understood to be defined on AdS. Inspection of the resulting representations then led them to construct analogous representations of “conformally coupled massive” free fields on AdS. From the construction of their examples and the results in [@BEM] it follows that Standing Assumptions (i)–(iv) hold. All of their examples are consistent with Theorem \[weaklocality2\].
In [@BEM] an axiomatic study is made of quantum fields on the covering space of AdS from the point of view of a suitable modification of the Wightman function approach to quantum field theory on Minkowski space. Explicit examples of two-point functions satisfying their assumptions are given there, which include Fronsdal’s examples. A subclass of those two-point functions restrict to AdS; it is of interest to note that such two-point functions are characterized by a certain uniformity property – cf. Section 6 in [@BEM]. Using the results of [@PuWo; @BEM] it is easy to show that free fields built upon those two-point functions satisfy our assumptions (i)-(iv); the elements of a subclass also satisfy assumption (NC). The authors of [@BEM] also observe that the locality property proven in Theorem \[weaklocality2\] obtains only in those free field models which restrict to AdS proper [@BEM p. 509].
[Bor]{}
S.J. Avis, C.J. Isham and D. Storey, Quantum field theory in anti-de Sitter space-time, [*Phys. Rev. D, **18***]{}, 3565–3576 (1978). M. Bertola, J. Bros, U. Moschella and R. Schaeffer, A general construction of conformal field theories from scalar anti-de Sitter quantum field theories, [*Nucl. Phys., **B 587***]{}, 619–644 (2000). H.-J. Borchers and D. Buchholz, Global properties of vacuum states in de Sitter space, [*Ann. Inst. Henri Poincaré, **70***]{}, 23–40 (1999). O. Bratteli and D.W. Robinson, [*Operator Algebras and Quantum Statistical Mechanics I*]{}, Berlin, Heidelberg, New York: Springer-Verlag, 1979. O. Bratteli and D.W. Robinson, [*Operator Algebras and Quantum Statistical Mechanics II*]{}, Berlin, Heidelberg, New York: Springer-Verlag, 1981. P. Breitenlohner and D.Z. Freedman, Stability in gauged extended supergravity, [*Ann. Phys., **144***]{}, 249–281 (1982). J. Bros, H. Epstein and U. Moschella, Towards a general theory of quantized fields on the anti–de Sitter space–time, [*Commun. Math. Phys., **231***]{}, 481–528 (2002). D. Buchholz, Product states for local algebras, [*Commun. Math. Phys., **36***]{}, 287–304 (1974). D. Buchholz and E.H. Wichmann, Causal independence and the energy-level density of states in local quantum field theory, [*Commun. Math. Phys., **106***]{}, 321–344 (1986). D. Buchholz, C. D’Antoni and K. Fredenhagen, The universal structure of local algebras, [*Commun. Math. Phys., **111***]{}, 123–135 (1987). D. Buchholz, O. Dreyer, M. Florig and S.J. Summers, Geometric modular action and spacetime symmetry groups, [*Rev. Math. Phys., **12***]{}, 475–560 (2000). D. Buchholz, M. Florig and S.J. Summers, The second law of thermodynamics, TCP and Einstein causality in anti-de Sitter space–time, [*Class. Quantum Grav., **17***]{}, L31–L37 (2000). D. Buchholz, J. Mund and S.J. Summers, Geometric modular action on Robertson–Walker space–times and transplantation of local nets, [ *Fields Inst. Commun., **30***]{}, 65–81 (2001). D. Buchholz and G. Lechner, Modular nuclearity and localization, preprint math-ph/0402072. A. Connes, Une classification des facteurs de type III, [*Ann. Scient. Éc. Norm. Sup., **6***]{}, 133–252 (1973). C. D’Antoni, R. Longo and F. Radulescu, Conformal nets, maximal temperature and models from free probability, [*J. Operator Theory, **45***]{}, 195–208 (2001). S. Deser and O. Levin, Accelerated detectors and temperature in (anti) de Sitter spaces, [*Class. Quantum Grav., **14***]{}, L163–L168 (1997). S. Doplicher and R. Longo, Standard and split inclusions of von Neumann algebras, [*Invent. Math., **75***]{}, 493–536 (1984). N.T. Evans, Discrete series for the universal covering group of the $3+2$ de Sitter group, [*J. Math. Phys., **8***]{}, 170–184 (1967). C. Fronsdal, Elementary particles in a curved space, [*Rev. Mod. Phys., **37***]{}, 221–224 (1965). C. Fronsdal, Elementary particles in a curved space, II, [*Phys. Rev. D, **10***]{}, 589–598 (1974). C. Fronsdal, Elementary particles in a curved space, III, [*Phys. Rev. D, **12***]{}, 3810–3818 (1975). C. Fronsdal, Elementary particles in a curved space, IV, [*Phys. Rev. D, **12***]{}, 3819–3830 (1975). J. Hamhalter, Statistical independence of operator algebras, [*Ann. Inst. Henri Poincaré, **67***]{}, 447–462 (1997). T. Jacobson, Comment on “Accelerated detectors and temperature in (anti) de Sitter spaces”, [*Class. Quantum Grav., **15***]{}, 251–253 (1998). R.V. Kadison and J.R. Ringrose, [*Fundamentals of the Theory of Operator Algebras*]{}, Volume II, Orlando: Academic Press, 1986. T. Kato, [*Perturbation Theory for Linear Operators*]{}, Berlin, Heidelberg, New York: Springer-Verlag, 1966. G. Lechner, On the existence of local observables in theories with a factorizing S-matrix, preprint math-ph/0405062. W. Pusz and S.L. Woronowicz, Passive states and KMS states for general quantum systems, [*Commun. Math. Phys., **58***]{}, 273–290 (1978). M. Reed and B. Simon, [*Methods of Modern Mathematical Physics*]{}, New York: Academic Press, 1972. K.-H. Rehren, Algebraic holography, [*Ann. Henri Poincaré, **1***]{}, 607–623 (2000). H. Roos, Independence of local algebras in quantum field theory, [*Commun. Math. Phys., **16***]{}, 238–246 (1970). R.F. Streater and A.S. Wightman, [*PCT, Spin and Statistics, and All That*]{}, Reading, Mass.: Benjamin, 1964. S.J. Summers, On the independence of local algebras in quantum field theory, [*Rev. Math. Phys., **2***]{}, 201–247 (1990). B. de Wit and H. Nicolai, Extended supergravity with local $SO(5)$ invariance, [*Nucl. Phys., **B 188***]{}, 98–108 (1981).
[^1]: We refer the interested reader to the SPIRES database, where a comprehensive list of articles on this topic can be retrieved.
[^2]: See Appendix \[grouptheory\] for our notation concerning ${\mbox{SO$_0(2,n\!-\!1)$}}$
[^3]: It is sufficient here to consider the subspace of “bosonic” states, so we shall not need to proceed to the covering group of the spacetime symmetry group.
[^4]: This proof is a straightforward adaptation of an argument given in [@BB].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Ken-Ichi [Aoki]{},[^1] Atsushi [Horikoshi]{},[^2] Masaki [Taniguchi]{}[^3]\
and Haruhiko [Terao]{}[^4]
title: |
Non-Perturbative Renormalization Group Analysis\
in Quantum Mechanics
---
Introduction
============
The non-perturbative renormalization group (NPRG) has been formulated through analyses of critical phenomena[@wk] and applied to non-perturbative studies of statistical mechanics and quantum field theories. It has been established as a powerful tool for analyses of non-perturbative dynamics in systems of many (infinite) degrees of freedom, because it allows for the evaluation of fluctuations without recourse to perturbation series. Several types of non-perturbative ([*exact*]{}) renormalization group equations have been derived by integration with scale decomposition and have been applied to various systems.[@aoki1; @mo; @jw; @erg1; @aoki2; @bb; @erg2; @btw] In this article, we apply the NPRG method to quantum mechanical systems, that is, systems of finitely many degrees of freedom[@aoki1; @aoki2; @ahtt] to analyze their non-perturbative dynamics.
Generally, there are two types of non-perturbative quantities. One corresponds to the summation of all orders of a perturbative series, which could be related to Borel resummation.[@zj] The other is an essential singularity with respect to a coupling constant $\lambda_0$, which has a structure like $e^{-\frac{1}{\lambda_0}}$.[@co; @kl] We are not able to expand such a singular contribution around $\lambda_0=0$. A singularity of this type appears in the case of quantum tunneling. For example, in a symmetric double well system, there are two degenerate energy levels at each minimum, which are mixed through tunneling to generate an energy gap $\Delta E \sim e^{-\frac{1}{\lambda_0}}$. The exponential factor is known to result from the free energy of topological configurations, i.e., instantons.
In this article, we first summarize how to analyze quantum mechanical systems using the concept of NPRG and check to what extent NPRG can be used to evaluate non-perturbative effects quantitatively. The NPRG equation we employ here is a local potential approximated Wegner-Houghton (LPA W-H) equation,[@wh; @ap] which we use to analyze quantum anharmonic oscillators and asymmetric double well systems. In contrast to the symmetric double well system, the standard instanton method does not work for an asymmetric potential, and the much more sophisticated method of the valley instanton has been developed for their treatment.[@ao1; @ao2] The NPRG method is found to work for asymmetric potentials as well as for symmetric potentials, because NPRG does not rely on parity symmetry.
We proceed to analyses of more complicated systems, supersymmetric quantum mechanics (SUSY QM) and many particle systems. SUSY QM is a toy model for dynamical SUSY breaking.[@wi1; @sv] Although, in general, there is no spontaneous symmetry breaking in systems with finitely many degrees of freedom, some extraordinary symmetries, such as SUSY, can be broken even in quantum mechanics. SUSY breaking is a highly non-perturbative phenomenon because of the non-renormalization theorem, and we will see that NPRG should be applicable for non-perturbative SUSY breaking.
In addition, analyses of quantum many particle systems have become very important with recent developments in nano-technology. Solving the problem of how the quantum coherence of a variable of a target system is affected by other variables (the environment) is quite important. For example, it is necessary for realization of [*qubit*]{} for quantum computers. However, standard methods that are well suited for treating systems of one degree of freedom, the Schrödinger equation, instanton, etc., do not work well in such complicated systems. We believe that NPRG is versatile enough to analyze such systems. As a first step, we analyze quantum tunneling phenomena in two particle quantum systems.
Non-perturbative renormalization group
======================================
In this section, we briefly summarize the formulation of NPRG with $D$-dimensional real scalar field theory.
Scale decomposition
-------------------
In the NPRG method, the theory is defined by the Wilsonian effective action $S_{\Lambda}[\phi]$. This is an effective theory with an ultraviolet energy cutoff $\Lambda$: $$\begin{aligned}
Z=\int {\cal D}\phi~e^{-S_\Lambda[\phi]}.\label{(1)}
\end{aligned}$$ We decompose the path integration variable $\phi (p)$ into two parts with respect to the momentum scale $p$ as $$\begin{aligned}
\phi (p)=\left\{
\begin{array}{ll}
\phi_{<} (p) &~~~~~~~~~~~~\!0\leq|p|< \Lambda -{\mit\Delta}\Lambda
~~:{\rm lower~modes},
\\
\phi_s ~(p) &~~~\Lambda-{\mit\Delta}\Lambda\leq|p|\leq
\Lambda ~~~~~~~~
~~:~{\rm shell~modes},
\end{array}
\right.
\end{aligned}$$ and transform the partition function $Z$ as follows: $$\begin{aligned}
Z&=&\int {\cal D}\phi_{<}~{\cal
D}\phi_{s}~e^{-S_{\Lambda}[\phi_{<}+\phi_{s}]}, \nonumber\\
&=&\int {\cal D}\phi_{<}~e^{-S_{\Lambda}[\phi_{<}]}
\int {\cal D}\phi_{s}~e^{-S_{\Lambda}[\phi_{s}]}~
e^{-S_{\Lambda}^{\rm ~\!int}[~\!\phi_{<},~\phi_{s}~\!]},\nonumber\\
&=&\int {\cal D}\phi_{<}~e^{-S_{\Lambda}[\phi_{<}]}
\left<~e^{-S_{\Lambda}^{\rm ~\!int}[~\!\phi_{<},~\phi_{s}~\!]}\right>
_{\phi_{s}},\nonumber\\
&=&\int {\cal D}\phi_{<}~e^{-S_{\Lambda}[\phi_{<}]}
~e^{- {\mit\Delta}S_{\Lambda}[\phi_{<}]},\nonumber\\
&=&\int {\cal D}\phi_{<}~e^{-S_{\Lambda-{\mit\Delta}\Lambda}[\phi_{<}]},\label{(2)}
\end{aligned}$$ where $$\begin{aligned}
{\mit\Delta}S_{\Lambda}[\phi_{<}]
\equiv-\log\left<~e^{-S_{\Lambda}^{\rm ~\!int}
[~\!\phi_{<},~\phi_{s}~\!]}\right>_{\phi_{s}}
\equiv\int {\cal D}\phi_{s}~e^{-S_{\Lambda}[\phi_{s}]}~
e^{-S_{\Lambda}^{\rm ~\!int}[~\!\phi_{<},~\phi_{s}~\!]}.\label{(3)}
\end{aligned}$$ We understand the shell mode path integral $\left<\cdot\cdot\cdot\right>_{\phi_{s}}$ as the renormalization transformation. If we evaluate it by perturbative expansion with respect to coupling constants, we obtain the so-called perturbative renormalization group equations.[@ps] Of course such equations are valid only in the weak coupling limit. Instead, we take the limit ${\mit\Delta}\Lambda\to0$ to define NPRG equation, which is the fundamental procedure.[@wh]
Derivation of the NPRG equation
-------------------------------
Taking the limit ${\mit\Delta}\Lambda\to0$, we can express the renormalization transformation as a differential equation, $$\begin{aligned}
\frac{\partial S_{\Lambda}}{\partial
\Lambda}=\lim_{{\mit\Delta}\Lambda \to 0}\frac{S_{\Lambda}-
S_{\Lambda-{\mit\Delta}\Lambda}}{{\mit\Delta}\Lambda}
=\lim_{{\mit\Delta}\Lambda \to 0}\frac{1}{{\mit\Delta}\Lambda}
~{\rm log}~\left<~e^{-S_{\Lambda}^
{\rm ~\!int}[~\!\phi_{<},~\phi_{s}~\!]}\right>
_{\phi_{s}}.\label{(4)}
\end{aligned}$$ Then, we can expand $S_{\Lambda}[\phi]$ as power series in $\phi _s$: $$\begin{aligned}
S_{\Lambda} \left[ \phi \right] &=& S_{\Lambda} \left[ \phi_{<}
\right]
+\int_{\rm shell} \left.{{\delta S_{\Lambda} \over \delta \phi \left( p \right)}}\right|
_{\phi_s=0} \!\!\!\!\!\!\cdot \phi_s \left( p \right) \nonumber\\
&& +~{1 \over 2}\int\int_{\rm shell}
\phi_s (p)\cdot\left.{{\delta ^2 S_{\Lambda} \over \delta \phi (p) \delta \phi (q)}}
\right| _{\phi _s =0} \!\!\!\!\!\!\cdot \phi_s(q)~
+~O({\mit\Delta}\Lambda ^2 ).\label{(5)}
\end{aligned}$$ Since an $O({\mit\Delta}\Lambda)$ calculation is required for the evaluation of the derivative (\[(4)\]), the shell mode path integral $\left<\cdot\cdot\cdot\right>_{\phi_{s}}$ can be evaluated [*exactly*]{} using a Gaussian integration. Then, the fundamental differential equation is derived as $$\begin{aligned}
\Lambda \frac{\partial S_{\Lambda}}{\partial\Lambda}
=~\frac{\Lambda}{2}\int_{\rm shell}
\left\{
-~{\rm log}\left(\left.{{\delta ^2 S_{\Lambda} \over \delta \phi_{p} \delta \phi_{-p}}}
\right|\right)
+\left.{{\delta S_{\Lambda} \over \delta \phi_{p}}}\right|
\left(\left.{{\delta ^2 S_{\Lambda} \over \delta \phi_{p} \delta \phi_{-p}}}
\right|\right)^{-1}\!\!\!
\left.{{\delta S_{\Lambda} \over \delta \phi_{-p}}}\right|
~\right\}.\label{(6)}
\end{aligned}$$ This is known as the Wegner-Houghton equation.[@wh] It represents exactly the cutoff $\Lambda$ dependence of the Wilsonian effective action $S_{\Lambda}$. Its right-hand side is generally called a [*$\beta$ functional*]{}.
Approximations
--------------
Although the Wegner-Houghton equation is exact, we cannot solve it without some approximation in practice. In this article we employ the local potential approximation (LPA), which means that we ignore corrections to derivative interactions. It can be considered the leading order of the derivative expansion of $S_{\Lambda}$, $$\begin{aligned}
S_{\Lambda}[\phi]=\int d^D x ~\left\{~V_{\Lambda}[\phi]
+\frac{1}{2}K_{\Lambda}[\phi]\!~
\partial_{\mu}\phi \partial_{\mu}\phi
+\cdot\cdot\cdot\right\}.\label{(7)}
\end{aligned}$$ To make this approximation, we substitute the zero mode for the lower mode $\phi_{<}(p)$: $$\begin{aligned}
\phi_{<}(p)\to\varphi ~(2\pi)^D\delta^D(p).\label{(8)}
\end{aligned}$$ The local potential approximated Wegner-Houghton (LPA W-H) equation is then obtained as follows: $$\begin{aligned}
\Lambda{{\partial V_{\Lambda} \over \partial \Lambda}}&=&
-~\!\frac{A_{D}}{2}~\!\Lambda^{D}
~\!\log{\left( 1+\frac{1}{\Lambda ^2}\frac{\partial ^2
V_{\Lambda}}{\partial \varphi ^2} \right) },\\ \label{(9)}
A_D &\equiv&\frac{\int d\Omega _{D}}{(2\pi)^D}.\label{(9a)}
\end{aligned}$$ This is a two-dimensional partial differential equation for $V_{\Lambda}(\varphi).$[@ap] Its right-hand side is called a [*$\beta$ function.*]{} We solve it mainly using numerical methods.
To obtain an intuitive understanding of this equation, we proceed to further approximation, the operator expansion. We expand $V_{\rm eff}$ as power series in $\varphi$: $$\begin{aligned}
V_{\Lambda} \left(\varphi\right) &=&
\sum_{n=0}^N\frac{a_n(\Lambda)}{n!}\varphi ^n.\label{(10)}\end{aligned}$$ The partial differential equation is then reduced to a set of ordinary differential equations for the coupling constants $\{{a}_n(\Lambda)\}$: $$\begin{aligned}
\Lambda{{d {a}_0 \over d \Lambda}}&=&-\frac{A_D}{2}~\!\Lambda^{D}~\!
\log\left(\frac{\Lambda^2+{a}_2}
{\Lambda^2}\right),\label{(11)}\\
\Lambda{{d {a}_1 \over d \Lambda}}&=&-\frac{A_D}{2}~\!\Lambda^{D}~\!
\left[{a}_3 \over
\Lambda^2+{a}_2\right], \label{(12)}\\
\Lambda{{d {a}_2 \over d \Lambda}}&=&-\frac{A_D}{2}~\!\Lambda^{D}~\!
{\left[ {{a}_4 \over \Lambda^2 +{a}_2}-
{{a}_3^2 \over {\left( \Lambda^2+{a}_2 \right) }^2} \right] }, \label{(13)}\\
\Lambda{{d {a}_3 \over d \Lambda}}&=&-\frac{A_D}{2}~\!\Lambda^{D}~\!
{\left[ {{a}_5 \over \Lambda^2+{a}_2}-{3{a}_4{a}_3
\over {\left( \Lambda^2+{a}_2 \right) }^2}
+{2{a}_3^3 \over {\left( \Lambda^2+{a}_2 \right) }^3} \right] }, \label{(14)}\\
\Lambda{{d {a}_4 \over d \Lambda}}&=&
-\frac{A_D}{2}~\!\Lambda^{D}~\!
\left[{{a}_6 \over \Lambda^2+{a}_2}-{4{a}_5{a}_3
\over {\left( \Lambda^2+{a}_2 \right) }^2}
-{3{a}_4^2 \over {\left( \Lambda^2+{a}_2 \right) }^2}\right.\nonumber\\
&&~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\left.+\frac{12{a}_4{a}_3^2}{{\left( \Lambda^2+{a}_2 \right) }^3}
-\frac{6{a}_3^4}{{\left( \Lambda^2+{a}_2 \right) }^4}
\right]
, \label{(15)}\\
&& \vspace{3cm} \vdots
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~. \nonumber
\end{aligned}$$ If the results of these equations converge as the order of the truncation, $N$, becomes large, we regard them as solutions of the LPA W-H equation.[@ka] This expansion allows us to treat differential equations more easily and to understand the origin and structure of quantum corrections physically. For example, a correction to odd $n$ couplings cannot be generated from even $n$ couplings only. This implies that if we choose an initial potential $V_0(\varphi)$ as $Z_2(\varphi\leftrightarrow-\varphi)$ symmetric, the solutions of the RG equations flow in $Z_2$ symmetric subspace; that is, the NPRG equation does not break the global symmetry of the system. However, we should note that in some extreme cases, the operator expansion is not good, and leads to pathological behavior, as seen in §4.
The constant part of $V_{\Lambda}$, $a_0$, is given by the vacuum bubble diagrams[^5] and is usually ignored. However, we retain it here, because it plays a crucial role in supersymmetric theories.
NPRG analysis of quantum mechanical systems
===========================================
Making use of the LPA W-H equation, we analyze systems in quantum mechanics, which is $D=1$ real scalar theory with a single dynamical variable, $x(\tau)$.[@ahtt]
Physical quantities
-------------------
The LPA W-H equation for quantum mechanics is given as follows: $$\begin{aligned}
\Lambda\frac{\partial V_{\Lambda}}{\partial \Lambda}&=&
-~\!\frac{1}{2\pi}~\!\Lambda
~\!\log{\left( 1+
\frac{1}{\Lambda ^2}\frac{\partial ^2
V_{\Lambda}}{\partial x ^2} \right) }.\label{(16)}
\end{aligned}$$ We solve it by lowering $\Lambda$ from the initial cutoff $\Lambda _0$, where the initial potential $V_{\Lambda_0}$ is given by the potential term $V_0(x)$ in the original action, $$\begin{aligned}
S[x]=\int d \tau
~\left\{~\frac{1}{2}\!~\dot{x}^2+V_0(x)
\right\}.\label{(17)}
\end{aligned}$$ In the infrared limit $\Lambda\to 0$, we obtain the effective potential $V_{\rm eff}(x)=\lim_{\Lambda\to 0}V_{\Lambda}(x)$, from which physical quantities are evaluated.
First, the expectation value of $x$ in the ground state ${\left| \Omega \right\rangle}$, $$\begin{aligned}
\langle x\rangle\equiv {\left\langle \Omega \right|}\hat{x}{\left| \Omega \right\rangle},\label{(18)}
\end{aligned}$$ is determined by the stationarity condition, $$\begin{aligned}
\left.\frac{d V_{\rm eff}}{d x}\right|_{x=\langle x\rangle}=0.\label{(19)}
\end{aligned}$$ The ground state energy of the quantum system is given by $$\begin{aligned}
E_0={\left\langle \Omega \right|}\hat{H}{\left| \Omega \right\rangle}=
V_{\rm eff}(x=\langle x\rangle).\label{(20)}
\end{aligned}$$ Also, we obtain the energy gap of the system through the following expressions of the two-point correlation function: $$\begin{aligned}
{\left\langle \Omega \right|}{T}\hat{x}(\tau)\hat{x}(0){\left| \Omega \right\rangle}
&=&
\int
\frac{dE}{2\pi}e^{iE\tau}\sum_{n=1}^{\infty}\frac{D_n}{E^2+(E_n-E_0)^2}
\stackrel{\tau\to\infty}{\propto } e^{-(E_1-E_0)\tau}, \label{(21)}\\
{\left\langle \Omega \right|}{T}\hat{x}(\tau)\hat{x}(0){\left| \Omega \right\rangle}
&\stackrel{LPA}{=}&
\int \frac{dE}{2\pi}e^{iE\tau}\frac{1}{E^2+m^2_{\rm eff}}
=\frac{1}{2m_{\rm eff}}
e^{-m_{\rm eff}\tau},\label{(22)}
\end{aligned}$$ where $D_{n}\equiv 2|C_n|^{2}(E_n-E_0),~C_{n}\equiv
{\left\langle n \right|}\hat{x}(0){\left| \Omega \right\rangle},~\sum_{n}D_n=1$ and $m_{\rm eff}^2$ is the curvature of the effective potential at the minimum. Comparing the damping factors of (\[(21)\]) and (\[(22)\]) in the $\tau\to\infty$ region, the energy gap $\Delta E=E_1-E_0$ is obtained as follows:[^6] $$\begin{aligned}
\Delta E =m_{\rm eff}=\sqrt{\left.\frac{\partial^2V_{\rm eff}}
{\partial x^2}\right|}_{x=\langle x\rangle}.\label{(23)}
\end{aligned}$$
Generally, the NPRG method does not evaluate the wave functions of the target system directly. Instead, it yields information regarding Green functions. For example, the $n$-th moment of $\hat{x}$ corresponds to information concerning the ground state wave function $\psi _0(x)$: $$\begin{aligned}
M_n &=&
\langle \Omega |~\hat{x}^n| \Omega \rangle
=\int \!dx ~x^n \left| \psi _0(x)\right|^2.\label{(24)}
\end{aligned}$$ We now give some examples of calculations of $M_n$ in a $Z_2$-symmetric (namely, $\langle x\rangle=0$) system. The two-point function $M_2$ is calculated as $$M_2=\langle \Omega |~\hat{x}^2| \Omega \rangle _c
\stackrel{LPA}{=}\int \frac{dE}{2\pi}\frac{1}{E^2+m^2_{\rm eff}}
=\frac{1}{2m_{\rm eff}},\label{(25)}$$ where the subscript $c$ denotes a connected function. In a similar way, the four-point function $M_4$, $$\begin{aligned}
M_4=\langle \Omega |~\hat{x}^4| \Omega \rangle _c
+3 M_{2}^{~2},\label{(26)}
\end{aligned}$$ is calculated by means of the LPA four-point coupling $~\lambda_{\rm eff}
\equiv\frac{\partial^{4}V_{\rm eff}}{\partial x^{4}}$: $$\begin{aligned}
\langle \Omega |~\hat{x}^4| \Omega \rangle _c
&\stackrel{LPA}{=}&
-4!~\lambda_{\rm eff}\int\frac{dE_1dE_2dE_3}{(2\pi)^3}
\nonumber \\
&&\times ~{1 \over E_1^2\!+m^2_{\rm eff}}
~{1 \over {\left( E_2-\!E_1 \right) }^2\!+m^2_{\rm eff}}
~{1 \over {\left( E_3-\!E_2 \right) }^2\!+m^2_{\rm eff}}
~{1 \over E_3^2\!+m^2_{\rm eff}},
\nonumber \\
&=&-{3\lambda_{\rm eff} \over 4m^5_{\rm eff}}.\label{(27)}
\end{aligned}$$ In this way, we are able to connect the effective couplings obtained from the NPRG with $M_n$ obtained from the ground state wave function. All of this information concerning the ground state taken together provides full information regarding the quantum system, including excited states.
Example: harmonic oscillator
----------------------------
To illustrate the characteristics of the NPRG analysis, we now consider the harmonic oscillator. The initial potential is chosen as $V_0(x)=\frac{1}{2}m^2 x^2$ at the initial cutoff scale $\Lambda_0$. In this case, we can solve the LPA W-H equation analytically, and we obtain $$\begin{aligned}
a_0 (\Lambda)&=&a_0(\Lambda_0)+
\frac{\sqrt{a_2(\Lambda_0)}}{2\pi}
\left[\hat{p}\log \frac{1+\hat{p}^2}{\hat{p}^2}+2\tan^{-1}\hat{p}
\right]^{\hat{p}
=\frac{\Lambda_0}{\sqrt{a_2(\Lambda_0)}}}_{\hat{p}
=\frac{\Lambda}{\sqrt{a_2(\Lambda_0)}}} , \label{(28)}\\
a_2(\Lambda)&=&a_2(\Lambda_0).\label{(29)}\end{aligned}$$ Since the initial conditions are $(a_0(\Lambda_0),a_2(\Lambda_0))=(0,m^2)$, if we take the simultaneous limit $\Lambda_0\to \infty$, $\Lambda\to 0$, $(a_0(\Lambda),a_2(\Lambda))=(\frac{m}{2},m^2)$ is obtained. Although $a_2 (\Lambda)$ is free from quantum corrections and does not run, $a_0$ runs and produces a zero-point energy $\frac{m}{2}$.
[r]{}[6.6cm]{}
Figure.\[fig:sirun\] plots the actual running of $a_0$ and shows that it is limited to a finite energy region that depends on the mass scale, $m$. Ultraviolet finiteness is a typical feature of quantum mechanical systems, and it implies that the theory is finite, even in the $\Lambda_0\to \infty $ limit. Contrastingly, the infrared finiteness in Figure.\[fig:sirun\] is related to the decoupling property that a heavy particle cannot propagate in the low energy region. Such ultraviolet finiteness and infrared finiteness enable us to obtain physical quantities even through numerical calculation within a finite energy scale region.
Analysis of anharmonic oscillators and double well systems
==========================================================
Symmetric single-well potential
-------------------------------
Now we proceed to analyze quantum mechanical anharmonic oscillators and double well systems. First, we consider a symmetric single-well potential, $$\begin{aligned}
V_0(x) =~~\lambda_0 x^4+\frac{1}{2} x^2.\label{(30)}\end{aligned}$$ Our interest is to compare our NPRG results with the perturbative series. First, the LPA W-H equation (\[(16)\]) is solved numerically, and we thereby obtain an effective potential $V_{\rm eff}$. The flow of $V_{\Lambda}$ is shown in Figure.\[fig:spote\]. Quantum corrections raise the potential and make its slope steeper. In Figure.\[fig:sspe\], we display the energy spectrum calculated with the relations (\[(20)\]) and (\[(23)\]). We refer to the results obtained by a numerical analysis of the Schrödinger equation as the “exact results.”
The perturbative series of $E_n$ is the asymptotic series $$\begin{aligned}
E_n\!=\![n+\frac{1}{2}]+\frac{3}{4}\lambda_0 [2n^2+2n+1]
-\frac{1}{8}\lambda_0 ^2 [34n^3+51n^2+59n+21]+\cdots .~~~~ \label{(31)}\end{aligned}$$ It diverges even in the weak coupling region. Note that the Borel resummation of the perturbative series works well in this case, and gives quantitatively good values. However, even in the lowest order approximation (LPA), the W-H equation can evaluate the energy spectrum almost perfectly. Therefore, we conclude that the NPRG does sum up all orders of the perturbative series in the correct manner.
We also display the two-point function $M_2 =\langle \Omega |~\hat{x}^2| \Omega \rangle$ in Figure.\[fig:sx2\] and the effective four-point coupling constant $\lambda_{\rm eff} $ in Figure.\[fig:seffLa\]. In both of these cases as well, the NPRG results give an almost perfect fit.
Symmetric double-well potential
-------------------------------
Next, we consider the $Z_2$-symmetric $x\leftrightarrow -x$ double-well potential, $$\begin{aligned}
V_0(x) =\lambda_0 x^4-\frac{1}{2}x^2. \label{(32)}\end{aligned}$$ In quantum mechanical systems, this $Z_2$ symmetry never breaks spontaneously, because the $x$ mode tunnels through the potential barrier, and the ground state is uniquely realized. In fact, in the NPRG evolution of the effective potential, the initial double-well potential finally becomes a single well, and an energy gap (effective mass) arises (Figure.\[fig:spote3\]).
In this system, there is no well-defined perturbation theory. A standard technique to obtain the energy gap $\Delta E$ is the dilute gas instanton calculation. This is a semi-classical method based on the one-instanton solutions $$\begin{aligned}
x_{\rm cl}(\tau)=\pm \frac{1}{2\sqrt{\lambda_0}}
\tanh \frac{(\tau-\tau_0)}{\sqrt{2}}. \label{(33)}\end{aligned}$$ The one-instanton contribution to the partition function $Z$ is $$\begin{aligned}
Z\simeq Te^{-S{\left[ x_{\rm cl} \right] }}
\sqrt{\frac{S{\left[ x_{\rm cl} \right] }}{2\pi}}
{\left\{ \frac{
\det ^{\prime}{\left( \left.
\frac{\delta ^2 S}{\delta x^2}
\right|_{x=x_{\rm cl}} \right) }}{
\det{\left( \left.
\frac{\delta ^2 S}{\delta x^2}
\right|_{x=\pm \frac{1}{2\sqrt{\lambda_0}}} \right) }
}
\right\} }^{-1/2} \equiv T\frac{\Delta _0}{2}, \label{(34)}\end{aligned}$$ where $T$ is an imaginary time volume. Assuming that instantons do not interact with each other, we can evaluate the multi-instanton contribution to $Z$ (the dilute gas instanton approximation), and we obtain the energy gap $$\begin{aligned}
\Delta E=\Delta _0=2\sqrt{\frac{2\sqrt{2}}{\pi \lambda_0}}
e^{-\frac{1}{3\sqrt{2}\lambda_0}}, \label{(35)}\end{aligned}$$ which has the structure of an essential singularity originating from the one-instanton action. The singularity coefficient obtained from the instanton method is known to be exact in the vanishing $\lambda_0$ limit.[@simon]
In Figure.\[fig:smass\] we display the energy gap evaluated using various methods. The NPRG results are very good in the strong coupling region, while the perturbation cannot be applied in this double-well system, and the dilute gas instanton method is not at all effective, because it is valid only in the very weak coupling region.[^7] Therefore, the NPRG method should provide a powerful tool for the analysis of tunneling, at least in such regions. However, our NPRG results deviate from the exact values as $\lambda _0 \to
0$, which corresponds to a very deep well. Because the $\beta$-function becomes singular in this region, the NPRG results become unreliable. We believe that the cause of the difficulty is the LPA approximation scheme that we adopt. It is important to note that the respective coupling regions in which the LPA W-H equation and the dilute gas instanton work well are separated, and therefore these two methods should be regarded as complementary.[@ahtt; @ggm]
We display the two-point function $M_2 =\langle \Omega |~\hat{x}^2| \Omega \rangle$ in Figure.\[fig:dx2\] and the effective four-point coupling $\lambda_{\rm eff} $ in Figure.\[fig:deffLa\]. As in the case of the energy gap, the NPRG results are excellent here, except in the extremely weak coupling region.
Flow diagrams
-------------
We now more carefully consider the difficulty arising in the weak coupling region for the double-well potential. We employ the operator expansion (\[(10)\]) and investigate the flows of the dimensionless coupling constants $\hat{a}_n\equiv a_n \Lambda^{-\frac{n+2}{2}}$. The flow diagrams elucidate the phase structure of the system. We display the flow diagrams for the $N=4,6,10$ truncated potentials (Figure.\[fig:sWHflowN4\], Figure.\[fig:sWHflowN6\], Figure.\[fig:sWHflowN10\]) and for the potential without an operator expansion (Figure.\[fig:sflow24\]).
These flow diagrams reveal that the phase structure of the theory with a truncated potential ($N=4,6,10$) is somewhat strange. As mentioned above, there is no spontaneous symmetry breaking in these quantum mechanical systems. When we truncate the potential as in (\[(10)\]), there appears a non-actual fixed point and false broken phases (Figures \[fig:sWHflowN4\], \[fig:sWHflowN6\] and \[fig:sWHflowN10\]). The flow starting from the weak coupling region ($\lambda_{0} \to 0$ i.e. $\hat{a_4}\to 0$) tends to be captured by the false broken phase, and we cannot obtain the correct result $m^2_{\rm eff}>0$. The region of the false broken phase becomes smaller as $N$ becomes larger, and then for the LPA exact (no truncation) calculation, the false broken phase disappears (Figure.\[fig:sflow24\]). However, even in the no truncation case, we cannot obtain reliable results for the flows that start from the weak coupling region, because singular behavior of the flow in the region near $\hat{a}_2=-1$ leads to large numerical errors. The results for the LPA W-H equation in Figure.\[fig:smass\] were obtained from numerical integration of the partial differential equation without any truncation.[@kt]
[r]{}[6.6cm]{}
Other methods
-------------
The NPRG equation we employ here is that with the local potential approximation. The results in the weak coupling region can be improved by upgrading the approximation. The LPA is the lowest order of the derivative expansion, and a higher-order calculation can be carried out.[@morris] In this quantum mechanical system, the second-order calculation of the Legendre flow equation does not improve the weak coupling results.[@ah] However, an analysis using the proper time renormalization group improves the LPA results considerably.[@za] Also, although it differs from the NPRG methods in its formulation, the density matrix renormalization group is useful for this system.[@mdsn] We exhibit in Figure.\[fig:various\] the results for the energy gap obtained with various renormalization group approaches.[^8]
Asymmetric double-well potential
--------------------------------
We proceed to consider the $Z_2$-asymmetric double-well potential $$\begin{aligned}
V_0(x) =~~\lambda_0 x^4-\frac{1}{2} x^2+h_0 x, \label{(36)}\end{aligned}$$ where the linear term $h_0x$ breaks the $Z_2$ symmetry explicitly. In this system there are a stable minimum and an unstable minimum.
How do we deal with the effect of such an asymmetric term? The NPRG method can treat this system in a manner that is quite similar to that for the symmetric system; it just changes the initial potential, while the LPA W-H equation does not change. Furthermore, when we apply the operator expansion (\[(10)\]), the situation becomes even simpler. The additional $h_0x$ term does not affect the running of other coupling constants, because the term $$\int d\tau \!~
h_0x(\tau)=h_0 x(E=0), \label{(37)}$$ consists entirely of the zero energy mode, and generates no quantum corrections. Therefore, the NPRG equations for the coupling constants are the same as those in the symmetric case.
By contrast, the standard instanton method cannot be applied to such an asymmetric system, because the term $\left.\frac{\delta^2 S}{\delta x\delta x}\right|_{x=x_{\rm cl}}$ in (\[(34)\]) has a negative eigenvalue in this case. For actually unstable systems, this negative eigenvalue is converted to a decay rate for the system. This is a typical prescription for the ‘bounce solution’ calculation. However, in the case of the bare potential (\[(32)\]), the true vacuum of the system is stable. The existence of a negative eigenvalue in actually stable systems is known as the problem of a fake instability. To overcome this problem, the valley method has been developed recently.[@ao1; @ao2] It is a generalization of the instanton method that is based on the valley structure in the configuration space.
As shown in Figure.\[fig:adpote\], an asymmetric bare potential leads to an asymmetric effective potential. We show in Figure.\[fig:adspe\] results for the energy gap in the cases of three values of $h_0$, from bottom to top, $h_0=0.02,\!~0.2,\!~0.4$. For any value of $h_0$, in the $\lambda_0\to 0$ limit, $\Delta E$ approaches $\sqrt{2}$. This is because in this limit the asymmetric double well approaches a single well. We employ the operator expansion and give the truncation $N=12,14,16$ results. We also plot the results obtained from the valley method with fourth and sixth order perturbations.[@ao4] A complementary relation between the NPRG and the valley method is observed, just as in the case of the symmetric potential.
As mentioned above, since in the $\lambda_0\to 0$ limit the potential approaches a single well, if we carry out the operator expansion at the potential minimum $x=x_{\rm min}$, the NPRG equations never become singular even in the $\lambda_0\to 0$ region, and we obtain $m_{\rm eff}\simeq\sqrt{2}$. We use this technique for analysis of SUSY QM in the next section.
We display results for other quantities in Figure.\[fig:asymXN\] and Figure.\[fig:asymXDN\] for three values of $h_0$, from top to bottom, $h_0=0.02,\!~0.2,\!~0.4$. The expectation value of $\hat{x}$, $M_1 =\langle \Omega |~\hat{x}| \Omega \rangle$, is shown in Figure.\[fig:asymXN\], and the variance of $\hat{x}$ is shown in Figure.\[fig:asymXDN\]. The NPRG results appear to be perfect on the strong coupling side, while they are incorrect in the weak coupling region.
Applications to various quantum systems
=======================================
We have seen that the NPRG method is very effective in analyses of non-perturbative dynamics in quantum mechanical systems. Here, we apply the NPRG method to more non-trivial quantum systems.
Supersymmetric quantum mechanics
--------------------------------
Here we analyze supersymmetric theory, in which the non-perturbative dynamics of the system are crucial. We consider the SUSY QM theory, which was introduced by Witten as a toy model for dynamical SUSY breaking.[@wi1; @sv] The Hamiltonian is given by $$\begin{aligned}
\hat{H}&=&\frac12\left[\hat{P}^2+\hat{W}^2(x)
+\sigma_3\frac{d\hat{W}(x)}{dx}\right]=
\left(\matrix{\hspace{-10mm}
\frac{1}{2}\hat{P}^2
+\hat{V}_{+}(x)~~~~~~~~~~0
\cr
~~~~~~~~0~~~~~~~~~~~~
\frac{1}{2}\hat{P}^2+\hat{V}_{-}(x)}\right) ,~~~~~~~~ \\ \label{(38)}
\hat{V}_{\pm}(x)&\equiv&\frac{1}{2}\hat{W}^2(x)\pm
\frac{1}{2}\frac{d\hat{W}(x)}{dx},\label{(38a)}\end{aligned}$$ where $\hat{W}(x)$ is called the SUSY potential. We define the super charges $$\begin{aligned}
\hat{Q}_1&=&\frac12(\sigma_1\hat{P}+\sigma_2\hat{W}(x)),\\ \label{(38b)}
\hat{Q}_2&=&\frac12(\sigma_2\hat{P}-\sigma_1\hat{W}(x)),\label{(38c)}\end{aligned}$$ and the Hamiltonian is written as $$\begin{aligned}
\hat{H}=\{\hat{Q}_1,\hat{Q}_1\}=\{\hat{Q}_2,\hat{Q}_2\}.\label{(38d)}\end{aligned}$$ This ensures that the vacuum energy is always non-negative: $$\begin{aligned}
E_0=\langle\Omega|
\hat{H}|\Omega\rangle
=2\left\Vert\hat{Q}_1|\Omega\rangle\right\Vert ^{2}
=2\left\Vert\hat{Q}_2|\Omega\rangle\right\Vert ^{2}\geq0.\label{(38e)}\end{aligned}$$
The vacuum energy $E_0$ is the order parameter of dynamical SUSY breaking; that is, $$\begin{aligned}
E_0=0 \quad \Rightarrow
&&~~~\hat{Q}_1|\Omega\rangle=0 , ~\hat{Q}_2|\Omega\rangle=0 \qquad
{\rm unbroken ~~SUSY}, \nonumber\\
E_0>0 \quad \Rightarrow
&&~~~\hat{Q}_1|\Omega\rangle\ne0 , ~\hat{Q}_2|\Omega\rangle\ne0
\qquad {\rm broken ~~SUSY}.\nonumber\end{aligned}$$ Furthermore, the perturbative corrections to $E_0$ are vanishing for any order of the perturbation. This is known as the non-renormalization theorem. In fact, with the SUSY potential $W(x)=gx^2-x$, the potential $V_{+}(x)$ becomes $$\begin{aligned}
V_+(x) =~~\frac{1}{2}g^2x^4-gx^3+\frac{1}{2}x^2+gx-\frac{1}{2}.\label{(39)}\end{aligned}$$ The perturbative corrections to the energy spectrum are calculated as $$\begin{aligned}
E_n=n&+&\frac38g^2[2n^2+2n+1]-\frac38g^2[10n^2+2n+1] \nonumber \\
&-&\frac{1}{32}g^4[34n^3+51n^2+59n+21]
+\cdots .\label{(40)}\end{aligned}$$ These corrections to $E_0$ are canceled out at each order of $g$, and thus there are no perturbative corrections. Hence, a non-vanishing $E_0$ is realized only through non-perturbative effects caused by the essential singularity at the origin of the coupling constant.
We analyze this system using the LPA W-H equation with $N=16$ for the operator expansion.[@ahtt] We calculate the effective potential for a wide range of values of the parameter $g$. The case of vanishing $g$ corresponds to the harmonic oscillator with a constant term $-\frac{1}{2}$, and SUSY does not break in this case ($E_0=\frac{1}{2}-\frac{1}{2}=0$). However, SUSY is dynamically broken for any non-vanishing $g$. Note that for small $g$, the bare potential is an asymmetric double-well, while for $g > \sqrt[4]{\frac{1}{108}}\simeq 0.31$, it is a single-well, and quantum tunneling is irrelevant (Figure.\[fig:bpote\]). For any value of $g$, the minimum of the bare potential $V_{+}$ is at $x=0$. Figure.\[fig:epote\] displays the result for $g$=0.24, where the effective potential evolves into a convex form, and its minimum turns out to be positive; that is, our NPRG method gives a positive $E_0$ correctly, and describes the dynamical SUSY breaking.
As is shown in Figure.\[fig:rspe\], the NPRG results are excellent in the weak coupling region and strong coupling region, but not in the region where the bare double-well potential becomes deep. In this intermediate region ($ 0.1{\raisebox{0.2ex}{$\ < \kern -1.05em \raisebox{-1.1ex}{$\sim$}\ $}}g{\raisebox{0.2ex}{$\ < \kern -1.05em \raisebox{-1.1ex}{$\sim$}\ $}}0.2$), we cannot obtain reliable results because of large numerical errors, while the valley method works very well, as shown in Figure.\[fig:vspe2\]. The valley method evaluates the ground state energy as $E_0=\frac{1}{2\pi}e^{-\frac{1}{3g^2}}$ and reproduces the exact value in the weak coupling region.[@ao2] However, it does not work in the strong coupling region ($g>0.31$), where the valley instanton is no longer a good approximate solution of the valley equation. Again, we find that the two methods are complementary.
Two particle systems
--------------------
Next, we apply the NPRG method to quantum many particle systems. As the simplest system, we analyze two particle ($\varphi_1,\varphi_2$) dynamics with the following potential $V_0(\varphi_1,\varphi_2)$: $$V_0(\varphi_1,\varphi_2)
=-\frac{1}{2}\varphi^2_1+\lambda_0 \varphi_1^4-\frac{1}{2}
\varphi_2^2+\lambda_0
\varphi_2^4+F(\varphi_1,\varphi_2). \label{(41)}$$ Without the interaction $F(\varphi_1,\varphi_2)$ between the two particles, the four degenerate ground states are mixed by tunneling, splitting into three $\varphi_1\leftrightarrow \varphi_2$ symmetric states and one anti-symmetric state. For the interaction $F(\varphi_1,\varphi_2)$, we now choose $\varphi_1\leftrightarrow \varphi_2$ symmetric interactions and investigate how this interaction affects the energy levels of three symmetric states.
The LPA W-H equation for two particles is written $$\Lambda{{\partial V_{\Lambda} \over \partial \Lambda}}
=-\frac{1}{2\pi}~\!\Lambda~\!
{\rm Tr}\log{\left( \delta_{ab}+{{\partial^2 {V}_{\Lambda
} \over \partial \varphi_{a} \partial \varphi_{b}}} \right) }, \label{(42)}$$ where “Tr” represents the trace over the subscripts $a,b$ which correspond to the two particles. We consider three types of interactions, $$F(\varphi_1,\varphi_2)=C\varphi_1\varphi_2~,
~C_2(\varphi_1-\varphi_2)^2~,~C_4(\varphi_1-\varphi_2)^4.\label{(43)}$$ In the cases of the second and third types, for $C_2, C_4 >0$ the interaction is attractive, and for $C_2, C_4 <0$ it is repulsive. Since we now treat only the ($\varphi_1\leftrightarrow \varphi_2$) state, it is convenient to convert the variables from ($\varphi_1,\varphi_2$) to ($x_1,x_2$) as follows: $$\left(\matrix{x_1 \cr x_2}\right)=\frac{1}{\sqrt{2}}\left(\matrix{1-1\cr1~~~~1
}\right)
\left(\matrix{\varphi_1 \cr \varphi_2}\right).\label{(44)}$$ The LPA W-H equation for $(x_1,x_2)$ has the same form as (\[(42)\]). The lowest energy splitting for symmetric state, $\Delta E=E_1-E_0$, is expressed in terms of the effective mass of $x_2$. Of course, in the $C=0$ case, this is equal to the effective mass in one particle system.
The bare potentials are written, corresponding to (\[(43)\]), as $$\begin{aligned}
V_0(x)&=&\frac{1}{2}(-1-C)x^2_1+\frac{\lambda_0}{2} x_1^4
+\frac{1}{2}(-1+C)x_2^2+\frac{\lambda_0}{2}
x_2^4+3\lambda_0 x_1^2x_2^2 ,\label{(45)}\\
V_0(x)&=&\frac{1}{2}(-1+4C_2)x^2_1+\frac{\lambda_0}{2} x_1^4
-\frac{1}{2}x_2^2+\frac{\lambda_0}{2}
x_2^4+3\lambda_0 x_1^2x_2^2 ,\label{(46)}\\
V_0(x)&=&-\frac{1}{2}x^2_1+\left(\frac{\lambda_0}{2}+4C_4\right) x_1^4
-\frac{1}{2}x_2^2+\frac{\lambda_0}{2}
x_2^4+3\lambda_0 x_1^2x_2^2.\label{(47)}
\end{aligned}$$ We analyze these systems for small $C,C_{2}$ and $C_{4}$. We set $\lambda_0=0.2$, which is in the parameter region where the NPRG works perfectly in previous analyses. The LPA W-H equation was solved numerically using the operator expansion with $N=12$. We also calculated $\Delta E$ from the first order perturbation theory with one particle Schrödinger wave functions.
The results for small $C,C_{2}$ and $C_{4}$ are shown in Figures \[fig:cm1\], \[fig:cm2\] and \[fig:cm4\]. We see that the NPRG results and the Schrödinger wave function results are almost the same in these small interaction regions. These results indicate that an attractive interaction ($C_2, C_4 >0$) causes $\Delta E$ to decrease, and a repulsive interaction ($C_2, C_4 <0$) causes it to increase.
Here we have shown that for multi-particle systems with interactions, the NPRG method can be applied equally without any change of formulation. We are now carrying out calculations to obtain non-trivial relations between particle interactions and tunneling enhancement/suppression. These results will be reported elsewhere.
Summary and Outlook
===================
We have applied the NPRG method to various quantum systems and used it to analyze non-perturbative physics. Even in the first stage of approximation, LPA, we successfully evaluated the non-perturbative quantities that should be given by the summation of all orders of the diverging perturbative series. We also found that for non-perturbative quantities characterized by an essential singularity, the LPA W-H equation again works very well in the region where the instanton-type method breaks down, i.e. the strong coupling region. However, NPRG is not effective in the weak coupling region, due to large numerical errors. In these regions, the approximation used to solve the NPRG equation should be improved in order to obtain correct results. To summarize, the NPRG method and the instanton (or valley) method play complementary roles. Also, from a practical point of view, the NPRG method is a useful new tool for analysis of various quantum systems in a wide parameter region. We have obtained good non-perturbative results for SUSY QM. We also showed that interacting quantum particles can be treated in a similar way.
In the flow diagrams, we observed singular behavior in the small coupling region, and found that it becomes more singular under low-order truncation of the operator expansion. The origin of the difficulty which we encounter in our NPRG analysis resides in the approximation scheme we employed. We must develop ‘better’ approximations, which may depend on the individual systems under study. We also need to study in detail how to extract physical information from the effective potential and the effective action.
First Pole Dominance
====================
Here we confirm the first pole dominance in the two-point function of anharmonic oscillators. In the local potential approximation, the two-point function is given by following $$\begin{aligned}
\sum_{n=1}^{\infty}\frac{D_n}{E^2+(E_n-E_0)^2}
&\stackrel{LPA}{=}&
\frac{1}{E^2+m^2_{\rm eff}}.\nonumber
\end{aligned}$$ This substitutes one pole for an infinite number of poles. Therefore, if the multi-pole contribution becomes significant, the correspondence (\[(23)\]) must be wrong.
We evaluated the first pole coefficient $D_1$ for a single-well (\[(30)\]) and a double-well (\[(32)\]) by solving the Schrödinger equation numerically. The results are displayed in Figures \[fig:D1s\] and \[fig:D1d\].
We should note that the relation $\sum_n D_n=1$ always holds. For the single-well potential, the first pole dominates almost completely. This corresponds to the fact that the results obtained with the LPA W-H equation reproduce the correct results. On the other hand, for the double-well potential, the first pole dominance begins to disappear in the region near $\lambda _0 = 0.1 - 0.15 $, where the results obtained with the LPA W-H equation become poor.
[99]{} K. G. Wilson and J. B. Kogut, Phys. Rep. [**12**]{} (1974), 75. K.-I. Aoki, Prog. Theor. Phys. Suppl. No.131 (1998), 129. T. R. Morris, Prog. Theor. Phys. Suppl. No.131 (1998), 395. D.-U. Jungnickel and C. Wetterich, Prog. Theor. Phys. Suppl. No.131 (1998), 495. , Proceedings of the First Conference on the Exact Renormalization Group (World Scientific, Singapore, 1999). K.-I. Aoki, Int. J. Mod. Phys. B [**14**]{} (2000), 1249. C. Bagnuls and C. Bervillier, Phys. Rep. [**348**]{} (2001), 91. Proceedings of the Second Conference on the Exact Renormalization Group, Int. J. Mod. Phys. A [**16**]{} (2001). J. Berges, N. Tetradis and C. Wetterich, Phys. Rep. [**363**]{} (2002), 223. K.-I. Aoki, A. Horikoshi, M. Taniguchi and H. Terao, in [*The Exact Renormalization Group*]{} (World Scientific, Singapore, 1999), 194 ; hep-th/9812050. ed. J. C. Le Guillou and J. Zinn-Justin, Large-Order Behaviour of Perturbation Theory (North-Holland, 1990). S. Coleman, Aspects of symmetry (Cambridge University Press, 1985). H. Kleinert, Path Integrals in Quantum Mechanics Statistics and Polymer Physics (World Scientific, 1995). F. Wegner and A. Houghton, Phys. Rev. A [**8**]{} (1973), 401. A. Hasenfratz and P. Hasenfratz, Nucl. Phys. B [**270**]{} (1986), 687. M. E. Peskin and D. V. Schroeder, An Introduction to Quantum Field Theory (Addison-Wesley, 1995). K.-I. Aoki, K. Morikawa, W. Souma, J.-I. Sumi and H. Terao, Prog. Theor. Phys. [**99**]{} (1998), 451. C. Wetterich, Phys. Lett. B [**301**]{} (1993), 90. B. Simon, Ann. Inst. Henri Poincare [**38-3**]{} (1983), 295 ; Ann. of Math. [**120**]{} (1984), 89 ; Ann. of Phys. [**158**]{} (1984), 415. P. Gosselin, B. Grosdidier and H. Mohrbach, Phys. Lett. A [**256**]{} (1999), 125. A. S. Kapoyannis and N. Tetradis, Phys. Lett. A [**276**]{} (2000), 225. T. R. Morris, Phys. Lett. B [**329**]{} (1994), 241. K.-I. Aoki and A. Horikoshi, unpublished. D. Zappala, Phys. Lett. A [**290**]{} (2001), 35. M. A. Martin-Delgado, G. Sierra and R. M. Noack, cond-mat/9903100. T. Kashiwa, Phys. Rev. D [**59**]{} (1999), 085002. R. Guida, K. Konishi and H. Suzuki, Ann. of Phys. [**241**]{} (1995), 152 ; Ann. of Phys. [**249**]{} (1996), 109. T. Hatsuda, T. Kunihiro and T. Tanaka, Phys. Rev. Lett. [**78**]{} (1997), 3229. T. Kunihiro Prog. Theor. Phys. Suppl. No.131 (1998), 459 ; Phys. Rev. D [**57**]{} (1998), 2035. H. Aoyama, H. Kikuchi, T. Harano, I. Okouchi, M. Sato and S. Wada, Prog. Theor. Phys. Suppl. No.127 (1997), 1. H. Aoyama, H. Kikuchi, I. Okouchi, M. Sato and S. Wada Phys. Lett. B [**424**]{} (1998), 93 ; Nucl. Phys. B [**553**]{} (1999), 644. H. Aoyama, private communications. E. Witten, Nucl. Phys. B [**188**]{} (1981), 513 ; Nucl. Phys. B [**202**]{} (1982), 253. P. Salomonson and J. W. van Holten, Nucl. Phys. B[**196**]{} (1982), 509.
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
[^3]: E-mail: [email protected]
[^4]: E-mail: [email protected]
[^5]: For other types of NPRG equations, such as the Legendre flow equation, which is derived by means of a cutoff function,[@we2] we cannot evaluate the vacuum bubble diagrams properly without the prescription of subtracting the contribution of the cutoff function from the constant part of $V_{\Lambda}$.
[^6]: For further details with regard to this relation, see Appendix A.
[^7]: It has long been known that the strong coupling expansion has a finite radius of convergence. Recently, variational perturbation theory has become highly developed, and very accurate results have been obtained.[@kl; @gks] The region of coupling constant values in which these approaches are good is estimated as $\lambda_0 {\raisebox{0.2ex}{$\ > \kern -1.05em \raisebox{-1.1ex}{$\sim$}\ $}}0.08$, which is almost coincident with the reliable region for our method. To elucidate the correspondence between the NPRG method and this improved perturbation theory is interesting.
[^8]: As for general non-perturbative methods, the auxiliary field method works very well both in the weak coupling and strong coupling regions.[@taro] Also, various improved perturbation theories have been applied to the anharmonic oscillator and the double well system, giving similar results.[@kl; @gks; @kuni1; @kuni2]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Momentum diffusion is a possible mechanism for driving macroscopic quantum systems towards classical behaviour. Experimental tests of this hypothesis rely on a precise estimation of the strength of this diffusion. We show that quantum-mechanical squeezing offers significant improvements, including when measuring position. For instance, with of mechanical squeezing, experiments would require a tenth of proposed free-fall times. Momentum measurement is better by an additional factor of three, while another quadrature is close to optimal. These have particular implications for the space-based MAQRO proposal—where it could rule out the spontaneous collapse theory due to Ghirardi, Rimini, and Weber—as well as terrestrial optomechanical sensing.'
author:
- Dominic Branford
- 'Christos N. Gagatsos'
- Jai Grover
- 'Alexander J. Hickey'
- Animesh Datta
bibliography:
- 'maqro.bib'
date: 2 September 2019
title: Quantum enhanced estimation of diffusion
---
Introduction
============
Finding a unified description of microscopic and macroscopic systems remains an enduring quest of fundamental physics. One class of proposed solutions are collapse models [@bassi_dynamical_2003; @bassi_models_2013; @bassi_uniqueness_2013; @bassi_gravitational_2017] which span [@pearle_combining_1989; @ghirardi_markov_1990; @toros_colored_2017], Karolyhazy [@karolyhazy_gravitation_1966], Diósi-Penrose [@diosi_universal_1987; @diosi_models_1989; @penrose_gravitys_1996; @bahrami_role_2014], and quantum gravity [@ellis_quantum_1989]; as well as collisional decoherence [@gallis_environmental_1990]. In the non-relativistic regime, they posit spatial decoherence due to diffusion in momentum. The outcome is a description of the evolution in terms of a phase-space density distribution obeying a Fokker-Planck diffusion equation [@ghirardi_unified_1986]. Experimental advances have now made the testing of this proposition a realistic prospect.
Mechanical systems have been used to bound the strength of such diffusive effects. Examples include gravitational-wave detectors [@carlesso_experimental_2016], the LISA pathfinder experiment [@carlesso_experimental_2016; @helou_lisa_2017; @carlesso_non-interferometric_2018], ultracold cantilevers [@vinante_improved_2017], and trapped ions [@li_detecting_2017]. Proposals for future experiments which could probe collapse models and further study macroscopic quantum states include the generation of macroscopic superpositions [@romero-isart_large_2011; @romero-isart_quantum_2011; @scala_matter-wave_2013; @wan_free_2016; @bose_spin_2017; @weaver_phonon_2018] and the space-based MAQRO mission [@kaltenbaek_macroscopic_2012; @kaltenbaek_macroscopic_2016] which formed a key focus of a recent ESA feasibility study [@european_space_agency_cdf_2018].
One simple experiment—which forms a part of the MAQRO mission [@kaltenbaek_macroscopic_2012; @kaltenbaek_macroscopic_2016]—to test collapse models is to let free particles evolve and measure the expanding width of the wavepacket. Once all classical noise sources have been ruled out, any excess wavepacket width must be attributed to momentum diffusion associated with collapse models. MAQRO aims to utilise ultracold nanoparticles and exploit the nano-gravity of space to observe free-fall over —enabling more precise sensing of momentum diffusion—as represented in Fig. \[fig:maqro\_schematic\].
![A pictorial representation of the measurement of wavepacket expansion which forms part of the MAQRO proposals [@kaltenbaek_macroscopic_2012; @kaltenbaek_macroscopic_2016]. (a) A particle is initially trapped, (b) then released, (c) the free particle wavefunction expands, more rapidly with a localisation term, (d) localisation rate can be inferred through position measurements. Expansion as depicted in two spatial detections is for illustrative purposes, we only analyse one independent spatial dimension.[]{data-label="fig:maqro_schematic"}](schematic_maqro)
Quantum techniques such as squeezing allow for more precise estimation [@toth_quantum_2014; @demkowicz-dobrzanski_quantum_2015]. Optical squeezing has been identified as valuable to fundamental physics, with squeezing-enhanced interferometry [@caves_quantum-mechanical_1981] set to enhance laser-interferometric gravitational-wave detectors [@the_ligo_scientific_collaboration_gravitational_2011; @grote_first_2013; @the_ligo_scientific_collaboration_enhanced_2013] and squeezing of optical vacuum reported [@vahlbruch_detection_2016]. It has also found application in photonic-force microscopy [@taylor_fundamental_2013; @taylor_subdiffraction-limited_2014], while microwave squeezing is being used in the search for axion dark matter [@malnou_squeezed_2019].
In this article, we show that quantum squeezing of the mechanical degree of freedom enables a more precise estimation of the strength of momentum diffusion. This enhancement is attainable with the currently proposed scheme of measuring the position of a particle. Quantum squeezing of mechanical degrees of freedom is beginning to be explored in thermal states [@pontin_squeezing_2014; @rashid_experimental_2016]. We conclude that squeezing can be used to achieve the same precision with reduced free-fall time or centre of mass cooling. This reduction could be ten-fold for a squeezing of . Thus, squeezing can compensate for reduced free-fall times, identified as one of the challenges for MAQRO [@kaltenbaek_macroscopic_2012; @kaltenbaek_macroscopic_2016] in a recent ESA CDF study [@european_space_agency_cdf_2018]. We further show that a momentum measurement is thrice as precise as that of position, while measurement of a more general quadrature is close to optimal. We briefly discuss the potential of the heterodyne and phonon counting measurements.
While our results will be presented in the context of collapse models, observing similar momentum diffusion processes could aid detection of certain dark-matter candidates [@riedel_direct_2013; @riedel_decoherence_2015; @riedel_decoherence_2017]. Since excess heating of wavepackets is also a consequence of momentum diffusion [@collett_wavefunction_2003; @diosi_testing_2015; @li_detecting_2017], our results imply a quantum enhanced estimation of heating. Finally, the ubiquitous phenomena of Brownian motion is also caused by diffusion. Our results can thus be applied in this very general scenario, as well as in particle tracking used to study biological systems [@ghislain_scanning-force_1993; @pralle_local_1998].
Before presenting our results, we note some recent works that have theoretically considered continuously monitoring a thermal state [@genoni_unravelling_2016] or squeezing a specific optomechanical coupling [@mcmillen_quantum-limited_2017], with the latter providing no attainable advantage from squeezing when measuring the optical subsystem. Previous works in quantum metrology have analysed quantum-limited estimation of related noise parameters including loss [@monras_optimal_2007; @adesso_optimal_2009], diffusion in phase shifts [@knysh_estimation_2013; @vidrighin_joint_2014] and displacements [@tsang_quantum_2019], and classical stochastic processes [@ng_spectrum_2016].
Background
==========
A particle of mass $ m $ in a harmonic potential has Hamiltonian $
{{\MakeUppercase{\mathcal{{\hat{H}}}}}}
=
{{\MakeUppercase{\mathcal{{\hat{p}}}}}}^2/2m
+
m \omega^2 {{\MakeUppercase{\mathcal{{\hat{x}}}}}}^2/2.
$ Dimensionless position and momentum operators are $ {\hat{x}} = \sqrt{m\omega}{{\MakeUppercase{\mathcal{{\hat{x}}}}}}/\sqrt{\hbar} $ and $ {\hat{p}} = {{\MakeUppercase{\mathcal{{\hat{p}}}}}} / \sqrt{\hbar m\omega} $ whose commutators are given by the matrix $ i\Omega $ where $$\Omega = -i \begin{pmatrix} {[{\hat{x}},{\hat{x}}]} & {[{\hat{x}},{\hat{p}}]} \\ {[{\hat{p}},{\hat{x}}]} & {[{\hat{p}},{\hat{p}}]} \end{pmatrix} = \begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix}.
\label{eq:commutator}$$
Quantum states of such a particle have a phase-space representation in terms of the Wigner function of an operator defined as [@ferraro_gaussian_2005 Chap. 1] $$W_{\rho}(x,p) = \frac{2}{\pi} \Trace{\rho {\hat{D}}\left(\frac{x+ip}{\sqrt{2}}\right) {\hat{\Pi}} {\hat{D}}^{\dagger}\left(\frac{x+ip}{\sqrt{2}}\right) },
\label{eq:wigner}$$ where $ {\hat{D}}(\alpha) = e^{\alpha{\hat{a}}^{\dagger}-\alpha^*{\hat{a}}} $ and $ {\hat{\Pi}} = e^{i\pi {\hat{a}}^{\dagger}{\hat{a}}} $. Gaussian states are those whose Wigner function is Gaussian and so determined by the averages—displacement vector $ \vec{d} $—and covariances—covariance matrix $ \sigma $—of the position and momentum operators. Examples include thermal, coherent, and squeezed states. A thermal state has covariance matrix $ \sigma = {\kappa_{\text{th}}}{\mathbb{1}}$, with $ \sigma = {\mathbb{1}}$ corresponding to the ground state.
We focus on the simplest setup to study momentum diffusion, that of a free particle as in Fig. \[fig:maqro\_schematic\]. Initially the particle is trapped in a harmonic potential with frequency $ \omega $ and cooled. Cooling of nano-particles has been reported to the order of phonons [@windey_cavity-based_2019; @delic_cavity_2019] with theory anticipating cooling much closer to the ground state [@romero-isart_quantum_2012; @gonzalez-ballestero_theory_2019]. After cooling the trapping potential is turned off. The particle then evolves freely under the Hamiltonian $ {{\MakeUppercase{\mathcal{{\hat{H}}}}}} = {{\MakeUppercase{\mathcal{{\hat{P}}}}}}^2/2m $ with Lindblad term $ \Lambda {[{{\MakeUppercase{\mathcal{{\hat{X}}}}}},{[{{\MakeUppercase{\mathcal{{\hat{X}}}}}},\rho]}]} $, whose strength $ \Lambda $ is our parameter of interest. The master equation for momentum diffusion for this system—in terms of the dimensionless position and momentum operators—is $$\frac{\partial \rho}{\partial {\tau}} = - \frac{i}{2} {[{\hat{p}}^2,\rho]} - \frac{1}{4} {\lambda}{[{\hat{x}},{[{\hat{x}},\rho]}]},
\label{eq:unitless_master}$$ where $ {\tau}= \omega t $ and $ {\lambda}= \Lambda / {\Lambda_{0}}$ are dimensionless parameters, and $ {\Lambda_{0}}= m\omega^2/(4\hbar) $. Being quadratic the master equation Eq. evolves Gaussian states to Gaussian states [@carmichael_statistical_1999; @nicacio_phase_2010; @serafini_quantum_2017].
Eq. can then be transformed to a Fokker-Planck equation [@barnett_methods_2002; @nicacio_phase_2010], in this case yielding $$\frac{\partial}{\partial {\tau}} W(x,p,{\tau}) = \left[ - p \frac{\partial}{\partial x} + \frac{1}{4} {\lambda}\frac{\partial^2}{\partial p^2} \right] W(x,p,{\tau}),
\label{eq:fokker}$$ which for Gaussian $ W $ can be mapped to the equations of motion of form [@carmichael_statistical_1999; @serafini_quantum_2017] $$\begin{aligned}
\frac{\partial\vec{\mu}}{\partial \tau} &= A \vec{\mu}, &
\frac{\partial\sigma}{\partial \tau} &= A \sigma + \sigma A^T + D,
\label{eq:eom_gaussian}
\end{aligned}$$ where $ \vec{\mu} $ and $ \sigma $ are the Gaussian’s moments. For an initial Gaussian state with moments $ \vec{d} $ and $ \sigma $ the evolved moments under Eq. become $$\begin{aligned}
\vec{d}({\tau}) &= \begin{pmatrix} 1 & {\tau}\\ 0 & 1 \end{pmatrix} \vec{d},
\label{eq:disp}\\
\sigma({\tau}) &= \begin{pmatrix} 1 & {\tau}\\ 0 & 1 \end{pmatrix} \sigma \begin{pmatrix} 1 & 0 \\ {\tau}& 1 \end{pmatrix} + {\lambda}\begin{pmatrix} {\tau}^3/3 & {\tau}^2/2 \\ {\tau}^2/2 & {\tau}\end{pmatrix}.
\label{eq:cov}\end{aligned}$$
Our results apply to estimation of diffusion in any scenario governed by Eq. for all values of ${\lambda}$ and $ {\tau}$. To estimate the strength of the momentum diffusion $ \Lambda $, we begin with a single-mode Gaussian state. Such a state can be described as a thermal state $ {\kappa_{\text{th}}}{\mathbb{1}}$ with a squeezing $ r \geq 0 $ of the quadrature $ {\hat{x}}\sin\phi+{\hat{p}}\cos\phi $ giving an initial covariance matrix $$\sigma = {\kappa_{\text{th}}}\begin{pmatrix}
\cosh 2r + \sinh 2r \cos 2\phi & \sinh 2r \sin 2\phi \\
\sinh 2r \sin 2\phi & \cosh 2r - \sinh 2r \cos 2\phi
\end{pmatrix},
\label{eq:covariance_initial}$$ with arbitrary displacements. The displacements do not begin with any parameter-dependence and do not gain any through the evolution given by Eq. and so their derivative with respect to the parameter satisfies $ \partial_\Lambda \vec{d} = 0 $. We will consider tuning $ \phi $ to maximise the precision for given thermal variance and squeezing magnitudes, with $ \phi=0 $ and $ \phi=\pi/2 $ corresponding to momentum and position squeezing respectively.
We will highlight special cases for $ {\lambda}\ll 1 $ and $ {\tau}\gg 1 $, which is the regime for MAQRO [@kaltenbaek_macroscopic_2012; @kaltenbaek_macroscopic_2016] as in Table \[tab:maqro\_values\]; and $ {\kappa_{\text{th}}}\sim 1 $ which is around the MAQRO regime.
-------------------------------------------------------------------- --
Localisation rate $ \Lambda $
Free-fall time $ t $
Mechanical frequency $ \omega $
Mass $ m $
Thermal occupation number $ n_{\text{th}} $
Thermal variance ($ {\kappa_{\text{th}}}= 2n_{\text{th}}+1 $)
Limiting localisation($ {\Lambda_{0}}= \frac{m\omega^2}{4\hbar} $)
Experiment timescale ($ {\tau}= \omega t $)
-------------------------------------------------------------------- --
: Parameter values based on @kaltenbaek_macroscopic_2016, primarily Table 1 therein.[]{data-label="tab:maqro_values"}
An estimator is required to estimate an unknown parameter from observed data. If limited to statistical noise the precision of the value produced by the estimator can be taken from the variance of that estimator. The lower bounds the variance of an unbiased estimator as [@kay_fundamentals_1998; @helstrom_quantum_1976; @holevo_probabilistic_2011; @paris_quantum_2009] $$(\Delta {\widetilde{\Lambda}})^2 \geq \frac{1}{\nu F(\Lambda)} \geq \frac{1}{\nu H(\Lambda)},
\label{eq:CRBs}$$ where $ \nu $ is the number of repetitions of an experiment, $ \widetilde{\Lambda} $ is an estimator of the parameter $ \Lambda $, and $ F(\Lambda) $ and $ H(\Lambda) $ are respectively the and . The is a function of the probability distribution [@kay_fundamentals_1998] $$F(\Lambda) = \int\! \mathrm{d}\mkern-1mu\vec{x} \, \frac{1}{P(\Pi_{\vec{x}}|\rho_{\Lambda})} \left( \frac{\partial P(\Pi_{\vec{x}}|\rho_{\Lambda})}{\partial \Lambda} \right)^2,
\label{eq:cfi}$$ where the probabilities $ P(\Pi_{\vec{x}}|\rho_{\Lambda}) $ are derived from applying the $ \boldsymbol{\Pi} $ to the state $ \rho_{\Lambda} $. The is a function of the state alone [@paris_quantum_2009; @toth_quantum_2014; @demkowicz-dobrzanski_quantum_2015] $$H(\Lambda) = \Trace{\rho_{\Lambda} L_{\Lambda}^2},
\label{eq:qfi}$$ where $ L_{\Lambda} $ is the defined by $ L_{\Lambda} \rho_{\Lambda} + \rho_{\Lambda} L_{\Lambda} = 2\partial_{\Lambda} \rho $.
These and provide the and , the first and second inequalities of Eq. respectively. The equalities in Eq. are obtained by an optimal measurement, where it exists, and an efficient estimator; we identify such a measurement and the maximum likelihood estimator is asymptotically efficient [@kay_fundamentals_1998].
For a Gaussian state (where $ \partial_{\Lambda} \vec{d} = 0 $) the can be evaluated explicitly as [@monras_phase_2013; @safranek_quantum_2015] $$H(\Lambda) = \frac{1}{2} (\partial_{\Lambda}\sigma|(\sigma\otimes\sigma-\Omega\otimes\Omega)^{-1}|\partial_{\Lambda}\sigma) ,
\label{eq:gaussian_qfi}$$ where the inner product is $ (A|B) = \Trace{A^T B} $.
Results
=======
Using Eqns. and , the can be calculated through Eq. to be $$(\Delta{\widetilde{\Lambda}})^2 \geq
\frac{
{\Lambda_{0}}^2\left[
\left({\kappa_{\text{th}}}^2 + {\tau}{\kappa_{\text{th}}}{\lambda}Z
+ \frac{{\tau}^4}{12} {\lambda}^2 \right)^2
- 1
\right]
}{
\frac{{\tau}^4}{12}\left( 1 - {\kappa_{\text{th}}}^2
+ {\tau}{\kappa_{\text{th}}}{\lambda}Z
+ \frac{{\tau}^4}{12} {\lambda}^2 \right)
+ \frac{{\tau}^2}{2}{\kappa_{\text{th}}}^2 Z^2
},
\label{eq:qcrb_squeeze}$$ where $
Z = \left(1+{\tau}^2/3\right)\cosh 2r + \left[ \left(1-{\tau}^2/3\right) \cos 2\phi + {\tau}\sin 2\phi \right]\sinh 2r.
$ The bound in Eq. behaves as $ (\Delta{\widetilde{\Lambda}})^2 \gtrsim \Lambda^2 $ to leading order in $ \Lambda $.
The in Eq. is minimised by squeezing or anti-squeezing (squeezing the orthogonal quadrature) with squeezing angle (See App. \[app:optimal\_squeezing\]) $$\phi =\arctan \left( \frac{ -3+{\tau}^2 - \sqrt{9+3{\tau}^2+{\tau}^4} }{ 3{\tau}} \right),
\label{eq:qcrb_optimal_squeeze}$$ which tends to $ 0 $ for $ {\tau}\gg 1 $, corresponding to squeezing of position or momentum. When squeezing at this angle in the regime of $ {\tau}\gg 1 $, with $ {\kappa_{\text{th}}}= 1 $, the simplifies to $$(\Delta{\widetilde{\Lambda}})^2 \gtrsim
{\Lambda_{0}}^2
\frac{8{\lambda}\left( e^{-2r} + \frac{{\tau}}{4}{\lambda}\right) \left( 1 + \frac{{\tau}^3}{6} e^{-2r} {\lambda}+ \frac{{\tau}^4}{24} {\lambda}^2 \right)}{
\frac{2{\tau}^3}{3} e^{-4r} + \frac{{\tau}^4}{3}e^{-2r}{\lambda}+ \frac{{\tau}^5}{12}{\lambda}^2
},$$ with the squeezing $ r $ not necessarily positive as anti-squeezing may be preferrable (see App. \[app:optimal\_squeezing\]).
Measurement of the particle’s position is a special case of homodyne detection which involves measuring a linear combination of the position and momentum quadratures [@adesso_continuous_2014; @serafini_quantum_2017]. Heterodyne allows for the simultaneous measurement of position and momentum, but with added noise [@shapiro_phase_1984; @leonhardt_measuring_1995]. The can be reached through projection onto eigenstates of the [@braunstein_statistical_1994] which, for a Gaussian system, entails performing some squeezing and displacement followed by measurement of Fock states [@monras_optimal_2007; @monras_phase_2013; @serafini_quantum_2017]. This additional squeezing is a resource applied to the system after the evolution as part of the measurement and does not improve the precision as an initial squeezing can. Further, in a mechanical system this involves measuring the number of phonons which remains experimentally demanding [@cohen_phonon_2015; @hong_hanbury_2017]. In the following, we calculate the performance of all these measurements for estimating $ \Lambda $.
Homodyne detection at an angle $ \theta $ measures the quadrature $
{\hat{q}}_{\theta} = {\hat{x}}\cos\theta+{\hat{p}}\sin\theta
$. When performed on a Gaussian state the homodyne statistics are Gaussian [@adesso_continuous_2014] and the moments are the appropriate marginal of the Wigner function. For a homodyne angle $ \theta $ the variance of the marginal is $$\begin{aligned}
\Sigma &=
{\kappa_{\text{th}}}\bigg[
\left[ (1 + {\tau}^2)\cos^2 \theta + {\tau}\sin 2\theta + \sin^2 \theta \right]\cosh 2r \\
&\mkern56mu+ \big\{
\left[ (1-{\tau}^2)\cos^2 \theta - {\tau}\sin 2\theta - \sin^2 \theta \right]\cos 2\phi \\
&\mkern112mu+ \left[ 2{\tau}\cos^2 \theta + \sin 2\theta \right]\sin 2\phi
\big\}\sinh 2r \\
&\mkern56mu+ {\lambda}\left( \frac{{\tau}^3}{3}\cos^2 \theta + \frac{{\tau}^2}{2}\sin 2\theta + {\tau}\sin^2 \theta \right)
\bigg],
\end{aligned}
\label{eq:homodyne_moments}$$ as the Wigner function’s mean is parameter-independent so is the marginal’s. The choices $ \theta = 0 $ and $ \theta = \pi/2 $ correspond to measurement of position and momentum respectively. We will consider the optimisation of $ \theta $, which more generally requires measuring a linear combination of the position and momentum operators.
For a Gaussian probability distribution with a parameter-independent mean, the is [@kay_fundamentals_1998 Chap. 3] $$F(\Lambda) = \frac{1}{2} \Trace{\Sigma^{-1} \partial_{\Lambda}\Sigma \Sigma^{-1} \partial_{\Lambda}\Sigma},
\label{eq:gaussian_cfi}$$ where $ \Sigma $ is the variance of the Gaussian distribution. Using Eqns. and , the for homodyne along an angle $ \theta $ is
$$\begin{aligned}
(\Delta{\widetilde{\Lambda}})^2 \geq
2{\Lambda_{0}}^2 \Bigg[
{\lambda}+ {\kappa_{\text{th}}}\Bigg(&
\frac{ {\tau}^2 \cos^2\theta + {\tau}\sin 2\theta + 1
}{
\frac{{\tau}^3}{3}\cos^2\theta + \frac{{\tau}^2}{2} \sin 2\theta + {\tau}\sin^2\theta
} \cosh 2r \\
&\mkern64mu-
\frac{
\left({\tau}^2 \cos^2\theta + {\tau}\sin 2\theta - \cos2\theta \right)\cos 2\phi
-
\left( 2{\tau}\cos^2 \theta + \sin 2\theta \right)\sin 2\phi
}{
\frac{{\tau}^3}{3}\cos^2\theta + \frac{{\tau}^2}{2} \sin 2\theta + {\tau}\sin^2\theta
} \sinh 2r
\Bigg)
\Bigg]^2.
\end{aligned}
\label{eq:crb_homodyne_squeezed}$$
To leading order in $ \Lambda $ this is $ (\Delta{\widetilde{\Lambda}})^2 \gtrsim 2\Lambda^2 $ which occurs when the first term in the square dominates, whereas when that can be neglected the bound is a $ \Lambda $-independent constant. The bound on estimating the diffusion $ \Lambda $ from position ($ \theta = 0 $) measurement is $$(\Delta{\widetilde{\Lambda}})^2 \geq
2{\Lambda_{0}}^2
\left[ {\lambda}+ {\kappa_{\text{th}}}\left(
\frac{
\left[ 1 + {\tau}^2 \right]\cosh 2r + \left\{ \left[ 1 - {\tau}^2 \right]\cos 2\phi + 2{\tau}\sin 2\phi \right\}\sinh 2r
}{{\tau}^3 / 3}
\right) \right]^2,
\label{eq:crb_position_squeezed}$$
which behaves as $$(\Delta{\widetilde{\Lambda}})^2 \gtrsim 2{\Lambda_{0}}^2 \left[ {\lambda}+ {\kappa_{\text{th}}}\frac{\cosh 2r - \sinh 2r \cos 2\phi}{{\tau}/3} \right]^2,
\label{eq:crb_position_omegat}$$ for $ {\tau}\gg 1 $. Instead for measuring the momentum ($ \theta = \pi/2 $) the bound on estimating the diffusion $ \Lambda $ is $$(\Delta{\widetilde{\Lambda}})^2 \geq
2{\Lambda_{0}}^2
\left[ {\lambda}+ {\kappa_{\text{th}}}\frac{\cosh 2r - \sinh 2r\cos 2\phi}{{\tau}}
\right]^2,
\label{eq:crb_momentum_squeezed}$$ which (neglecting squeezing) matches the large $ {\tau}$ limit of position measurements when $ {\lambda}\gg {\kappa_{\text{th}}}/ {\tau}$ and is a factor of 9 better when $ {\lambda}\ll {\kappa_{\text{th}}}/ {\tau}$.
The optimal input squeezing angle $ \phi $ can in general be found by minimising the coefficient of $ \sinh 2r $ in Eq. which gives $$\phi = - \arctan \left( \frac{1}{{\tau}+ \tan\theta} \right).
\label{eq:optimal_hom_squeezing_angle}$$ For momentum measurements ($ \theta = \pi/2 $) this squeezing angle is $ \phi = 0 $ (squeezing of momentum). While for position measurements ($ \theta = 0 $) this is $ \phi = - \arctan(1/{\tau}) $ tending to $ \phi = -\pi/2 $ for $ {\tau}\ll 1 $, and $ \phi = 0 $ for $ {\tau}\gg 1 $.
In general the squeezing angle in Eq. produces a precision $$(\Delta{\widetilde{\Lambda}})^2 \geq
2{\Lambda_{0}}^2 \left[
{\lambda}+ {\kappa_{\text{th}}}e^{-2r} \chi({\tau},\theta)
\right]^2,
\label{eq:crb_homodyne_optimally_squeezed}$$ from which the unsqueezed case ($ r = 0 $) can also be extracted, where $$\chi({\tau},\theta) =
\frac{ {\tau}^2 \cos^2\theta + {\tau}\sin 2\theta + 1}{ \frac{{\tau}^3}{3}\cos^2\theta + \frac{{\tau}^2}{2} \sin 2\theta + {\tau}\sin^2\theta }.$$ One effect of squeezing is equivalent to an effective reduction of $ {\kappa_{\text{th}}}$ by $ e^{-2r} $, unlike reducing the centre-of-mass motion which reaches $ {\kappa_{\text{th}}}= 1 $ at absolute zero this squeezing allows an unlimited reduction in the second term. For $ {\tau}\gg 1 $ (as $ \chi \sim 1/{\tau}$) the same squeezing could instead be considered as an effective increase in $ \tau $ by a factor of $ e^{2r} $ to obtain the same precision from a much shorter free-fall time.
When the quadrature given by Eq. is squeezed the homodyne angle which minimises the bound in Eq. is $$\theta = - \arctan \left( \frac{3+2{\tau}^2+\sqrt{9+3{\tau}^2+{\tau}^4}}{3{\tau}} \right),
\label{eq:optimal_homodyne}$$ which tends to $ \theta \approx -\pi / 2 + 1 / {\tau}$ for $ {\tau}\gg 1 $. Measuring the quadrature given by Eq. with squeezing as Eq. gives a precision $$(\Delta{\widetilde{\Lambda}})^2 \geq
2{\Lambda_{0}}^2 \left[
{\lambda}+ {\kappa_{\text{th}}}e^{-2r}
\frac{3+{\tau}^2-\sqrt{9+3{\tau}^2+{\tau}^4}}{{\tau}^3/2}
\right]^2.
\label{eq:crb_optimal_homodyne_optimally_squeezed}$$
Measuring the quadrature of Eq. does not in general attain the . When $ {\lambda}$ dominates, the behaves as $ \Lambda^2 $ while any homodyne terms tend to $ 2\Lambda^2 $. In the $ \tau \gg 1 $ regime, one could improve on the precision by no more than a factor of 2 using heterodyne detection (see App. \[app:heterodyne\]). Fig. \[fig:homodyne\_heterodyne\_optimality\] suggests that heterodyne otherwise shows little promise.
Phonon counting—in combination with displacement and squeezing operations—can in principle attain the for all $ {\lambda}$ and $ {\tau}$ as the is a quadratic operator in the quadrature operators [@monras_phase_2013; @serafini_quantum_2017] and so has eigenstates which are squeezed-displaced Fock states. The additional squeezing required to attain the is derived in full generality in Appendix \[app:sld\_projectors\]. For MAQRO, this squeezing seems nugatory, with $ \SI{79}{\dB} $ required to attain the for $ \Lambda = \SI{1e20}{\m^{-2}\s^{-1}} $ which would only improve precision by a factor of $ \sqrt{2} $, to $ \SI{158}{\dB} $ for $ \Lambda=\SI{1e10}{\m^{-2}\s^{-1}} $, where the improvement on position measurements would be more pronounced. In other scenarios, however, this could be worthwhile. For $ {\tau}\ll 1 $ and $ {\lambda}{\tau}^2 \lesssim 1 $ the squeezing needed is only $ e^{2z} \approx 1 + {\tau}\approx 1 $, while for $ {\tau}\gg 1 $ and $ {\lambda}{\tau}\gtrsim 1 $ this goes to $ e^{2z} \approx 2{\tau}/\sqrt{3} $.
Discussion
==========
![Precision of estimating momentum diffusion from wavepacket expansion for MAQRO parameters (Tab. \[tab:maqro\_values\]). Dashed lines denote a squeezing of . The optimal homodyne and fundamental limit lines overlap until around $ \Lambda \sim \SI{1e20}{\m^{-2}\s^{-1}} $. Three years data collection with $ t = \SI{100}{\s} $ yields $ \nu \sim \num{1e6} $ repetitions. []{data-label="fig:maqro_precision"}](maqro_plot)
Fig. \[fig:maqro\_precision\] shows the potential improvement in precision for estimating diffusion via momentum or homodyne measurements, or through squeezing, for MAQRO parameters as given in Tab. \[tab:maqro\_values\]. For reference, position measurement is the present proposal. We propose squeezing of the momentum quadrature which offers a substantial improvement across much of the pertinent $ \Lambda $ range for both measurement of position and momentum, with $ \SI{10}{\dB} $ enabling an order of magnitude higher resolution of $ \Lambda $. Measuring the quadrature described by Eq. allows further improvement keeping within a factor of two of the across the whole regime.
Our bounds can be mapped to the wealth of diffusive processes whose parameters enter into the observed diffusion rate $ \Lambda $. In the case of (mass-proportional) the two parameters of interest are $ {\lambda^{\mathrm{CSL}}}$ and $ {r_\mathrm{C}}$—the time and length scales in the model. The observed diffusion rate $ \Lambda $ for a free sphere of mass $ m $ and radius $ {r_\mathrm{s}}$ is—as a function of $ {\lambda^{\mathrm{CSL}}}$ and $ {r_\mathrm{C}}$—given by [@collett_wavefunction_2003; @kaltenbaek_macroscopic_2016] $$\Lambda = \frac{{\lambda^{\mathrm{CSL}}}}{4{r_\mathrm{C}}^2} \left( \frac{m}{m_0} \right)^2 f\!\left( \frac{{r_\mathrm{s}}}{{r_\mathrm{C}}} \right),$$ where $ m_0 $ is a reference (nucleon) mass and $
f(x) =
\frac{6}{x^4} \left[ 1 - \frac{2}{x^2} + \left( 1 + \frac{2}{x^2} \right) e^{-x^2} \right] .
$ From this bounds on $ {\lambda^{\mathrm{CSL}}}$ as a function of $ {r_\mathrm{C}}$ can be calculated using $$\Delta {\lambda^{\mathrm{CSL}}}= 4 {r_\mathrm{C}}^2 \left[ \left( \frac{m}{m_0} \right)^2 f\left( \frac{{r_\mathrm{s}}}{{r_\mathrm{C}}} \right) \right]^{-1} \Delta \Lambda.$$
To describe the minimal discernable $ {\lambda^{\mathrm{CSL}}}$ for measurement of a mechanical quadrature we take the limit of the single-shot $ {\lambda^{\mathrm{CSL}}}_0 = \lim\limits_{{\lambda^{\mathrm{CSL}}}\to 0} \Delta {\lambda^{\mathrm{CSL}}}$. Allowing for $ \nu $ independent repetitions the uncertainty can be reduced to $ \Delta {\lambda^{\mathrm{CSL}}}\approx \sqrt{\frac{1}{\nu}} \left( \sqrt{\frac{1}{\nu}} + 1 \right) {\lambda^{\mathrm{CSL}}}_0 $ at $ {\lambda^{\mathrm{CSL}}}\approx {\lambda^{\mathrm{CSL}}}_0 / \sqrt{\nu} $. To ensure any deviation can be recognised with statistical significance we take the minimum detectable collapse rate $ {\lambda^{\mathrm{CSL}}}_{\mathrm{min}} $ to be $ {\lambda^{\mathrm{CSL}}}_{\mathrm{min}} \sim \frac{2}{\sqrt{\nu}} {\lambda^{\mathrm{CSL}}}_0 $. Thus, for a quadrature measurement the minimum resolvable $ {\lambda^{\mathrm{CSL}}}$ we take to be given by $ {\lambda^{\mathrm{CSL}}}_{\mathrm{min}} = \frac{2}{\sqrt{\nu}} \lim\limits_{{\lambda^{\mathrm{CSL}}}\to 0} \Delta{\lambda^{\mathrm{CSL}}}$ in Eq. .
![ Minimum detectable collapse rate for three years of observation with a radius sphere of mass , with other parameters as Tab. \[tab:maqro\_values\]. The minimum required collapse rate given is based on the criteria of Ref. [@toros_colored_2017] to ensure macroscopic objects rapidly collapse to classical states. The magenta dot represents the values originally proposed by @ghirardi_unified_1986. []{data-label="fig:maqro_csl"}](maqro_csl)
For MAQRO such bounds can be seen in Fig. \[fig:maqro\_csl\] for the position, momentum, and optimal quadratures. For position or momentum measurements with up to squeezing the bounds are competive across , below X-ray emission data begins to provide a tighter bound [@piscicchia_csl_2017] while above LISA Pathfinder data is tighter [@carlesso_experimental_2016; @carlesso_non-interferometric_2018]. Additional squeezing can of course further reduce the undertainty, with of squeezing sufficient to match the theoretical minimum collapse rate to above . This would include testing the original parameters suggested by @ghirardi_unified_1986.
The optimal quadrature identified in Eq. meanwhile could yield a conclusive test of the conventional model at a precision of six orders of magnitude more than the theoretical lower bound on [@toros_colored_2017]. Attaining the can offer further improvements, however this would be of little value to MAQRO if the optimal homodyne sensitivity can be reached.
In conclusion, we have shown that squeezing could be used to compensate for reduced free-fall times, an aspect which a recent ESA CDF study [@european_space_agency_cdf_2018] has identified as one of the more demanding of the original proposals [@kaltenbaek_macroscopic_2012; @kaltenbaek_macroscopic_2016]. As—for both Eq. and Eq. —the precision is constant for $ e^{2r}{\tau}$ being constant, longer effective free-fall times can be generated through mechanical squeezing. We have also shown the efficacy of momentum and general quadrature measurements over the proposed position measurement.
We thank Rainer Kaltenbaek, Hendrik Ulbricht, Matteo Carlesso, and Francesco Albarelli for illuminating discussions. This study has been supported by the European Space Agency’s Ariadna scheme (Study Ref. 17-1201a), the UK EPSRC (EP/K04057X/2), and the UK National Quantum Technologies Programme (EP/M01326X/1, EP/M013243/1). D.B. has received support for travel and attendance at workshops from QTSpace (COST Action CA15220).
Appendix {#appendix .unnumbered}
========
Appendix \[app:optimal\_squeezing\] calculates the necessary squeezing angle to maximise precision for the fundamental limit and quadrature measurements. Appendix \[app:heterodyne\] calculates the of heterodyne measurements. Appendix \[app:sld\_projectors\] derives the necessary squeezing required to then project onto the eigenstates of the by phonon counting. Appendix \[app:relative\_precisions\] compares performance of the fundamental limit, optimal quadrature, and heterodyne measurements. Appendix \[app:csl\] translates the bounds on the observed diffusion rate $ \Lambda $ to the parameters of .
Optimal squeezing {#app:optimal_squeezing}
=================
Fundamental limit {#app:qcrb_squeezing}
-----------------
The is $$(\Delta\Lambda)^2 \geq
B =
{\Lambda_{0}}^2
\frac{
\left({\kappa_{\text{th}}}^2 + {\tau}{\kappa_{\text{th}}}{\lambda}Z + \frac{{\tau}^4}{12} {\lambda}^2 \right)^2
- 1
}{
\frac{{\tau}^4}{12}\left( {\kappa_{\text{th}}}^2
+ {\tau}{\kappa_{\text{th}}}{\lambda}Z
+ \frac{{\tau}^4}{12} {\lambda}^2 \right)
+\frac{{\tau}^4}{12}\left( 1-2{\kappa_{\text{th}}}^2 \right) + \frac{1}{2}{\kappa_{\text{th}}}^2 {\tau}^2 Z^2
},
\label{eq:app:qcrb_squeeze}$$ where $$Z = \left(1+\frac{{\tau}^2}{3}\right)\cosh 2r + \left[ \left(1-\frac{{\tau}^2}{3}\right) \cos 2\phi + {\tau}\sin 2\phi \right]\sinh 2r.$$ Minima with respect to the squeezing angle of the bound in Eq. are either solutions of $ \frac{\partial B}{\partial Z} = 0 $ or $ \frac{\partial Z}{\partial \phi} = 0 $ as $$\frac{\partial B}{\partial \phi} = \frac{\partial B}{\partial Z} \frac{\partial Z}{\partial \phi},$$ and the second derivative $$\frac{\partial^2 B}{\partial \phi^2} = \frac{\partial^2 B}{\partial Z^2} \left(\frac{\partial Z}{\partial \phi}\right)^2
+ \frac{\partial B}{\partial Z} \frac{\partial^2 Z}{\partial \phi^2},$$ distinguishes minima and maxima. The stationary points of $ B(Z) $ are $$Z_{\pm} = \frac{
144(1-{\kappa_{\text{th}}}^4) + 24{\lambda}^2{\tau}^4(1-2{\kappa_{\text{th}}}^2) + {\lambda}^4{\tau}^8 \pm \left|12(1-{\kappa_{\text{th}}}^2)+{\lambda}^2{\tau}^4\right| \sqrt{[12(1+{\kappa_{\text{th}}}^2)+{\lambda}^2{\tau}^4]^2-48{\lambda}^2{\kappa_{\text{th}}}^2{\tau}^4}
}{
288 {\lambda}{\kappa_{\text{th}}}^3 \tau
},$$ where the negative root is not possible with $ r > 0 $ and for the positive root $
\frac{\partial^2 B}{\partial Z^2} < 0,
$ means that the minimum of $ B $ is found for $ \frac{\partial Z}{\partial \phi} = 0 $. The stationary points of $ Z(\phi) $ are $$\phi_{\pm} = \arctan \left( \frac{-3+{\tau}^2 \pm \sqrt{9+3{\tau}^2+{\tau}^4} }{ 3{\tau}} \right),$$ where we have $$(\phi_+ - \phi_-) \mod \pi = \frac{\pi}{2}$$ as $ \tan (\phi_+) \tan (\phi_-) = -1 $. Hence we recognise that squeezing the quadrature $ {\hat{x}}_{\phi_+} $ is equivalent to anti-squeezing of the orthogonal quadrature $ {\hat{x}}_{\phi_-} = {\hat{x}}_{\phi_+ + \frac{\pi}{2}} $. This follows as $ r > 0 $ and $ \phi \in [0,\pi] $ and $ r \in \mathbb{R} $ and $ \phi \in [0,\pi/2] $ are equivalent parameterisations of the same squeezings—squeezing a quadrature $ {\hat{x}}_{\phi} $ is equivalent to anti-squeezing the quadrature $ {\hat{x}}_{\phi+\pi/2} $.
As $ B(Z_+) $ is a maximum and $ Z_- < 0 $ is outside the range of $ Z(\phi) $ at least one of $ \phi_{\pm} $ is a minimum of $ B(\phi) $. We therefore find the global minimum of $ B(\phi) $ by finding the smaller of $ B(\phi_+) $ and $ B(\phi_-) $. For $ Z(\phi_{\pm}) $ $$Z(\phi_{\pm}) =
\left[ \left( 1+\frac{{\tau}^2}{3} \right) \cosh 2r \pm \frac{\sqrt{9+3{\tau}^2+{\tau}^4}}{3} \sinh 2r \right],$$ where we note that exchanging $ \phi_+ \to \phi_- $ is equivalent to $ r \to -r $.
For these squeezing angles ($ \phi_{\pm} $) the bound (Eq. ) is $$\begin{aligned}
(\Delta\Lambda)^2 \geq
{\Lambda_{0}}^2 &
\left(
\left\{{\kappa_{\text{th}}}^2 + {\tau}{\kappa_{\text{th}}}{\lambda}\left[ \left( 1+\frac{{\tau}^2}{3} \right) \cosh 2r \pm \frac{\sqrt{9+3{\tau}^2+{\tau}^4}}{3} \sinh 2r \right] + \frac{{\tau}^4}{12} {\lambda}^2 \right\}^2
- 1
\right) \\
&\mkern32mu\times
\Bigg(
\frac{{\tau}^4}{12}\left\{ {\kappa_{\text{th}}}^2
+ {\tau}{\kappa_{\text{th}}}{\lambda}\left[ \left( 1+\frac{{\tau}^2}{3} \right) \cosh 2r \pm \frac{\sqrt{9+3{\tau}^2+{\tau}^4}}{3} \sinh 2r \right]
+ \frac{{\tau}^4}{12} {\lambda}^2 \right\} \\
&\mkern96mu +\frac{{\tau}^4}{12}\left( 1-2{\kappa_{\text{th}}}^2 \right) + \frac{{\tau}^2}{2} {\kappa_{\text{th}}}^2 \left[ \left( 1+\frac{{\tau}^2}{3} \right) \cosh 2r \pm \frac{\sqrt{9+3{\tau}^2+{\tau}^4}}{3} \sinh 2r \right]^2
\Bigg)^{-1},
\end{aligned}
\label{eq:app:qcrb_optimal_squeeze}$$ which can be written as $$\frac{(a \pm b)^2 - 1}{c \pm d},$$ where $$\begin{aligned}
a &= {\kappa_{\text{th}}}^2 + {\kappa_{\text{th}}}{\lambda}{\tau}\left( 1 + \frac{{\tau}^2}{3} \right) \cosh 2r + \frac{{\tau}^4}{12} {\lambda}^2, \\
b &= {\kappa_{\text{th}}}{\lambda}{\tau}\frac{\sqrt{9+3{\tau}^2+{\tau}^4}}{3} \sinh 2r, \\
c &= \frac{{\tau}^4}{12} \left[ 1 - {\kappa_{\text{th}}}^2 + {\kappa_{\text{th}}}{\lambda}{\tau}\left( 1 + \frac{{\tau}^2}{3} \right) \cosh 2r + \frac{{\tau}^4}{12} {\lambda}^2 \right] + \frac{{\tau}^2}{2} {\kappa_{\text{th}}}^2 \left[ \left( 1 + \frac{{\tau}^2}{3} \right)^2 \cosh^2 2r + \left( 1 + \frac{{\tau}^2}{3} + \frac{{\tau}^4}{9} \right) \sinh^2 2r \right], \\
d &= {\kappa_{\text{th}}}{\lambda}\frac{{\tau}^5}{12} \frac{\sqrt{9+3{\tau}^2+{\tau}^4}}{3} \sinh 2r + {\kappa_{\text{th}}}^2 {\tau}^2 \left( 1 + \frac{{\tau}^2}{3} \right) \frac{\sqrt{9+3{\tau}^2+{\tau}^4}}{3} \cosh 2r \sinh 2r,\end{aligned}$$ where we have $ a $, $ b $, $ c $, and $ d $ all positive as well as $ a > b+1 $ and $ c > d $. The squeezing angle $ \phi_+ $ therefore offers a better precision for $$c < d \left( \frac{a^2+b^2-1}{2ab} \right),$$ which in this case is $$\begin{aligned}
0 >
&{\lambda}{\tau}\left[ - {\kappa_{\text{th}}}^4 \left( 1 + 3\frac{{\tau}^2}{4} + \frac{{\tau}^4}{9} \right) + \frac{{\tau}^2}{12} \left( 1 + \frac{{\lambda}^2{\tau}^4}{12} \right)^2 + {\kappa_{\text{th}}}^2 \frac{{\tau}^2}{6} \left( 1 - \frac{{\lambda}^2{\tau}^4}{12} \right) \right] \\
&+ {\kappa_{\text{th}}}\left( 1 + \frac{{\tau}^2}{3} \right) \left[ 1 - {\kappa_{\text{th}}}^4 + \frac{{\lambda}^2{\tau}^4}{6} (1-2{\kappa_{\text{th}}}^2) + \left( \frac{{\lambda}^2{\tau}^4}{12} \right)^2 \right] \cosh 2r
- \frac{{\tau}^3}{6} {\kappa_{\text{th}}}^4 {\lambda}\cosh 4r.
\end{aligned}$$
Homodyne detection
------------------
The for homodyne measurement of the quadrature $ {\hat{x}}\cos\theta + {\hat{p}}\sin\theta $ is $$\begin{aligned}
(\Delta\Lambda)^2 \geq
2{\Lambda_{0}}^2 \Bigg[
{\lambda}+ {\kappa_{\text{th}}}\Bigg(&
\frac{
\left[ 1 + {\tau}^2 \right]\cos^2\theta + {\tau}\sin 2\theta + \sin^2\theta
}{
\frac{1}{3}{\tau}^3\cos^2\theta + \frac{1}{2}{\tau}^2 \sin 2\theta + {\tau}\sin^2\theta
} \cosh 2r \\
&+
\frac{
\left\{ \left[ 1-{\tau}^2 \right]\cos^2\theta - {\tau}\sin 2\theta - \sin^2\theta \right\}\cos 2\phi
+
\left\{ 2{\tau}\cos^2 \theta + \sin 2\theta \right\}\sin 2\phi
}{
\frac{1}{3}{\tau}^3\cos^2\theta + \frac{1}{2}{\tau}^2 \sin 2\theta + {\tau}\sin^2\theta
} \sinh 2r
\Bigg)
\Bigg]^2.
\end{aligned}
\label{eq:app:crb_homodyne_squeezed}$$
### Optimal squeezing {#optimal-squeezing}
The bound is minimised with respect to the squeezing angle $ \phi $ by minimising the coefficient of $ \sinh 2r $ $$\left[ \left( 1-{\tau}^2 \right)\cos^2\theta - {\tau}\sin 2\theta - \sin^2\theta \right]\cos 2\phi
+
\left[ 2{\tau}\cos^2 \theta + \sin 2\theta \right]\sin 2\phi,$$ which has minima $$\phi = - \arctan \left( \frac{1}{{\tau}+ \tan \theta} \right),$$ for which squeezing angle the becomes $$(\Delta\Lambda)^2 \geq
2{\Lambda_{0}}^2 \left[
{\lambda}+ e^{-2r} {\kappa_{\text{th}}}\left(
\frac{
\left[ 1 + {\tau}^2 \right]\cos^2\theta + {\tau}\sin 2\theta + \sin^2\theta
}{
\frac{1}{3}{\tau}^3\cos^2\theta + \frac{1}{2}{\tau}^2 \sin 2\theta + {\tau}\sin^2\theta
}
\right)
\right]^2.
\label{eq:app:crb_homodyne_optimally_squeezed}$$
The optimal homodyne detection can then be recognised as the angle $ \theta $ $$\theta = - \arctan \left( \frac{3+2{\tau}^2+\sqrt{9+3{\tau}^2+{\tau}^4}}{3{\tau}} \right),
\label{eq:app:optimal_homodyne}$$ when this homodyne angle is used the optimal squeezing angle is $$\varphi = \arctan \left( \frac{3{\tau}}{3-{\tau}^2 + \sqrt{9+3{\tau}^2+{\tau}^4}} \right).$$
### Position and Momentum squeezing
Squeezing of position and momentum can be evaluated with $ \phi = 0 $, with $ r > 0 $ corresponding to squeezing of momentum while $ r < 0 $ is a squeezing $ |r| $ of position. For $ \phi = 0 $ the (Eq. ) becomes $$(\Delta\Lambda)^2 \geq
2{\Lambda_{0}}^2 \Bigg[
{\lambda}+ {\kappa_{\text{th}}}\left(
\frac{e^{2r} \cos^2 \theta + e^{-2r}\left({\tau}^2\cos^2 \theta+ {\tau}\sin 2\theta + \sin^2 \theta\right)
}{
\frac{1}{3}{\tau}^3\cos^2\theta + \frac{1}{2}{\tau}^2 \sin 2\theta + {\tau}\sin^2\theta
}
\right)
\Bigg]^2.$$ The optimal homodyne quadrature is then $$\theta = - \arctan \left( \frac{3e^{4r} + 2{\tau}^2 + \sqrt{9e^{8r}+3e^{4r}{\tau}^2+{\tau}^4}}{3{\tau}} \right),$$ which gives a precision $$(\Delta\Lambda)^2 \geq
2{\Lambda_{0}}^2 \left[
{\lambda}+ {\kappa_{\text{th}}}\left( \frac{2(3e^{2r} + e^{-2r}{\tau}^2 - \sqrt{9e^{4r} + 3{\tau}^2 + e^{-4r}{\tau}^4})}{{\tau}^3} \right)
\right],$$ where squeezing of position ($ r < 0 $) is beneficial for $ {\tau}< \sqrt{3} $ while squeezing of momentum (anti-squeezing of position, $ r > 0 $) is beneficial for $ {\tau}> \sqrt{3} $.
Heterodyne detection {#app:heterodyne}
====================
Heterodyne detection is the projection onto the overcomplete basis of Gaussian states which amounts to sampling from the Husimi Q-function [@shapiro_phase_1984; @leonhardt_measuring_1995]. The Q-function can be extracted from the Wigner function as [@ferraro_gaussian_2005] $$Q(x,p)=\frac{1}{\pi} \int\mathrm{d}x'\mkern-1mu\mathrm{d}p'\mkern-1mu W(x',p') \exp \left[ -(x-x')^2-(p-p')^2 \right].$$ which is a convolution and so for a Gaussian Wigner function with moments $ \vec{d} $ and $ \sigma $ the Q function will be Gaussian with moments $ \vec{d} $ and $ \sigma + \mathbb{1} $ [@serafini_quantum_2017 Chap. 5].
The mean of the distribution again contains no parameter dependence and so Eq. can also be applied here. The covariances from heterodyne detection are $$\Sigma({\tau}) =
\begin{pmatrix}
1 + \Sigma_{xx} + 2{\tau}\Sigma_{xp} + {\tau}^2 \Sigma_{pp} + \frac{1}{3} {\lambda}{\tau}^3 &
\Sigma_{xp} + {\tau}\Sigma_{pp} + \frac{1}{2} {\lambda}{\tau}^2 \\
\Sigma_{xp} + {\tau}\Sigma_{pp} + \frac{1}{2} {\lambda}{\tau}^2 &
1 + \Sigma_{pp} + {\lambda}{\tau}\end{pmatrix},
\label{eq:heterodyne_covariance}$$ giving a of $$(\Delta\Lambda)^2 \geq
\frac{12{\Lambda_{0}}^2|\Sigma({\tau})|^2}{{\tau}^4 |\Sigma({\tau})| + 6 {\tau}^2 \left( 1 + \Sigma_{xx} + {\tau}\Sigma_{xp} + \frac{{\tau}^2}{3}\Sigma_{pp} \right)^2 + 2 {\tau}^4 \left[ 1 + \Sigma_{xx} - \Sigma_{pp} - \Sigma_{xx}\Sigma_{pp} + \Sigma_{xp}^2 + \frac{{\tau}^2}{3}\left( 1-\Sigma_{pp} \right) \right] },
\label{eq:app:heterodyne_crb}$$ where $ |\Sigma| $ is the determinant, and $ \Sigma_{xx} $, $ \Sigma_{xp} $, and $ \Sigma_{pp} $ are the initial variances and covariance of the position and momentum operators. Without mechanical squeezing ($ r=0 $) this is $$(\Delta\Lambda)^2 \geq
\frac{
6 {\Lambda_{0}}^2 \left[ (1+{\kappa_{\text{th}}})^2 + {\kappa_{\text{th}}}{\tau}^2 + {\lambda}(1+{\kappa_{\text{th}}}){\tau}\left(1+\frac{{\tau}^2}{3}\right) + \frac{{\tau}^4 }{12} {\lambda}^2 \right]^2
}{
\frac{{\tau}^2}{3} \left[ (1+{\kappa_{\text{th}}}^2)(9+3{\tau}^2+{\tau}^4) + {\kappa_{\text{th}}}(18 + 6{\tau}^2-{\tau}^4) \right] + \frac{{\tau}^4}{2} \left[ (1+{\kappa_{\text{th}}})^2 + {\kappa_{\text{th}}}{\tau}^2 + {\lambda}(1+{\kappa_{\text{th}}}){\tau}\left(1+\frac{{\tau}^2}{3}\right) + \frac{{\tau}^4}{12} {\lambda}^2 \right]
}.$$
Optimal measurement {#app:sld_projectors}
===================
For a Gaussian system the is a hermitian operator, quadratic in the quadrature operators [@monras_phase_2013; @serafini_quantum_2017]. Any such hermitian operator, quadratic in the quadrature operators, can be transformed through some squeezing and displacement to an operator diagonal in the Fock basis [@monras_phase_2013; @serafini_quantum_2017].
The is primarily defined through identification of $ L^{(2)} $ which is given by [@monras_phase_2013; @serafini_quantum_2017] $$\sigma L^{(2)} \sigma + \Omega L^{(2)} \Omega = \partial \sigma,
\label{eq:app:ltwo}$$ which for a state with constant zero displacements $ \vec{d} = 0 $ then gives the [@monras_phase_2013; @serafini_quantum_2017] $$L_{\rho_{\Lambda}} = \begin{pmatrix} {\hat{x}} & {\hat{p}} \end{pmatrix} L^{(2)} \begin{pmatrix} {\hat{x}} \\ {\hat{p}} \end{pmatrix} - \frac{1}{2}\Trace{L^{(2)}\sigma}.$$ The covariance matrix which we wish to solve for Eq. is Eq. , which gives $ L^{(2)} $ as $$L^{(2)} =
\frac{1}{{\Lambda_{0}}(|\sigma({\tau})|^2-1)}
\begin{pmatrix}
l^{(2)}_{xx} & l^{(2)}_{xp} \\
l^{(2)}_{xp} & l^{(2)}_{pp}
\end{pmatrix},$$ where $$\begin{aligned}
l^{(2)}_{xx} &= {\tau}+{\tau}\sigma_{xp}({\tau})^2 - {\tau}^2 \sigma_{xp}({\tau})\sigma_{pp}({\tau}) + {\tau}^3 \sigma_{pp}({\tau})^2, \\
l^{(2)}_{xp} &= -{\tau}\sigma_{xx}({\tau}) \sigma_{xp}({\tau}) + \frac{{\tau}^2}{2} \left( \sigma_{xx}({\tau}) \sigma_{pp}({\tau}) + \sigma_{xp}({\tau})^2-1 \right) - \frac{{\tau}^3}{3} \sigma_{xp}({\tau})\sigma_{pp}({\tau}), \\
l^{(2)}_{pp} &= {\tau}\sigma_{xx}({\tau})^2 - {\tau}^2 \sigma_{xx}({\tau})\sigma_{xp}({\tau}) + \frac{{\tau}^3}{3}(1+\sigma_{xp}({\tau})^2).\end{aligned}$$ Then $ L^{(2)} $ has eigenvalues $$\alpha \pm \sqrt{\alpha^2 - {\tau}^2 \left( \sigma_{xx}({\tau}) - {\tau}\sigma_{xp}({\tau}) + \frac{{\tau}^2}{3} \sigma_{pp}({\tau}) \right)^2 - \frac{{\tau}^4}{12}\left( |\sigma({\tau})|-1 \right)^2},$$ with $$\alpha({\tau}) = \frac{{\tau}}{2}\left[ 1 + \sigma_{xx}({\tau})^2 - \sigma_{xp}({\tau})(\sigma_{xx}({\tau})+\sigma_{pp}({\tau})) {\tau}+ \frac{{\tau}^2}{3}(1+\sigma_{pp}({\tau})^2) + \sigma_{xp}({\tau})^2 \left( 1 + \frac{{\tau}^2}{3} \right) \right].$$ In order for phonon-number resolving detection to become optimal we then seek the symplectic transformation which gives the Williamson normal form of $ L^{(2)} $. For a single-mode system this can be recognised by first diagonalising $ L^{(2)} $ with a phase shift $
\begin{pmatrix} \cos\psi & \sin\psi \\ -\sin\psi & \cos\psi \end{pmatrix}
$, followed by a squeezing $ \operatorname{diag}(e^{z} , e^{-z}) $. The phase shift diagonalises $ L^{(2)} $, which has eigenvalues $ D_1 $ and $ D_2 $. The symplectic eigenvalue of $ L^{(2)} $ is then $ \sqrt{D_1D_2} $ and so the squeezing $ z $ required to bring $ L^{(2)} $ into its normal form is $ e^{2z} = e^{\frac{1}{2}|\ln D_1 - \ln D_2|} $.
Thus the required squeezing is $$e^{2z} = \sqrt{\frac{
1 + \sqrt{1 - \frac{1}{\alpha^2} \left[ {\tau}^2 \left( \sigma_{xx}({\tau}) - {\tau}\sigma_{xp}({\tau}) + \frac{{\tau}^2}{3} \sigma_{pp}({\tau}) \right)^2 + \frac{{\tau}^4}{12}\left( |\sigma({\tau})|-1 \right)^2 \right]}
}{
1 - \sqrt{1 - \frac{1}{\alpha^2} \left[ {\tau}^2 \left( \sigma_{xx}({\tau}) - {\tau}\sigma_{xp}({\tau}) + \frac{{\tau}^2}{3} \sigma_{pp}({\tau}) \right)^2 + \frac{{\tau}^4}{12}\left( |\sigma({\tau})|-1 \right)^2 \right]}
}}.$$
Optimality of detection schemes {#app:relative_precisions}
===============================
Our bounds cover a range of settings with $ {\Lambda_{0}}^2 $ pre-factoring the bounds and their ratios being a function of only $ {\lambda}$, $ {\tau}$, $ {\kappa_{\text{th}}}$, and squeezing $ re^{i\phi} $ (with parameters such as homodyne angle $ \theta $ representing different measurement choices rather than properties of the system). This allows comparison of our bounds in terms of these parameters alone, perhaps the simplest case being where we assume trapping allows us to take $ {\kappa_{\text{th}}}= 1 $ and that no external squeezing is applied
Homodyne
--------
For $ {\kappa_{\text{th}}}= 1 $ and $ r = 0 $ we can easily compare the with the optimal homodyne numerically across the $ {\lambda}$ and $ {\tau}$ variables in Fig. \[fig:homodyne\_optimality\].
![Ratio of quantum Fisher information against classical Fisher information for optimal homodyne quadrature ($ F(\Lambda;\theta_{\textrm{opt}})/H(\Lambda) $), plotted for $ {\kappa_{\text{th}}}= 1 $ and $ r = 0 $. The red rectangle is representative of the MAQRO parameter regime[]{data-label="fig:homodyne_optimality"}](homodyne_optimality)
The analytic form of the ratio is $$R = \frac{ {\tau}^4 \left\{ \left[ {\lambda}{\tau}\left( 1 + \frac{{\tau}^2}{3} + \frac{{\tau}^3}{12}{\lambda}\right) + 1 \right]^2 - 1 \right\}
}{
72 \left( 1 + \frac{{\tau}^2}{3} + \frac{{\tau}^3}{6} {\lambda}- \frac{\sqrt{9+3{\tau}^2+{\tau}^4}}{3}\right)^2\left[ \left( 1 + \frac{{\tau}^2}{3} + \frac{{\tau}^3}{12} {\lambda}\right)^2 - \frac{1}{2} \left( 1 + \frac{{\tau}^2}{3} \right) \left( 1 + \frac{{\tau}^2}{3} + \frac{{\tau}^3}{6} {\lambda}\right) \right]
}.
\label{eq:app:hom_q}$$
Heterodyne
----------
For $ {\kappa_{\text{th}}}= 1 $ and $ r = 0 $ we can easily compare the with the heterodyne numerically across the $ {\lambda}$ and $ {\tau}$ variables in Fig. \[fig:heterodyne\_optimality\].
![Ratio of quantum Fisher information against classical Fisher information for heterodyne detection ($ F(\Lambda)/H(\Lambda) $), plotted for $ {\kappa_{\text{th}}}= 1 $ and $ r = 0 $. The red rectangle is representative of the MAQRO parameter regime[]{data-label="fig:heterodyne_optimality"}](heterodyne_optimality)
The analytic form of the ratio is $$R=
\frac{\left\{\left[{\lambda}{\tau}\left(1+\frac{{\tau}^2}{3}+{\lambda}\frac{{\tau}^3}{12}\right)+1\right]^2-1\right\} \left[\left(1+\frac{{\tau}^2}{3}+{\lambda}\frac{{\tau}^3}{12}\right)^2+\left(1+\frac{{\tau}^2}{6}\right)^2\right]
}{
16 (1 + \frac{{\tau}}{2}{\lambda})^2 \left(1+\frac{{\tau}^2}{4}+\frac{{\lambda}{\tau}^3}{24}\right)^2 \left[\left(1+\frac{{\tau}^2}{3}+\frac{{\lambda}{\tau}^3}{12}\right)^2-\frac{1}{2} \left(1+\frac{{\tau}^2}{3}\right) \left(1+\frac{{\tau}^2}{3}+\frac{{\lambda}{\tau}^3}{6}\right)\right]
}.
\label{eq:app:het_q}$$
Homodyne and Heterodyne
-----------------------
In the same $ {\kappa_{\text{th}}}= 1 $ and $ r = 0 $ case we can compare the optimal homodyne against the heterodyne numerically across the $ {\lambda}$ and $ {\tau}$ variables in Fig. \[fig:homodyne\_heterodyne\_optimality\].
![Ratio of classical Fisher information for heterodyne detection against classical Fisher information for homodyne detection of the optimal quadrature, plotted for $ {\kappa_{\text{th}}}= 1 $ and $ r = 0 $. The red rectangle is representative of the MAQRO parameter regime[]{data-label="fig:homodyne_heterodyne_optimality"}](homodyne_heterodyne_optimality)
This demonstrates no more than a factor of two advantage for heterodyne in the $ {\tau}\gg 1 $ and $ {\lambda}\gg 1 $, while in the $ {\lambda}\ll 1 $ regime homodyne has a near unbounded advantage.
The analytic form of the ratio (which can be seen from Eqs. and ) is $$R=
\frac{
9 \left[\left(1+\frac{{\tau}^2}{3}+{\lambda}\frac{{\tau}^3}{12}\right)^2+\left(1+\frac{{\tau}^2}{6}\right)^2\right]
\left( 1 + \frac{{\tau}^2}{3} + \frac{{\tau}^3}{6} {\lambda}- \frac{\sqrt{9+3{\tau}^2+{\tau}^4}}{3}\right)^2
}{
2 {\tau}^4(1 + \frac{{\tau}}{2}{\lambda})^2 \left(1+\frac{{\tau}^2}{4}+\frac{{\lambda}{\tau}^3}{24}\right)^2
}.
\label{eq:app:hom_het}$$
Tests of Continuous Spontaneous Localisation {#app:csl}
============================================
For MAQRO the minimum resolvable $ {\lambda^{\mathrm{CSL}}}$ for position and momentum can be seen in Fig. \[app:fig:csl\_conventional\], plotted for a $ {r_\mathrm{s}}= \SI{100}{\nm} $ sphere of mass with values otherwise as Tab. \[tab:maqro\_values\] in the main text, where the black line is based on the minimum required strenght proposed in @toros_colored_2017.
![ Bounds plotted for a $ {r_\mathrm{s}}= \SI{100}{\nm} $ sphere of mass with values otherwise as Tab. \[tab:maqro\_values\] in the main text. The minimum required collapse rate given is based on the criteria of Ref. [@toros_colored_2017]. The magenta dot represents the values originally proposed by @ghirardi_unified_1986. []{data-label="app:fig:csl_conventional"}](csl_posmom)
This plot shows the potential improvments, with MAQRO already competitive in , squeezing allows a test down to the lower bound for $ {r_\mathrm{C}}< \SI{1e-7}{\m} $ and significant improvement on reported results up to $ {r_\mathrm{C}}= \SI{1e-5}{\m} $.
This is plotted in Fig. \[app:fig:csl\_optimal\], plotted again for a $ {r_\mathrm{s}}= \SI{100}{\nm} $ sphere of mass with values otherwise as Tab. \[tab:maqro\_values\] in the main text.
![ Bounds plotted for a $ {r_\mathrm{s}}= \SI{100}{\nm} $ sphere of mass with values otherwise as Tab. \[tab:maqro\_values\] in the main text. The minimum required collapse rate given is based on the criteria of Ref. [@toros_colored_2017]. The magenta dot represents the values originally proposed by @ghirardi_unified_1986. []{data-label="app:fig:csl_optimal"}](csl_opt)
As might be guessed from the significant gap in Fig. \[fig:maqro\_precision\] of the main text the optimal quadrature allows for a categorical test of . This bound can be reduced through squeezing and the fundamental limit given by the will further allow a superior precision through a saturating measurement. Such improvements however offer little significance, as the will give a lower bound no less than that of the optimal quadrature.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Parameter tuning is a common issue for many tracking algorithms. In order to solve this problem, this paper proposes an online parameter tuning to adapt a tracking algorithm to various scene contexts. In an offline training phase, this approach learns how to tune the tracker parameters to cope with different contexts. In the online control phase, once the tracking quality is evaluated as not good enough, the proposed approach computes the current context and tunes the tracking parameters using the learned values. The experimental results show that the proposed approach improves the performance of the tracking algorithm and outperforms recent state of the art trackers. This paper brings two contributions: (1) an online tracking evaluation, and (2) a method to adapt online tracking parameters to scene contexts.'
author:
- Duc Phu Chau
- Julien Badie
- François Brémond
- |
Monique Thonnat\
STARS team, INRIA, France\
2004 route des Lucioles, 06560 Valbonne, France\
[{Duc-Phu.Chau, Julien.Badie, Francois.Bremond, Monique.Thonnat} @inria.fr]{}
bibliography:
- 'avss\_refs.bib'
title: Online Tracking Parameter Adaptation based on Evaluation
---
Introduction
============
Many studies have been proposed to track the movements of objects in a scene [@yilmaz; @dpchauVisapp11; @benfold11]. However the selection of a tracking algorithm for an unknown scene becomes a hard task. Even when the tracker has already been determined, it is difficult to tune its parameters to get the best performance due to the variations of scene context (scene illumination, object occlusion level, 2D object sizes).
Some approaches have been proposed to address these issues. The authors in [@kuo10] propose an online learning scheme based on Adaboost to compute a discriminative appearance model for each mobile object. However the online Adaboost process is time consuming. The author in [@hall] proposes two strategies to regulate the parameters for improving the tracking quality. In the first strategy, the parameter values are determined using an enumerative search. In the second strategy, a genetic algorithm is used to search for the best parameter values. This approach does not require human supervision and parameter knowledge for controlling its tracker. However, it is computationally expensive because of the parameter optimization stage performed in the online phase.
In the other hand, some approaches integrate different trackers and then select the convenient tracker depending on video content. For example, the authors in [@prost; @yoon12] present tracking frameworks which are able to control a set of different trackers to get the best performance. The system runs the tracking algorithms in parallel. At each frame, the best tracker is selected to compute the object trajectories. These two approaches require the execution of different trackers in parallel which is expensive in terms of processing time. In [@dpchauIcdp11], the authors propose a tracking algorithm whose parameters can be learned offline for each tracking context. However the authors suppose that the context within a video sequence is fixed over time. Moreover, the tracking context is selected manually.
These studies have obtained relevant results but show strong limitations on the online processing time and the self-adaptation capacity to the scene variations. In order to solve these problems, we propose in this paper a new method to adapt the tracking algorithms to the scene variations. The principle of the proposed approach is the automatic parameter tuning of tracking algorithms over time during the online process. This parameter tuning relies on an offline learning process and an online tracking evaluation method. The proposed tracking evaluation is responsible for detecting the tracking errors and activating the parameter tuning process if necessary. The parameter tuning relies entirely on the offline learned database, this helps to avoid slowing down the processing time of the tracking task. The variation of scene over time during the online phase is also addressed in the proposed approach.
This paper is organized as follows. Sections 2 and 3 present in detail the proposed approach. Section 4 shows the results of the experimentation and validation. A conclusion as well as future work are presented in section 5.
Offline Learning
================
{width="0.96\linewidth"}
The objective of the learning phase is to create a database which supports the control process of a tracking algorithm. This database contains satisfactory parameter values of the tracking algorithm for various contexts. This phase takes as input training video sequences, annotated objects, annotated trajectories, a tracking algorithm including its control parameters. The term “control parameters” refers to parameters which are considered in the control process (i.e. to look for satisfactory values in the learning phase and to be tuned in the online phase). At the end of the learning phase, a learned database is created. A learning session can process many video sequences. Figure \[fig\_learning\_scheme\] presents the proposed scheme for building the learned database.
The notion of “context” (or “tracking context”) in this work represents elements in the videos which influence the tracking quality. More precisely, a context of a video sequence is defined as a set of six features: density of mobile objects, their occlusion level, their contrast with regard to the surrounding background, their contrast variance, their 2D area and their 2D area variance. The offline learning is performed as follows.
First, for each training video, the **“contextual feature extraction”** step computes the contextual feature values from annotated objects for all video frames.
Second, in the **“context segmentation and code-book modeling”** step, these context feature values are used to segment the training video in a set of consecutive chunks. Each video chunk has a stable context. The context of a video chunk is represented by a set of six code-books (corresponding to six context features).
Third, the **“tracking parameter optimization”** is performed to determine satisfactory tracking parameter values for the video chunks using annotated trajectories. These parameter values and the set of code-books are then inserted into a temporary learned database.
After processing all training videos as three above steps, a **”clustering”** step, which is composed of two sub-steps **“context clustering”** and **“parameter computation for context clusters”**, is performed. In the first sub-step, the contexts are clustered using a QT clustering. In the second one, for each context cluster, its satisfactory tracking parameter values are defined in function of the tracking parameters learned for each element context. The context clusters and their satisfactory tracking parameters are then inserted into the learned database.
Online Control
==============
In this section, we present in detail how the tracking algorithm is controlled to adapt itself to the contextual variations. This controller takes as input the video stream, the list of detected objects at every frame, the offline learned database, the object trajectories and gives as output the satisfactory tracking parameter values to parameterize the tracker if necessary (see figure \[fig\_online\_control\]).
At each frame, the tracking quality is estimated online. When a tracking error is detected, the proposed controller computes the context of the $n$ latest frames. This context is then used for finding the best matching context cluster in the offline learned database. If such a context cluster is found, the tracking parameters associated with this context cluster are used. In the following sections, we describe the three steps of this phase: online tracking evaluation, context computation and parameter tuning.
![The online control[]{data-label="fig_online_control"}](online_scheme_3.png){width="8.65cm"}
Online Tracking Evaluation
--------------------------
In this paper, we propose a method to estimate online the tracking quality. The main advantage of this approach is that it can be used for evaluating any tracking algorithm. This method takes as input the current object trajectories, the processing video stream, and gives as output at each frame an alarm of tracking quality if necessary.
The principle of this evaluation method relies on the following hypothesis: a tracked object is supposed to have a coherence (low variation) on some appropriate descriptors. The selection of appropriate object descriptors is crucial. These descriptors have to satisfy three criteria: they have to be representative of the tracked object, discriminative enough for distinguishing with the other objects, and can take into account the popular tracking errors (ID switch, ID lost). Regarding these criteria, we use the following five descriptors to evaluate the tracking quality of a mobile object: 2D bounding box, speed, direction, color histogram and color covariance [@dpchauIcdp11].
Using these descriptors, we define two values representing the tracking risks for each object at each frame. The first one is the object interaction score which takes into account the occlusion and density between the consider object and its spatial neighbors. The second value is called “object tracking error score” that evaluates the variations of the last four above object descriptors over time. A high tracking error score (near to 1) alerts a tracking problem such as ID switch or ID lost. In the following sections, we present in detail how to compute these two scores.
### Object Interaction Score
The object interaction score is computed at every frame and for each object. It represents the interaction possibility between mobile objects (spatial overlap, cross each other). This score takes into account the density of mobile objects at the current instant and the object occlusion levels in the last two frames.
Given an object at instant $t$, denoted $o_t^i$, we can find its neighbors, denoted $\mathfrak{N} (o_t^i)$, which are the spatially close objects. The density score for the object $o_t^i$ is defined as follows: $$d (o_t^i) = \frac{union (o_t^i, \mathfrak{N} (o_t^i)) } { cover (o_t^i, \mathfrak{N} (o_t^i)) }$$
where $union(o_t^i, \mathfrak{N} (o_t^i))$ is the union of 2D areas occupied by object $o_t^i$ and its neighbors $\mathfrak{N} (o_t^i)$; $cover (o_t^i, \mathfrak{N} (o_t^i))$ is the area of the smallest rectangular which covers $o_t^i$ and $\mathfrak{N} (o_t^i)$.
In order to compute the occlusion level of an object, we define first the occlusion level between two objects $o_t^i$ and $o_t^j$ as follows:
$$\mathcal{O} (o_t^i, o_t^j )\ =\ \frac {a_t^{ij}} {min (a_t^i, a_t^j)}$$
where $a_t^i$ is 2D area of object $i$ at time $t$, $a_t^{ij}$ is the overlap area of objects $i$ and $j$ at $t$.
Second, the occlusion level between object $o_t^i$ and its neighbors, denoted $\mathcal{O}_t (o_t^i)$, is defined as the $max \{ \mathcal{O} (o_t^i, o_t^j ) \ | \ j \in \mathfrak{N} (o_t^i) \} $. In the same way, we compute the occlusion level between object $o_t^i$ and its neighbors detected at $t-1$, denoted $\mathcal{O}_{t-1} (o_t^i)$.
The interaction score of the object $o_t^i$, denoted $I(o_t^i)$, is defined as the mean value of its density score and the two occlusion level scores $\mathcal{O}_{t-1} (o_t^i)$, $\mathcal{O}_{t} (o_t^i)$: $$I(o_t^i) = \frac{d(o_t^i) + \mathcal{O}_{t-1} (o_t^i) + \mathcal{O}_t (o_t^i) }{3}$$
### Object Tracking Error Score
The object tracking error score is computed at every frame and for each object. It represents the potential error on the tracking quality of the considered tracked object. This scores takes into account the variations of the four object descriptors: object speed, direction, histogram color and color covariance over time. The 2D bounding box descriptor is not used because it is very dependent on the detection quality. For each object descriptor at instant $t$, we compute the mean and standard deviation values, denoted $\mu_t^k$ and $\delta_t^k$, where $k$ representing the considered descriptor ($k = 1..4$). The tracking error score of an object at $t$ is defined as follows: $$E_t = \frac{ \sum_{\alpha=1}^{4} \frac {\delta_t^\alpha} {\mu_t^\alpha} } { 4 }$$
### Object Tracking Error Alarm {#sec_error_alarm}
At instant $t$, a tracked object is considered as “erroneous” if its interaction score and tracking error score are greater than a same threshold $Th_1$; and its tracking error score increases by a predefined threshold $Th_2$ compared to its tracking error score computed at $t-1$. If there exists such a tracked object, the tracking evaluation task sends a tracking error alarm to the context computation task to improve the tracking performance.
Context Computation {#sec_context_computation}
-------------------
The context computation task is only activated when the tracker fails. The objective of this step is to find the context cluster stored in the offline learned database to which the context of the current processing video belongs. This step takes as input for every frame, the list of the current detected objects and the processing video stream. First, we compute the six context feature values (density, occlusion level, contrast, contrast variance, 2D area and 2D area variance of mobile objects) of the video chunk corresponding to the last $n$ frames ($n$ is a predefined parameter). The set of these feature values is denoted $\mathfrak{C}$. Second, let $\mathfrak{D}$ represent the offline learned database, a context feature set $\mathfrak{C}$ belongs to a cluster $C_i$ if both conditions are satisfied:
$$\label{eq_cond_belonging_cluster_thrld}
contextDistance(\mathfrak{C},\ C_i)\ <\ Th_3 \\$$
$$\label{eq_cond_belonging_cluster_min}
\begin{array}{llr}
\forall C_j\ \in \ \mathfrak{D}, j \neq i: \\
\ contextDistance(\mathfrak{C}, C_i) \leq contextDistance(\mathfrak{C}, C_j)
\end{array}$$
where $Th_3$ is a predefined threshold; $contextDistance(\mathfrak{C},\ C_i)$ represents the distance between a context feature set $\mathfrak{C}$ and a context cluster $C_i$. This distance relies on the number of times where the context feature values belonging to $\mathfrak{C}$ matches to code-words in $C_i$.
Parameter Tuning
----------------
If such a context cluster $C_i$ is found, the satisfactory tracking parameters associated with $C_i$ are used for parameterizing the tracking of the current video chunk. Otherwise, the tracking algorithm parameters do not change, the current video chunk is marked to be learned offline later.
Experimental Results
====================
Parameter Setting and Object Detection Algorithm
------------------------------------------------
The proposed control method has four predefined parameters. The first two parameters are thresholds $Th_1$ and $Th_2$, presented at section \[sec\_error\_alarm\], are respectively set to 0.2 and 0.15. The third parameter is the distance threshold $Th_3$ (section \[sec\_context\_computation\]) is set to $0.5$. The last parameter is the number of frames $n$ to compute the context, presented at section \[sec\_context\_computation\], is set to $50$. These parameter values are unchanged for all the experiments presented in this paper. A HOG-based algorithm [@corvee10] is used for detecting people in videos.
Tracking Evaluation Metrics
---------------------------
In this experimentation, we use the following tracking evaluation metrics. Let $GT$ be the number of trajectories in the ground-truth of the test video. The first metric **$MT$** computes the number of trajectories successfully tracked for more than 80% divided by GT. The second metric **$PT$** computes the number of trajectories that are tracked between 20% and 80% divided by GT. The last metric **$ML$** is the percentage of the left trajectories.
Controlled Tracker {#sec_controlled_tracker}
------------------
In this paper, we select an object appearance-based tracker [@dpchauIcdp11] to test the proposed approach. This tracker takes as input a video stream and a list of objects detected in a predefined temporal window. The object trajectory computation is based on a weighted combination of five object descriptor similarities: 2D shape ratio, 2D area, RGB color histogram, color covariance and dominant color. For this tracker, the five object descriptor weights $w_k$ ($k$ = 1..5, corresponding to the five above descriptors) are selected for testing the proposed control method. These parameters depend on the tracking context and have a significant effect on the tracking quality.
Training Phase {#sec_tracker_1_training_phase}
--------------
In the training phase, we use 15 video sequences belonging to different context types (i.e. different levels of density and occlusion of mobile objects as well as of their contrast with regard to the surrounding background, their contrast variance, their 2D area and their 2D area variance). These videos belong to four public datasets (ETISEO, Caviar, Gerhome and PETS), to two European projects (Caretaker and Vanaheim). They are recorded in various places: shopping center, buildings, home, subway stations and outdoor.
Each training video is segmented automatically in a set of context segments. Each object descriptor similarity can be considered as a weak classifier for linking two objects detected within a temporal window. Therefore in the tracking parameter optimization process, we use an Adaboost algorithm to learn the object descriptor weights for each context segment. The Adaboost algorithm has a lower complexity than the other heuristic optimization algorithms (e.g. genetic algorithm, particle swam optimization). Also, this algorithm avoids converging to the local optimal solutions. After segmenting the 15 training videos, we obtain 72 contexts. By applying the clustering process, 29 context clusters are created.
Testing Phase
-------------
All the following test videos do not belong to the set of the 15 training videos.
### Subway video
The first tested video sequence belongs to the Caretaker European project whose video camera is installed in a subway station (see the left image of the figure \[fig\_dataset\]). The length of this sequence is 5 minutes. It contains 38 mobile objects.
$
\begin{array}{ccc}
\includegraphics[width=5cm, height = 3cm]{img_caretaker/caretaker_frame_468.png} &
\includegraphics[width=5cm, height = 3cm]{img_pets/pets_frame_779.jpg} &
\includegraphics[width=5cm, height = 3cm]{img_tud/tud_frame_26.jpg} \\
\end{array}$
Figure \[fig\_caretaker\] illustrates the output of the controlled tracking process. We consider the tracking result of the two persons on the left images. At the frame 125, these two persons with respectively ID 254 (the left person) and ID 215 (the right person) are correctly tracked. Person 254 has a larger bounding box than person 215. At the frame 126, due to an incorrect detection, the left person has a quite small bounding box. By consequence, the IDs of these two persons are switched because the tracking algorithm currently uses object 2D area as an important descriptor. Now the online tracking evaluation sends an alarm on tracking error to the context computation task. The context cluster associated to the following parameters are selected for tuning the tracking parameters: $w_1 = 0$, $w_2 = 0$, $w_3 = 0.72$, $w_4 = 0$ and $w_5 = 0.28$ (see section \[sec\_controlled\_tracker\] for the meaning of these parameters). The color histogram which is selected now as the most important descriptor ($w_3 = 0.72$). The 2D area descriptor is not used ($w_2 = 0$). At the frame 127, after the tracking parameter tuning, the two considered objects take the correct IDs as in frame 125.
$
\begin{array}{ccc}
\includegraphics[width=5cm, height=2.5cm]{img_caretaker/frame_125.png} &
\includegraphics[width=5cm, height=2.5cm]{img_caretaker/frame_126.png} &
\includegraphics[width=5cm, height=2.5cm]{img_caretaker/frame_127.png} \\
\end{array}$
Table \[tab\_caretaker\_result\] presents the tracking results of the tracker [@dpchauIcdp11] in two cases: without and with the proposed controller. We find that the proposed controller helps to improve significantly the tracking performance. The value of $MT$ increases 52.7% to 84.2% and the value of $ML$ decreases 18.4% to 10.5%.
Methods MT(%) PT(%) ML(%)
-------------------------------------------------------------- ------- ------- -------
Chau et al. [@dpchauIcdp11] without the proposed controller 52.7 28.9 18.4
**Chau et al. [@dpchauIcdp11] with the proposed controller** 5.3
: \[tab\_caretaker\_result\]Tracking results of the subway video. The proposed controller improves significantly the tracking performance. The best values are printed in .
### PETS 2009 Dataset
In this test, we select the sequence S2\_L1, camera view 1, time 12.34 belonging to the PETS 2009 dataset for testing because this sequence is experimented in several state of the art trackers (see the middle image of the figure \[fig\_dataset\]). This sequence has 794 frames, contains 21 mobile objects and several occlusion cases. In this test, we use the CLEAR MOT metrics presented in [@petsmetric] to compare with other tracking algorithms. The first metric is MOTA which computes Multiple Object Tracking Accuracy. The second metric is MOTP computing Multiple Object Tracking Precision. We also define a third metric $\overline{M}$ representing the average value of MOTA and MOTP. All these metrics are normalized in the interval $[0,\ 1]$. The higher these metrics, the better the tracking quality is.
For this sequence, the tracking error alarms are sent six times to the context computation task. For all these six times, the context cluster associated to the following tracking parameters is selected for tracking objects: $w_1 = 0$, $w_2 = 0.14$, $w_3 = 0.12$, $w_4 = 0.13$ and $w_5 = 0.61$ (see section \[sec\_controlled\_tracker\] for the meaning of these parameters). The dominant color descriptor ($w_5$) is selected as the most important descriptor for tracking objects. This selection is reasonable. This descriptor can well handle the object occlusion cases (see [@dpchauIcdp11] for more details) which happen frequently in this video. Table \[tab\_pets\_result\] presents the metric results of the proposed approach and four recent trackers from the state of the art. With the proposed controller, the tracking result increases significantly. We also obtain the best values in all the three metrics.
Methods MOTA MOTP $\overline{M}$
-------------------------------------------------------------- ------ ------ ----------------
Berclaz et al. [@berclaz11] 0.80 0.58 0.69
Shitrit et al. [@shitrit11] 0.81 0.58 0.70
Henriques et al. [@henriques11] **** 0.69 0.77
Chau et al. [@dpchauIcdp11] without the proposed controller 0.62 0.63 0.63
**Chau et al. [@dpchauIcdp11] with the proposed controller** **** **** ****
: \[tab\_pets\_result\]Tracking results on the PETS sequence S2.L1, camera view 1, time 12.34. The best values are printed in .
Methods MT(%) PT(%) ML(%)
-------------------------------------------------------------- ------- ---------- ----------
Kuo et al. [@kuo11] 60 30.0 ****
Andriyenko et al. [@andriyenko11] 60.0 30.0 ****
Chau et al. [@dpchauIcdp11] without the proposed controller 50.0 30.0 20.0
**Chau et al. [@dpchauIcdp11] with the proposed controller** **** **10.0** **20.0**
: \[tab\_tud\_result\]Tracking results for the TUD-Stadtmitte sequence. The best values are printed in .
### TUD Dataset
For the TUD dataset, we select the TUD-Stadtmitte sequence. This video contains only 179 frames and 10 objects but is very challenging due to heavy and frequent object occlusions (see the right image of the figure \[fig\_dataset\]). Table \[tab\_tud\_result\] presents the tracking results of the proposed approach and three recent trackers from the state of the art. We obtain the best $MT$ value compared to these two trackers.
Computational Cost
------------------
All experiments presented in this paper have been performed in a machine of Intel(R) CPU @ 2.60GHz and 8GB RAM. The average processing time of the tracking process for all test videos is 13 fps while using the proposed controller, and is 15 fps without the controller. We find that the controller increases only slightly the computational cost.
Conclusion and Future Work
==========================
In this paper, we have presented a new control approach to adapt the tracking performance to various tracking context. While using the proposed online tracking evaluation, tracking errors are detected quickly. The parameter tuning is then activated to improve immediately the tracking quality. The experiments show a significant improvement of the tracking performance when the proposed controller is used. Although we only address the parameter tuning problem, the proposed approach can also be applied to select online trackers to adapt better the context variations. In future work, the tracking parameters will be learned by an unsupervised method to remove completely the human knowledge from training phase.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by The European Vanaheim, Panorama and Support projects.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the one-loop effective action at finite temperature for a scalar field with quartic interaction has the same renormalized expression as at zero temperature if written in terms of a certain classical field $\phi_c$, and if we trade free propagators at zero temperature for their finite-temperature counterparts. The result follows if we write the partition function as an integral over field eigenstates (boundary fields) of the density matrix element in the functional Schrödinger field-representation, and perform a semiclassical expansion in two steps: first, we integrate around the saddle-point for fixed boundary fields, which is the classical field $\phi_c$, a functional of the boundary fields; then, we perform a saddle-point integration over the boundary fields, whose correlations characterize the thermal properties of the system. This procedure provides a dimensionally-reduced effective theory for the thermal system. We calculate the two-point correlation as an example.'
author:
- 'A. [Bessa]{}$^{1,2}$[^1], F. T. Brandt$^{2}$[^2], C. A. A. [de Carvalho]{}$^{3}$[^3] and E. S. [Fraga]{}$^{3}$[^4]'
title: |
**Quantum statistical correlations in thermal field theories:\
boundary effective theory**
---
Introduction
============
Effective actions are generating functionals for the one-particle irreducible vertices of a field theory. From the vertices, one obtains all the correlations of the theory, which may be used to compute physical quantities, after suitable renormalization. In finite-temperature field theories [@FTFT-books], the effective action is the thermal equilibrium Gibbs free energy of the system of relativistic quanta of the theory. Effective actions are, therefore, natural quantities to compute at finite temperature. For instance, effective action calculations [@Blaizot:2000fc; @Peshier:1998rz; @Berges:2004hn; @Berges:2005hc; @Arrizabalaga:2006hj; @Borsanyi:2007bf] are among the techniques used to improve the poor convergence of the perturbative series in the bosonic sector, plagued by infrared divergencies.
In this paper we build an effective field theory for quantum statistical correlations using the boundary fields as the relevant degrees of freedom, which we denote by [*boundary effective theory*]{} (BET). To do so, we start from the partition function of our field theory in equilibrium at temperature $T=1/\beta$ written as a functional integral of the diagonal element of the density matrix (the Boltzmann operator for the field theory Hamiltonian) in the Schrödinger field-representation. The integral is performed over the eigenvalues of the field operator in that representation, which are the stochastic variables of the problem, and which we will henceforth call boundary fields.
The density matrix element itself may also be written as a functional integral, one that describes an Euclidean time $\tau$ evolution from a given boundary field configuration at $\tau=0$, to that same configuration at $\tau=\beta$. The integral described in the previous paragraph is then performed over boundary fields. The quantum statistical correlations of the boundary fields should fully characterize thermal equilibrium.
The use of the density matrix functional leads to the derivation of the effective quantum statistical description for an underlying field theory, given by the effective action constructed from its microscopic Hamiltonian. The correlations of the stochastic fields (boundary fields, which only depend on spatial coordinates) express the thermalization encoded in the density matrix.
The density matrix functional was already used in a similar context in [@deCarvalho:2001xv], where it led to the construction of dimensionally-reduced effective actions. Likewise, in [@Bessa:2007vq], it was used up to one-loop order to investigate the thermodynamics of scalar fields. In the latter reference, the boundary fields (considered as fluctuations around a homogeneous background) played a very nontrivial role in the calculation of the partition function of scalar fields. In quantum mechanics, this formalism was developed in Refs. [@deCarvalho:1998mv; @deCarvalho:1999fi; @deCarvalho:2001vk] and led to the functional density matrix formulation of quantum statistics in Ref. [@Bessa:2008xj].
Our goal, in the present article, is to use the density matrix functional in a scalar field theory with quartic interaction to demonstrate a simple relationship between effective actions at zero and finite temperature, computed up to one-loop order. The relation provides a natural bridge between correlations, as well as renormalization conditions, at zero and finite temperature. In practice, it yields a recipe to read off the renormalized finite temperature result from its renormalized zero-temperature counterpart.
The effective action that we obtain at finite temperature is a functional of the expectation value of the boundary (stochastic) fields. It is the generator of the one-particle irreducible vertices for the stochastic fields, and therefore allows us to reconstruct all stochastic correlations which define physical quantities in thermal equilibrium. In practice, our relation establishes a connection among those and zero-temperature physical quantities, such as particle masses and couplings.
This article is structured as follows: in Sec. \[densitymatrix\], we write both the density matrix and the partition function as functional integrals, define the generating functionals for quantum statistical correlations, and associate them to thermodynamic free energies; in Sec. \[fluctuations\], we integrate over the dynamical fields of the underlying theory, keeping the boundary fields fixed; in Sec. \[bcfluctuation\], we integrate over the (stochastic) boundary fields to obtain the effective action; in Sec. \[renormalization\], we go through the renormalization procedure to compute one-loop renormalized correlations; finally, in Sec. \[conclusions\], we present our conclusions.
Density matrix for scalar theories {#densitymatrix}
==================================
In quantum statistical mechanics, the partition function for a system in contact with a thermal reservoir at temperature $T$ ($\beta=1/ T$) is expressed as a sum (integral) over a stochastic variable, whose probabilistic weight is given by the density matrix. For a system described by a (self-interacting) scalar field theory, the stochastic variable is the field eigenvalue in the functional Schrödinger field-representation, $\phi_0 (\x)$, which satisfies $\hat\phi|\phi_0 (\x)\rangle=\phi_0 (\x)|\phi_0 (\x)\rangle$.
The partition function is then a functional integral over the field eigenvalue of the diagonal element of the density matrix functional $\rho_{\beta}$, $$\label{Z1}
Z(\beta)=\int [{\cal D}\phi_0(\x)]\,\,\rho_{\beta}[\phi_0(\x),\phi_0(\x)]\;,$$ $$\label{phi}
\rho_{\beta}[\phi_0(\x),\phi_0(\x)]\equiv\langle\phi_0 (\x)|\exp(-\beta \hat{H})|\phi_0 (\x)\rangle\;.$$ The Hamiltonian in the preceding formula specifies the underlying dynamics. In the present case, it is the dynamics of a (self-interacting) scalar field $\phi(\tau, \x)$.
As is well known [@Feynman-Hibbs], $\rho_{\beta}$ can also be expressed as a functional integral over dynamical fields $\phi(\tau, \x)$ defined for Euclidean time $\tau$, subject to the boundary conditions $\phi(0,\x)=\phi(\beta,\x)=\phi_0(\x)$, $$\label{rho}
\rho_{\beta}[\phi_0(\x),\phi_0(\x)]=\int\limits_{\phi(0,\x)=\phi(\beta,\x)=\phi_0(\x)}
[{\cal D}\phi(\tau, \x)]\, e^{-S[\phi(\tau, \x)]}\;.$$ The boundary conditions relate the stochastic variable of the integral in Eq. to the dynamical fields that are integrated over in Eq. . The Hamiltonian $$H[\Pi,\phi]= \frac{1}{2}\Pi^2 + \frac{1}{2}({\bf{\nabla}} \phi)^2 + \frac{m_0^2}{2}\, \phi^2 + U(\phi)\;,$$ which involves the time-dependent field $\phi(\tau, \x)$ and the conjugate momentum $\Pi(\tau,\x)$, leads to the Euclidean action (we assume $d$ spatial dimensions), $$S[\phi]=\int_0^{\beta} (d^Dx)_{_E}\, {\cal L}[\phi]\;,$$ $${\cal L}[\phi]=\frac{1}{2}(\partial_{\tau}\phi)^2+ \frac{1}{2}({\bf{\nabla}} \phi)^2 + \frac{m_0^2}{2}\, \phi^2 + U(\phi)\;.$$ In this paper, we use the shorthands $x \equiv (\tau,\x)$ and $\int (d^Dx)_{_E} \equiv \int_0^\beta d\tau\,\int d^d\x$, with $D=d+1$.
The density matrix is a functional of $\phi_0(\x)$ only. We may write $$\begin{aligned}
\label{sd}
\rho_{\beta}[\phi_0(\x),\phi_0(\x)]\, = \, e^{-S_d[\phi_0(\x)]}\;,\end{aligned}$$ $S_d$ being a certain temperature-dependent dimensionally-reduced action. The field $\phi_0(\x)$, the argument of $S_d$, depends only on the $d$ spatial coordinates; all the $\tau$ dependence of the original $(d+1)$-theory has been eliminated through the $\phi$ integration. The remaining integral over $\phi_0(\x)$, required to obtain the partition function, is unrestricted (except for the vacuum boundary conditions that must be imposed at spatial infinity).
The fields $\phi_0(\x)$ are the natural degrees of freedom of the reduced theory. Any thermal observable can be constructed by integrating the appropriate functional of $\phi_0(\x)$ over the fields $\phi_0(\x)$ weighed with the corresponding diagonal element of the density matrix. The Euclidean time evolution can be viewed as an intermediate step which calculates the weights.
Before proceeding, notice that the density matrix is [*not*]{}, in general, of the form $e^{-\beta H_d}$ with $H_d$ being independent of $\beta$ as in ordinary quantum statistical mechanics; its $\beta$ dependence is far more complicated. This had already been pointed out in the analogous discussion of the transfer matrix carried out in Ref. [@wilsonkogut]. The density matrix provides a direct but alternative way of deriving a dimensionally-reduced theory.
In order to construct generating functionals for the dimensionally-reduced theory, we couple the field $\phi_0(\x)$ to an external current $j(\x)$, obtaining a modified action, $$\begin{aligned}
\label{I[phi0]}
I[\phi_0,j] = S_d[\phi_0] - \beta \int d^d\x \,j(\x)\phi_0(\x)\;.\end{aligned}$$ The associated partition function, $$\begin{aligned}
\label{Z[j,alpha]}
Z[j] = \int[D\phi_0(\x)]\;e^{-I[\phi_0,j]}\end{aligned}$$ is the generating functional of quantum statistical correlation functions.
The quantum statistical connected correlation functions are obtained as functional derivatives of the (Helmholtz) free energy functional $F[j(\x)]$, defined as $$\begin{aligned}
F[j] \,=\, -\frac{1}{\beta}\,\lim_{V \rightarrow \infty}\,\log Z[j] \;.\end{aligned}$$ In particular, the expectation value of the field $\phi_0$ is given by $$\begin{aligned}
\langle \phi_0 (\x)\rangle_j\, =\, -\frac{\delta F[j]}{\delta j(\x)} \;.\end{aligned}$$ The label $j$ in $\langle \phi_0 \rangle_j$ is to stress that such an expectation value is a response of the system to an external current. In the sequel, we will drop the index $j$ to simplify the notation.
Another important quantity in the theory is the effective action $\Gamma$, $$\begin{aligned}
\Gamma [\langle \phi_0(\x)\rangle ] = F[j(\x)] + \int d^d\x \,j(\x)\,\langle \phi_0(\x)\rangle\;,\end{aligned}$$ the Legendre transform of $F[j]$. The argument of the effective action is the expectation value of the field, an intrinsic characteristic of the system. $\Gamma$ is the generating functional of the one-particle irreducible quantum statistical correlations. We shall use it to impose renormalization conditions that connect with the physical (zero-temperature) parameters of the underlying field theory, and to implement a renormalization procedure. From a thermodynamical viewpoint, this effective action is the Gibbs free energy functional, essentially the pressure of the system as a function of $\langle \phi_0(\x)\rangle$. In magnetic systems, for instance, the effective action gives the dependence of the pressure on the magnetization.
Fluctuations at fixed boundary {#fluctuations}
==============================
The standard one-loop computation of the effective action for a position dependent background is performed in detail in Ref. [@amit] (Appendix $6-1$). It makes use of the steepest-descent method, and can be thought as the first term of a semiclassical series. In the present case, one has a double integration to perform, which will require an adaptation of the standard techniques.
The first integral over $\phi(\tau, \x)$, in Eq. , will be dominated by configurations in the vicinity of classical solutions $\phi_c$ satisfying
\[eqmotion\] $$\begin{gathered}
\square_{_E} \phi_c + m_0^2\, \phi_c + U^{\prime}(\phi_c) = 0\;,\\
\phi_c(0,\x) = \phi_c(\beta,\x) = \phi_0(\x)\;,\end{gathered}$$
where $\square_{_E} = -(\partial_\tau^2 + \bs \nabla^2)$ is the Euclidean D’Alembertian operator. Let us, for simplicity, assume that $U$ is a single-well potential. Then, a unique solution $\phi_c(\tau,\x)$ will exist for each boundary configuration $\phi_0(\x)$. The classical solution is a functional $\phi_c[\phi_0]$ of the boundary field. We write $$\begin{gathered}
\nonumber
\phi(\tau,\x) = \phi_c(\tau,\x) +\eta(\tau,\x)\;, \\\label{decomposephic}
\eta(0,\x) = \eta(\beta,\x) = 0\;,\end{gathered}$$ and expand the action around $\phi_c$ to quadratic order in $\eta$, $$\begin{aligned}
S[\phi] = S[\phi_c] + \delta^{(1)}S[\phi_c,\eta] + \delta^{(2)}S[\phi_c,\eta] + {\cal O}(\eta^3)\;.\end{aligned}$$ The first-order variation $\delta^{(1)}S$ is given by $$\begin{aligned}
\label{eq:delta1S}
\delta^{(1)}S
=
\int (d^Dx)_{_E}\;
\frac{1}{2}\left[\square_{_E}\phi_c(x)+m_0^2\,\phi_c(x)+U^\prime(\phi_c(x))
\right] \eta(x)
+
\int d^d\x \left[\eta(x)\,\partial_{\tau}\phi_c(x)\right]_{0}^{\beta}\; ,\end{aligned}$$ where we used the notation $\left[A(\tau)\right]_{0}^{\beta} = A(\beta)- A(0)$. The integrand in the first term of the right hand side of vanishes identically because $\phi_c(x)$ obeys . The second term – a boundary term – vanishes due to the property of the fluctuation $\eta$. Therefore, $\phi_c$ is an extremum of $S$ for $\eta$-like variations (fixed boundary).
The second-order variation $\delta^{(2)}S$ is $$\begin{aligned}
&&
\delta^{(2)}S
=\frac{1}{2}\,
\int (d^Dx)_{_E}
\Big[\Big(\partial_\mu\eta(x)\Big)\Big(\partial^\mu\eta(x)\Big)
\!+\!
m_0^2\,\eta^2(x)
\!+\!
U^{\prime\prime}(\phi_c(x))\,\eta^2(x))\Big]
\nonumber\\
&&
\qquad=\frac{1}{2}\,
\int (d^Dx)_{_E}\;
\partial_\mu\Big[\eta(x)\,\partial^\mu\eta(x)\Big]
\nonumber\\
&&
\qquad\qquad
+\frac{1}{2}\,\int (d^Dx)_{_E}\;
\eta(x)\Big[\square_{_E}+m_0^2+U^{\prime\prime}(\phi_c(x))\Big]\,\eta(x)\; .\end{aligned}$$ Using again , we obtain, up to quadratic order in $\eta$, $$\begin{aligned}
S[\phi] \approx S[\phi_c] + \frac{1}{2}\,
\int (d^Dx)_{_E}\;
\eta(x)\,\Big[\square_{_E} + m_0^2+ U^{\prime\prime}(\phi_c(\x))\Big]\,\eta(x)\;.\end{aligned}$$ It is convenient to introduce the Green’s function
\[eq:prop\] $$\begin{aligned}
\left[\square_{_E} + m_0^2 + U^{\prime\prime}\left(\phi_c(x)\right)\right]
G[\phi_c](x;x^\prime)
=\delta^{(4)}(x-x^\prime)\\
G[\phi_c](\tau,\x;0,\x^\prime)=G[\phi_c](\tau,\x;\beta,\x^\prime)=0\;.\end{aligned}$$
In terms of $G[\phi_c]$, we obtain $$\label{intovereta}
\int\limits_{\phi(0,\x)=\phi(\beta,\x)=\phi_0(\x)}\!\!\!\!\!\!\!\!\!\!\!\![D\phi(\tau,\x)]
\,e^{-S[\phi]}\; \approx \;e^{-S[\phi_c]}\; (\det G[\phi_c])^{1/2}\;.$$ For single-well potentials it can be shown that $\delta^{(2)}S > 0$, a necessary condition for $\phi_c[\phi_0]$ to be a minimum of $S$.
Building the classical solution {#building}
-------------------------------
In Appendix A, we obtain the following recursive relation for the classical solution $\phi_c[\phi_0]$, $$\begin{aligned}
&&
\phi_c(\tau,\x)
=
\int d^d\x^\prime\;
\phi_0(\x^\prime)\;
\Big[
\partial_{\tau^\prime}G_0(\tau,\x;\tau^\prime,\x^\prime)
\Big]_{0}^{\beta}
\nonumber\\
&&
\qquad\qquad\qquad
-\int_{0}^{\beta}d\tau^\prime\int d^d\x^\prime\;
G_0(\tau,\x;\tau^\prime,\x^\prime)\;
U^\prime(\phi_c(\tau^\prime,\x^\prime))
\; ,
\label{eq:green1}\end{aligned}$$ where $G_0(x,x^\prime)$ is a Green’s function of the free operator
\[eq:bound\_prop\] $$\begin{gathered}
\left[\square_{_E} +m_0^2\right]
G_0(\tau,\x;\tau^\prime,\x^\prime)
=\delta^{(4)}(x-x^\prime)\;, \\
G_0(\tau,\x;0,\x^\prime)
=
G_0(\tau,\x;\beta,\x^\prime)
=0\; .\end{gathered}$$
Diagrammatically, the classical solution $\phi_c(x)$ can be represented as the sum of all the tree diagrams terminated by the boundary field $\phi_0(\x)$. It is also demonstrated in Appendix A that $$\begin{aligned}
\label{jacobian}
\frac{\delta \phi_c(\tau,\x)}{\delta \phi_0(\y)} &= \left[\partial_{\tau^{\prime}} G[\phi_c](\tau,\x;\tau^{\prime},\y)\right ]_0^\beta\;.\end{aligned}$$ From the previous equation, if one knows the classical solution ${\overline{\phi}_c}$ associated to some boundary configuration ${\overline{\phi}_0}$, the approximate classical solution associated to the following fluctuation of the boundary condition, $$\begin{aligned}
\phi_0(\x) = \overline{\phi}_0(\x) +\xi(\x)\,\end{aligned}$$ is given by $$\begin{aligned}
\label{phicofphi0}
\phi_c[\phi_0](x) = {\overline{\phi}_c}(x) + \int d^d\y\, \xi(\y)\,\left[\partial_{\tau^{\prime}} G[{\overline{\phi}_c}](x;\tau^{\prime},\y)\right ]_0^\beta\;+\;{\cal O}(\xi^2)\;,\end{aligned}$$ as found in [@Bessa:2007vq]. Another useful relation is $$\begin{aligned}
\label{jacobian2}
\frac{\delta \left [\partial_\tau \phi_c(\tau,\x)\right ]_0^\beta}{\delta \phi_0(\y)} &= \left[\partial_{\tau}\partial_{\tau^{\prime}} G[\phi_c](\tau,\x;\tau^{\prime},\y)\right ]_0^\beta\;.\end{aligned}$$
Finally, notice that $\phi_c[\phi_0]$ is, in general, a non-periodic function of $\tau$. As long as the only condition on the configurations that enter in the calculation of $Z[j]$ is to coincide at $\tau=0$ and $\beta$, non-periodic configurations are allowed.
Fluctuating the boundary field {#bcfluctuation}
==============================
From the first functional integration (see Eq. ) comes out an approximation for $S_d$ defined in , $$\begin{aligned}
S_{d}[\phi_0] = S[\phi_c[\phi_0]] + \frac{1}{2}\hbox{ Tr} \log G^{-1}[\phi_c[\phi_0]]\;.\end{aligned}$$ $S_d$ works as the classical action of the theory. Using again the recipe to obtain the one-loop quantum effective action (now for the expectation value of the field $\phi_0(\x)$), we arrive at $$\begin{aligned}
\label{gammasd}
\beta\,\Gamma[\langle \phi_0 \rangle] &= S_{d}[\langle \phi_0\rangle] + \frac{1}{2} \hbox{ Tr} \log \delta^{(2)}S_d[\langle\phi_0 \rangle]\;,\end{aligned}$$ where $\delta^{(2)}S_d[\langle\phi_0 \rangle]$ is a shorthand for the second-order variation of $S_d$ around $\langle \phi_0 \rangle$. However, to be consistent with the one-loop approximation, $\delta^{(2)}S_d[\langle\phi_0 \rangle]$ has to be replaced by $\delta^{(2)}S[\langle\phi_0 \rangle]$. In fact, the expression for $\Gamma$ as given in contains only part of the higher-order terms. Therefore, we are left with the following one-loop effective action, $$\begin{aligned}
\label{gammasd2}
\beta\,\Gamma [\langle \phi_0\rangle]&= S[\phi_c] + \frac{1}{2} \hbox{ Tr} \log G^{-1}[\phi_c] + \frac{1}{2} \hbox{ Tr} \log \left [\frac{\delta^2 S[\phi_c]}{\delta \langle \phi_0 \rangle^2}\right ]\;,\end{aligned}$$ where $\phi_c$ in this case is $\phi_c[\langle \phi_0 \rangle]$.
It is convenient to integrate the classical action by parts, obtaining $$\begin{aligned}
\label{actionintparts}
S[\phi_c] &= \int (d^Dx)_{_E} \, \left (\frac{1}{2} \phi_c \square_{_E} \phi_c + \frac{1}{2}m_0^2\, \phi_c^2 + U(\phi_c)\right ) + \frac{1}{2}\int d^d\x \,\phi_0(\x)\,\left [\partial_\tau \phi_c (\tau,\x)\right]_{0}^{\beta}\;.\end{aligned}$$ The presence of a boundary term is a direct consequence of the non-periodicity of $\phi_c[\phi_0]$, pointed out in the previous section. Such boundary contribution introduces nonstandard terms in the calculations.
At one-loop order, the expectation value of $\phi_0$ is given by the saddle-point of the modified action $I[\phi_0,j]$ (see Eq. ). Let us call ${\overline{\phi}_0}$ such an extremal configuration. Fluctuating the boundary field as $$\begin{aligned}
\label{boundaryfluct}
\phi_0(\x) = \overline{\phi}_0(\x) + \xi(\x)\;,\end{aligned}$$ and denoting $\phi_c[{\overline{\phi}_0}]$ by ${\overline{\phi}_c}$, the first-order variation of $I[\phi_0,j]$ is given by $$\begin{aligned}
\nonumber
\delta^{(1)} I[\phi_0,j] & = \int (d^Dx)_{_E} \,\left [ \square_{_E} {\overline{\phi}_c}(x) + m_0^2\, {\overline{\phi}_c}(x) + U^{\prime}({\overline{\phi}_c}(x))\right ]\,(\phi_c(x)-{\overline{\phi}_c}(x))\\
& \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; +\int d^d\x \,\xi(\x)\,\left \{ [\partial_\tau {\overline{\phi}_c}(\tau,\x) ]_0^\beta -j(\x) \right \}\;.\label{delta1I}\end{aligned}$$ By construction, $\phi_c[{\overline{\phi}_0}]$ satisfies the equation of motion , so that the first term on the r.h.s. of vanishes. From , we obtain the saddle-point condition [^5] for ${\overline{\phi}_0}$: $$\begin{aligned}
\label{saddlej}
[\partial_\tau \phi_c[{\overline{\phi}_0}](\tau,\x) ]_0^\beta = j(\x)\;.\end{aligned}$$ Notice that Eq. is a condition on the time derivative of $\phi_c$. Therefore, whenever $j(\x) \neq 0$ (even for constant currents) the classical configuration $\phi_c[{\overline{\phi}_0}]$ will depend on $\tau$ (even in the free case).
For the second-order variation, we obtain $$\begin{aligned}
\delta^2 I[\phi_0] = \delta^2 I^{(a)}[\phi_0] + \delta^2 I^{(b)}[\phi_0]\;,\end{aligned}$$ where $$\begin{aligned}
\nonumber
\delta^2 I^{(a)}[\phi_0] = \frac{1}{2}\int (d^Dx)_{_E} (d^Dx)_{_E}^{\prime} (d^Dz)_{_E}\, &\left[\partial_{\tau^{\prime}} G[{\overline{\phi}_c}](z;\tau^{\prime},\x)\right ]_0^\beta \xi(\x) \left [ \square_{_E} + m_0^2 + U^{\prime\prime}({\overline{\phi}_c}(z))\right ] \\
&\hspace{3cm}\times\, \left[\partial_{\tau^{\prime}} G[{\overline{\phi}_c}](z;\tau^{\prime},\x^{\prime})\right ]_0^\beta \xi(\x^{\prime})\end{aligned}$$ and $$\begin{aligned}
\label{delta2I}
\delta^2 I^{(b)}[\phi_0] = \frac{1}{2}\int d^d\x\, d^d\x^{\prime}\, \xi(\x)\, \left [\partial_\tau\partial_{\tau^{\prime}}G[{\overline{\phi}_c}](\tau,\x;\tau^{\prime},\x^{\prime})\right]_0^\beta\,\xi(\x^{\prime})\;.\end{aligned}$$ However, only the term $\delta^2 I^{(b)}[\phi_0]$ contributes, because $$\begin{aligned}
\label{dazero!}
\left [ \square_{_E} + m_0^2 + U^{\prime\prime}({\overline{\phi}_c}(z))\right ] \left[\partial_{\tau^{\prime}} G[{\overline{\phi}_c}](z;\tau^{\prime},\x^{\prime})\right ]_0^\beta =0\;.\end{aligned}$$ Using that the second-order variation of $S[\phi_c[\phi_0]]$ and $I[\phi_0,j]$ coincide, we can write $$\begin{aligned}
\label{sndvariation}
\frac{\delta^2 S[\phi_c[\phi_0]]}{\delta \phi_0(\x)\delta \phi_0(\x^{\prime})}\bigg |_{\phi_0=\overline{\phi}_0}&= \left [\partial_\tau\partial_{\tau^{\prime}}G[{\overline{\phi}_c}](\tau,\x;\tau^{\prime},\x^{\prime})\right]_0^\beta\,.\end{aligned}$$ We now have: $$\begin{aligned}
\label{aux:gammaA}
\beta\,\Gamma[{\overline{\phi}_0}] &= S[{\overline{\phi}_c}] + \frac{1}{2} \hbox{ Tr} \log G^{-1}[{\overline{\phi}_c}] + \frac{1}{2} \hbox{ Tr} \log\left( \left [\partial_\tau\partial_{\tau^{\prime}}G[{\overline{\phi}_c}]\right ]_0^\beta\right)\,.\end{aligned}$$ Using Eq. written in terms of $G_0$ (see Eq. ), we are led to $$\begin{aligned}
\left [\partial_\tau\partial_{\tau^{\prime}}G[{\overline{\phi}_c}]\right ]_0^\beta & = \left[\partial_\tau\partial_{\tau^{\prime}}\left\{ G_0\left( 1+ G_0 U^{\prime\prime}({\overline{\phi}_c})\right)^{-1}\right\}\right]_0^\beta\;.\end{aligned}$$ We may sum the series to obtain $$\begin{aligned}
\left [\partial_\tau\partial_{\tau^{\prime}}G[{\overline{\phi}_c}]\right ]_0^\beta
&= \left[\partial_\tau\partial_{\tau^{\prime}} G_0\right]_0^\beta + \left[\partial_\tau G_0\right]_0^\beta U^{\prime\prime}({\overline{\phi}_c})G_0^{-1}G[{\overline{\phi}_c}]\left[\partial_{\tau^{\prime}} G_0\right]_0^\beta \,.\end{aligned}$$ When we take the logarithm, we will have $$\begin{aligned}
\log \left [\partial_\tau\partial_{\tau^{\prime}}G[{\overline{\phi}_c}]\right ]_0^\beta= \log {\cal G}_0 + \log \left[1 + {\cal G}_0^{-1}\left(\partial_\tau G_0 U^{\prime\prime}({\overline{\phi}_c})G_0^{-1}G[{\overline{\phi}_c}]\partial_{\tau^{\prime}} G_0\right)\right],\end{aligned}$$ where we have denoted $\left[\partial_\tau\partial_{\tau^{\prime}} G_0\right]_0^\beta$ as ${\cal G}_0$. Taking the trace and using that $\hbox{ Tr}\, (AB)^n = \hbox{ Tr}\, (BA)^n$ to reorganize the series, we obtain $$\begin{aligned}
\hbox{ Tr} \log \left [\partial_\tau\partial_{\tau^{\prime}}G[{\overline{\phi}_c}]\right ]_0^\beta
&= \hbox{ Tr} \log {\cal G}_0 + \hbox{ Tr}\log \left[1 + G_0^{-1}G[{\overline{\phi}_c}]\left (\partial_{\tau^{\prime}} G_0{\cal G}_0^{-1}\partial_\tau G_0 \right ) U^{\prime\prime}({\overline{\phi}_c})\right]\,.\end{aligned}$$ We may write $$\begin{aligned}
\log \left ( G^{-1}[{\overline{\phi}_c}]\right ) = \log \left (G_0^{-1} \right ) \, +\,\log \big (G^{-1}[{\overline{\phi}_c}]G_0\big)\;.\end{aligned}$$ Therefore, $$\begin{aligned}
\nonumber
\hbox{ Tr} \log\left( G^{-1}[{\overline{\phi}_c}]\right) &+ \hbox{ Tr}\log\left( \left [\partial_\tau\partial_{\tau^{\prime}}G[{\overline{\phi}_c}]\right ]_0^\beta\right) = \hbox{ Tr}\log \left( G_0^{-1}\right ) + \hbox{ Tr}\log {\cal G}_0 \,+ \\\label{aux:nada}
&\;\;\;\;\;\;\;\;\;\;\;\;+\hbox{ Tr}\log \left \{1 + \left (G_0 + \partial_\tau G_0 {\cal G}_0^{-1}\partial_{\tau^{\prime}}G_0 \right) U^{\prime\prime}({\overline{\phi}_c}) \right \}\;.\end{aligned}$$ In Appendix B we show that the combination which appears multiplying $U^{\prime\prime}({\overline{\phi}_c})$ in is the usual thermal propagator $\Delta_F$ which is given, in spatial Fourier space, by $$\begin{gathered}
\label{aux:thermalprop}
\Delta_F(\k;\tau,\tau^\prime) = \frac{1}{2\omega_\k}\left [(1+n(\omega_\k))e^{-\omega_k|\tau-\tau^\prime|} + n(\omega_\k)e^{\omega_k|\tau-\tau^\prime|} \right ]\;,\end{gathered}$$ where $\omega_\k = \sqrt{\k^2+ m_0^2}$ and $n(\omega_\k)$ is the Bose-Einstein distribution. In terms of $\Delta_F$, one finally obtains $$\begin{aligned}
\label{Gammafinal}
\beta\,\Gamma[{\overline{\phi}_0}] &= S[{\overline{\phi}_c}] + \frac{1}{2}\hbox{ Tr} \log \left (G_0^{-1}\right ) + \frac{1}{2}\hbox{ Tr}\log {\cal G}_0 +\frac{1}{2} \hbox{ Tr}\log \left (1 + \Delta_F\,U^{\prime\prime}({\overline{\phi}_c}) \right )\;.\end{aligned}$$ Eq. can be further simplified to $$\begin{aligned}
\label{Gammafinalb}
\beta\,\Gamma[{\overline{\phi}_0}] &= S[{\overline{\phi}_c}] + \frac{1}{2}\hbox{ Tr} \log \left (G_\beta^{-1}[{\overline{\phi}_c}]\right )\;,\end{aligned}$$ where, using that $\hbox{ Tr} \log \left (G_0^{-1}\right ) + \hbox{ Tr}\log {\cal G}_0 = \hbox{ Tr} \log \Delta_F^{-1}$ (see Appendix B), $$\begin{aligned}
G_\beta[\phi_c]^{-1} = \Delta_F^{-1} + U^{\prime\prime}(\phi_c)\;.\end{aligned}$$ Notice that the dependence of $\Gamma$ on the physical field $\phi_0$ comes from the nontrivial functional dependence of the classical configuration $\phi_c$ on $\phi_0$.
Renormalization procedure for the single-well quartic interaction (D=4) {#renormalization}
=======================================================================
The effective action in Eq. is written in terms of non-renormalized parameters. We conclude from the above calculation that $\Gamma$ has the usual zero-temperature expression when viewed as a function of the classical field $\phi_c$. This fact suggests that the renormalization procedure should follow the standard recipe. In this section, we verify that for the single-well quartic interaction in d = 3 spatial dimensions.
We start by introducing a cutoff $\Lambda$ for integrations over momenta and adding to $\Gamma$ couterterms in order to obtain a renormalized effective action, $$\begin{aligned}
\label{Gammafinalc}
\beta\,\Gamma_R[{\overline{\phi}_0}] &= \beta\, \Gamma[{\overline{\phi}_0}] - \frac{C_1}{2}\,\int (d^4x)_{_E} \,{\overline{\phi}_c}^2(x) - \frac{C_2}{4}\,\int (d^4x)_{_E} \,{\overline{\phi}_c}^4(x) - \frac{C_3}{2}\,\int (d^4x)_{_E} \,\left (\partial_\mu {\overline{\phi}_c}\right)^2(x)\;.\end{aligned}$$ Suitable choices of $C_1$, $C_2$ and $C_3$ should render $\Gamma_R$ finite. Renormalization conditions can be fixed using correlation functions of the ${\overline{\phi}_c}$ fields. Each choice of counterterms under that recipe will correspond to a definite set of renormalization conditions in terms of physical correlations involving $\phi_0$. That translation demands, in principle, the full knowledge of $\phi_c$ as a function of ${\overline{\phi}_0}$.
Let us call $v_0(\x)$ the saddle-point boundary configuration corresponding to $j(\x)\equiv 0$. Functional derivatives of $\Gamma_R$ with respect to ${\overline{\phi}_0}(\x)$ evaluated at $v_0(\x)$ lead to the $n$-point $1PI$ renormalized vertex function. For single-well potentials, one finds $v_0(\x) = 0$, and $\Gamma_R[{\overline{\phi}_0}]$ admits the expansion $$\begin{aligned}
\Gamma_R[{\overline{\phi}_0}] = \sum_{n=1}^{\infty}\frac{1}{n!} \int \Gamma_R^{(n)}(\x_1,\ldots,\x_n)\,{\overline{\phi}_0}(\x_1)\,\ldots\,{\overline{\phi}_0}(\x_n)\, d^{3}\x_1\ldots d^{3}\x_n \;,\end{aligned}$$ where $$\begin{aligned}
\label{gammaR}
\Gamma_R^{(n)}(\x_1,\ldots,\x_n) = \frac{\delta^{(n)}\Gamma_R}{\delta {\overline{\phi}_0}{}_1 \ldots \delta {\overline{\phi}_0}{}_n}\bigg |_{{\overline{\phi}_0}=0}\;.\end{aligned}$$ One can also expand $\Gamma_R[{\overline{\phi}_0}]$ in powers of ${\overline{\phi}_c}[{\overline{\phi}_0}]$ (using that $\phi_c[0]=0$), $$\begin{aligned}
\label{npointusual}
\Gamma_R[{\overline{\phi}_0}] = \sum_{n=1}^{\infty}\frac{1}{n!} \int \widetilde{\Gamma}_R^{(n)}(x_1,\ldots,x_n)\,{\overline{\phi}_c}(x_1)\,\ldots\,{\overline{\phi}_c}(x_n)\, d^{4}x_1\ldots d^{4}x_n \;,\end{aligned}$$ where $$\begin{aligned}
\label{gammaRc}
\widetilde{\Gamma}_R^{(n)}(x_1,\ldots,x_n) = \frac{\delta^{(n)}\Gamma_R}{\delta {\overline{\phi}_c}{}_1 \ldots \delta {\overline{\phi}_c}{}_n}\bigg |_{{\overline{\phi}_c}=0}\;.\end{aligned}$$ A standard calculation yields $$\begin{aligned}
\frac{1}{(2\pi)^4\delta^{(4)}(p_1+p_2)}\widetilde{\Gamma}_R^{(2)}(p_1,p_2) = p_1^2 + m_0^2 + \frac{\lambda}{2}\hbox{ Tr} \Delta_F - C_1 -C_3\,p_1^2 \end{aligned}$$ and $$\begin{aligned}
\frac{1}{(2\pi)^4\delta^{(4)}( \sum_{i}^4 p_i)}\widetilde{\Gamma}_R^{(4)}(p_1,p_2,p_3,p_4) = \lambda - \frac{\lambda^2}{2}I_1 -6C_2 \;,\end{aligned}$$ where $$\begin{aligned}
I_1 = \int^{\Lambda} \frac{d^4k}{(2\pi)^4}\left [\Delta_F(k)\Delta_F(k+p_3+p_4) +\Delta_F(k)\Delta_F(k+p_1+p_4) +\Delta_F(k)\Delta_F(k+p_3+p_1) \right ] \,.\end{aligned}$$
For massive theories, we use the following zero-temperature renormalization conditions, $$\begin{aligned}
&\frac{1}{(2\pi)^4\delta^{(4)}(0)}\widetilde{\Gamma}_R^{(2)}(0,0) = m_0^2\;,\\
&\frac{1}{(2\pi)^4\delta^{(4)}(0)}\frac{d\widetilde{\Gamma}_R^{(2)}}{dp_1^2}(0,0) = 1\;, \\
&\frac{1}{(2\pi)^4\delta^{(4)}(0)}\widetilde{\Gamma}_R^{(4)}(0,\ldots,0) = \lambda_R\,.\end{aligned}$$ In this case, the counterterms are $$\begin{aligned}
\label{CT1}
C_1 = \frac{\lambda}{2}\int^{\Lambda}\frac{d^4k}{(2\pi)^4}\Delta_F^0(k)\;,\;\;\;\;\;\;\;C_2 = -\frac{\lambda^2}{4}\int^{\Lambda} \frac{d^4k}{(2\pi)^4}(\Delta_F^0(k))^2\;\;\;\;\;\;\;\hbox{and}\;\;\;\;\;C_3 =0\;,\end{aligned}$$ where $$\begin{aligned}
\Delta_F^0(k) = \frac{1}{k^2 + m_0^2}\end{aligned}$$ is the zero-temperature free propagator in 4-dimensional Euclidean Fourier space.
Alternatively, the renormalization conditions can be expressed in terms of correlations of the fields ${\overline{\phi}_0}$. In principle, the general n-point function $\Gamma_R^{(n)}$ can be obtained from if one writes $\phi_c[\phi_0]$ in a power series of $\phi_0$, as given by the implicit relation . In particular, $\Gamma_R^{(4)}$ is a complicated combination of $\tilde{\Gamma}_R^{(2)}$ and $\tilde{\Gamma}_R^{(4)}$. On the other hand, the 2-point function admits a simple expression, $$\begin{aligned}
\frac{\beta}{(2\pi)^3\delta(\p_1+\p_2)}\,\Gamma_R^{(2)}(\p_1,\p_2) = {\cal G}_0(\p_1)
+ \left(\frac{\lambda}{2}\,\hbox{ Tr} \Delta_F-C_1- C_3 \p_1^2 \right)\,\int d\tau \,\left[\partial_{\tau^{\prime}} G_0(\tau,\p_1)\right ]_0^\beta\left[\partial_{\tau^{\prime}} G_0(\tau,-\p_1)\right ]_0^\beta\;.\end{aligned}$$ In particular, using the counterterms given in Eq. , we obtain, in Fourier space, $$\begin{aligned}
\nonumber
&\frac{\beta}{(2\pi)^3\delta(\p_1+\p_2)}\,\Gamma_R^{(2)}(\p_1,\p_2) = 2\omega_{\p_1}\,\tanh \beta \omega_{\p_1}/2 +\frac{\lambda_R}{24\beta}\,\frac{\tanh \beta \omega_{\p_1}/2}{\beta\omega_{\p_1}} \, \left (1+ \frac{\beta\omega_{\p_1}}{\sinh \beta \omega_{\p_1}}\right )\;.\end{aligned}$$ When the zero-temperature theory is massless, we adopt the following set of renormalization conditions,
$$\begin{aligned}
&\frac{1}{(2\pi)^4\delta^{(4)}(0)}\widetilde{\Gamma}_R^{(2)}(\mu,-\mu) = \mu^2\;,\\
&\frac{1}{(2\pi)^4\delta^{(4)}(0)}\frac{d\widetilde{\Gamma}_R^{(2)}}{dp_1^2}(\mu,-\mu) = 1\;, \\
&\frac{1}{(2\pi)^4\delta^{(4)}(0)}\widetilde{\Gamma}_R^{(4)}(p_1,\ldots,p_i)\bigg |_{\mu} = \lambda_R\;,\end{aligned}$$
where the arguments of the four-point function are such that $\sum_i p_i=0$, $p_i^2=\mu^2$ and $(p_i+p_j)^2=4\mu^2/3$ for $i\neq j$. The quantity $\mu$ is the renormalization scale. We obtain the counterterms $$\begin{aligned}
\label{CT}
C_1 = \frac{\lambda}{2}\int^{\Lambda}\frac{d^4k}{(2\pi)^4}\Delta_F^0(k)\;,\;\;\;\;\;\;\;C_2 = -\frac{\lambda^2}{4}\int^{\Lambda} \frac{d^4k}{(2\pi)^4}\Delta_F^0(k)\Delta_F^0(k+\mu)\;\;\;\;\;\;\;\hbox{and}\;\;\;\;\;C_3 =0\;.\end{aligned}$$ The renormalized coupling constant $\lambda_R$ satisfies the following renormalization group equation, $$\begin{aligned}
\mu \,\frac{d\lambda_R}{d\mu} = \frac{3\lambda_R^2}{16\pi^2}\;+\;{\cal O}(\lambda_R^3)\;.\end{aligned}$$ As a consequence, $\lambda_R$ runs with the renormalization scale $\mu$ as $$\begin{aligned}
\lambda_R(\mu) = \lambda_R(2\pi /\beta) + \frac{3\lambda_R^2(2\pi/\beta)}{16 \pi^2}\,\log\frac{\mu}{2\pi/\beta}\,+\,{\cal O}(\lambda_R^3)\;.$$ The function $\Gamma_R^{(2)}$ in the massless case is given by $$\begin{aligned}
\label{aux:gamma2definem2}
&\frac{\beta}{(2\pi)^3\delta(\p_1+\p_2)}\,\Gamma_R^{(2)}(\p_1,\p_2;\mu) = 2|\p_1|\,\tanh \beta |\p_1|/2 +\frac{\lambda_R(\mu)}{24\beta}\,\frac{\tanh \beta |\p_1|/2}{\beta|\p_1|} \, \left (1+ \frac{\beta|\p_1|}{\sinh \beta |\p_1|}\right )\;.\end{aligned}$$ The second term on the r.h.s. of plays the role of a mass term for the boundary theory (the first term is the kinetic part of that theory). The corresponding dynamical mass comes from the interaction of the boundary field with the $\eta$ fields and strongly depends on the external momentum, as shown in Fig. \[Fig:dynamass\].
One can think of an effective mass of the boundary field as the zero external momentum limit of the two-point function, $$\begin{aligned}
\frac{1}{(2\pi)^3\delta(\p_1+\p_2)}\,\Gamma_R^{(2)}(0,0;\mu) \;=\; \frac{\lambda_R(\mu)}{24\beta^2}\;.\end{aligned}$$ This effective mass coincides with the first-order perturbative result for the thermal mass ($m_{PT}$). The inverse of $\Gamma_R^{(2)}$ can be thought as the 3-dimensional propagator $G^{(2)}(\p_1,\p_2)$ of the field $\phi_0$. Numerical calculation shows the asymptotic behavior of $G^{(2)}(|x_1-x_2|)$ as $$\begin{aligned}
G^{(2)}(|\x|) \propto \frac{\exp^{-m_{PT} |x|}}{|x|}\;,\end{aligned}$$ which reinforces the role played by $m_{PT}$ as the mass scale for the boundary theory.
Conclusions
===========
Naive perturbation theory applied to finite-temperature field theory is doomed to failure due to severe bosonic infrared divergences [@FTFT-books]. These are related to zero modes in the Matsubara frequencies, and introduce new mass scales associated with collective modes of the system. Incorporating this fact in the description of a thermal system in equilibrium naturally leads to the possibility of reorganizing the perturbative series, to different optimizations in the diagrammatic formulation, and ultimately to the construction of effective field theories [@Kraemmer:2003gd; @Andersen:2004fp]. The process will be successful in the measure that one finds appropriate quasiparticles or degrees of freedom to anchor the effective theory.
In this paper we have built an effective field theory that describes the quantum statistical correlations in thermal field theories using the expectation value of the boundary fields, which are the effective degrees of freedom, our natural stochastic variables and zero modes. This [*boundary effective theory*]{} (BET) is dimensionally-reduced by construction, as well as the thermal correlation functions it provides, which can be directly associated with the equilibrium thermal system.
To illustrate our formulation of the BET, we considered a scalar self-interacting field to one loop, but the procedure can in principle be extended to gauge fields and we believe that the result holds for higher loop orders. In the scalar case, we proved that the one-loop effective action at finite temperature has the same renormalized expression as at zero temperature provided it is written in terms of $\phi_{c}$ and zero-temperature propagators are substituted by their thermal versions. The classical solution $\phi_c(\tau,\x)$ is a functional of the boundary configuration $\phi_0(\x)$ and carries, in the underlying statistical mechanical problem, the dynamical information associated to the quantum functional integral formalism. Using that connection, one can compute directly all the renormalized thermal vertex functions and read off the renormalization group running of all parameters.
Even at one-loop order, the functional dependence of $\phi_c$ on $\phi_0$ led to a highly non-perturbative effective action for the boundary field. Our result for $\Gamma_R^{(2)}$ shows that, although the dynamical mass of the new degrees of freedom — the boundary fields — have a complicated momentum dependence, its zero external momentum limit coincides with the first-order perturbative result for the thermal mass and that $G^{(2)}$ has the correct asymptotic behavior as a function of $|x|$, an effect of dynamical screening. These are non-obvious consistency checks of the procedure.
The rich structure encoded in the d-dimensional vertex functions obtained from the new degrees of freedom of BET will, of course, affect the thermodynamic functions and seem to provide a more direct and natural way to compute quantities such as the pressure. These results will be presented in a future publication [@future].
This work was partially supported by CAPES, CNPq, FAPERJ, FAPESP and FUJB/UFRJ.
In this Appendix, we calculate the derivative of the functional $\phi_c[\phi_0]$. First of all, we multiply Eq. by the classical field $\phi_c(\tau^\prime,\x^\prime)$, and integrate over $\tau^\prime$ and $\x^\prime$. This gives $$\label{phic=aux}
\phi_c(\tau,\x)=
\int_{0}^{\beta}d\tau^\prime\int d^d\x^\prime\;
\phi_c(\tau^\prime,\x^\prime)
\left[\square_{_E} +m_0^2\right]G_0(\tau,\x;\tau^\prime,\x^\prime)\; .$$ Now, multiply the equation of motion (Eq. ) for $\phi_c(\tau^\prime,\x^\prime)$ by the Green’s function $G_0(\tau,\x;\tau^\prime,\x^\prime)$ and integrate over $\tau^{\prime}$ and $\x^\prime$: $$\begin{aligned}
\label{eqmotiontimesG}
0 = \int_{0}^{\beta}d\tau^\prime\int d^d\x^\prime\,G_0(\tau,\x;\tau^{\prime},\x^{\prime})\,\left [(\square_{_E}+m_0^2)\phi_c(\tau,\x^{\prime}) + U^{\prime}(\phi_c(\tau^{\prime},\x^{\prime})) \right ]\;.\end{aligned}$$ Subtracting from leads to $$\begin{aligned}
&&
\phi_c(\tau,\x)
=
\int_{0}^{\beta}d\tau^\prime\int d^d\x^\prime\;
G_0(\tau,\x;\tau^\prime,\x^\prime)
\Big[
\stackrel{\leftarrow}{\square}_{_E}
-
\stackrel{\rightarrow}{\square}_{_E}
\Big]
\phi_c(\tau^\prime,\x^\prime)
\nonumber\\
&&
\quad
\;\;\;\;\;\;\;\;\;-\int_{0}^{\beta}d\tau^\prime\int d^d\x^\prime\;
G_0(\tau,\x;\tau^\prime,\x^\prime)\;
U^\prime(\phi_c(\tau^\prime,\x^\prime))
\; ,
\nonumber\\
&&
\label{eq:green}\end{aligned}$$ where the arrows on the differential operators on the first line indicate on which side they act. The first line can be rewritten as a boundary term, by noting that $$A\Big[\stackrel{\rightarrow}{\partial_\mu^2}
-
\stackrel{\leftarrow}{\partial_\mu^2}\Big]B
=
\partial^\mu \left\{
A\Big[\stackrel{\rightarrow}{\partial_\mu}
-
\stackrel{\leftarrow}{\partial_\mu}\Big]B
\right\}\; .$$ In eq. (\[eq:green\]), the boundary in the spatial directions does not contribute to the classical field at the point $\x$ because the free propagator decreases fast enough when the spatial separation increases. Thus, we are left with only a contribution from the boundaries in time. At this point, since the boundary conditions for $\phi_c$ consist in specifying the value of the field at $\tau^\prime=0,\beta$, while its first time derivative is not constrained, it is very natural to choose a Green’s function $G_0$ that obeys the following conditions, $$G_0(\tau,\x;0,\x^\prime)
=
G_0(\tau,\x;\beta,\x^\prime)
=0\; .
\label{eq:bound_propb}$$ With this choice of the propagator, we obtain formula for $\phi_c(\tau,\x)$. From , we calculate $$\begin{aligned}
\nonumber
\frac{\delta \phi_c(\tau,\x)}{\delta \phi_0(\y)} &= \int d^d\x^{\prime}\,\delta(\x-\y)\,\left [\partial_{\tau^{\prime}}G_0(x;\tau^{\prime},\x^{\prime}) \right]_0^\beta \\
&\;\;\;\;\;\;\;\;\;\;\;\;\; -\int (d^Dx)_{_E}^{\prime}G_0(x,x^{\prime})U^{\prime\prime}(\phi_c(x^{\prime}))\,\frac{\delta \phi_c(\tau^{\prime},\x^{\prime\prime})}{\delta \phi_0(\y)}\\
&= \left [\partial_{\tau^{\prime}}G_0(x;\tau^\prime,\y) \right]_0^\beta - \int (d^Dx)_{_E}^{\prime}G_0(x,x^{\prime})U^{\prime\prime}(\phi_c(x^{\prime}))\,\frac{\delta \phi_c(\tau^{\prime},\x^{\prime\prime})}{\delta \phi_0(\y)}\end{aligned}$$ Multiplying the previous equation by $G_0^{-1}(z,x)$ and integrating over $x$, we obtain $$\begin{aligned}
\nonumber
\int (d^Dx)_{_E}\, G_0^{-1}(z,x)\,\frac{\delta \phi_c(\tau,\x)}{\delta \phi_0(\y)} &=
\int (d^Dx)_{_E}\, G_0^{-1}(z,x)\,\left [\partial_{\tau^{\prime}}G_0(x;\tau^\prime,\y) \right]_0^\beta \\
&\;\;\;\;\;\;\;- \int (d^Dx)_{_E} (d^Dx)_{_E}^{\prime} \,G_0^{-1}(z,x)G_0(x,x^{\prime})U^{\prime\prime}(\phi_c(x^{\prime}))\,\frac{\delta \phi_c(\tau^{\prime},\x^{\prime\prime})}{\delta \phi_0(\y)}\;.\end{aligned}$$ Using that $G_0^{-1}(z,x) = \delta^{(4)}(z-x) \,(\square_{_E} + m_0^2)$, we can write $$\begin{aligned}
(\square_{_E} + m_0^2)\,\frac{\delta \phi_c(z)}{\delta \phi_0(\y)} + U^{\prime\prime}(\phi_c(z))\frac{\delta \phi_c(z)}{\delta \phi_0(\y)} = \int (d^Dx)_{_E}\, G_0^{-1}(z,x)\,\left [\partial_{\tau ^{\prime}} G_0(\x,\tau ^{\prime};\y) \right]_0^\beta\;.\end{aligned}$$ From , we have $$\begin{aligned}
G^{-1}[\phi_c](z^{\prime},z) = \delta^{(4)}(z^{\prime}-z)\,G_z^{-1}[\phi_c]\;,\end{aligned}$$ where $G_z^{-1}[\phi_c] = \square_{_E} + m_0^2 + U^{\prime\prime}(\phi_c(z))$. Therefore, $$\begin{aligned}
G^{-1}_z[\phi_c]\,\frac{\delta \phi_c(z)}{\delta \phi_0(\y)} = \int (d^Dx)_{_E}\, G_0^{-1}(z,x)\,\left [\partial_{\tau ^{\prime}} G_0(\x,\tau ^{\prime};\y) \right]_0^\beta\;.\end{aligned}$$ Denoting $y$ by $(\tau^{\prime},\y)$ and $z$ by $(\tau^{\prime\prime},\z)$, notice that $$\begin{aligned}
\partial_{\tau^{\prime}} \int (d^Dx)_{_E}\, G_0^{-1}(z,x)\,G_0(x;\tau^{\prime},\y) = \partial_{\tau^{\prime}} \left [ \delta^{(3)}(\z-\y) \delta(\tau^{\prime\prime} -\tau^{\prime})\right]_{0}^{\beta}\;,\end{aligned}$$ so that $$\begin{aligned}
G_z^{-1}[\phi_c]\,\frac{\delta \phi_c(z)}{\delta \phi_0(\y)} = \delta^{(3)}(\z-\y)\, \left [\partial_{\tau^{\prime}} \delta(\tau^{\prime\prime} -\tau^{\prime})\right]_0^{\beta}\;.\end{aligned}$$ Multiplying both sides by $G[\phi_c](x,z^{\prime})\,\delta^{(4)}(z^{\prime}-z)$ and integrating over $(d^Dz)_{_E} (d^Dz)_{_E}^{\prime}$, we obtain $$\begin{aligned}
\frac{\delta \phi_c(x)}{\delta \phi_0(\y)} &= \int (d^Dz)_{_E} (d^Dz)_{_E}^{\prime}\,G[\phi_c](x,z^{\prime})G^{-1}[\phi_c](z^{\prime},z)\,\frac{\delta \phi_c(z)}{\delta \phi_0(\y)} \\
&= \int d\tau^{\prime\prime} d^d\z\, G[\phi_c](x;\tau^{\prime\prime},\z) \, \delta^{(3)}(\z-\y)\, \left [ \partial_{\tau^{\prime}} \delta(\tau^{\prime\prime} -\tau^{\prime})\right]_0^\beta\;\\
&= \int d\tau ^{\prime\prime} \,G[\phi_c](x;\tau^{\prime\prime},\y)\,\left [ \partial_{\tau^{\prime}} \delta(\tau^{\prime\prime} -\tau^{\prime})\right]_0^\beta\;.\end{aligned}$$ Integrating by parts, Eq. follows.
The free propagator (solution of Eq. ) can be explicitly calculated. It is given in Fourier space by $$\begin{aligned}
\label{eq:bound_prop2}
G_0(\tau,\tau^{\prime};\k) = \frac{\sinh [\omega_\k(\tau_{>}-\beta)]\,\sinh (\omega_\k\tau_{<})}{\omega_\k\,\sinh (\beta \omega_\k)}\;,\end{aligned}$$ where $\tau_>(\tau_<) = $ max (min) $\{\tau,\tau^{\prime}\}$. From Eq. we obtain $$\begin{aligned}
\left [\partial_{\tau^{\prime}}G_0(\tau,\tau^{\prime};\k)\right ]_{0}^\beta = \frac{\cosh [\omega_\k(\beta/2-\tau)]}{\cosh (\omega_\k\beta/2)}\end{aligned}$$ and $$\begin{aligned}
\label{fancyG0}
{\cal G}_0(\tau,\tau^{\prime};\k) = \left [\partial_\tau\partial_{\tau^{\prime}}G_0(\tau,\tau^{\prime};\k)\right ]_{0}^\beta = 2\omega_\k \tanh (\omega_\k \beta /2)\;.\end{aligned}$$ It is a simple matter to verify that $$\begin{aligned}
\nonumber
&\frac{1}{2\omega_\k}\left [(1+n(\omega_\k))e^{-\omega_\k|\tau-\tau^{\prime}|} + n(\omega_\k)e^{\omega_\k|\tau-\tau^{\prime}|}\right ] = \frac{\sinh [\omega_\k(\tau_{>}-\beta)]\,\sinh (\omega_\k\tau_{<})}{\omega_\k\,\sinh (\beta \omega_\k)}\\
& \hspace{3cm}+ \frac{\cosh [\omega_\k(\beta/2-\tau)]}{\cosh (\omega_\k\beta/2)} \frac{\coth \omega_\k \beta/2}{2\omega_\k} \frac{\cosh [\omega_\k(\beta/2-\tau^{\prime})]}{\cosh (\omega_\k\beta/2)}\;,\end{aligned}$$ which is the expression in three dimensional Fourier space for $$\begin{gathered}
\Delta_F = G_0 + \partial_\tau G_0 \,{\cal G}_0^{-1}\,\partial_{\tau^{\prime}}G_0\;.\end{gathered}$$
It is possible to show that $$\begin{aligned}
\hbox{ Tr} \log G_0^{-1} \,=\, \int \frac{d^d\k}{(2\pi)^3} \, \log \frac{\sinh \beta \omega_\k}{ \omega_\k} \;.\end{aligned}$$ Using and the trigonometric identity $$\begin{aligned}
2\tanh \frac{x}{2} \, \sinh x = e^x \,(1-e^{-x})^2\;,\end{aligned}$$ we obtain $$\begin{aligned}
\frac{1}{2}\hbox{ Tr} \log G_0^{-1} + \frac{1}{2}\hbox{ Tr} \log {\cal G}_0 \,&=\, \int \frac{d^d\k}{(2\pi)^3} \,\left [\frac{\beta \omega_\k}{2} + \log (1-e^{-\beta\omega_\k})\right ]\\
&=\;\frac{1}{2} \hbox{ Tr} \log \Delta_F^{-1} + C\;,\end{aligned}$$ where $C$ is an unimportant infinity constant.
[99]{}
M. Le Bellac, [*Thermal Field Theory*]{} (Cambridge University Press, 2000). J. I. Kapusta and C. Gale, [*Finite-Temperature Field Theory: Principles and Applications*]{} (Cambridge University Press, 2006). A. Das, [*Finite Temperature Field Theory*]{} (World Scientific, 1997). J. Zinn-Justin, [*Quantum Field Theory and Critical Phenomena*]{} (Oxford University Press, Oxford, 1993).
J. P. Blaizot, E. Iancu and A. Rebhan, Phys. Rev. D [**63**]{}, 065003 (2001). A. Peshier, B. Kampfer, O. P. Pavlenko and G. Soff, Europhys. Lett. [**43**]{}, 381 (1998).
J. Berges, S. Borsanyi, U. Reinosa and J. Serreau, Phys. Rev. D [**71**]{} (2005) 105004.
J. Berges, S. Borsanyi, U. Reinosa and J. Serreau, Annals Phys. [**320**]{} (2005) 344.
A. Arrizabalaga and U. Reinosa, Nucl. Phys. A [**785**]{} (2007) 234.
S. Borsanyi and U. Reinosa, Phys. Lett. B [**661**]{} (2008) 88.
C. A. A. de Carvalho, J. M. Cornwall and A. J. da Silva, Phys. Rev. D [**64**]{}, 025021 (2001). A. Bessa, C. A. A. de Carvalho, E. S. Fraga and F. Gelis, JHEP [**0708**]{}, 007 (2007). C. A. A. de Carvalho, R. M. Cavalcanti, E. S. Fraga and S. E. Jorás, Annals Phys. [**273**]{}, 146 (1999). C. A. A. de Carvalho, R. M. Cavalcanti, E. S. Fraga and S. E. Joras, Phys. Rev. E [**61**]{}, 6392 (2000). C. A. A. de Carvalho, R. M. Cavalcanti, E. S. Fraga and S. E. Joras, Phys. Rev. E [**65**]{}, 056112 (2002). A. Bessa, C. A. A. de Carvalho and E. S. Fraga, Phys. Rev. E [**81**]{}, 011103 (2010). R. P. Feynman and Hibbs, [*Quantum Mechanics and Path Integrals*]{} (McGraw-Hill, 1965).
K. G. Wilson and J. B. Kogut, Phys. Rept. [**12**]{} (1974) 75. D. J. Amit, [*Field Theory, the Renormalization Group, and Critical Phenomena*]{} (McGraw-Hill, 1978)
U. Kraemmer and A. Rebhan, Rept. Prog. Phys. [**67**]{}, 351 (2004). J. O. Andersen and M. Strickland, Annals Phys. [**317**]{}, 281 (2005). A. Bessa, F. T. C. Brandt, C. A. A. de Carvalho and E. S. Fraga (unpublished).
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
[^4]: [email protected]
[^5]: When $j=0$, condition finally implies that $\phi_c$ is a periodic function of $\tau$ with period $\beta$ (in fact, the trivial solution), but this property is not shared by classical solutions in the presence of nontrivial currents.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Weyl fermions, which are fermions with definite chiralities, can give rise to anomalous breaking of the symmetry of the physical system which they are a part of. In their $(3+1)$-dimensional realizations in condensed matter systems, *i.e.*, the so-called Weyl semimetals, this anomaly gives rise to topological electromagnetic response of magnetic fluctuations, which takes the form of non-local interaction between magnetic fluctuations and electromagnetic fields. We study the physical consequences of this non-local interaction, including electric field assisted magnetization dynamics, an extra gapless magnon dispersion, and polariton behaviors that feature “sibling" bands in small magnetic fields.'
author:
- 'Jimmy A. Hutasoit'
- Jiadong Zang
- Radu Roiban
- 'Chao-Xing Liu'
bibliography:
- 'References.bib'
title: Weyl fermions induced Magnon electrodynamics in Weyl semimetal
---
In the 1980s, the study of anomalous behaviors of classically conserved currents in systems with Weyl fermions revealed a deep connection between this physical phenomena and the underlying topology of the systems. In particular, it was realized that these anomalies are deeply related to the skewness of the zero mode structure of the Dirac operators, which in turn, using index theorems, can then be related to characteristic classes, which are topological invariants [@Nielsen1977445; @Nielsen1978475; @alvarez1984topological; @alvarez1985structure]. Recently, with the advancement of realizations of topologically ordered condensed matter systems, the interest on the connection between topology and anomaly has been revived. Not only the study of anomalies might give rise to a way to classify topological phases in matters in the presence of interactions [@PhysRevB.85.045104], but it can also lead to topological responses, which are physical manifestations of the underlying topological nature, of these topologically ordered systems [@PhysRevB.85.045104; @PhysRevB.84.014527; @PhysRevB.85.184503; @PhysRevB.88.115307].
In this letter, we study topological aspects of the Weyl semimetal, a topologically protected semimetal with Weyl fermions. Weyl semimetals can be regarded as a three-dimensional cousin of graphene, where pairs of bands cross at certain points in the momentum space, *i.e*, the Weyl points. For a short introduction to Weyl semimetals, see for example Ref. . Some material realizations of Weyl semimetals consist of topological insulator heterostructures that contain magnetic materials or magnetic dopants [@Burkov:2011zr; @cho2011possible]. An advantage of this realization is that magnetic texture and fluctuations inherit some physical properties that reflect the underlying topological nature of this system. In particular, magnetic fluctuations are coupled to Weyl fermions as an axial vector field [@Liu:2012ly] and therefore, magnon excitations in this system possess topologically non-trivial electromagnetic responses from the axial anomaly.
Our main result takes the form of a non-local interaction between magnons and electromagnetic fields in Weyl semimetals, dictated by the effective action Eq. (\[eq:eff\]) below. The non-locality of the interaction arises from the fact that the mediators of this interaction are gapless excitations of Weyl fermions. The modifications of the Landau-Lifshitz (LL) equation and Maxwell equation due to this non-local interaction will give rise to two physical consequences, which reflect the underlying topological nature of Weyl semimetals. Firstly, in Weyl semimetals, electric fields can couple to the local magnetic moments through gapless Weyl fermions, leading to an additional magnon excitation. Compared to the conventional spin wave in ferromagnet, this new magnon branch is gapless and linear, inheriting the nature of Weyl fermions. Secondly, the non-local coupling between magnons and electromagnetic fields can induce a magnon-polariton excitations in Weyl semimetals, which exhibit a quite different spectrum from the usual polariton spectrum. In particular, in small values of magnetic fields, there exists a band with finite width that bifurcates into a pair of “sibling" bands with well-defined quasiparticles. Let us start by first describing the Weyl fermion sector of the Weyl semimetal and how it is coupled to the magnetic fluctuations. We start with a 4-band model H\_0 = (
[M]{} +M & 0 & - i L\_1 k\_3 & i L\_2 k\_-\
0 & [M]{} - M & - i L\_2 k\_+ & - i L\_1 k\_3\
i L\_1 k\_3 & i L\_2 k\_- & - [M]{} + M & 0\
- i L\_2 k\_+ & i L\_1 k\_3 & 0 & - [M]{} - M
), \[eq:4band\] where &=& M\_0 + M\_1 k\_3\^2 + M\_2 k\_\^2,\
k\_ &=& k\_1 i k\_2, and $|M| > |M_0|$ to ensure the existence of Weyl points [@Liu:2012ly]. Here, there are two Weyl fermions and expanding around the Weyl points, we see that the Weyl fermion sector is described by S = i d\^4x | \^ (\_ - i e A\_ - i g \_5 a\_ ) , where we have written the two Weyl fermions as a Dirac fermion $\psi$, $A_{\mu}$ is the electromagnetic gauge field and $a_{\mu}$ is the axial vector field whose space-like components $\mathbf{a}$ are identified as magnetic fluctuations [@Liu:2012ly]. Here, we have adopted the convention where the metric is mostly positive.
Let us start by considering a topological insulator doped with magnetic impurities and assume that magnetic moments are magnetized along the growth direction, which we will take to be the $\hat{x}_3$-direction. This system can be realized in for example, Cr doped Bi$_2$Te$_3$ [@ADMA:ADMA201203493]. When magnetization is large enough, this model exhibits Weyl nodes, at which the effective excitations are two Weyl fermions with a relativistically-invariant dispersion relation. Thus, this system provides a natural description of Weyl semimetals using the 4-band model [@Liu:2012ly], the details of which are given in Appx. \[appx:4\]. It turns out that in this system, magnetic fluctuations of magnetic moments are coupled chirally to Weyl fermions [@Liu:2012ly], and the effective action describing the interaction between Weyl fermions, electromagnetic fields and magnetic fluctuations is given by S = i d\^4x | \^ (\_ - i e A\_ - i g \_5 a\_ ) , \[action\] where two Weyl fermions have been written together as a single Dirac fermion $\psi$, $A_{\mu}$ is the electromagnetic gauge field and $a_{\mu}$ is an axial vector field whose space-like components $\mathbf{a}$ are identified as magnetic fluctuations [@Liu:2012ly]. Our convention for the $\gamma$ matrices \^ =
0 & \^\
|\^ & 0\
; \^ = (\_[22]{}, ), |\^ = (\_[22]{}, -),\
and the metric follows closely Ref. , where the metric is mostly positive. In the following, we will consider only the case where the axial vector field strength $f_{\mu \nu} = \partial_{\mu} a_{\nu} - \partial_{\nu} a_{\mu}$ vanishes [^1], which is the case when there is no magnetic domain wall in the system [@Liu:2012ly]. Even though we are not going to use this fact here, it is worth noting for $f_{\mu \nu} =0$, the axial vector field can be written as $a_{\mu}=\partial_{\mu}\theta$, where $\theta$ has the physical meaning of axion fields [@PhysRevB.78.195424].
To completely define this quantum field theory, it is necessary to specify a regularization scheme. This is particularly important here as the chiral nature of the interactions implies that the theory exhibits an anomaly [@PhysRev.177.2426; @bell1969pcac], which appears as a violation of current conservation in the three-point function $\Gamma^{\mu\nu\rho}=\langle j^\mu(p)\, j^{\nu}(q) \,j^{\rho 5}(-p-q) \rangle$, where $j^\mu={\bar\psi}\gamma^\mu\psi$ and $j^{\mu5}={\bar\psi}\gamma^\mu\gamma_5\psi$ are the $U(1)$ vector and axial current, respectively. The anomaly is a reflection of the impossibility of simultaneously preserving the vector and axial symmetries in the presence of any regulator. Since the vector symmetry characterizes the interaction of fermions and electromagnetic fields, the correct definition of the theory must include a regularization scheme that respects the vector symmetry, which is nothing but the gauge invariance of electromagnetism. An example of such scheme is the dimensional regularization scheme of ’t Hooft and Veltman [@t1972regularization], and the calculation of $\Gamma^{\mu \nu\rho}$ using this scheme was done in Ref. . One can also calculate this three-point function using Cutkosky rules and the dispersion relation, as was done in Ref. . The result is \^= - \^ p\_ q\_ , \[eq:3pt\] where $\varepsilon^{\alpha \beta \gamma \delta}$ is the totally antisymmetric Levi-Civita tensor.
It is easy to see that this three-point function satisfies the conservation of the vector current, $p_\mu \, \Gamma^{\mu\nu\rho}=0 = q_\nu \, \Gamma^{\mu\nu\rho} $, but violates axial current conservation, $(-p-q)_\rho \, \Gamma^{\mu\nu\rho} = \frac{i e^2 g}{2 \pi^2} \, \varepsilon^{\alpha \mu \beta \nu} \, p_{\alpha} \, q_{\beta} \ne 0$. We note that since anomalies are infrared phenomena (see for example, Ref. and references within), we can expect the topological electromagnetic response of magnons to be insensitive to the details of the model away from the Weyl points as long as the electromagnetic gauge invariance is not broken. In a classic (particle physics) example, a similar anomaly is responsible for the decay of a neutral pion into two photons independently of the high energy completion of the theory of strong interactions that does not break the electromagnetic gauge invariance. For example, the pion decay is independent of the QCD quark masses [@PhysRev.177.2426; @bell1969pcac]. Nevertheless, it will be interesting to study the non-topological electromagnetic response of magnons from the high energy sector of Weyl semimetals and such study will be taken up elsewhere. For the rest of this letter, we will focus on studying the physical consequences of the anomalous term Eq. (\[eq:3pt\]).
To that end, we construct the effective action of the topological electromagnetic response of magnons as follows \[eq:eff\] &&S\_[top]{} = \^A\_(p) A\_(q) a\_(-p-q)\
&=& - d\^4x d\^4y \^ F\_(x) F\_(x) a\^ (y), where $F_{\mu \nu} = \partial_{\mu} A_{\nu} - \partial_{\nu} A_{\mu}$ is the (vector) field strength, $G(x-y)$ is the Green function of the d’Alembertian $\Box = \partial_{\mu} \partial^{\mu}$ and it obeys $\Box_x \, G(x-y) = \delta^4 (x-y)$.
We note that in the limit of a constant axial vector $a^\mu$ we recover the result of Refs. [@PhysRevB.86.115133; @PhysRevB.87.161107]. For details, see Appx. \[appx:const\]. Furthermore, using the definition $j^{\alpha} = \delta S/\delta A_{\alpha}$, we can obtain the anomalous Hall response j\^(x) = - \^ F\_(x) \_d\^4y a\^ (y), which, in the limit of a constant axial vector $a^\mu$, reduces to the known result $j^{\alpha} = -\frac{e^2 g}{2 \pi^2}\, \varepsilon^{\alpha \beta \gamma \delta} a_{\beta}\,F_{\gamma \delta}$ of Ref. .
As another non-trivial check, we can also compare the effective action Eq. (\[eq:eff\]) with the result from 4-band model of Ref. at uniform magnetic field $\mathbf{B} = B \,\hat{x}_3$, akin to the calculation done in Ref. . In this case, we have Landau levels and we can ask how the system responses to an applied electric field $\mathbf{E} = E \, \hat{x}_3$ and a perturbation due to magnetization. The result agrees with Eq. (\[eq:eff\]) and for details, see Appx. \[appx:4\].
We can approximate the response function by neglecting the contribution from higher Landau levels (,q \_3) &\~& \_[n,n’]{} \_[k\_3]{} | n’,k\_3+q| n, k\_3 |\^2\
& & , where $n_F$ is the Fermi distribution and the factor $e\,B$ comes from the degeneracy of Landau levels. For vanishing temperature and chemical potential, and for small $q$, we have (q,q \_3) &\~& (), \[eq:response\] where we have absorbed the parameter dependent coefficient into the definition of $g$. The Lagrangian density in momentum space is then given by (see Supplemental Material at for details) = E a\_3 = EB () a\_3, which upon Fourier transforming back to real space, reads = - E(x)B(x) \_y G(x-y) (y), in agreement with Eq. (\[eq:eff\]).
We are now ready to study the modifications of LL and Maxwell equations caused by the topological response of Eq. (\[eq:eff\]). Assume an easy axis anisotropy is present such that the magnetic moments are uniformly polarized along the $\hat{x}_3$ direction in equilibrium. The magnon excitations are investigated by considering the magnetization dynamics of the following Hamiltonian: H\_[magnet]{} = (J ( )\^2 + m\^2 | \_|\^2 ) - ,\[eq:magnon1\] where $\mathbf{M}$ is the magnetization, $\parallel$ denotes the in-plane direction, $\mathbf{B}$ is the magnetic field, and $m^2$ is the easy axis anisotropy. Let $\mathbf{M} = M\, \hat{x}_3 +
\mathbf{a}$, with $M \gg a_i$. Substituting Eqs. (\[eq:eff\]) and (\[eq:magnon1\]) into the LL equation = , we then have &=& - d\^4y \^ F\_(y) F\_(y) \_2 G(x-y)\
& & + B\_2 + (J \^2 - m\^2 - B\_[3]{}/M )a\_2, \[eq:LL1\]\
&=& d\^4y \^ F\_(y) F\_(y) \_1 G(x-y)\
& & - B\_1 - (J \^2 - m\^2 - B\_[3]{}/M ) a\_1, \[eq:LL2\] where $\gamma$ is the product of the gyromagnetic ratio, Bohr magneton and permeability of vacuum. It is interesting to note that a spatial-dependent term contributed from the Weyl fermion enters the magnetization dynamics. It plays the same role as the in-plane magnetic field $B_1$ or $B_2$. The magnetic moments experience this spatially modulated effective field such that the magnon dispersion can be significantly changed. As the electromagnetic field strength $F_{\alpha\beta}$ contains electric fields as its component, it is quite interesting to see that the electric field can dramatically modify magnetization dynamics. To illustrate this more clearly, let us consider Weyl semimetals in a magnetic field along the $\hat{x}_3$ direction ${\bf B}=B_3\hat{x}_3$. For an oscillating electric field ${\mathbf E}=E_3 \,\exp\left( i \omega t- i {\mathbf{q}_{\parallel} \cdot \mathbf{x}_{\parallel}}\right)\,\hat{x}_3$, we obtain \_(, \_) &=& -\
& & , \[eq:magnon\] where $q = |\mathbf{q}_{\parallel}|$. Here, we have inserted back the Fermi velocity of the Weyl fermions in order to differentiate it with the speed of light in the medium, which we are taking to be unity.
Two poles of $\mathbf{a}_{\parallel}(\omega, \mathbf{q}_{\parallel})$ suggest the existence of two magnon branches in this system. In addition to the usual spin wave $\omega = \gamma \,M\,(J \,q^2+m^2+ B_{3}/M)$, a magnon with $\omega = v_F\,|q|$ is present. This novel dispersion is determined solely by a property of the Weyl fermions, namely their Fermi velocity. More importantly, this new branch is gapless, leading to a long-range correlation of spin excitations. Physically, this magnon excitation can be understood as a direct result of the coupling between two magnetic moments mediated by Weyl fermions. The gapless nature of Weyl fermions leads to long-range correlation of this magnon excitation. Therefore, this new magnon dispersion is a distinct feature of Weyl semimetals. One can then employ neutron scattering experiments to test our prediction.
Let us now look at the Maxwell equation in the presence of topological response of Eq. (\[eq:eff\]). It is given by - \_F\^ - \^F\_ d\^4y &=& 0.\
\[eq:eomA\] By using the identity $\varepsilon_{\mu \nu \rho \sigma} \partial^{\rho} F^{\mu \nu} = 0$ and keeping only the linear term in $\mathbf{E}$, we get - \^2\_t + \^2 - d\^4y \^2\_t G(x-y) ( )(y)\
+ ( ) d\^4y G(x-y) ( )(y) &=& 0.\
\[eq:eomlinearE\] For concreteness, let us again consider applying a uniform magnetic field $B_3$ along the $\hat{x}_3$ direction. If we shine a light with the electric field ${\mathbf E}=E_3 \,\exp\left( i \omega t- i {\mathbf{q}_{\parallel} \cdot \mathbf{x}_{\parallel}}\right)\,\hat{x}_3$, we have &&E\_3(,)=0.\
\[eq:polariton\]
---------------------------------------
{width="6.5in"}
---------------------------------------
The solutions of the equation above correspond to the poles of the polariton modes, which can then be detected using various spectroscopy techniques, such as angle-resolved electron energy-loss spectroscopy. In a typical magnetic material, in the long wavelength regime, the effective magnon velocity $2 \gamma M J q$ is significantly smaller than the Fermi velocity of the Weyl fermions. Thus, the dispersion of the magnon can be neglected. The typical behavior of the real part of the poles are plotted in Fig. \[fig:spec\]. See also Appx. \[appx:green\].
One can find four bands in total in Fig. \[fig:spec\]. These solutions represent the hybridization between electric fields and magnetic moments due to the non-local coupling induced by Weyl fermions. At a non-vanishing magnetic field, the top and bottom bands are non-degenerate. Furthermore, the imaginary parts of their respective poles vanish and therefore, their spectral density is given by a Dirac $\delta$-function. The intermediate band, however, acquires a non-vanishing imaginary part of its pole and therefore, feature broadened spectral density. This broadening is due to its ability to emit Weyl fermions, which results in it acquiring complex self-energy. This is not unlike the physics of plasmon, see for example Ref. . At a low magnetic field, this band bifurcate into a pair of “sibling" bands, whose spectral densities are given by Dirac $\delta$-functions, where the threshold for emitting Weyl fermions is beyond the energetics. As the magnetic field increases, this bifurcation disappears. The value of magnetic field at which this happens scales as $g^{-2}$ and for $g \sim 0.1$, this value is of order 10 T.
At a small $q$, the top and bottom bands scale as $\omega \sim q^0$ and $q^2$, respectively, while the intermediate band scales like $\omega \sim q^{1/2}$. We note that for the intermediate band, there is a regime where the velocity of the latter exceeds the speed of light in the Weyl semimetal. This “tachyonic" regime needs to be excised, similar to the case of surface optical phonon for a polar crystal such as NaCl [@mahan2010condensed].
The polariton spectrum also features an energy range at which there exists no polariton modes. This “forbidden" band is particularly manifest at larger values of magnetic field. Therefore, the incident light will be totally reflected if its frequency lies within the forbidden band. Such forbidden band is predicted to be a generic feature of topological magnetic insulator [@li2010dynamical], however, sibling bands are particular to the Weyl semimetal.
In order to probe the polariton, it is crucial that the energy dumped into the system is spent to excite the polariton and not the Weyl fermions. In other words, the observability of the polariton spectrum depends heavily on how much it overlaps with the single particle excitation regimes of the Weyl fermions. We find indeed that this overlap is negligible as the typical minimum energy needed to excite the Weyl fermions is about $10$ meV (see Appx. \[appx:4\] for details) while the typical magnon gap $\omega_m=\gamma (M\,m^2+
B_{3})$ is about $0.1$ meV [@popova2001structure].
*Summary* – In this article, we have shown that the topological response of magnons in Weyl semimetal is given by a non-local interaction between magnons and electromagnetic fields. This non-local interaction manifests itself in term of electric-field-induced magnetization dynamics that results in gapless magnon excitations. It also gives rise to resonant behavior in the form of magnon polariton featuring sibling bands and forbidden band.
*Acknowledgements* – We would like to thank Gerald Mahan, Xiaoliang Qi, Cenke Xu and Jainendra Jain for insightful discussions. J. H. is supported by NSF grant DMR-1005536 and DMR-0820404 (Penn State MRSEC). J. Z. is supported by the Theoretical Interdisciplinary Physics and Astrophysics Center and by the U.S. Department of Energy, Office of Basic Energy Sciences, Division of Materials Sciences and Engineering under Award DEFG02-08ER46544. R. R. is supported by the U.S. Department of Energy under contract DE-SC0008745.
4-band model calculations \[appx:4\]
====================================
Let us start with the 4-band model of Ref. H = H\_0 + H\_1, where \[4band\] H\_0 = (
[M]{} +M & 0 & - i L\_1 k\_3 & i L\_2 k\_-\
0 & [M]{} - M & - i L\_2 k\_+ & - i L\_1 k\_3\
i L\_1 k\_3 & i L\_2 k\_- & - [M]{} + M & 0\
- i L\_2 k\_+ & i L\_1 k\_3 & 0 & - [M]{} - M
) , \[eq:4band\] with &=& M\_0 + M\_1 k\_3\^2 + M\_2 k\_\^2,\
k\_ &=& k\_1 i k\_2 ; and H\_1 = [diag]{}( \_3,- \_3, \_3,- \_3) = \_3 \_3 \_[2 2]{}. Here, we have magnetized the system along the $\hat{x}_3$-direction with magnetization $M$ and for simplicity, allow magnetic fluctuations only along that same direction, where $\tilde{g} \ll 1$. All the material related parameters are defined in Ref. .
For $|M| > |M_0|$, this model exhibits Weyl points. Expanding around these Weyl points, one can obtain the low energy effective theory of Weyl fermions coupled chirally to the magnetic fluctuations as in Eq. (\[action\]). In particular, the axial vector field can then be related to the magnetic fluctuations as a\_3 = - \_3. For details, see Ref. .
Let us now turn on the external uniform magnetic field $\mathbf{B} = B \,\hat{x}_3$. The conjugate momenta on the direction perpendicular to the magnetic field become $k_+ \rightarrow \Pi_+ = \sqrt{2} \,a/ \ell_c$ and $k_- \rightarrow \Pi_- = \sqrt{2} \,a^{\dagger}/ \ell_c$, where $a$ and $a^{\dagger}$ are the annihilation and creation operators for the Landau levels, respectively, and $\ell_c$ is the magnetic length. Since $a \, \phi_n = \sqrt{n} \, \phi_{n-1}$ and $a^{\dagger} \, \phi_{n} = \sqrt{n+1} \, \phi_{n+1}$, writing the wave function as = (
f\_1\^n \_[n-1]{}\
f\_2\^n \_n\
f\_3\^n \_[n-1]{}\
f\_4\^n \_n\
), the Hamiltonian then can be written as H\_[LL]{} &=& (
[M]{}\_n+M & 0 & - i L\_1 k\_3 &\
0 & [M]{}\_n-M& - & - i L\_1 k\_3\
i L\_1 k\_3 & & - [M]{}\_n+M & 0\
- & i L\_1 k\_3 & 0 & - [M]{}\_n-M
), where \_n = M\_0+M\_1 k\_3\^2 + (n-). Diagonalizing this Hamiltonian, we can then obtain the Landau levels for $n>0$. For $n=0$, this Hamiltonian is reduced to half in size H\_[LLL]{} &=& (
[M]{}\_0-M & - i L\_1 k\_3\
i L\_1 k\_3 & - [M]{}\_0-M
). The resulting Landau levels are plotted in Fig. \[fig:landau\].
![Landau levels for $M_0 = -0.005$ eV, $M_1 = 0.342$ eV$\cdot$$\mathring{\rm A}^2$, $M_2 = 18.225$ eV$\cdot$$\mathring{\rm A}^2$, $L_1 = 1.3$ eV$\cdot$$\mathring{\rm A}$, $L_2 = 2.82$ eV$\cdot$$\mathring{\rm A}^2$ and $B = 5$T. $n = 0$ Landau levels are depicted in red, while the higher Landau levels are depicted in blue.[]{data-label="fig:landau"}](landau.pdf){width="5in"}
We can now perturb the above Hamiltonian by applying an external electric field $\mathbf{E} = E \, \hat{x}_3$ and ask what the response of the system to the axial vector field is. The response function is given by \_[aE]{} (,q \_3) &=&\
&=& \_[n,n’LL]{} \_[k\_3]{} (i e)n’,k\_3+q| \_[44]{} | n, k\_3\
& & (- )n, k\_3 | \_3 \_[2 2]{} | n’,k\_3+q . Here, $n_F$ is Fermi distribution, $\ell_3$ is the length of the system in $\hat{x}_3$ direction and the factor $e\,B$ comes from the degeneracy of Landau levels. We can approximate this by neglecting the contribution from higher Landau levels \_[aE]{} (,q \_3) &=& \_[n,n’]{} \_[k\_3]{} | n’,k\_3+q| n, k\_3 |\^2,\
where we have projected the matrix elements of the interaction Hamiltonian $H_1$ (which is $4\times4$) into the LLL space (which is $2\times2$). For $T=0$, $\mu = 0$ and small $q$, we therefore have \_[aE]{} (q,q \_3) && \^[-]{}\
& & |M\_1|L\_1\
&& (). \[eq:response\] The Lagrangian density in momentum space is then given by = \_[a E]{} E a\_3 = EB () a\_3, which upon Fourier transforming back to real space, reads = - E(x)B(x) \_y G(x-y) (y), in agreement with Eq. (\[eq:eff\]).
Next, let us look at the polarization operator in the presence of the external magnetic field. The regime where its imaginary part is non-vanishing corresponds to the regime of single particle excitations (SPE) of the Weyl fermions. Since we are interested in comparing it to the spectrum of the polariton, we are going to focus on the case where the momentum is perpendicular to the direction of the magnetic field $\mathbf{q} = \mathbf{q}_{\parallel}$. The polarization operator is then given by \_[EE]{} (,\_) &=&\
&=& \_[n,n’]{} \_[k\_3]{} | n’,k\_3| n, k\_3 |\^2. We note that the right hand side does not depend on $\mathbf{q}_{\parallel}$. Furthermore, the bottom boundary of the SPE regime is the smallest gap between the filled part of the lowest Landau level and the second Landau level. As can be seen from Fig. \[fig:landau\], it is of order $0.03$ eV.
The Constant Vector Limit \[appx:const\]
========================================
Let us start by putting our theory, Eq. (\[eq:eff\]), in a finite volume by introducing a finite volume regulator $a^{\mu}(y) \rightarrow a^{\mu} (y) \exp\left[-|\hat{a}\cdot y|/\Lambda\right]$, such that S\_[top]{} =- d\^4x d\^4y \^ F\_(x) F\_(x) a\^ (y) e\^[-]{}, \[eq:eff’\] where $\Lambda \gg L$ and $L$ is the typical size of the system. We note that when the axial vector field goes to a constant vector limit, the field strength $\partial_{\mu} \left[a_{\nu} (y) \exp\left(-|\hat{a}\cdot y|/\Lambda\right) \right]- \partial_{\nu} \left[ a_{\mu} (y) \exp\left(-|\hat{a}\cdot y|/\Lambda\right)\right]$, which includes the curl $\mathbf{\nabla} \times [\mathbf{a}\, \exp(-|\hat{a}\cdot y|/\Lambda)]$, remains vanishing while the divergence $\partial_{\mu} [a^{\mu}\, \exp(-|\hat{a}\cdot y|/\Lambda)]$ remains non-zero. Therefore, even at the constant vector limit, the magnon is Helmholtz decomposed into the curl-free term only.
In order to obtain the constant vector limit of Eq. (\[eq:eff’\]), we write $a^\mu \exp(-|\hat{a}\cdot y|/\Lambda) = g^{\mu \nu} \, \partial_\nu \left[a\cdot y\, \exp(-|\hat{a}\cdot y|/\Lambda)\right] + {\cal O}(1/\Lambda)$. Substituting it in Eq. and integrating by parts, we obtain S\_[top]{} &=& - d\^4x d\^4y \^ F\_(x) F\_(x) \_\
& & + d\^4x d\^4y \^ F\_(x) F\_(x) \_y G(x-y) ay e\^[-]{}. The first term is the surface term that vanishes due to the regulator and using the definition of the Green’s function we recover the action in the Ref. S\_[top]{} d\^4x \^ F\_(x) F\_(x) (ax).
The Polariton Green Function\[appx:green\]
==========================================
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![The spectral density at small magnetic field for different values of momenta $q/q_c = 0.2$, $0.5$, $0.75$ and $0.99$ for the blue, purple, yellow and green lines (from left to right), respectively. Here, $q_c \approx 2 e g B_3 /\left[\pi \gamma (M m+\sqrt{M B_3})\right]$ is the momentum at which the bifurcation into the “sibling" bands starts and $\omega_m=\gamma (M m^2+B_3)$ is the magnon gap.[]{data-label="fig:dens"}](density.pdf "fig:"){width="5in"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Setting ${\mathbf B} = B \hat{x}_3$ and then applying an electric field along the $\hat{x}_3$ direction, the Landau-Lifshitz and Maxwell equations can be written as (,q)
\
E\_3\
= 0, \[eq:eom\] where $|\mathbf{q}_{\parallel}| = q$ and (,q) =
- (J q\^2+m\^2 + )\^2 &\
- & \^2 - q\^2\
. The Green function for the polariton then must satisfy (,q) [G]{}(,q) = \_[22]{}, and its singularities are given by the singularities of ${\cal E}^{-1}$, which are the zeroes of $\det {\cal E}$. We note that the solutions to $\det {\cal E}=0$ are identical to the solutions of Eq. (\[eq:polariton\]). Furthermore, we can obtain the spectral density from $\rho (\omega, q) = \Im[ {\cal G} (\omega,q)]$. The spectral densities of the top and bottom band are trivial as they correspond to well-defined quasiparticles, while the spectral density of the intermediate band exhibits finite width. In Fig. \[fig:dens\], we plot the spectral density of the intermediate band at small magnetic field for different values of momenta up to $q=0.99 q_c$, where $q_c$ is the momentum at which bifurcation into the “sibling" bands occur.
[^1]: When $f_{\mu \nu} =0$, the solution to Dirac equation is given by $\psi=0$. In this letter, we would like to obtain the effective interaction between magnons and electromagnetic fields by integrating out Weyl fermion fluctuations around the vacuum solution $\psi=0$. However, when the flux of the axial vector field strength takes non-zero integer values, the solution to Dirac equation consists of additional $(1+1)$-dimensional Weyl fermions [@Liu:2012ly]. The topological response obtained by integrating out fermionic fluctuations around this non-trivial background will be studied elsewhere.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have carried out first principles all-electron calculations of the (001)-projected 2D electron momentum density and the directional Compton profiles along the \[100\], \[001\] and \[110\] directions in the Fe-based superconductor LaOFeAs within the framework of the local density approximation. We identify Fermi surface features in the 2D electron momentum density and the directional Compton profiles, and discuss issues related to the observation of these features via Compton scattering experiments.'
author:
- |
Y. J. Wang$^1$, Hsin Lin$^1$, B. Barbiellini$^1$, P.E. Mijnarends$^{1,2}$, \
S. Kaprzyk$^{1,3}$, W. Al-Sawai$^1$, R.S. Markiewicz$^1$ and A. Bansil$^1$
title: 'High resolution Compton scattering as a Probe of the Fermi surface in the Iron-based superconductor LaO$_{1-x}$F$_{x}$FeAs'
---
Introduction
============
Since the discovery [@Kamihara] of superconductivity in a family of iron-based superconductors (pnictides), there have been a large number of studies of their electronic properties that have revealed similarities between pnictides and cuprates. The theoretical prediction of a striped antiferromagnetic spin-density-wave (SDW) ground state [@Yildirim] was confirmed by neutron scattering [@Clarina]. Superconductivity is found in LaOFeAs with either hole doping [@Hai] or electron doping [@Kamihara; @Chen]. First principles calculations [@Boeri; @Singh1] find that the density of states (DOS) near the Fermi level ($E_F$) is predominantly due to Fe-$d$ orbitals. Owing to the approximate S$_4$ symmetry of the FeAs tetrahedra, these Fe-$d$ orbitals split into lower lying e$_{g}$ (d$_{x^{2}-y^{2}}$,d$_{3z^{2}-r^{2}}$) and higher lying t$_{2g}$ states (d$_{xy}$,d$_{yz}$,d$_{zx}$)[@Chao; @Singh1]. Theoretical calculations [@Boeri; @Singh1] suggest that superconductivity may not be caused by electron-phonon coupling. Just as in the cuprates, the antiferromagnetic instability, which is suppressed by doping, is one candidate to explain unconventional superconductivity. Angle-resolved photoemission spectroscopy (ARPES) experiments have recently been carried out in BaFe$_2$As$_2$ [@Evtushinsky; @HDing], a related pnictide. The superconducting gap in LaFeAsO$_{1-x}$F${_x}$ with $x \approx
10\%$ has been determined from the optical reflectance in the far-infrared region.[@Chen]
X-ray scattering spectroscopy in the deeply inelastic (Compton) regime provides a direct probe of the correlated many-body ground state in bulk materials while avoiding the surface sensitivity of ARPES. The use of modern synchrotron sources [@cooper] makes it possible to investigate complex materials via the measurement of directional Compton profiles[@Laukkanen].
In this paper, we report first-principles computations of the 2D-projected electron momentum density (2D-EMD) and Compton profiles (CPs) in the iron-based superconductor LaOFeAs. We discuss Fermi surface (FS) images in the 2D-EMD and its anisotropy defined by subtracting a smooth isotropic function from the spectrum. Our analysis of the CPs reveals that FS features related to hole- as well as electron-pockets are more prominent in the CP when the momentum transfer vector lies along the \[100\] rather than the \[110\] direction.
Methods
=======
{width="\hsize"}
Our electronic structure calculations are based on the local density approximation (LDA) of density functional theory. An all-electron fully charge self-consistent semi-relativistic Korringa-Kohn-Rostoker (KKR) method is used[@ABkkr]. The compound LaO$_{1-x}$F$_x$FeAs has a simple tetragonal structure (space-group P4/nmm). We used the experimental lattice parameters[@Qiu2008] of LaO$_{0.87}$F$_{0.13}$FeAs in which no spin-density-wave order was observed in neutron-scattering. A non-spinpolarized calculation was performed and the magnetic structure was neglected. Self-consistency was obtained for $x$=0 and the effects of doping $x$ were treated within a rigid band model by shifting the Fermi energy to accommodate the proper number of electrons. The convergence of the crystal potential was approximately $10^{-4}$ Ry. The electron momentum density (EMD) $\rho(p_x,p_y,p_z)$ was calculated on a momentum mesh with step ($\Delta p_x,\Delta p_y,\Delta p_z) = (1/16a,1/16a,1/16c)2\pi$. The total number of points was $14.58\times10^{6}$ within a sphere of radius 12.8 a.u. in momentum space. The 2D-EMD $\rho^{2d}(p_x,p_y)$ was calculated as $$\rho^{2d}(p_x,p_y)= \int \rho(p_x,p_y,p_z)
dp_z$$ while the Compton profile $J(p_z)$ is given by $$J(p_z) = \int \int \rho(p_x,p_y,p_z)
dp_xdp_y.$$
Results and Discussions
=======================
In Fig. \[fig:EMD\](a), we show the LDA band structure of LaOFeAs. For $x$=0, three bands cross the $E_F$ around the $\Gamma$ point, forming hole-like FSs \[marked by green dots in (c)\] while two bands cross $E_F$ around $M$($\pi$,$\pi$), forming electron-like FSs \[marked by blue dots in (c)\]. As electrons are added, the $\Gamma$ centered FSs shrink and completely disappear around $x$=0.13. The bands near $E_F$ are dominated by the Fe $d$ orbitals. The FeAs layers are separated by insulating LaO layers, with the result that the dispersion of these bands along $\Gamma-Z$ is small and, apart from a small $\Gamma$-centered 3D hole pocket, the FSs are quasi two-dimensional. Based on the fully three-dimensional computations, we take advantage of this quasi two-dimensionality and investigate quantities in the $k_x-k_y$ plane by integrating out the $k_z$ component.
Figure \[fig:EMD\](b) shows a map of the theoretical 2D-EMD[@matsumoto2001]. This distribution can be described by an inverted bell shape with fourfold symmetry. The peak is at the zone center with tails extending over several unit cells. The dense contours around high symmetry points are signatures of the FS discontinuities. All these features are hidden behind the large inverted bell shaped signal. In order to investigate the Fermi surface topology in detail, we have employed both the 2D Lock-Crisp-West (LCW) folding[@lcw1973; @matsumoto2001] and the 2D-EMD anisotropy.
The 2D-LCW folding of the projected 2D-EMD $\rho^{2d}(p_x,p_y)$ is defined by $$n(k_x,k_y)=\sum_{G_x,G_y}\rho^{2d}(k_x+G_x,k_y+G_y),$$ where $n(k_x,k_y)$ gives the number of occupied states at the point $(k_x,k_y)$ in the first Brillouin zone by summing over all projected reciprocal lattice vectors $(G_x,G_y)$. The effect of the matrix element can be eliminated via the 2D-LCW folding process of Eq. (3), which thus provides a tool for focusing on the FS features. The theoretical 2D-LCW folding shown in Fig. \[fig:EMD\](c) has been smoothed using a Gaussian function with $\Delta p=0.17$ a.u., which is typical of the resolutions available in high resolution Compton scattering experiments. The positions and sizes of the FS pockets of the undoped parent compound LaOFeAs found by our KKR band calculation are shown as green dots for hole pockets and as blue dots for electron pockets. Before the application of the aforementioned Gaussian broadening, $n(k_x,k_y)$ shows a maximum $n_{max}$=29.4 and a minimum $n_{min}$=24.5. The difference $n_{max}-n_{min}$=4.9 is consistent with five bands crossing the Fermi level in the LDA calculation. Even after including experimental resolution, the FS features are still quite visible as seen in Fig. \[fig:EMD\](c). The maximum of $n(k_x,k_y)$ at $M(\pi,\pi)$ is associated with the electron pockets; the minimum at $\Gamma$ ($0$,$0$) is related to the hole pockets.
![(Color online) Calculated 2D-EMD anisotropy in the parent compound LaOFeAs. Letters A and B label $\Gamma$ points for odd and even $k$-space sublattices of the hole-pockets as discussed in the text. The red lines are centered at $M$ points. The rapid changes in the momentum density along these lines are FS signatures of the electron-pockets. []{data-label="fig:alternating"}](fig2.eps){width="\hsize"}
Figure \[fig:alternating\] shows the 2D-EMD anisotropy, found by subtracting a smooth isotropic function from the 2D-EMD. FS features show up as closely spaced contours around $\Gamma$($0$,$0$) and $M$($\pi$,$\pi$) points. The momentum density around the $\Gamma(or M)$ points in the higher zones is seen to be lower (or higher) than the average due to the presence of hole pockets (or electron pockets). The zone-to-zone variation of intensity of these features can be understood as a matrix element effect associated with the symmetry of the hole pockets at $\Gamma$ and electron pockets at $M$. For instance, the weak signal at the origin ($0$,$0$) can be understood since the bands crossing the Fermi level are predominantly $d$ orbitals, whereas only an $s$ orbital yields a significant contribution to the momentum density at the origin. Owing to interference effects, the FS features display a marked modulation from zone to zone. The Fe atoms in the unit cell are located at high symmetry positions, Fe1 (0,0,0) and Fe2 (0.5,0.5,0) (in units of lattice constants). The wavefunctions of these two Fe atoms show a constructive and destructive interference in momentum space, which can be represented by the structure factor $S_{\bf G}=1+e^{-i\pi(m+n)}$, where ${\bf G}=(m\hat{x}+n\hat{y})[2\pi/a]$ is a reciprocal lattice vector. Whereas $S_{\bf G}$ is largest when $(m+n)$ is even, $S_{\bf G}$ vanishes when $(m+n)$ is odd. Therefore, the FS features show the alternating pattern seen in Fig. \[fig:alternating\]. For the FS hole-pockets centered at $\Gamma$, the FS features at B for odd ($m+n$) are much weaker than those at A for even ($m+n$). Deviations from this rule are an indication of hybridization of the Fe orbitals with other orbitals. For the FS electron-pockets centered at $M$, a similar pattern is found. In Fig. \[fig:alternating\] we show red lines crossing the B sites. The change of the momentum density along the direction in which $\Delta m = \Delta n$ is more rapid than changes along a direction for which $\Delta m = -\Delta n$.
The theoretical CPs along \[100\], \[110\] and \[001\] are shown in Fig. \[fig:cp\](a). The CP along \[001\] is a smooth curve, since there are no Fermi breaks along this direction. Thus, we can use this profile for highlighting FS features from the other profiles.
![(Color online) (a) Directional Compton profiles(CPs) $J(p)$ along \[100\], \[001\] and \[110\] for $x$=0. (b)-(e) Differences $\Delta J(p)$ between various pairs of CPs for undoped ($x=0$) and doped ($x=0.15$) cases as indicated in the figures. Vertical arrows mark FS crossings with the electron pockets centered at $M$ (blue) and the hole pockets at $\Gamma$ (green) as discussed in the text. []{data-label="fig:cp"}](fig3.eps){width="\hsize"}
In Fig. \[fig:cp\](b), we subtract the \[001\] from the \[100\] CP. The resulting periodic patterns occur at the $\Gamma$ and $M$ points and are associated with hole pockets (green arrows) and electron pockets (blue arrows) respectively. The CP has a dip within the hole pocket regions and a hump within the electron pocket regions. The FS breaks are clearly visible and should be amenable to exploration via high resolution Compton scattering experiments with statistics high enough to numerically differentiate the difference profiles.
The same strategy is applied to the \[110\] CP in Fig. \[fig:cp\](c); however, the FS features are not as clear as in the \[100\] direction. The main reason is that the contributions of the hole- and electron-like FSs projected on \[110\] overlap each other and tend to cancel out. The interference pattern acts to amplify this effect as follows. When the EMD is projected to form the CP, the projections of A and B are distinct points along \[110\]. For hole pockets (centered at $\Gamma$), A (B) has strong (weak) signals associated with FS breaks. For electron pockets \[centered at $(\pi,\pi)$\], the red lines in the EMD are parallel to \[110\] at the projection of A, while they are perpendicular to \[110\] at the projection of B. We notice that the FS signals are strong only along the red lines. Therefore, at the projection of A (B), the FS signals are strong (weak) for both hole and electron pockets. In simple terms, at A, strong hole pocket signals cancel strong electron pocket signals, while at B weak hole pocket signals cancel weak electron pocket signals. Thus \[110\] is not a suitable direction for studying the Fermi breaks.
Next, we discuss how FS breaks disappear with electron doping. The Fermi level for $x$=0.15 shown in Fig. 1(a) at 0.075eV is obtained by a rigid band shift. At this doping level, all hole pockets around $\Gamma$ are removed. As indicated by blue arrows in Fig. \[fig:cp\](d), the FS breaks associated with the FS electron pockets remain in the \[100\] CP. Compared to Fig. \[fig:cp\](b), the dips associated with the hole pockets have completely disappeared (green arrows). In Fig. \[fig:cp\](e), we subtract the \[100\] CP with $15$% doping from the \[100\] CP with $0$% doping. An interesting pattern of periodic maxima and minima appears around the $\Gamma$ and $M$ points which are identical with the positions of the hole and electron pockets, respectively. This may prove the most promising method of detecting FS signatures.
Conclusions
===========
In conclusion, we have identified FS signatures in the momentum density of LaOFeAs, finding alternating intensity patterns in the 2D-EMD due to the symmetry of the crystal. FS signatures for both hole- and electron pockets are shown to be relatively strong in the \[100\] CP in comparison to the \[110\] CP. We thus conclude that the \[100\] direction is the favorable one for studying FS signatures. Our analysis further indicates that a doping-dependent experimental study should be able to determine at which doping level the hole-like FS’s disappear. The present work sets a baseline for future experimental Compton scattering studies in the pnictides.
Acknowledgments
===============
This work is supported by the US Department of Energy, Office of Science, Basic Energy Sciences contract DE-FG02-07ER46352, and benefited from the allocation of supercomputer time at NERSC and Northeastern University’s Advanced Scientific Computation Center (ASCC). It was also sponsored by the Stichting Nationale Computer Faciliteiten (NCF) for the use of supercomputer facilities, with financial support from the Nederlandse Organisatie voor Wetenschappelijk Onderzoek (Netherlands Organization for Scientific Research).
[99]{} Y. Kamihara, T. Watanabe, M. Hirano, and H. Hosono, J. Am. Chem. Soc. [**130**]{}, 3296 (2008).
T. Yildirim, Phys. Rev. Lett. [**101**]{}, 057010 (2008).
C. de la Cruz, Q. Huang, J.W. Lynn, J. Li, W. Ratcliff II, J.L. Zarestky, H.A. Mook, G.F. Chen, J.L. Luo, N.L. Wang and P. Dai, Nature [**453**]{}, 899 (2008).
H.-H. Wen, G. Mu, L. Fang, H. Yang and X. Zhu, Europhys. Lett. [**82**]{}, 17009 (2008).
G.F. Chen, Z. Li, G. Li, J. Zhou, D. Wu, J. Dong, W.Z. Hu, P. Zheng, Z.J. Chen, H.Q. Yuan, J. Singleton, J.L. Luo and N.L. Wang, Phys. Rev. Lett. [**101**]{}, 057007 (2008).
L. Boeri, O.V. Dolgov, and A.A. Golubov, Phys. Rev. Lett. [**101**]{}, 026403 (2008).
D.J. Singh and M.-H. Du, Phys. Rev. Lett. [**100**]{}, 237003 (2008).
C. Cao, P.J. Hirschfeld and H.-P. Cheng, Phys. Rev. B [**77**]{}, 220506 (2008).
D.V. Evtushinsky, D.S. Inosov, V.B. Zabolotnyy, A.Koitzsch, M. Knupfer, B. Büchner, Viazovska, M.S., G.L. Sun, V. Hinkov, A.V. Boris, C.T. Lin, B. Keimer, A. Varykhalov, A.A. Kordyuk and S.V. Borisenko, Phys. Rev. B [**79**]{}, 054517 (2009)
H. Ding, P. Richard, K. Nakayama, K. Sugawara, T. Arakane, Y. Sekiba, A. Takayama, S. Souma, T. Sato, T. Takahashi, Z. Wang, X. Dai, Z. Fang, G.F. Chen, J.L. Luo and N. L. Wang, Europhysics Letters [**83**]{}, 47001 (2008).
M.J. Cooper, P.E. Mijnarends, N. Shiotani, N. Sakai and A. Bansil (editors), *X-Ray Compton Scattering*, Oxford University Press (2004).
J. Laukkanen, K. Hämäläinen, S. Manninen, A. Shukla, T. Takahashi, K. Yamada, B. Barbiellini, S. Kaprzyk and A. Bansil, J. Phys. Chem. Solids [**62**]{}, 2249 (2001). (2004).
A. Bansil, S. Kaprzyk, P.E. Mijnarends and J. Toboła, Phys. Rev. B [**60**]{}, 13396 (1999).
Y. Qiu, M. Kofu, W. Bao, S.-H. Lee, Q. Huang, T. Yildirim, J.R.D. Copley, J.W. Lynn, T. Wu, G. Wu, and X.H. Chen, Phys. Rev. B [**78**]{}, 052508 (2008).
I. Matsumoto, J. Kwiatkowska, F. Maniawski, M. Itou, H. Kawata, N. Shiotani, S. Kaprzyk, P.E. Mijnarends, B. Barbiellini and A. Bansil, Phys. Rev. B [**64**]{}, 045121 (2001).
D.G. Lock, V.H.C. Crisp, and R.N. West, J. Phys. F: Met. Phys. [**3**]{}, 561 (1973).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Integer factorization is a very hard computational problem. Currently no efficient algorithm for integer factorization is publicly known. However, this is an important problem on which it relies the security of many real world cryptographic systems.
I present an implementation of a fast factorization algorithm on MapReduce. MapReduce is a programming model for high performance applications developed originally at Google. The quadratic sieve algorithm is split into the different MapReduce phases and compared against a standard implementation.
author:
- Javier Tordable
title: MapReduce for Integer Factorization
---
Introduction
============
The security of many cryptographic algorithms relies on the fact that factoring large integers is a very computationally intensive task. In particular RSA [@1] would be vulnerable if there was an efficient algorithm to factor semiprimes (products of two primes). This could have severe consequences, as RSA is one of the most widely used algorithms in electronic commerce applications [@2].
There are many algorithms for integer factorization [@3]. From the trivial trial division to the classical Fermat’s factorization method [@4] and Euler’s factoring method [@5] to the modern algorithms, the quadratic sieve [@6] and the number field sieve [@7]. In particular the number field sieve algorithm was used in 1996 to factor a 512 bit integer [@8], the lowest integer length used in commercial RSA implementations. There have been several other big integers factored over the course of the last decade. I would like to point out that in those cases the feat was accomplished with tremendous effort developing the software and a very considerable investment in hardware [@9],[@10].
In what follows I will expose how MapReduce, a distributed computational framework, can be used for integer factorization. As an example I will show an implementation of the quadratic sieve algorithm. I will also compare in terms of performance and cost a conventional implementation with the MapReduce implementation.
MapReduce
=========
I claim no participation in the development of the MapReduce framework. This section is basically a short extract of the original MapReduce paper by Jeff Dean and Sanjay Ghemawat [@11]. MapReduce is a programming model inspired in computational programming. Users can specify two functions, *map* and *reduce*. The *map* function processes a series of (key, value) pairs, and outputs intermediate (key, value) pairs. The system automatically orders and groups all (key, value) pairs for a particular key, and passes them to the reduce function. The reduce function receives a series of values for a single key, and produces its output, which is sometimes a synthesis or aggregation of the intermediate values.
The canonical example of a MapReduce computation is the construction of an inverted index. Let’s take a collection of documents $\mathit{\mathcal{D}}=\left\{ D_{0},D_{1},...,D_{N}\right\} $ which are composed of words $D_{0}=\left(d_{0,0},d_{0,1},...,d_{0,L_{0}}\right),D_{1}=\left(d_{1,0},d_{1,1},...,d_{1,L_{1}}\right)$ and so on. We define a map function the following way:
$$map:(i,D_{i})\rightarrow\left\{ \left(d_{i,0},\left(i,0\right)\right),\left(d_{i,1},\left(i,1\right)\right),...,\left(d_{i,L_{i}},\left(i,L_{i}\right)\right)\right\}$$
that is, for a given document it processes each word in the document and outputs an intermediate pair. The key is the word itself, and the value is the location in the corpus, indicated as (document, position). The reduce function is defined as:
$$reduce:\left\{ \left(d,\left(i_{1},j_{1}\right)\right),...,\left(d,\left(i_{L},j_{L}\right)\right)\right\} \rightarrow\left(d,\left\{ \left(i_{1},j_{1}\right),...,\left(i_{L},j_{L}\right)\right\} \right)$$
For a collection of pairs with the same key (the same word), it outputs a new pair, in which the key is the same, and the value is the aggregation of the intermediate values. In this case, the set of locations (document and position in the document) in which the word can be found in the corpus.
The MapReduce implementation automatically takes care of the parallel execution in a distributed system, data transmission, fault tolerance, load balancing and many other aspects of a high performance parallel computation. The MapReduce model escales seamlessly to thousands of machines. It is used continously for a multitude of real world applications, from machine learning to graph computations. And most importantly the effort required to develop a high performance parallel application with MapReduce is much lower than using other models, like for example MPI [@12].
Quadratic Sieve
===============
The Quadratic Sieve algorithm was conceived by Carl Pomerance in 1981. A detailed explanation of the algorithm can be found in [@13]. Here we will just review the basic steps. Let $N$ be the integer that we are trying to factor. We will attempt to find $a,b$ such that: $N\mid\left(a^{2}-b^{2}\right)\Rightarrow N\mid\left(a+b\right)\left(a-b\right)$. If $\left\{ \left(a+b,N\right),\left(a-b,N\right)\right\} \neq\left\{ 1,N\right\} $ then we will have a factorization of $N$.
Lets define:$$Q\left(x\right)=x^{2}-N$$
if we find $x_{1},x_{2},...x_{K}$ such that $\prod_{i=1}^{K}Q\left(x_{i}\right)$ is a perfect square, then: $$N\mid\prod_{i=1}^{K}Q\left(x_{i}\right)-\left(\prod_{i=1}^{K}x_{i}\right)^{2}=\prod_{i=1}^{K}\left(x_{i}^{2}-N\right)-x_{1}^{2}x_{2}^{2}...x_{K}^{2}$$
Finding Squares
---------------
Let’s take a set of integers $x_{1},...,x_{L}$ which are $B$-smooth (all $x_{i}$ factor completely into primes $\leq B$). One way to look for $i_{1},i_{2},...,i_{M}$ such that $\prod_{j=1}^{M}x_{i_{j}}$ is a square is as follows. Let’s denote $p_{i}$ the i-th prime number. $\prod_{j=1}^{M}x_{i_{j}}=p_{j_{1}}^{a_{1}}p_{j_{2}}^{a_{2}}...p_{j_{L}}^{a_{L}}$ is a square if and only if $2\mid a_{k}$ for all $k$ $\Leftrightarrow a_{k}\equiv0\, mod\,(2)$. For each $x_{i}$ we will obtain a vector $v^{i}=v\left(x_{i}\right)$ where $v_{j}^{i}=max\left\{ k:p_{j}^{k}\mid x_{i}\right\} \, mod\,\left(2\right)$. That is, each component $j$ of $v^{i}$ is the exponent of $p_{j}$ in the factorization of $x_{i}$ modulo $2$. For example, for $B=4$:
$$\begin{aligned}
x_{1}=6,v^{1}=\left(1,1,0,0\right)\\
x_{2}=45,v^{2}=\left(0,0,1,0\right)\\
x_{3}=75,v^{3}=\mbox{\ensuremath{\left(0,1,0,0\right)}}\end{aligned}$$
It is immediate that:
$$v\left(\prod_{j=1}^{M}x_{i_{j}}\right)=\sum_{j=1}^{M}v\left(x_{i_{j}}\right)$$
Then$$\prod_{j=1}^{M}x_{i_{j}}\mbox{is a square}\Leftrightarrow v\left(\prod_{j=1}^{M}x_{i_{j}}\right)=\overrightarrow{0}$$
In conclussion, in order to find a subset of $x_{1},...,x_{L}$ which is a perfect square, we just need to solve the linear system:$$\left(\begin{array}{ccccccc}
v^{1} & \mid & v^{2} & \mid & \ldots & \mid & v^{L}\end{array}\right)\left(\begin{array}{c}
e_{1}\\
e_{2}\\
\vdots\\
e_{L}\end{array}\right)\equiv\overrightarrow{0}\, mod\,(2)$$
Sieving for smooth numbers
--------------------------
Back to the original problem, we just need to find a convenient set $\left\{ x_{1},x_{2},...,x_{L}\right\} $ such that $\left\{ Q\left(x_{1}\right),Q\left(x_{2}\right),...,Q\left(x_{L}\right)\right\} $ are $B$-smooth numbers for a particular $B$. First of all, lets notice that we don’t need to consider every prime number $\leq B$. If a prime $p$ verifies: $p\mid Q(x)$ for some $x$ then:
$$p\mid Q(x)\Leftrightarrow p\mid x^{2}-N\Leftrightarrow x^{2}\equiv N\, mod\,(p)\Leftrightarrow\left(\frac{N}{p}\right)=1$$
Because $N$ is a quadratic residue modulo p if and only if the Legendre symbol of n over p is 1. We will take a set of primes which verifies that property and we will call it *factor base*.
In order to consider smaller values of $Q(x)$ we will take values of $x$ around $\sqrt{N},$ i.e. $x\in\left[\lfloor\sqrt{N}\rfloor-M,\lfloor\sqrt{N}\rfloor+M\right]$ for some $M.$ Both $B$ above and $M$ here are chosen as indicated in [@13].
In order to factor all the $Q(x_{i})$ we will use a method called *sieving* which is what gives the quadratic sieve its name. Notice that $p\mid Q(x)\Rightarrow p\mid Q(x+kp)=x^{2}+2kpx+k^{2}p^{2}-N=\left(x^{2}-N\right)+p\left(2kx+k^{2}p\right)$. Then$$Q(x)\equiv0\, mod\,(p)\Rightarrow\forall k\in\mathbb{N},Q(x+kp)\equiv0\, mod\,(p)$$
We can solve the equation $Q(x)\equiv0\, mod\,(p)\Leftrightarrow x^{2}-N\equiv0\, mod\,(p)$ efficiently and obtain two solutions $s_{1},s_{2}$ [@14]. If we take: $$z_{p,\left\{ 1,2\right\} }=min\left\{ x\in\left[\lfloor\sqrt{N}\rfloor-M,\lfloor\sqrt{N}\rfloor+M\right]:x\equiv s_{\left\{ 1,2\right\} }\, mod\,(p)\right\}$$ then all $Q\left(z_{p,\{1,2\}}+kp\right),k\in\left[0,K\right]$ are divisible by $p$. We can divide each one of them by the highest power of $p$ possible. For example:
$$\begin{aligned}
\left(x_{i}\right)= & \left(\ldots,6,7,8,9,10,\ldots\right)\\
\left(Q\left(x_{i}\right)\right)= & \left(\ldots,-41,-28,-13,4,23,\ldots\right)\\
& \left(\frac{77}{2}\right)=1\mbox{ as }77\equiv1\equiv1^{2}\, mod\,(2)\\
& x^{2}-77\equiv0\, mod\,(2)\mbox{ yields }1,3,5,7,9,...\\
& \left(\ldots,-41,-7,-13,1,23,\ldots\right) & \mbox{after sieving by }2\end{aligned}$$
After sieving for every appropriate $p$, all the $Q(z)$ that are equal to $1$ are smooth over the factor base.
Method
======
I developed a basic implementation of the Quadratic Sieve MapReduce which runs on Hadoop [@15]. Hadoop is an open source implementation of the MapReduce framework. It is made in Java and it has been used effectively in configurations ranging from one to a few thousand computers. It is also available as a commercial cloud service [@16].
This implementation is simply a proof of concept. It relies too heavily on the MapReduce framework and it is severy bound by IO. However the size and complexity of the implementation are several orders of manitude lower than many competing alternatives.
The 3 parts of the program are :
- *Controller*: Is the master job executed by the platform. It runs before spawning any worker job. It has two basic functions: first it generates the factor base. The factor base is serialized and passed to the workers as a counter. Second it generates the full interval to sieve. All the data is stored in a single file in the distributed Hadoop file system [@17]. It then relies on the MapReduce framework to automatically split it in an adequate number of shards and distribute it to the workers
- *Mapper*: The mappers perform the sieve. Each one of them receives an interval to sieve, and they return a subset of the elements in that input sieve which are smooth over the factor base. All output elements of all mappers share the same key
- *Reducer*: The reducer receives the set of smooth numbers and attempts to find a subset of them whose product is a square by solving the system modulo 2 using direct bit manipulation. If it finds a suitable subset, it tries to factor the original number, N. In general there will be many subsets to choose from. In case that the factorization is not succesful with one of them, it proceeds to use another one. The single output is the factorization
In order to compare performance I developed another implementations of the Quadratic Sieve algorithm in Maple. Both implementations are basic in the sense that they implement the basic algorithm described above and the code has not been heavily optimized for performance. There are many differences between the two frameworks used that could impact performance. Because of that a direct comparison of running times or memory space may not be meaningful. However it is interesting to notice how each of the implementations scales depending on the size of the problem. The source code is available online at http://www.javiertordable.com/research.
Results
=======
Figures 1 and 2 show the results both in absolute terms and normalized. Figure 3 shows the disk usage of the MapReduce implementation. To test both implementations I took a set of numbers of different sizes[^1]. The number of decimal digits $d$ is indicated in the first column of each table. In order to contruct those numbers I took two factors close to $10^{\frac{d}{2}}$, with their product slightly over $10^{d}$.
In each table sieve size indicates the number of elements that the algorithm analyzed in the sieve phase. For the MapReduce application the time result is taken from the logs, and the memory result is obtained as the maximum memory used by the process. For the Maple implementation both time and memory data are taken from the on screen information in the Maple environment. Finally disk usage data for the MapReduce is taken as the size of the file that contains the list of numbers to sieve. The Maple program runs completely in memory for the samples analyzed.
Decimal Sieve
--------- ----------- ------------ --------------- ------------ ---------------
Digits Size Time ($s$) Memory ($MB$) Time ($s$) Memory ($MB$)
$10$ $5832$ $2.0$ $149.6$ $0.1$ $7.5$
$15$ $85184$ $3.0$ $397.1$ $3.5$ $15.5$
$20$ $970299$ $35.0$ $463.1$ $116.0$ $100.8$
$25$ $7529536$ $495.0$ $670.0$ $3413.7$ $894.0$
: Absolute performance of the MapReduce and Maple implementations
Decimal Sieve
--------- ---------- --------- -------- ----------- ---------
Digits Size Time Memory Time Memory
$10$ $1.0$ $1.0$ $1.0$ $1.0$ $1.0$
$15$ $14.6$ $1.5$ $2.7$ $35.0$ $2.1$
$20$ $166.4$ $17.5$ $3.1$ $1160.0$ $13.4$
$25$ $1291.1$ $247.5$ $4.5$ $34137.0$ $119.2$
: Normalized performance of the MapReduce and Maple implementations
Decimal Absolute Sieve Relative Sieve Absolute Relative
--------- ---------------- ---------------- ------------- -------------
Digits Size Size Disk ($MB$) Disk ($MB$)
$10$ $5832$ $1.0$ $0.1$ $1.0$
$15$ $85184$ $14.6$ $2.1$ $14.6$
$20$ $970299$ $166.4$ $29.4$ $166.4$
$25$ $7529536$ $1291.1$ $275.3$ $1291.1$
: Disk usage of the MapReduce implementation
Discussion
==========
The MapReduce implementation has a relatively big setup cost in time and memory when compared with an application in a conventional mathematical environment. However it scales better with respect to the size of the input data.
MapReduce is optimized to split and distribute data form disk. If an application handles a significant volume of data, IO capacity and performance can be a limiting factor. In our case disk usage is directly proportional to the size of the sieve set, which grows exponentially on the number of digits.
Both MapReduce and Maple implementations are similar in terms of development effort. The Maple implementation seems more adequate for small-sized problems while the MapReduce application is more efficient for medium-sized problems. Also it will be easier to scale in order to solve harder problems.
[17]{} Rivest, R.; A. Shamir; L. Adleman. 1978. A Method for Obtaining Digital Signatures and Public-Key Cryptosystems. Communications of the ACM 21 (2): 120126.
Nash, A., Duane, W., and Joseph, C. 2001. Pki: Implementing and Managing E-Security. McGraw-Hill, Inc.
Donald Knuth. 1997. The Art of Computer Programming, Volume 2: Seminumerical Algorithms, Third Edition. Addison-Wesley. ISBN 0-201-89684-2. Section 4.5.4: Factoring into Primes, pp. 379417
Israel Kleiner. 2005. Fermat: The Founder of Modern Number Theory. Mathematics Magazine, Vol. 78, No. 1 (Feb., 2005), pp. 3-14
McKee, James. 1996. Turning Euler’s Factoring Method into a Factoring Algorithm; in Bulletin of the London Mathematical Society; issue 28 (volume 4); pp. 351-355
Pomerance, C. 1985. The quadratic sieve factoring algorithm. In Proc. of the EUROCRYPT 84 Workshop on Advances in Cryptology: theory and Application of Cryptographic Techniques. Springer-Verlag New York. 169-182.
Lenstra, A. K., Lenstra, H. W., Manasse, M. S., and Pollard, J. M. 1990. The number field sieve. In Proceedings of the Twenty-Second Annual ACM Symposium on theory of Computing. ACM, New York, NY, 564-572.
Cowie, J., Dodson, B., et al. 1996. A World Wide Number Field Sieve Factoring Record: On to 512 Bits. In Proceedings of the international Conference on the theory and Applications of Cryptology and information Security. Lecture Notes In Computer Science, vol. 1163. Springer-Verlag, London, 382-394.
Golliver, R. A., Lenstra, A. K., and McCurley, K. S. 1994. Lattice sieving and trial division. In Proceedings of the First international Symposium on Algorithmic Number theory. L. M. Adleman and M. A. Huang, Eds. Lecture Notes In Computer Science, vol. 877. Springer-Verlag, London, 18-27.
S. Cavallar and W. M. Lioen and H. J. J. Te Riele and B. Dodson and A. K. Lenstra and P. L. Montgomery and B. Murphy Et Al and Mathematisch Centrum. 2000. Factorization of a 512-bit RSA modulus. Proceedings of Eurocrypt 2000. Springer-Verlag. 1-18.
Dean, J. and Ghemawat, S. 2004. MapReduce: Simplified Data Processing on Large Clusters. OSDI’04: Sixth Symposium on Operating System Design and Implementation, San Francisco, CA, December, 2004, 137-150
Gropp, W., Lusk, E., and Skjellum, A. 1994. Using Mpi: Portable Parallel Programming with the Message-Passing Interface. MIT Press. 257-260
Carl Pomerance. 1996. A Tale of Two Sieves, Notices of the AMS, 1473-1485
Niven, I. and Zuckerman, H.S. and Montgomery, H.L. 1960. An introduction to the theory of numbers. John Wiley and Sons, Inc. 110-115
http://hadoop.apache.org/
http://aws.amazon.com/elasticmapreduce/
Borthakur, D. 2007. The hadoop distributed file system: Architecture and design. http://svn.apache.org/repos/asf/hadoop/core/tags/release-0.15.3/docs/hdfs\_design.pdf
[^1]: 1164656837, 117375210056563, 10446257742110057983, 1100472550655106750000029
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Swift J1357.2-0933 is one of the shortest orbital period black hole X-ray transients (BHTs). It exhibited deep optical dips together with an extremely broad H$\alpha$ line during outburst. We present 10.4-m GTC time-resolved spectroscopy during quiescence searching for donor star absorption features. The large contribution of the accretion flow to the total luminosity prevents the direct detection of the companion. Nevertheless, we constrain the non-stellar contribution to be larger than $\sim 80\%$ of the total optical light, which sets new lower limits to the distance ($d > 2.29\, \rm{kpc}$) and the height over the Galactic plane ($z>1.75\, \rm{kpc}$). This places the system in the galactic thick disc. We measure a modulation in the centroid of the H$\alpha$ line with a period of $P=0.11\pm0.04\, \rm{d}$ which, combined with the recently presented FWHM-$K_2$ correlation, results in a massive black hole ($M_1>9.3 \, \rm{M_\odot}$) and a $\sim$ M2V companion star ($M_2\sim 0.4\, \rm{M_\odot}$). We also present further evidence supporting a very high orbital inclination ($i\gtrsim 80^\circ$).'
author:
- |
D. Mata Sánchez $^{1,2}$[^1], T. Muñoz-Darias$^{1,2}$, J. Casares$^{1,2}$, J. M. Corral-Santana $^{3}$ and T. Shahbaz$^{1,2}$\
$^{1}$Instituto de Astrofísica de Canarias, 38205 La Laguna, Tenerife, Spain\
$^{2}$Departamento de astrofísica, Univ. de La Laguna, E-38206 La Laguna, Tenerife, Spain\
$^{3}$ Instituto de Astrofísica, Pontificia Universidad Católica de Chile, Casilla 306, Santiago 22, Chile\
bibliography:
- 'MiBiblio.bib'
date: 'Accepted 2015 September 8. Received 2015 September 7; in original form 2015 August 3 '
title: 'Swift J1357.2-0933: a massive black hole in the galactic thick disc'
---
\[firstpage\]
accretion, accretion discs – X-rays: binaries – stars: black holes
Introduction {#intro}
============
Galactic X-ray binaries offer the best opportunity to detect stellar mass black holes (BHs) through the accretion luminosity produced by material from the companion star falling into its gravitational well. The vast majority of BHs are detected in transient low-mass X-ray binaries (LMXBs), which spend most part of their lives in a faint, quiescent state (about $\rm 10^{30-34}\, erg\, s^{-1}$; see e.g. @Armas-Padilla2014b). They are discovered during occasional outburst, where their X-ray luminosity increases above $\sim 10$ per cent of the Eddington limit. There are only 17 dynamically confirmed BHTs, as well as $\sim 33$ candidates (see @Casares2014; @Corral-Santana2015). Swift J1357.2-0933 (hereafter J1357) is a LMXB X-ray transient with galactic coordinates $l=328.702 {\,}^{\circ}$ and $b=+50.004{\,}^{\circ}$ [@Krimm2011], whose peak X-ray luminosity place it in the very faint regime (see @Armas-Padilla2013 for an X-ray analysis during outburst). The detection of optical dips both in outburst (@Corral-Santana2013, hereafter CS13) and quiescence [@Shahbaz2013] suggests J1357 is a very high inclination system. The orbital period is among the shortest of its class ($P= 2.8\, \rm{h}$, CS13) and the mass function is constrained to be $f(M_1)> 3.0\, \rm{M_\odot}$, where the radial velocity of the donor ($K_2$) was estimated from the double-peak separation of the H$\alpha$ profile [@Orosz1994; @Orosz1995]. This strongly advocates for the presence of a BH. @Rau2011 proposed a tentative distance to the system of $d\sim 1.5\,\rm{kpc}$ considering the donor star to be the dominant source of quiescent light in the optical and near-infrared regime. However, the quiescent spectral energy distribution (SED) is best described by a single power-law model, which suggests little, if any, thermal contribution from the secondary star [@Shahbaz2013]. This implies that only lower limits to the distance can be obtained unless the donor star contribution is properly characterized.
Observations
============
{width="168mm"}
J1357 was observed with the Optical System for Imaging and low-Intermediate-Resolution Integrated Spectroscopy (OSIRIS) located in the Nasmyth-B focus of the 10.4-m GTC, in La Palma (Spain). We used the R500R optical grism combined with a 1.0” slit ($\rm R=352,\, 4.88$ [Å]{}/pix), covering the spectral range $\rm 4800-10000$ [Å]{}. Observations were obtained in four different nights within April to June 2014.
Eleven consecutive spectra per night were acquired at airmasses in the range $\sim 1.3 - 1.7$, with individual exposure times of $\rm 875\, s$ ($\rm \sim 3\, h$ per night) sampling the proposed $\rm 2.8\, h$ orbital period. We used IRAF[^2] standard routines for bias and flat-field corrections. The slit was rotated to $\rm{PA}=-53.16^\circ$ in order to allow for simultaneous observations of the object and a nearby early-type dwarf star located at $2.06\, \rm{arcmin}$ NE of the target. The wavelength calibration was obtained from observations of calibration arcs and instrumental flexure was calculated from the drift of the \[\] $\rm 5577$ [Å]{} sky line and subsequently corrected in every individual spectrum.
The spectrum of the calibration star is classified as a F8V after comparison with templates from the *MILES* spectral library (@Sanchez-Blazquez2006, @Falcon-Barroso2011) using $\chi^2$ minimization routines within MOLLY (e.g. @Casares1996). In order to maximize the chances of detecting donor star absorption features in J1357, we produced a set of telluric spectra by removing the F8V spectral lines, which were subsequently substracted from the target to produce spectra free of telluric features (e.g. @Beekman2000).
The above strategy yielded four sets of eleven spectra. Points deviating more than $\rm 3\sigma$ above or below the continuum, probably caused by bad pixels or residuals from the sky subtraction, were interpolated. One spectrum was discarded on each of the two first nights due to very poor signal-to-noise ratio (S/N) caused by the presence of clouds. Absorption features from the donor star are not evident in any of the spectra, nor in the combined spectrum (see Fig. \[figaverage\]).
Results
=======
Using the 42 GTC spectra we have determined the veiling factor, the orbital period and the main properties of the H$\alpha$ line in quiescence.
Search for companion star features: Skew-mapping
------------------------------------------------
The skew-mapping technique has been applied in cataclysmic variables to detect weak companion star features (e.g. @Vande2003, @Smith2005). Our analysis compares a template M3V star spectrum with different Doppler-corrected averaged spectra created from our database. Each averaged spectrum is produced by considering a specific zero phase in the range 0-1 (we inspect all values with a 0.05 step) and velocity of the companion star between $K_2= 0-2000\, \rm{km\, s^{-1}}$, in $100\, \rm{km\, s^{-1}}$ steps. Individual spectra are shifted and coadded assuming the orbital period of CS13. The cross-correlation with the M3 template star is expected to reveal a significant peak above noise at the correct parameters for the companion star motion. This technique did not reveal any preferred value in the *$K_2$ - zero phase* parameter space, not even after considering several different periods around the one proposed in CS13 ($0.09-0.15\, \rm{d^{-1}}$, steps of $0.001\, \rm{d^{-1}}$). Therefore, we conclude that no companion star features are present in our spectra.
Veiling factor {#veiling}
--------------
Magnitudes in the OSIRIS $r'$ filter were obtained from the acquisition images of each night: $r'_1=21.71\pm 0.08;\quad r'_2=21.45\pm 0.06;\quad r'_3=20.84\pm 0.08 ;\quad r'_4=21.00\pm 0.11$. The first two values (April 29th and 30th) are similar to those reported in @Shahbaz2013 [$r'=21.54\pm 0.35$]. On the other hand, the latest two nights (June 2nd and 28th) reveal a somehow brighter system (but still in quiescence). We note that although this might be caused by the known strong short-term variability displayed by the system [@Shahbaz2013], the spectra of the first two nights are clearly noisier under similar sky conditions, reflecting a true drop in brightness.
The absence of companion star features imposes a minimum constraint to the veiling factor ($X$), defined as the fractional contribution of the accretion related luminosity ($L^{r'}_{\rm{acc}}$) to the total flux ($L^{r'}_{\rm{acc}}+L^{r'}_2$, where $L^{r'}_2$ is the donor star luminosity) in the OSIRIS $r'$ filter wavelength range ($\rm 5500-7400$ [Å]{}). On the other hand, by considering the relation between the orbital period and the mean density presented in @Faulkner1972, and tabulated values for main-sequence stars [@Cox2000], the spectral type of the companion star is constrained to be M2V or later. We compared our spectrum with templates of main-sequence stars from the *MILES* Spectral Library. For each spectral type, we measured the normalized flux of the deepest photospheric absorption line present in the template spectrum. The veiling necessary to make these features shallower than the noise level ($3\sigma$) within the corresponding spectral region of the J1357 spectrum is a lower limit to the veiling factor (see @MataSanchez2015 for a similar analysis of NIR spectra).
We find that the required veiling factor for a $\rm M2V$ star to be swamped by the accretion flux is $X>0.81$. This limit is also valid if the donor is evolved and its spectral type is later than $\rm M2V$ (we have explored spectral types as late as $\rm M6V$ ). This result is obtained using the averaged quiescent spectrum of the first two nights where the system is fainter, and therefore, a higher contribution of the companion star is expected.
We have explored spectral types as late as $\rm M6V$, obtaining similar results than for $\rm M2V$. We note that using slightly different tabulated values for dwarf star mean densities (e.g. @Carroll2007), does not affect our results.
Orbital period from a two-Gaussian model fit {#period}
--------------------------------------------
We fit the H$\alpha$ double-peaked emission profile with a model consisting of two Gaussians with equal FWHM. The height of the Gaussians and offset with respect to the H$\alpha$ rest wavelength ($\rm 6562.78$ [Å]{}) are left as free parameters.
![Orbital evolution of several H$\alpha$ parameters obtained from our two-Gaussian model fit (three top panels) and from a diagnostic diagram (bottom panel) folded with the best fit period obtained from periodogram analysis. Panel a: intensity ratio between red and blue peak. Best fit is represented as a black, solid line. Panel b: radial velocities of the non-detrended centroid as measured by the two-Gaussian model. In both panels, each night is represented by different symbols and colours as follows: red circles (April 29th), blue triangles (April 30th), green upside-down triangles (June 2nd) and orange squares (June 28th). The best sinusoidal fit to the non-detrended data is also shown. Panel c: radial velocities of the line centroid obtained from $10$ phase bin averaged spectra after de-trending. The best sinusoidal fit is shown as a blue, solid line. Panel d: same as previous panel but using radial velocity curves from the diagnostic diagram (Gaussian separation of $a=4500\, \rm{km\,s^{-1}}$). Data are plotted over two phase cycles for clarity. []{data-label="figperfold"}](Period_fold_art_totabcd.ps){width="85mm"}
### H$\alpha$ centroid velocity {#halpha-centroid-velocity .unnumbered}
The mean of the Gaussians offsets is expected to trace the motion of the accretion disc around the centre of mass of the system. We note that the mean offset value for each night was significantly different (see Fig. \[figperfold\]b). This behaviour has been reported in other BHTs, for example in XTE J1118+480 it is explained as a result of disc precession [@Torres2002]. Recently, @Torres2015 detected variations in the systemic velocity of J1357. They did not consider this result to support the precessing disc scenario, but interpreted as systematic effects. However, our longer database suggests that instead this is a consequence of a physical process taking place in the system. In order to fit the whole sample, we then decided to de-trend the mean nightly offsets.
We performed a Lomb-Scargle normalised periodogram of the (de-trended) mean offset variations and the result is presented in Fig. \[figpgram\]. We used the *period* time-series analysis package [@Dhillon2001]. A Gaussian fit to the broad peak with the highest power yields a period of $P=0.11\pm 0.04\, \rm{d}$ ($1\, \sigma$). This is consistent with the one measured by CS13 using a longer database with higher S/N during outburst ($P=0.117\pm 0.013\,\rm{d}$). We find two families of peaks consistent with the outburst period. Gaussian fits result in $P=0.109\pm 0.005\,\rm{d}$ and $P=0.121\pm 0.006\,\rm{d}$, respectively. The former is favoured by its highest amplitude (depicted as a red, stripped band in Fig. \[figpgram\]). Nevertheless, further observations are necessary to confirm this result. We note that comparable results are obtained by performing sinusoidal fits and minimise chi-square.
![Periodogram obtained with the Lomb-Scargle method. Red, solid line shows a Gaussian fit to the highest group of peaks. The red, dotted horizontal line represents the $1\sigma$ uncertainty ($P=0.11\pm 0.04\,\rm{d}$). We inspected frequencies up to $60\, \rm{d^{-1}}$, but the power spectrum for frequencies over $30\, \rm{d^{-1}}$ is nearly flat. The red, dashed vertical line refers to the strongest peak consistent with the outburst period proposed by CS13. Grey band depicts the constraint to the period ($P=0.117\pm 0.013\,\rm{d}$) from outburst data (CS13). The red, stripped band refers to the standard deviation obtained from a Gaussian fit to the family of peaks inside the grey band ($P=0.109\pm 0.005\,\rm{d}$). []{data-label="figpgram"}](Pgram_art.ps){width="85mm"}
The data were folded on the previously obtained period and are shown in Fig. \[figperfold\]b, where different colours and symbols are used to mark different systemic velocities observed on each night. We subsequently performed a non-linear least squares fit to the data using the following sinusoidal function: $$V=\gamma + K_1 \sin{(2\pi(\phi-\phi_{~0}))}$$ We obtained $K_1=44\pm 3 \, \rm{km\, s^{-1}}$, $\gamma=75\pm 2 \, \rm{km\, s^{-1}}$ and $\phi_{~0} = 0.71\pm 0.01$. Even if it might be qualitatively acceptable, the high reduced chi-square value ($\chi_\mathrm{~r}^2$=17.26) shows that the non de-trended data does not properly fit to a sinusoid, and de-trending is necessary for an optimal analysis. Subsequently, the original spectra were combined into $10$ phase bins after removing each night’s mean value. The fit results in $\phi_{~0} = 0.720\pm 0.009$ and $K_1=61\pm 4 \, \rm{km\, s^{-1}}$ for $\chi_\mathrm{~r}^2$=1.30 (see Fig. \[figperfold\]c).
### Intensity ratio and double peak separation {#intensity-ratio-and-double-peak-separation .unnumbered}
The intensity ratio of the peaks of the H$\alpha$ profile (defined as $I_{\rm{red}}/I_{\rm{blue}}$) exhibits a periodicity similar to that observed in the centroid offsets. Folding these data onto the previously obtained period and fitting a sinusoid results in $\chi_\mathrm{~r}^2$=5.12 (see Fig. \[figperfold\]a). @Torres2015 reported asymmetric H$\alpha$ profiles, but no recurrent variability was observed. The detection of this periodic evolution of the profile along four different nights suggest an scenario with an asymmetric outer disc structure, perhaps a hot-spot on a tidal arm. This would also explain the deviation observed from a perfect fit, since other processes affecting the intensity ratio might be at play. The double-peak separation does not show significant variability, and it is consistent with a constant value. The weighted average of the double-peak separation is $D_p=2430\pm 50 \, \rm{km\, s^{-1}}$, significantly larger than the one measured in outburst by CS13 ($D_p=1790\pm 67 \, \rm{km\, s^{-1}}$). This implies that the outer disc velocity ($v_d=1215\pm 25\, \rm{km\, s^{-1}} $) is higher in quiescence, which is an expected behaviour since the disc expands during outburst, reaching areas with lower velocities. The reported value is consistent within $2\sigma$ with that reported by @Torres2015 [$D_p=2340\pm 20 \, \rm{km\, s^{-1}}$].
The diagnostic diagram {#dgmethod}
----------------------
The broad wings of the H$\alpha$ profile are supposed to trace the motion of the compact object. We apply the diagnostic diagram (aka double Gaussian technique @Shafter1986), aiming at constraining radial velocity variations of the H$\alpha$ wings. We inspected several Gaussian separations ($ a=2500-5000\,\rm{km\, s^{-1}}$) to perform a diagnostic diagram analogous to that presented in CS13. Unfortunately, our limited database and the lower S/N ratio prevents us from obtaining conclusive determinations for the $K_1$ and $\gamma$ values. A wide plateau on the evolution of the fitted parameters from $ a=4000\,\rm{km\, s^{-1}}$, combined with the inspection of the individual radial velocity curves, suggests $ a=4500\,\rm{km\, s^{-1}}$ as the preferred value (see the folded, radial velocity curve for this separation in Fig. \[figperfold\]d).
Mean values of the compact object orbital velocity ($K_1=40\pm 12 \,\rm{km\, s^{-1}}$) and systematic velocity of the system ($\gamma = -79 \pm 13 \,\rm{km\, s^{-1}}$) are obtained from sinusoidal fits to radial velocity curves in the range $ a=4000-5000\,\rm{km\, s^{-1}}$ (uncertainties refer to standard deviations). This result is consistent with previous determinations of the systemic velocity in quiescence ($\gamma = -130 \pm 50 \,\rm{km\, s^{-1}}$; @Torres2015) and the compact object’s velocity reported in outburst ($K_1=43\pm 2 \,\rm{km\, s^{-1}}$; CS13). However, the systemic velocity value obtained from the outburst periodogram ($\gamma \sim -150 \,\rm{km\, s^{-1}}$, CS13) does not seem to be in agreement with our results.
We note that, in contrast to the two-Gaussian profile modelling (see Section \[period\]), nigthly de-trending is not necessary for the radial velocity curves obtained with the double Gaussian technique. This can be explained by considering that the two-Gaussian modelling traces the core of the H$\alpha$ double peak profile, and therefore, outer parts of the disc where disc preccesion effects might be important. However, the diagnostic diagram is sensitive to the emission line wings, which traces inner parts of the disc less affected by companion star tidal forces.
H-alpha full-width-at-half-maximum
----------------------------------
We have measured the full-width-at-half-maximum (FWHM) of the H$\alpha$ emission profile in each individual spectrum. After substracting quadratically the instrumental resolution, the average value is $\rm{FWHM}=4152\pm 209 \, \rm{km\,s^{-1}}$, where the uncertainty reflects the standard deviation of the 42 measures. This value is larger than the one measured in outburst ($\rm{FWHM}\sim 3300 \, \rm{km\,s^{-1}}$, CS13) and consistent with the quiescence value reported in [@Torres2015] and @Casares2015, i.e. $\rm{FWHM} = 4025 \pm 110 \, \rm{km\,s^{-1}}$ and $\rm{FWHM} = 4085 \pm 328 \, \rm{km\,s^{-1}}$ respectively. The previously reported broadest FWHM corresponds to XTE J1118+480 ($\rm{FWHM}\sim 2500 \, \rm{km\,s^{-1}}$, @Torres2004).
Discussion
==========
Observations taken during the outburst decay have sometimes resulted in wrong orbital period determinations (e.g. V404 Cygni; @Casares1992). However, in our case, the orbital period measured from quiescent spectra is fully consistent with the outburst results, supporting CS13 conclusions. @Casares2015 has recently presented a correlation between the FWHM of H$\alpha$ and the orbital velocity of the companion star $K_2 \simeq 0.233(13)\, \rm{FWHM}$. For the case of J1357 this results in a donor star velocity of $K_2 = 967\pm 49 \, \rm{km\,s^{-1}}$. Note that this value is consistent with the independent empirical relation derived for quiescent BHT between the outer disc velocity ($v_d$) and $K_2$ [@Orosz1994; @Orosz1995], which results in $K_2>806\, \rm{km\,s^{-1}}$ (see also @Torres2015). This, combined with CS13 orbital period (fully consistent with our results) yields $f(M_1)=11.0 \pm 2.1 \, \rm{M_\odot}$ (i.e. $M_{BH}>8.3\, \rm{M_\odot}$ at 90% confidence). If we include the BH orbital velocity estimated in CS13 ($K_1=43\pm 2 \, \rm{km\,s^{-1}}$), the obtained mass ratio ($q=M_2/M_1 \sim 0.04$) provides more restrictive limits to both $M_{BH}$ and $M_2$. We find $M_{BH}\geq 9.3 \, \rm{M_\odot}$ and $M_2\geq 0.4\, \rm{M_\odot}$. Here, we use the limit value for the orbital inclination $i=90^\circ$. This conservative result places the system as the most massive LMXB BH together with GRS 1915+105 ($\rm 10.1\pm 0.6\, M_\odot$, @Steeghs2013). Only Cyg X-1, a high-mass X-ray binary, exceeds the BH mass presented here ($\sim 15 \, \rm{M_\odot}$, @Orosz2011). The donor star mass is consistent with the constraint on the spectral type (later than M2V; see Sec. \[veiling\]), which requires $M_2 \lesssim 0.4 \, \rm{M_\odot}$.
Distance and height over the Galactic plane
-------------------------------------------
The constraint on the veiling factor, combined with the spectral type (mass) of the donor star, imposes a more restrictive lower limit on the distance to the system. Photometric values were obtained from acquisition images (one per night) in the $r'$-band. We use the relation between $r'$, $B$ and $V$ magnitudes described in @Fukugita1996, combined with $M_V$ and $B-V$ tabulated values of main-sequence stars [@Cox2000] to obtain absolute magnitudes, $M_{r'}$. If we consider the latest spectral type for the donor star proposed by CS13 ($\rm {M5V}$, $M_{r'}=11.61$), we obtain a conservative lower limit to the distance of the system. This assumes a somewhat evolved donor from the @SmithDhillon1998 empirical relation, instead of a main sequence star. We measure $r'=21.6\pm 0.2$ from the two nights when the system is faintest, which combined with $X>0.81$ results in a conservative constraint to the distance of $d > 2.29\, \rm{kpc}$. If we consider the companion star spectral type obtained in the previous section ($\rm {M2V}$, $ M_{r'}=9.28$), the system is required to be further than $d > 6.7\, \rm{kpc}$, a value which is close to the upper limit reported in @Shahbaz2013.
The above constraint to the distance sets new lower limits to the X-ray luminosity during both outburst and quiescence. @Armas-Padilla2014b [@Armas-Padilla2013] obtained $L_{\rm{quies}}=8\times 10^{29}\, \rm{erg\, s^{-1}}$ and $L_{\rm{peak}}=10^{35}\, \rm{erg\, s^{-1}}$ considering a distance of $d=1.5\,\rm{kpc}$. Our constraint implies: $L_{\rm{quies}}>1.9\times 10^{30}\, \rm{erg\, s^{-1}}$ and $L_{\rm{peak}}>2.33\times 10^{35}\, \rm{erg\, s^{-1}}$. These values are still consistent with J1357 being a very faint X-ray transient and perhaps the faintest stellar mass BH known in quiescence (see @Armas-Padilla2014b for a discussion on the topic).
The lower limit to the distance, combined with the high Galactic latitude ($b=50.004^{\rm{\circ}}$), places the system at $z>1.75\, \rm{kpc}$ above the Galactic plane. Other BHT that might be members of the Galactic thick disc population ($z \gtrsim 1\, \rm{kpc}$, @Gilmore1983) are: SWIFT J1357.2-0933 (this work and CS13), MAXI J1659-152 (BH candidate, @Kuulkers2013), XTE J1118+480 [@Uemura2000; @Gelino2006], XTE J1859+226 [@Corral-Santana2011], H1705-250 (@Remillard1996, @Jonker2004), GS 1354-64 [@Casares2004; @Casares2009] and SWIFT J1753.5-0127 (BH candidate, @Zurita2008); see @Corral-Santana2015.
On the orbital inclination
--------------------------
The $\rm 5876$ [Å]{} emission line exhibits a particular profile with a sharp, deep absorption core, even reaching $0.95$ the continuum value in some spectra. This has to be treated with caution since the $\rm 5890$ [Å]{} sky line is placed close to the core of the double-peaked profile, and therefore a deficient sky subtraction could affect the profile. In order to minimize this effect we combined the data of the two nights were the object was brighter, because this allows for a better correction. The final extracted spectrum (see Fig. \[figeclipse\]) still exhibits a sharp core profile, reaching $\sim 1.05$ times the continuum level. The residual sky substraction of the nearest and equally intense sky line \[\] $\rm 5577$ [Å]{} is reduced to almost noise level. This result suggests that the deep absorption core is a real feature.
Deep line absorption cores have only been previously observed in eclipsing, high inclination ($i\gtrsim 75^\circ$, @Schoembs1983) cataclysmic variables in quiescence e.g. Z Cha, [@Marsh1987], which has $i=81^\circ$. Obscuring material above the plane of the accretion disc has been proposed as the origin of the deep, central absorption observed in He I and Balmer hydrogen lines (see @Rayne1981). Therefore, it should not be surprising to find similar behaviour in high inclination BHTs. However, only systems with inclinations up to $i \sim 70^\circ$ have been discovered and detected in quiescence so far (e.g. XTE J1118+480, @Gelino2006). None of them shows such deep absorption cores. @Torres2015 already noticed depth variations in the core of of H$\alpha$ profile. We also observe variations in the normalized flux of the H$\alpha$ core in the range 1.3 to 2.1 times the continuum value.
It has been suggested by @Armas-Padilla2014a and @Torres2015 that the X-ray properties of the source (e.g. absence of emission/absorption lines) and the low extinction, comparable to that expected from interstellar material in the line-of-sight, argue against the very high inclination ($i \gtrsim 80^\circ$) scenario proposed by CS13. Given the lack of eclipsing BHs, the closest comparison can be made with high inclination neutron star systems. These only show absorption lines during thermal, soft states (@Ponti2014; @Ponti2015). However, it should be noted that J1357 never abandoned the hard state (see @Munoz-Darias2014 for a direct comparison between neutron star and BH states). On the other hand, the two eclipsing accretion disc corona sources X1822-371 and 2S 0921-63, show extinction values consistent with an interstellar origin (i.e no intrinsic extinction) when fitted with standard hard state spectral models (@Iaria2001; @Kallman2003). This suggests that X-ray photons emitted in the direction of the outer disc rim do not reach the observer, and only those radiated above/below the rim do it, solely interacting with the interstellar material in the line-of-sight. Therefore, we still consider $i \geq 80^\circ$ as the most plausible value. The detection of line cores and the already large BH mass (for $i=90^\circ$) presented here further support this conclusion.
![Spectra centred on the ($\rm 5876$ [Å]{}) emission line. Upper spectrum: normalised, averaged J1357 spectrum of the two brightest nights after sky correction. Lower spectrum: Averaged spectrum of the sky multiplied by a constant factor in order to compare with the upper spectrum. []{data-label="figeclipse"}](eclipse.ps){width="85mm"}
Conclusions {#conclusion}
===========
We have used 10.4-m GTC optical spectra of Swift J1357.2-0933 to constrain the accretion related contribution to the optical emission in quiescence. This yields more restrictive lower limits to the distance ($d > 2.29\, \rm{kpc}$) and height over the Galactic plane ($z>1.75\, \rm{kpc}$), placing the system in the Galactic thick disc. We detect variability in the H$\alpha$ profile modulated with a period $P=0.11\pm0.04\, \rm{d}$, confirming the period detected in outburst. Using the recently presented $\rm FWHM-K_2$ correlation, new constraints to the fundamental parameters are derived. In particular, we obtain $M_{BH}>9.3 \, \rm{M_\odot}$. We also favour $q\sim 0.04$ and $M_2\sim 0.4\, \rm{M_\odot}$. This indicates that Swift J1357.2-0933 harbours one of the most massive BHs known in our Galaxy.
Acknowledgements {#acknowledgements .unnumbered}
================
We are thankful to Gabriele Ponti for useful discussion on the X-ray properties of the source. DMS acknowledges Fundación La Caixa for the financial support received in the form of a PhD contract. JMC-S acknowledge financial support to CONICYT through the FONDECYT project No. 3140310. We also acknowledge support by the Spanish Ministerio de Economía y competitividad (MINECO) under grant AYA2013-42627. MOLLY software developed by T. R. Marsh is gratefully acknowledged.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: IRAF is distributed by National Optical Astronomy Observatories, operated by the Association of Universities for Research in Astronomy, Inc., under contract with the National Science Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Chemical abundances and abundance ratios measured in galaxies provide precious information about the mechanisms, modes and time scales of the assembly of cosmic structures. Yet, the nucleogenesis and chemical evolution of elements heavier than helium are dictated mostly by the physics of the stars and the shape of the stellar mass spectrum. In particular, estimates of CNO isotopic abundances in the hot, dusty media of high-redshift starburst galaxies offer a unique glimpse into the shape of the stellar initial mass function (IMF) in extreme environments that can not be accessed with direct observations (star counts). Underlying uncertainties in stellar evolution and nucleosynthesis theory, however, may hurt our chances of getting a firm grasp of the IMF in these galaxies. In this work, we adopt new yields for massive stars, covering different initial rotational velocities. First, we implement the new yield set in a well-tested chemical evolution model for the Milky Way. The calibrated model is then adapted to the specific case of a prototype submillimeter galaxy (SMG). We show that, if the formation of fast-rotating stars is favoured in the turbulent medium of violently star-forming galaxies irrespective of metallicity, the IMF needs to be skewed towards high-mass stars in order to explain the CNO isotope ratios observed in SMGs. If, instead, stellar rotation becomes negligible beyond a given metallicity threshold, as is the case for our own Galaxy, there is no need to invoke a top-heavy IMF in starbursts.'
author:
- |
Donatella Romano,$^{\! 1}$[^1] Francesca Matteucci,$^{\! 2, 3, 4}$ Zhi-Yu Zhang,$^{\! 5, 6}$ Rob J. Ivison$^{\, 5, 6}$ and Paolo Ventura$^{\, 7}$\
$^{1}$ INAF, Osservatorio di Astrofisica e Scienza dello Spazio, Via Gobetti 93/3, I-40129 Bologna, Italy\
$^{2}$ Dipartimento di Fisica, Sezione di Astronomia, Universit[à]{} di Trieste, Via Tiepolo 11, I-34131 Trieste, Italy\
$^{3}$ INAF, Osservatorio Astronomico di Trieste, Via Tiepolo 11, I-34131 Trieste, Italy\
$^{4}$ INFN, Sezione di Trieste, Via Valerio 2, I-34127 Trieste, Italy\
$^{5}$ Institute for Astronomy, University of Edinburgh, Royal Observatory, Blackford Hill, Edinburgh, EH9 3HJ, UK\
$^{6}$ European Southern Observatory, Karl-Schwarzschild-Strasse 2, D-85748, Garching bei München, Germany\
$^{7}$ INAF, Osservatorio Astronomico di Roma, Via Frascati 33, I-00040 Monte Porzio Catone, Roma, Italy
bibliography:
- '/Users/donatella/Papers/pap-rotation/R19\_rot\_arXiv.bib'
date: 'Accepted . Received ; in original form 2019 July 19'
title: 'The evolution of CNO isotopes: the impact of massive stellar rotators'
---
\[firstpage\]
nuclear reactions, nucleosynthesis, abundances – galaxies: abundances – galaxies: evolution – galaxies: ISM – stars: abundances – stars: rotation.
Introduction {#sec:intro}
============
After hydrogen and helium, oxygen, carbon and nitrogen are the most abundant elements in the universe. Their seven stable isotopes, $^{12}$C, $^{13}$C, $^{14}$N, $^{15}$N, $^{16}$O, $^{17}$O, and $^{18}$O, are synthesised in different stellar sites through different processes [@1957RvMP...29..547B]: (i) the main isotopes of carbon and oxygen come from He burning in stars of all masses ($^{12}$C) or in massive stars only ($^{16}$O); (ii) the cold CNO cycle, that takes place in the H-burning zones of main sequence and giant branch stars in the presence of carbon- (and oxygen-)seed nuclei, leads to the production of $^{14}$N and, possibly, $^{13}$C and $^{17}$O (at a lower pace); (iii) in the external layers of novae and supernovae (SNe), the activation of the hot CNO cycle explains the formation of $^{13}$C, $^{15}$N, and $^{17}$O in H-rich zones, and (iv) of $^{15}$N and $^{18}$O in He-rich ones.
While $^{12}$C and $^{16}$O are purely primary elements, i.e., they form directly from a hydrogen-helium mixture through a succession of nuclear burnings, all other isotopes may have a primary and/or secondary nature, depending on whether the metal seeds necessary for their formation are produced inside the star, or are already present on the zero-age main sequence. Whether it is the primary contribution or the secondary one that prevails, it depends primarily on the initial mass and metallicity of the star. For instance, fast-rotating intermediate- and high-mass stars produce large amounts of primary $^{14}$N at very low metallicities, via production channels that lose their effectiveness at higher metallicities . This primary N production was hypothesised long ago and explains observations of \[N/O\][^2] abundance ratios in metal-poor Galactic dwarfs and ionized HII regions in our own and other galaxies .
In chemical evolution studies, it is customary to use abundance ratios to gain insights into the mechanisms that shape galaxies, and their time scales. As matter is more and more strongly processed by succeeding generations of stars, one might naively expect an increase in the abundances of secondary elements relative to those of primary elements. However, elements that are produced via primary nucleosynthetic processes in low-mass stars will behave as secondary elements from the point of view of chemical evolution, owing to the long lifetimes of their progenitors [e.g. @1989MNRAS.239..885M]; similarly, primary elements originating in the outer layers of massive stars may exhibit a pseudo-secondary character, if their production is boosted at higher metallicities because of enhanced mass loss .
With their multifaceted stellar production sites and nucleosynthesis paths, CNO isotopes have repeatedly attracted the attention of chemical evolution modelers. Following pioneering works by , and @1978ApJ...223..557D, many authors have attempted to explain the origin and evolution of (one or more) CNO isotope ratios in the framework of increasingly sophisticated models. These, however, have mostly dealt with the Milky Way , with a few exceptions .
Nowadays, although important uncertainties remain, the constant improvement of stellar evolution and nucleosynthesis theory on the one hand, and the advent of revolutionary instrumentation of unprecedented sensitivity on the other, allow an effective usage of specific CNO isotope ratios as sensible constraints on general models of galaxy formation and evolution. In @2017MNRAS.470..401R, we have extended a successful Galactic chemical evolution (GCE) model, that reproduces satisfactorily well the CNO abundance data for the Milky Way, to other galaxies and shown that the predicted $^{16}$O/$^{18}$O abundance ratio may decrease by orders of magnitude if the assumed stellar initial mass function (IMF) in galaxy-wide starbursts varies from a canonical to a top-heavy one; the same IMF changes have only a minor effect on the $^{12}$C/$^{13}$C ratio[^3]. Building on this, in @2018Natur.558..260Z we read the remarkably uniform $^{13}$C/$^{18}$O $\simeq$ 1 ratio measured by us in four strongly lensed submillimeter galaxies (SMGs) observed with the Atacama Large Millimeter/submillimeter Array (ALMA) at redshift $z \sim$ 2–3 as a clear-cut evidence that the IMF is biased towards massive stars in the most extreme starburst events in the universe, and propose to use this ratio as a powerful diagnostic of the IMF shape in dust-enshrouded galaxies.
Several independent, more or less direct pieces of observational evidence [@2005MNRAS.364L..23N; @2009ApJ...705..112B; @2011MNRAS.415.1647G; @2013ApJ...764..155L; @2018Sci...359...69S], as well as theoretical arguments , support our finding that massive star formation is favoured in extreme environments, such as dense star-forming regions of high temperature and pressure permeated by intense radiation fields. However, new grids of stellar yields have been published recently [@2018ApJS..237...13L], which might challenge our interpretation of the low $^{13}$C/$^{18}$O ratios observed in SMGs. Indeed, even small differences in the yields, when weighted by an appropriate IMF, may result in non-negligible effects on galactic-scale model predictions . Therefore, a reappraisal of our models in the light of the newly published nucleosynthesis results is urgently needed.
This paper is organized as follows: in Section \[sec:gce\], we briefly outline the main assumptions and ingredients of our chemical evolution models, with special emphasis on the adopted nucleosynthesis prescriptions. In Section \[sec:res\], we present the model results in comparison to up-to-date data for the Milky Way and for the SMGs studied by @2018Natur.558..260Z. We discuss our findings in Section \[sec:dis\]; the conclusions follow (Section \[sec:end\]).
Chemical evolution models {#sec:gce}
=========================
The Milky Way model
-------------------
As in @2017MNRAS.470..401R, we adopt the two-infall chemical evolution model for the Milky Way originally developed by @1997ApJ...477..765C [@2001ApJ...554.1044C], where a detailed description of the basic equations and assumptions can be found. In the next paragraphs, we briefly summarize its main features, strengths, and limitations.
The model embraces a multi-zone scheme and divides the Galactic disc in several concentric annulii 2 kpc wide. It assumes that the inner halo and thick-disc components form out of a first episode of accretion of matter of primordial chemical composition. In the first $\sim$1 Gyr of evolution, the fast gas consumption resulting from the efficient star formation burst that generates the most ancient stellar populations makes the gas density quickly drop below a critical threshold, under which the star formation stops. The occurrence of such a sudden decrease in star formation was first seen in the \[Fe/O\] versus \[O/H\] data by @1996ASPC...92..307G [see also ]. The thin disc forms later, out of a second, almost independent infall episode on time scales that range from less than 3 Gyr for the inner disc to nearly a Hubble time for the outer regions [equation 2 in @2000ApJ...539..235R]. This ‘inside-out’ formation of the disc ensures the establishment of radial abundance gradients [@1976MNRAS.176...31L; @1989MNRAS.239..885M]. The adopted star formation rate is proportional to both the gas and the total surface mass density [see @2001ApJ...554.1044C their equation 2]. The efficiency of star formation is maximum during the halo/thick-disc phase; during the thin-disc phase, it dwindles and becomes a function of the distance from the Galactic centre [see @2015ApJ...802..129S and references therein]. The adopted IMF is that of @2002ASPC..285...86K, with a slope $x = 1.7$ in the high-mass domain ($x = 1.35$ for the extrapolated @1955ApJ...121..161S law), normalized to unity in the 0.1–100 M$_\odot$ stellar mass range[^4]. Hereinafter, this IMF will be referred to as the ‘canonical IMF’.
The two-infall model has proven able to meet the minimal set of observational constraints which should be honored by any successful Galactic chemical evolution model, namely: the surface densities of gas, stars and total matter in the solar vicinity, as well as their radial profiles; the present-day rates of star formation, infall, core-collapse and type Ia SNe in the Galaxy; the fractional contribution of metal-poor stars to the total stellar mass in the solar neighbourhood; the solar abundances; the trends of the abundances of different elements with respect to Fe as a function of \[Fe/H\]; the degree of deuterium astration in the solar vicinity; the radial abundance gradients of all the major chemical species .
The adopted model, like all other pure chemical evolution models in the literature, makes use of simple, heuristic recipes to treat complex processes, such as gas accretion, cooling, star formation, and thermal feedback from stars. It does not include complex dynamical processes, such as radial stellar migration [but see @2015ApJ...802..129S], and does not deal self-consistently with chemical inhomogeneities. However, it is extremely efficient in terms of computational costs and allows a quick exploration of the free parameter space. Therefore, since our main scope here is to test the effect of different nucleosynthesis prescriptions on the model predictions, and since we are mostly concerned with average behaviours, similarly to other authors we regard the pure chemical evolution model as the optimal tool to comply with our requests.
The SMG model
-------------
----------- --------------- ----------------- ------------------------------ ------------------------------ --------------- ------- --
LIMS Super-AGB stars Massive stars $\vel_{\rm{rot}}$ Hypernovae Novae
Milky Way Prototype SMG (km s$^{-1}$)
Karakas (2010) Doherty et al. (2014a,b) Nomoto et al. (2013) 0
Karakas (2010) – Nomoto et al. (2013) 0
Nomoto et al. (2013) 0
Nomoto et al. (2013) 0
Limongi & Chieffi (2018)$^b$ 300
Limongi & Chieffi (2018)$^b$ 150
Limongi & Chieffi (2018)$^b$ 0
Karakas (2010) Doherty et al. (2014a,b) Limongi & Chieffi (2018)$^b$ 300
Karakas (2010) Doherty et al. (2014a,b) Limongi & Chieffi (2018)$^b$ 150
Karakas (2010) Doherty et al. (2014a,b) Limongi & Chieffi (2018)$^b$ 0
Limongi & Chieffi (2018)$^c$ var$\,^d$
Karakas (2010) Doherty et al. (2014a,b) Limongi & Chieffi (2018)$^c$ var$\,^d$
----------- --------------- ----------------- ------------------------------ ------------------------------ --------------- ------- --
\[tab:nuc\]
*Notes.* $^a$We identify a particular model with the notation XXX-YY, where XXX individuates the galaxy (MWG: Milky Way galaxy, SMG: submillimeter galaxy) and YY is a number referring to the selected yield set combination. All different nucleosynthesis prescriptions have been tested against the Milky Way data, whilst only a subset of prescriptions has been used in SMG models. $^b$We adopt their recommended yield set (set R). $^c$Their set R is used for \[Fe/H\] $< -$1, while for \[Fe/H\] $\ge -$1 the mass range for full collapse to black holes is reduced (60–100 M$_\odot$). $^d \vel_{\rm{rot}} = 300$ km s$^{-1}$ for \[Fe/H\] $< -$1, $\vel_{\rm{rot}} = 0$ for \[Fe/H\] $\ge -$1.
We adopt the one-zone model for the typical SMG discussed in @2017MNRAS.470..401R and @2018Natur.558..260Z. The model relies on the evolutionary sequence advocated by @2014ApJ...782...68T, according to which SMGs are the precursors of passively-evolving massive elliptical galaxies. Fresh gas is accreted at early times on a short time scale at an exponentially decreasing rate, $$\frac{{\rm d}\mathscr{M}_{\rm inf}}{{\rm d}t} \propto {\rm e}^{-t/\tau},$$ where $\mathscr{M}_{\rm{inf}} = 4 \times 10^{11}$ M$_\odot$ is the total baryonic mass accreted by the system and $\tau =$ 50 Myr is the infall time scale. The gas forms stars following a @[email protected] law, $$\psi(t) = \nu\,\mathscr{M}_{\rm{gas}}^k(t),$$ where $\nu$ is the star formation efficiency, $k = 1$ and $\mathscr{M}_{\rm{gas}}(t)$ is the mass of neutral gas available for star formation at each time [@1959ApJ...129..243S; @1998ApJ...498..541K]. The star formation is halted when a stellar mass $\mathscr{M}_{\rm{stars}} \simeq 2
\times 10^{11}$ M$_\odot$ is attained; a strong galactic wind triggered by SN explosions and AGN activity is supposed to clean the galaxy of its residual gas at this point. We consider a rather continuous or a bursty star formation regime, as well as a canonical or a top-heavy IMF ($x = 1.1$ in the high-mass domain), and vary the duration of the star formation episode(s). We compare the model outputs in the different cases.
Nucleosynthesis prescriptions
-----------------------------
In our computations the instantaneous recycling approximation is relaxed, i.e. we take into account in detail the finite stellar lifetimes. This is necessary in order to treat properly elements that are produced on different time scales by stars of different initial masses and various chemical compositions. The adopted nucleosynthesis prescriptions are summarized in Table \[tab:nuc\] and briefly discussed in the remainder of this section; a more thorough description can be found in the original papers.
### Single stars
The stellar yields for low- and intermediate-mass stars (LIMS; 1 $\lesssim
m/{\rm M}_{\odot} \lesssim$ 6) are either from @2010MNRAS.403.1413K or from @2013MNRAS.431.3642V. The latter authors also provide detailed yields for super-asymptotic giant branch (AGB) stars (6 $\lesssim m/{\rm M}_{\odot} \lesssim$ 8–9) and (unpublished) yields for super-solar metallicity stars; both ingredients are often missing from GCE calculations. The nucleosynthetic yields for super-AGB stars computed by @2014MNRAS.437..195D [@2014MNRAS.441..582D] are used to complement the LIMS grid of @2010MNRAS.403.1413K, apart from one model where this super-AGB contribution is set to zero (model MWG-02). Although the stellar mass range covered by super-AGB stars is quite modest, they are found to contribute non-negligible amounts of $^{14}$N on a galactic scale (see Section \[sec:res\]).
For massive stars, we adopt the grid of yields suggested by , largely based on work published in @2006ApJ...653.1145K [@2011MNRAS.414.3231K] and extending to super-solar metallicities (models from MWG-01 to MWG-04, and model SMG-01). These yields account for some important physical processes that impact deeply the nucleosynthetic outcome of a star, such as mass loss, the ‘mixing and fallback’ process [@2002ApJ...565..385U and references therein], and the occurrence of highly energetic explosions [the so-called hypernovae, according to the terminology first introduced by @1998ApJ...494L..45P]. In particular, all stars above 20 M$_\odot$ are assumed to explode as hypernovae in model MWG-04, while they end up as ordinary core-collapse SNe in the other models. The important effects of stellar rotation, however, are not accounted for in the ’s grid. Therefore, we also implement in our GCE code @2018ApJS..237...13L’s [[email protected]] recommended yield set (set R), where: (i) mass loss and (ii) rotation are taken into account; (iii) stars with initial mass in the range 13–25 M$_\odot$ eject 0.07 M$_\odot$ of $^{56}$Ni after going through mixing and fallback in the inner SN regions; (iv) more massive objects fully collapse to black holes, which results in a null injection of Fe from $m >$ 30 M$_\odot$ stars[^5].
All the adopted yields are dependent on the initial mass and metallicity of the stars. Since they are necessarily computed for a limited number of points in the ($m$, $Z$) space, some interpolation in mass and metallicity has to be performed between published adjacent values. In particular, although we implement in the code up-to-date nucleosynthesis prescriptions for super-AGB stars, a most uncertain interpolation of the yields is still required in the mass range pertaining to low-mass core-collapse SNe ($\sim$9–12 M$_\odot$). On top of that, when adopting the yield sets for metallicities $Z \neq 0$, extrapolation is needed from 40 to 100 M$_\odot$. Possible spurious effects can be introduced as artifacts of the interpolation/extrapolation procedures, and one should keep this in mind when comparing GCE model predictions to observations. Denser grids of stellar yields are the only way to overcome this problem.
### Binary systems {#sec:bs}
The nucleosynthetic outcome of binary systems ending their lives as type Ia SNe is included in our models: we adopt the single-degenerate scenario [@2001ApJ...558..351M and references therein] for the progenitors and nucleosynthetic yields from @1999ApJS..125..439I.
While type Ia SNe contribute negligible amounts of CNO nuclei to the interstellar medium, classical nova explosions could significantly impact the evolution of the abundances of $^{13}$C, $^{15}$N and $^{17}$O on a Galactic scale [see @2007JPhG...34..431J for a review].
The nova contribution to the synthesis of $^{13}$C, $^{15}$N and $^{17}$O is included in two of our Milky Way models (models MWG-11 and MWG-12) by assuming average yields per nova system, similarly to what is done in @2017MNRAS.470..401R [see next paragraphs]. Models SMG-11 and SMG-12, that formally share the same nucleosynthesis prescritions with, respectively, model MWG-11 and model MWG-12, do not include a nova contribution in practice: in fact, before a newborn nova system can give rise to strong enough outbursts, it is necessary to wait at least 1 Gyr to ensure that the white dwarf has sufficiently cooled down (see below) and no SMG is expected to be forming stars on time scales longer than this. Therefore, there is basically no room for any CNO isotope pollution from novae in SMGs.
In models MWG-11, MWG-12, SMG-11 and SMG-12 nova nucleosynthesis is implemented following the prescriptions of , as briefly recalled hereunder; the interested reader is referred to the original papers for more details:
The rate of formation of nova systems at a given time $t$ is computed as a fraction $\alpha$ of the white dwarf birth rate at a previous time $t - \Delta t$, where $\Delta t =$ 1 Gyr is a suitable average time delay that guarantees that the white dwarfs cool down to temperatures low enough to ensure strong outbursts. The free parameter $\alpha$ is assumed to be constant in time [but see @2019arXiv190609130G] and set to reproduce the current nova outburst rate in the Galaxy.
Each nova system is assumed to experience 10$^4$ powerful eruptions on an average [@1978MNRAS.183..515B].
The average masses ejected in the form of $^{13}$C, $^{15}$N and $^{17}$O in each outburst are fixed by the request that the observed declining trends of $^{12}$C/$^{13}$C, $^{14}$N/$^{15}$N and $^{18}$O/$^{17}$O in the last 4.5 Gyr in the solar vicinity are reproduced. For model MWG-11 (MWG-12), these masses are: $M^{\mathrm{ejec}}_{\mathrm{^{13}C}} = 4$ $(2) \times 10^{-7}$ M$_\odot$, $M^{\mathrm{ejec}}_{\mathrm{^{15}N}} = 5$ $(6.5) \times 10^{-8}$ M$_\odot$ and $M^{\mathrm{ejec}}_{\mathrm{^{17}O}} = 3$ $(1.5) \times 10^{-8}$ M$_\odot$. These quantities are in reasonable agreement with those emerging from hydrodynamic nova models, though a high variability is found in the theoretical nova yields in dependence of the parameters adopted to produce the outburst [see figure 6 in @2017MNRAS.470..401R and the relevant discussion therein].
Results {#sec:res}
=======
In the following, we first present our GCE model results in comparison with up-to-date CNO measurements in the Milky Way and select the stellar yields that guarantee the best fit to the Galactic data (Section \[sec:mw\]). In the second place, we analyse a reduced set of models in comparison to CNO data for a sample of SMGs caught at the peak of their star formation activity and make some considerations about the shape of the IMF in these extreme starbursts (Section \[sec:smg\]).
{width="15.92cm"}
{width="7.8cm"} {width="7.8cm"}
{width="7.8cm"} {width="7.8cm"}
![ \[O/Fe\]–\[Fe/H\] diagram for solar neighbourhood stars. Data are from , , , @2014AJ....147..136R [*empty triangles*], and . The model predictions *(solid lines)* are color-coded according to the adopted nucleosynthesis prescriptions (see Table \[tab:nuc\] and explanatory legend in Fig. \[fig:CNO\]).[]{data-label="fig:O"}](R19MW_f2.eps){width="7.8cm"}
Milky Way galaxy {#sec:mw}
----------------
### Abundance ratios of solar neighbourhood stars {#sec:aburat}
Low-mass stars have lifetimes comparable to the age of the universe. Therefore, they are fossils bearing the imprints of the chemical enrichment processes that shaped their host galaxies from the very beginning.
It is well documented, though, that small mass field stars climbing the upper red giant branch undergo some important abundance changes. The original composition of their outer layers is modified markedly: the abundance of $^{14}$N increases, at the expenses of that of $^{12}$C, and the $^{12}$C/$^{13}$C ratio decreases to values close to, though higher than, the equilibrium value . Unluckily, limitations imposed by current observational capabilities prevent any firm conclusion about the oxygen isotope ratios . Unusually high C abundances in some cases may be attributable to mass transfer from a post-AGB companion across a binary system via Roche lobe overflow [@2005ApJ...625..825L and references therein]. Last but not least, some low-metallicity stars currently found in the solar neighbourhood have likely been accreted from dwarf satellites and, thus, reflect distinct evolutionary paths. If observed stellar samples are not cleaned off for these effects, the comparison between GCE model predictions and observations via widely-used diagnostic diagrams, such as \[C/Fe\] versus \[Fe/H\] or \[N/O\] versus \[O/H\], can easily lead to misinterpretation. It is, thus, advisable to rely on statistically significant samples of single, Sun-like stars, formed *in situ* and, possibly, analysed homogeneously.
#### \[X/Fe\] versus \[Fe/H\] plots.
The grey dots in Fig. \[fig:CNO\], right-hand panels, show the \[C/Fe\] and \[N/Fe\] abundance ratios (upper and lower panel, respectively) of solar neighbourhood stars in the SAGA database [@2008PASJ...60.1159S] and Hypatia catalog [@2014AJ....148...54H]. For illustration purposes, we plot stars in different evolutive stages, as well as binary stars: as expected, a large dispersion is seen in the data at all metallicities. No attempts are made to homogenize the data, that come from different sources, except for a rescaling to the same solar abundances (taken from in this work).
Symbols in different shades of grey represent measurements for unevolved stars from , , , @2014AJ....147..136R [triangles], and .
have determined precise C (and O) abundances for 151 F and G main-sequence stars with metallicities in the range $-$1.6 $<$ \[Fe/H\] $<$ 0.5, classified into four different populations: thin-disc, thick-disc, high-$\alpha$ halo and low-$\alpha$ halo stars (shown, respectively, as small dark-grey, small light-grey, medium-sized light-grey and empty circles in Fig. \[fig:CNO\], upper panels). These authors have used plane parallel (1D) model atmospheres for their analysis and applied non-LTE corrections to the abundances derived from the CI $\lambda \lambda$5052,5380 lines and $\lambda$7774 OI triplet. As seen from Fig. \[fig:CNO\], upper right-hand panel, these precise abundances define a tight trend in the \[C/Fe\]–\[Fe/H\] plane. Under the reasonable assumption that the low-$\alpha$ halo stars formed in dwarf galaxies that were later accreted by the Milky Way and, thus, do not trace the local enrichment history, the observed trend for \[Fe/H\] $> -$1.5 is fitted at best by models MWG-07 and MWG-10, that assume the yields by [@2018ApJS..237...13L] for non-rotating massive stars (see Table \[tab:nuc\]). Marginal agreement is obtained by the models that adopt the yields by [@2018ApJS..237...13L] for rotating massive stars (models MWG-05, MWG-06, MWG-08, MWG-09) and by models MWG-01 and MWG-02, that assume the yields by for (non-rotating) massive stars without hypernovae, as well as the yields by @2010MNRAS.403.1413K for LIMS (C production from super-AGB stars as issued in @2014MNRAS.437..195D [@2014MNRAS.441..582D] turns out to be negligible).
The behaviour of C (and O) abundances at lower metallicities has been reassessed recently by . Basing on 3D, non-LTE results for 39 turn-off stars with \[Fe/H\] $\le -$1.5, they find that \[C/Fe\] is almost solar and stays nearly flat with metallicity. Interestingly, the abundances of three stars previously identified as carbon-enhanced stars with \[C/Fe\] in excess of 1 are significantly revised downwards, to \[C/Fe\] $\simeq$ 0.1–0.3 . When compared with our model predictions, these new measurements seem to point to the need for some hypernova pollution at low metallicities (cfr. the predictions from model MWG-04, in which all stars above 20 M$_\odot$ end up with a hypernova explosion, with those from model MWG-03, which is the same as model MWG-04, but with all massive stars exploding as ordinary SNe; Fig. \[fig:CNO\], upper right-hand panel).
We now examine the behaviour of \[N/Fe\] versus \[Fe/H\]. An inspection of Fig. \[fig:CNO\], lower right-hand panel, immediately shows that a significant dispersion characterizes the data at low-metallicities, even when only unevolved stars are considered . Below \[Fe/H\] $\sim -$1.5, some observations can be explained only by invoking a significant primary N production from massive stars . However, a few low \[N/Fe\] values point to a milder N enrichment. The N yields of massive stars depend critically on the stellar rotational velocity, therefore, a large dispersion in the N abundances would naturally arise during the Galactic halo assembly, because of the highly inhomogeneous evolution: regions polluted by non-rotating massive stars will display a N content significantly lower than regions where matter was processed through one or more fast rotators. Assessing the evolution of the CNO elements in a more realistic, inhomogeneous medium would be extremely valuable, and can be done by means of detailed hydrodynamical simulations. A quantitative explanation of the observed scatter, however, is beyond the scope of the present work and will be addressed in a future paper. It suffices to note here that the dispersion is sensibly reduced at disc metallicities, where reliable N abundances can be derived for statistically significant samples of dwarf stars from the $\lambda$3360 NH molecular band . Overall, we find that models MWG-05 and MWG-06, with yields from [@2018ApJS..237...13L] for rotating massive stars and from @2013MNRAS.431.3642V for LIMS and super-AGB stars, provide an excellent fit to the average \[N/Fe\] ratios over the full range of metallicities.
The run of \[O/Fe\] as a function of \[Fe/H\] is shown in Fig. \[fig:O\]. All models provide an adequate fit to the observed trend, apart from models MWG-07 and MWG-10, that are computed with the yields from non-rotating massive stars by @2018ApJS..237...13L [their set R], which severely underestimate the \[O/Fe\] ratios of disc stars. If the 3D oxygen abundances by have to be given a higher weight, it is possible that in the early Galaxy a large fraction of high-mass stars exploded as hypernovae, with an energy release substantially higher than that of standard SNe (cfr. the predictions from model MWG-04, in which all stars above 20 M$_\odot$ end up as hypernovae, with those from model MWG-03, in which the same stars explode as normal core-collapse SNe).
#### \[X/O\] versus \[O/H\] plots.
The \[C/O\]–\[O/H\] and \[N/O\]–\[O/H\] diagrams shown in Fig. \[fig:CNO\], left-hand panels, are particularly useful. At low metallicities, they trace the enrichment from massive stars with minimum linkage to one of the most uncertain parameters of stellar evolution, namely, the location of the mass cut (in most extant one- and two-dimensional models of the latest stages of massive star evolution, this is the boundary between the collapsed core and the ejected mantle); by contrast, the choice of the mass cut influences considerably the theoretical trends in the \[X/Fe\] versus \[Fe/H\] planes. At disc metallicities, the rise in \[C/O\] and \[N/O\] can be related to the late contribution to C and N production from low-metallicity, low-mass stars, and directly compared to the release of C and N from relatively more metal-rich massive stars, which occurs on much shorter time scales.
Moreover, from a theoretical point of view, the trends with metallicity as gauged by oxygen abundance are free from the uncertainties that arise from the inclusion of type Ia SNe in the models (these systems produce a large fraction of the solar iron, but negligible amounts of oxygen). However, from a spectroscopist’s standpoint, the abundance estimates are less robust for O than for Fe. Oxygen abundances are affected by non-negligible non-LTE corrections when derived from the OI triplet, and by 3D effects (as well as a NiI blend) when derived from the forbidden \[OI\] line at 6300 Å; the latter line is also too weak to be detectable in the majority of halo stars . Furthermore, the O abundances of metal-poor stars derived from near-UV OH lines are found to be sistematically higher than the ones obtained from other indicators (see Fig. \[fig:O\]).
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
With all the above-mentioned possible shortcomings in mind, we compare the trends predicted by our GCE model under different nucleosynthesis prescriptions (colored lines in Fig. \[fig:CNO\]) to the observational data from , , , @2014AJ....147..136R, [^6] and discussed in the previous paragraphs. Overall, the most appropriate yields appear to be those of [@2018ApJS..237...13L] for rotating massive stars coupled to those of @2013MNRAS.431.3642V for LIMS and super-AGB stars, i.e. models MWG-05 and MWG-06. In particular, the relatively low N yields by @2013MNRAS.431.3642V guarantee a better fit to the observed \[N/O\] ratios for \[O/H\] $> -$1.2, that is when the contributions from super-AGB stars first, and LIMS right after, start to become relevant.
### CNO isotope ratios in space and time in the Galaxy
The determination of isotopic abundances in stars is limited to a handful of chemical species. For CNO elements, it can be made mostly in brilliant, giant stars. These, however, can have the original composition of their atmospheres altered by internal processes. It is thus convenient to resort to measurements of gas-phase isotopic abundances in interstellar clouds, though they provide only snapshots in time and, hence, no information about the past evolution that led to the observed configuration.
Radio observations are the prime tools for exploring the present-day gradients of $^{12}$C/$^{13}$C, $^{14}$N/$^{15}$N, $^{16}$O/$^{18}$O, $^{18}$O/$^{17}$O and $^{13}$C/$^{18}$O in the disc of our Galaxy. For the purpose of comparison with the predictions of our GCE model, we take the data from: (i) the compilation of , based on results from molecular transitions of carbon monoxide and formaldehyde (small filled triangles in Fig. \[fig:MW\], upper and middle right-hand panels); (ii) @2005ApJ...634.1126M, based on $N =$ 1$-$0 transitions of the CN radical (large empty triangles in Fig. \[fig:MW\], upper right-hand panel); (iii) , which analysed the ground state rotational lines of the OH isotopologues (large empty circles in Fig. \[fig:MW\], middle right-hand panel); (iv) , providing a) the $^{13}$CO/C$^{18}$O gradient, from simultaneously observed $^{13}$CO and C$^{18}$O $J =$ 1$-$0 and $J =$ 2$-$1 transitions (respectively, small filled and small empty circles in Fig. \[fig:MW\], middle and lower right-hand panels; the original data have been combined with the $^{12}$C/$^{13}$C gradient by @2005ApJ...634.1126M to trace the $^{16}$O/$^{18}$O gradient) and b) the $^{18}$O/$^{17}$O gradient, from $J =$ 1$-$0, 2$-$1 and 3$-$2 transitions of C$^{18}$O and C$^{17}$O (small light-grey circles in Fig. \[fig:MW\]); (v) @2016RAA....16...47L, which derived C$^{18}$O/C$^{17}$O abundance ratios for 13 sources spanning the Galactocentric distance range 3–16 kpc (small filled diamonds in Fig. \[fig:MW\]); and (vi) @2018MNRAS.478.3693C, which targeted the $J =$ 1$-$0 rotational transitions of hydrogen isocyanide and hydrogen cyanide and derived the $^{14}$N/$^{15}$N ratios for both molecules (Fig. \[fig:MW\], the small filled and small empty pentagons are for HNC and HCN, respectively). To these we add some solid-state $^{12}$C/$^{13}$C data by and preliminary $^{12}$C/$^{13}$C and $^{18}$O/$^{17}$O measurements and lower limits for outer disc targets from Zhang et al. (2019, in preparation, large grey pentagons in Fig. \[fig:MW\]).
While the shapes of the present-day gradients of CNO isotope ratios in our Galaxy are reasonably well-known (with the exception of the outer disc, where only a few measurements are available; but see Zhang et al. 2019, in preparation), the knowledge of the evolution of these ratios in the solar neighbourhood is tightly limited. For the comparison with our model predictions, we rely on solar ratios from @2013ApJ...765...46A [$^{12}$C/$^{13}$C $= 91.4 \pm
1.3$, $^{16}$O/$^{18}$O $= 511 \pm 10$, $^{18}$O/$^{17}$O $= 5.36 \pm
0.34$] and @2011Sci...332.1533M [$^{14}$N/$^{15}$N $= 441 \pm
6$]. These are shown as Sun symbols in Fig. \[fig:MW\], left-hand panels. It should be noted, however, that the Sun might have migrated to its actual position from an inner birthplace and, thus, its chemical composition might not reflect that of the local medium 4.5 Gyr ago. For the local interstellar medium, we adopt average values (large open squares in Fig. \[fig:MW\], left-hand panels) from the studies discussed above, namely: $\langle ^{12}$C/$^{13}$C$\rangle_{\rm{LISM}} = 68 \pm 15$ [@2005ApJ...634.1126M], $\langle ^{16}$O/$^{18}$O$\rangle_{\rm{LISM}} = 395 \pm 56$ , $\langle^{18}$O/$^{17}$O$\rangle_{\rm{LISM}} = 4.16 \pm 0.09$ . For $\langle^{14}$N/$^{15}$N$\rangle_{\rm{LISM}}$, both the low value from @2012ApJ...744..194A [$290 \pm 40$] and the high value from @2018MNRAS.478.3693C [$\sim 400$] are displayed.
The predictions from models MWG-01, MWG-02 and MWG-03 are in good agreement with the available $^{12}$C/$^{13}$C data both for the solar vicinity and across the disc. Should a substantial contribution to $^{13}$C production come from nova systems (see next section), models MWG-05, MWG-06, MWG-07, MWG-08, MWG-09 and MWG-10 could be made consistent with the observations, but the fit would worsen for models from MWG-01 to MWG-03.
Actually, in order to explain the observed decrease of the $^{14}$N/$^{15}$N ratio in the local disc over the last 4.5 Gyr, as well as its positive gradient across the disc, it is necessary that a large amount of $^{15}$N is supplied by some up to now neglected stellar factory. Thermonuclear runaways occurring in nova outbursts stand out as possible candidates [but see @1999ApJ...512L.143C; @2015ApJ...808L..43P].
As regards the O isotope ratios, models in which the yields from fast-rotating massive stars are implemented during the full Galactic evolution – namely, models MWG-05, MWG-06, MWG-08 and MWG-09 – produce gradients at variance with the observations. Indeed, some theoretical arguments suggest that low-metallicity stars rotate faster than high-metallicity stars, on an average. For this reason, @2018MNRAS.476.3432P introduced an initial distribution of stellar rotational velocities (IDROV) in their GCE model, in analogy with the adoption of a stellar IMF. Because of the extremely loose constraints that can be imposed on the actual IDROV, however, their choice is far from unique. In the following section, we discuss the results of our GCE model when a simple step function is adopted for the IDROV: we assume the yields corresponding to stellar models computed with $\vel_{\rm{rot}} = 300$ km s$^{-1}$ for \[Fe/H\] $< -$1 and the yields of non-rotating models at higher metallicities. In the latter case, we also set to 60–100 M$_\odot$, rather than 35–100 M$_\odot$, the mass range for full collapse to black holes, which guarantees a better fit to the observed O abundances of local disc stars in the framework of our model.
### A possible, comprehensive evolutive scenario {#sec:scen}
{width="15.92cm"}
{width="7.8cm"} {width="7.8cm"}
{width="7.8cm"} {width="7.8cm"}
![ Same as Fig. \[fig:O\], but shown are predictions from models MWG-11 and MWG-12 (the latter, concealed behind model MWG-11 predictions).[]{data-label="fig:Onov"}](R19MW_f5.eps){width="7.8cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
From Figs. \[fig:CNO\] and \[fig:O\] discussed above, one sees that the adoption of massive star yields computed with either $\vel_{\rm{rot}} = 300$ km s$^{-1}$ or $\vel_{\rm{rot}} = 150$ km s$^{-1}$ results in comparable \[C/Fe\], \[C/O\] and \[O/Fe\] ratios, that differ by 0.15 dex at most, during the entire solar neighbourhood evolution. Larger differences (up to 0.5 dex at the lowest metallicities) are found between the theoretical \[N/Fe\] and \[N/O\] ratios predicted for halo stars in the two cases, while the differences cancel out at disc metallicities.
The largest contrasts arise anyway between the evolutive tracks obtained in the rotating case (whatever the value of $\vel_{\rm{rot}}$) and the non-rotating one. While nothing can be said from direct observations about the distribution of the rotational velocities of low-metallicity, high-mass stars, there is strong, indirect evidence from the chemical composition of unevolved halo stars that below \[Fe/H\] $= -$2 most SNII progenitors must have been fast rotators: indeed, fast rotation leads to the synthesis of large amounts of primary $^{14}$N at low metallicities, which explains the relevant observations .
Let us now throw the minor CNO isotopes into the mix (Fig. \[fig:MW\]): the situation gets more complicated! In fact, novae – not included in models from MWG-01 to MWG-10 – may inject significant amounts of $^{13}$C, $^{15}$N and $^{17}$O in the Galactic medium on long time scales [see @2003MNRAS.342..185R and references therein]. If implemented in models including fast rotators, this contribution may, in principle, reverse the trend at odds with the observations predicted for $^{14}$N/$^{15}$N and improve the agreement between predicted and observed $^{12}$C/$^{13}$C ratios. Yet, the $^{16}$O/$^{18}$O ratio is immune to changes caused by an implementation of this nucleosynthetic channel (neither $^{16}$O nor $^{18}$O are produced in appreciable amounts in nova outbursts).
From Fig. \[fig:MW\], middle panels, it can be seen that GCE models without fast rotators produce $^{16}$O/$^{18}$O ratios in remarkable agreement with local and inner disc data. From the more uncertain abundance determinations in the outer disc, however, it can not be excluded that a fraction of massive stars are rotating fast at large Galactocentric distances. Our GCE model predicts the current metallicity in the inner (outer) Milky Way disc to be $Z \simeq$ 0.016–0.028 (0.005–0.009) at $R_{\mathrm{G}} =$ 4 (16) kpc (the exact value depending on the choice of the stellar yields). Thus, the overall picture is one in which the formation of fast-rotating massive stars is inhibited in environments with solar or super-solar metallicity, while it can not be excluded that a fraction of massive stars are rotating fast in regions with metallicities close to those typical of the Magellanic Clouds. Likewise from N observations, and as expected from theory , from the $^{16}$O/$^{18}$O ratios we get an indication that fast-rotating massive stars are more common in less evolved environments. It goes without saying that more precise measurements of the $^{16}$O/$^{18}$O ratios in more populous samples of low-metallicity molecular clouds are crucial to better understand the transition from the fast-rotation to the null(or slow)-rotation regime. It is worth recalling at this point that the $^{16}$O/$^{18}$O gradient in the Milky Way is determined from $^{13}$CO and C$^{18}$O transitions via the adoption of a $^{12}$C/$^{13}$C gradient that is poorly defined in the outer disc (see Fig. \[fig:CNO\], upper right-hand panel). Thus, better $^{16}$O/$^{18}$O determinations call for better estimates of the $^{12}$C/$^{13}$C ratio. We will address this issue in detail in a forthcoming paper (Zhang et al. 2019, in preparation). Grids of yields computed for rotational velocities in between those considered here are also in demand.
{width="\textwidth"}
{width="\textwidth"}
In the remainder of this section we use what is at hand and discuss the results of two GCE models that: (i) assume a simple step function for the IDROV of core-collapse SN progenitors and (ii) take into account the production of rare isotopes during nova outbursts.
As regards the massive stars, for \[Fe/H\] $= -$3 and \[Fe/H\] $= -$2 we adopt the recommended yields by @2018ApJS..237...13L [their set R] computed with $\vel_{\rm{rot}} =$ 300 km s$^{-1}$. For \[Fe/H\] $= -$1 and \[Fe/H\] = 0, we use the yields for non-rotating stars and reduce the mass range for full collapse to black holes from 30–100 M$_\odot$ to 60–100 M$_\odot$, which improves the predictions about the behaviour of \[O/Fe\] versus \[Fe/H\] – and, hence, the run of \[C/O\] and \[N/O\] as a function of \[O/H\] – at disc metallicities in our models. The yields for LIMS and super-AGB stars are taken from @2013MNRAS.431.3642V [model MWG-11] or from @2010MNRAS.403.1413K and @2014MNRAS.437..195D [@2014MNRAS.441..582D model MWG-12]. Linear interpolation in mass and metallicity is performed between adjacent values of the yields. In both model MWG-11 and model MWG-12, nova nucleosynthesis is implemented following the prescriptions of .
The results of models MWG-11 and MWG-12 are illustrated in Figs. \[fig:CNOnov\]–\[fig:MWnov\]. From Figs. \[fig:CNOnov\] and \[fig:Onov\], it can be seen that for the main CNO isotopes the agreement with the observations is from satisfactory to very good during most of the solar neighbourhood evolution. The predicted overabundances of carbon and (to a lower extent) oxygen in the early Galaxy have to be expected and could be rectified by the inclusion of a fraction of hypernovae at low metallicities (see the discussion in Section \[sec:aburat\]). Such an energy effect would also have the advantage of redressing the well-known discrepancies between predicted and observed cobalt and zinc abundances in low-metallicity Galactic halo stars .
Stars with initial masses $m \le$ 9 M$_\odot$ do not produce $^{16}$O (rather, they partly destroy it in their interiors), but synthesise about one third of the solar $^{12}$C (30 per cent for model MWG-11, and more than 35 per cent for model MWG-12), as well as most of the solar $^{14}$N (65 per cent in the case of model MWG-11, and more than 70 per cent for model MWG-12). As already stressed by @2005ApJ...623..213C, the fractional contribution from LIMS (and super-AGB stars) and massive stars is strongly dependent on time, on the galactic region under scrutiny and, of course, on the adopted GCE model and stellar yields. Therefore, one should beware of any daring extrapolations of the rates reported above to other galaxies. In the next section, we further stress this point by discussing the results for a prototype SMG observed at high redshift.
The enrichment of $^{13}$C coarsely tracks that of $^{14}$N, with more than 50 and 70 per cent of the solar $^{13}$C being produced by LIMS and super-AGB stars in models MWG-11 and MWG-12, respectively. In particular, according to @2010MNRAS.403.1413K and @2014MNRAS.437..195D [@2014MNRAS.441..582D], low-metallicity intermediate- to high-mass AGB stars are effective $^{13}$C producers (see Fig. \[fig:CylowZ\]), which makes the predicted carbon isotope ratio to reach down to $^{12}$C/$^{13}$C $\simeq$ 130 at the end of the halo/thick-disc formation in model MWG-12 (Fig. \[fig:MWnov\], upper left-hand panel). Model MWG-11, that assumes the much lower $^{13}$C yields by @2013MNRAS.431.3642V for the same stars, barely reaches down to $^{12}$C/$^{13}$C $\sim$ 400, making it harder to conciliate its predictions to observations of post first dredge-up halo stars . At solar metallicities, the $^{13}$C yields from both sets of stellar models align with each other and the stars with $m >$ 4 M$_\odot$ efficiently destroy $^{12}$C in their interiors (see Fig. \[fig:CyZsun\]), which leads to converging GCE results (the slightly more pronounced decline characterising the late evolution of the carbon isotope ratio in model MWG-11 compared to model MWG-12 is due to the slightly larger $^{13}$C nova yield necessary to bring the predicted solar abundance of $^{13}$C in agreement with the observed one in model MWG-11).
The minor isotope of nitrogen, $^{15}$N, is destroyed in non-rotating stellar models, but it is efficiently synthesised in the presence of rotation, owing to the diffusion of matter between the H- and He-burning zones [@2018ApJS..237...13L; @2018MNRAS.476.3432P]. However, this production is not sufficient, alone, to explain the Galactic $^{14}$N/$^{15}$N data. In this work, we fit well measurements of nitrogen isotope ratios in Milky Way targets by resorting to nova pollution (see Fig. \[fig:MWnov\]), but we ought to mention that unusually high $^{15}$N abundances are found in some Galactic AGB stars [@2013ApJ...768L..11H] and planetary nebulae [@2018Natur.564..378S]. These are currently unexplained by stellar evolution theory and might call for some substantial revision of stellar physics.
The $^{17}$O yields are overall positive, apart from those from $m >$ 20–25 M$_\odot$ star models, that turn negative[^7] for \[Fe/H\] $\ge -$1. As a consequence, the predicted $^{18}$O/$^{17}$O ratio decreases with time in the local disc, except in the last 4.5 Gyr, when it would mildly increase without the addition of a nova contribution. Our GCE models, including such contribution, account very well for the local evolution of this isotope ratio, as well as its Galactic gradient (see Fig. \[fig:MWnov\], second left- and right-hand panels down, respectively).
As regards $^{18}$O, its yields are largely positive for stars in the mass interval 13–25 M$_\odot$, while out of this range the stars either destroy $^{18}$O, or provide smaller amounts of it. From Fig. \[fig:MWnov\], middle panels, it is seen that the evolution of the $^{16}$O/$^{18}$O isotope ratio in the Milky Way is reproduced fairly well by both models MWG-11 and MWG-12. However, in order to assess the validity of the models at relatively low metallicities, it is mandatory to obtain firmer measurements of this ratio in the external disc (see arguments set out in previous paragraphs).
Prototype SMG {#sec:smg}
-------------
[Model]{} SFH IMF $^{12}$C/$^{13}$C $^{14}$N/$^{15}$N $^{16}$O/$^{18}$O $^{18}$O/$^{17}$O $^{13}$C/$^{18}$O
----------- ------------------------------------------------- ----------- ------------------- ------------------- ------------------- ------------------- -------------------
Continuous, $\Delta t_{\mathrm{burst}} = 1$ Gyr Canonical 118 $4.5 \times 10^4$ $2.9 \times 10^3$ 2.5 6.2
One-shot, $\Delta t_{\mathrm{burst}} = 20$ Myr Canonical $1.7 \times 10^3$ $1.8 \times 10^6$ $3.9 \times 10^4$ 5.3 1.6
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Canonical 132 $3.8 \times 10^4$ $2.3 \times 10^3$ 2.4 5
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Top-heavy 64 453 57 3.2 0.09
Continuous, $\Delta t_{\mathrm{burst}} = 1$ Gyr Canonical 162 400 79 81 0.2
One-shot, $\Delta t_{\mathrm{burst}} = 20$ Myr Canonical 765 $3.9 \times 10^3$ 447 126 0.1
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Canonical 176 424 85 65 0.2
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Top-heavy 173 $1.7 \times 10^4$ 334 6 0.9
Continuous, $\Delta t_{\mathrm{burst}} = 1$ Gyr Canonical 175 214 126 45 0.3
One-shot, $\Delta t_{\mathrm{burst}} = 20$ Myr Canonical 970 695 451 99 0.1
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Canonical 193 220 131 37 0.3
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Top-heavy 209 $2.4 \times 10^3$ 480 4 1.1
Continuous, $\Delta t_{\mathrm{burst}} = 1$ Gyr Canonical 521 $2.7 \times 10^3$ 344 1.4 0.23
One-shot, $\Delta t_{\mathrm{burst}} = 20$ Myr Canonical $1.3 \times 10^3$ $1.3 \times 10^4$ 819 168 0.13
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Canonical 546 $3.0 \times 10^3$ 372 1 0.3
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Top-heavy 307 $6.3 \times 10^4$ $2.8 \times 10^3$ 0.01 3.6
Continuous, $\Delta t_{\mathrm{burst}} = 1$ Gyr Canonical 184 $3.9 \times 10^3$ 380 27 0.95
One-shot, $\Delta t_{\mathrm{burst}} = 20$ Myr Canonical $1.3 \times 10^3$ $1.3 \times 10^4$ 815 168 0.13
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Canonical 201 $3.7 \times 10^3$ 419 18 1
Gasping, $\Delta t_{\mathrm{burst}} = 50$ Myr Top-heavy 232 $1.3 \times 10^5$ $4 \times 10^3$ 0.5 7
*Note.* $^a$At star formation halt (see text).
\[tab:endrat\]
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
{width="7.8cm"} {width="6.435cm"}
In this section, we discuss the results of our one-zone models for the typical SMG, namely models SMG-01, SMG-08, SMG-09, SMG-11 and SMG-12 (the numbers refer to the adopted nucleosynthesis prescriptions, see Table \[tab:nuc\]). Each model is run four times, by assuming a set of three distinct star formation histories (SFHs) and two different IMFs. Specifically, we adopt: (i) a continuous star formation, mildly decreasing in time and lasting 1 Gyr, with a star formation rate of hundreds of solar masses per year, and a canonical IMF; (ii) an extremely vigorous (SFR$_{\rm{peak}} \sim$ 15000 M$_\odot$ yr$^{-1}$), short ($\Delta t_{\rm{burst}}$ = 20 Myr) starburst, forming stars according to a canonical IMF; (iii) a gasping SFH, which consists of several short star formation bursts ($\Delta t_{\rm{burst}}$ = 50 Myr each) interspersed with similarly long quiescent periods, with either a canonical or (iv) a top-heavy IMF (see Table \[tab:endrat\]). In all cases, when the star formation stops, $\mathscr{M}_{\rm{stars}} \simeq 2 \times 10^{11}$ M$_\odot$ are in place.
The resulting theoretical tracks are displayed in Figs. \[fig:SMGco\] and \[fig:SMGn\], where the solid and short-dashed lines in the left-hand panels refer to case (i) and (ii), respectively, and the solid and dotted lines in the right-hand panels refer to case (iii) and (iv), respectively. The model results are further color-coded according to the adopted nucleosynthesis prescriptions: green is for model SMG-01, yellow for model SMG-08, orange for model SMG-09, purple for model SMG-11 and lilac for model SMG-12. The values of the ratios at the time of the halt of the star formation are reported in Table \[tab:endrat\].
The results of model SMG-01 are essentially the same as the ones discussed in @2018Natur.558..260Z. In that paper, because of the adoption of stellar yields that did not include the effects of rotation, $^{13}$C/$^{18}$O ratios of about unity in SMGs could only be explained by assuming an IMF skewed towards massive stars. Indeed, from Fig. \[fig:SMGco\], lower panels, and Table \[tab:endrat\], last column, it is seen that model SMG-01, that does not include the effects of stellar rotation on the yields, always predicts $^{13}$C/$^{18}$O $>$ 1, unless the IMF is skewed towards massive stars. The slope of $x = 1.1$ assumed here for the top-heavy IMF in the high-mass domain is exaggerated – less extreme values would fit better the observations – but we prefer to show extreme variations, both for illustration purposes and to be consistent with the choice of @2018Natur.558..260Z. Alternatively, a ratio around unity could be reached if very high star formation rates, of the order of tens of thousands of solar masses per year, were assumed, which we deem unrealistic [see, however, @2019MNRAS.483.3060G for a different perspective].
In the hot, turbulent medium of intense, gas-rich starburst galaxies forming stars at high redshifts, it may well be that the formation of fast rotators is favoured with respect to the local conditions. For this reason, we compute models SMG-08 and SMG-09, in which all massive stars are supposed to rotate fast at birth, with $\vel_{\rm{rot}} =$ 300 and 150 km s$^{-1}$, respectively, irrespective of their initial metallicity. In this case, a top-heavy IMF is needed once again to bring the predicted $^{13}$C/$^{18}$O ratio into agreement with the observations (see Fig. \[fig:SMGco\], lower panels, yellow and orange lines, and Table \[tab:endrat\], last column).
As a last case study, we consider the possibility that massive stars are born fast rotators until a metallicity of \[Fe/H\] = $-1$ is reached, above which the probability of gaining a high initial rotational velocity quickly drops to zero. This is exactly the picture that provides the best fit to the CNO data for the Milky Way (see Section \[sec:mw\]). In this case, if a prolonged star formation (either bursty or continuous) is considered together with the yields of @2010MNRAS.403.1413K and @2014MNRAS.437..195D [@2014MNRAS.441..582D] for LIMS and super-AGB stars, a ratio $^{13}$C/$^{18}$O $\simeq$ 1 in SMGs can be obtained without any alteration of the IMF from its canonical form. If the yields of @2013MNRAS.431.3642V are assumed in place of those by @2010MNRAS.403.1413K and @2014MNRAS.437..195D [@2014MNRAS.441..582D], or if a one-shot, short star formation event is assumed to build up the bulk of the stellar population, a $^{13}$C/$^{18}$O ratio approaching unity can be recovered, again, only by flattening the IMF in the high-mass domain (though, probably, a slope $x = 1.1$ as adopted here is too much; see the entries referring to models SMG-11 and SMG-12 in Table \[tab:endrat\], last columm).
So in summary: When considering the effects of stellar rotation on the yields of massive stars, it is possible to explain a value of the ratio $^{13}$C/$^{18}$O $\simeq$ 1 in SMGs without resorting to a top-heavy IMF only under very specific circumstances: (i) stars have to rotate fast only below a given metallicity threshold; (ii) the star formation must last long enough for a significant amount of $^{13}$C to be ejected by AGB stars; (iii) even in the latter case, the canonical IMF solution only holds for a particular choice of the AGB yield set. At star formation halt, in the case of a continuous star formation lasting 1 Gyr, LIMS and super-AGB stars have provided about 75 (model SMG-11) and more than 90 per cent (model SMG-12) of the total $^{13}$C content of the interstellar medium in our models. These percentages are higher than those reported with regard to the solar chemical composition in Section \[sec:scen\], pursuant to the more intense star formation activity characterizing the SMGs.
Clearly, observing different molecules to infer other isotope ratios is a crucial step towards a definite answer. In Table \[tab:endrat\] we provide predictions, to be tested by future observations. In particular, from Table \[tab:endrat\] it can be seen that simultaneous measurements of $^{14}$N/$^{15}$N and $^{18}$O/$^{17}$O isotope ratios would be particularly useful to break the degeneracies.
Discussion {#sec:dis}
==========
It is well known that stellar rotation significantly alters the outcome of stellar nucleosynthesis computations, especially so for the so-called spinstars, the first generations of extremely low-metallicity fast rotators in galaxies .
Recently, a large, homogeneous grid of massive star yields covering an extended range of initial masses, metallicities and rotational velocities has been published, including the contributions from the presupernova evolution, as well as from the final, explosive stages [@2018ApJS..237...13L]. The inclusion of these yields in GCE models, however, requires the adoption of a most uncertain quantity, the IDROV [@2018MNRAS.476.3432P], similar in spirit to the stellar IMF, but by far more uncertain.
As pointed out by [@2009ApJ...700..844P], as a matter of fact our ability to establish the true distribution of equatorial rotational velocities from direct observations of massive stars is severely hampered by the reduced sample sizes. One may hope to infer the Galactic IDROV indirectly, from a fit to the chemical properties of Galactic stars of different ages/metallicities. However, as stressed by @2018MNRAS.476.3432P, this is a risky procedure, since different combinations of IDROVs and IMFs may provide equally good results. Moreover, the Galactic IDROV may not apply to other galaxies.
Indeed, @2007AJ....133.1092W confirm that massive stars formed in high-density regions lack the numerous slow rotators seen for their counterparts born in low-density regions and the field, with the metallicity also likely playing a role . But alas, the most updated census of massive stars in the Galaxy and Magellanic Clouds does not show evidence for significant differences in early-type stellar rotational velocities with metallicity !
Owing to all the above-mentioned uncertainties, we prefer to study some extreme cases here. We rather crudely assume that all stars rotate fast, or do not rotate at all or, in some cases, we consider the existence of a metallicity threshold determining an abrupt transition between these opposite regimes. We do not pretend to depict faithfully the complex reality of galaxies, but expect their true evolution to sit somewhere in between the extreme cases discussed in this paper.
Conclusions {#sec:end}
===========
In this paper, we discuss the results we find when implementing in our GCE code a new, large grid of yields for massive stars covering a wide range of initial masses and metallicities and including the effects of different stellar initial rotational velocities on the nucleosynthesis [@2018ApJS..237...13L].
The grid is first tested against available CNO abundance data for stars and molecular gas in the solar neighbourhood and Milky Way disc and then used to model the evolution of CNO isotopes in the typical SMG. In our sample of four strongly-lensed SMGs seen by ALMA, the observations suggest extremely low $^{13}$C/$^{18}$O ratios, that are thought to arise from an excess of massive stars in starbursts [@2018Natur.558..260Z see also [@2019ApJ...879...17B], for recent work on local analogs]. The following key conclusions can be drawn:
1. In qualitative agreement with the results of the study by @2018MNRAS.476.3432P, that are based on an analysis of the chemical abundances of solar neighbourhood stars, we find that our model predictions for the Galaxy can be reconciled with the observations on the assumption that most stars rotate fast until a metallicity threshold is reached above which the majority of the stars have small or null rotational velocities; the value of this metallicity threshold happens to coincide with the one characterizing the end of the halo phase in our model.
2. If the formation of fast rotators is favoured in the hot, turbulent medium of dusty starburst galaxies at high redshifts irrespective of metallicity, a value around unity for the $^{13}$C/$^{18}$O ratio as observed in SMGs needs a top-heavy IMF in order to be explained.
3. If, instead, a metallicity threshold prevents the formation of massive fast rotators at relatively high metallicities, similarly to what is suggested for the Milky Way, it is possible to explain the low $^{13}$C/$^{18}$O ratios observed in SMGs with a universal IMF; however, in this case the star formation in SMGs must last long ($\sim$1 Gyr) and some special choice has to be made as for the yields of LIMS and super-AGB stars.
Overall, when the yields of rotating massive stars are included in GCE models, the issue of whether the IMF varies (becoming flatter) during the most violent starburst events in the universe appears unsettled, unless a long-lasting SFH (either gasping or continuous) can be definitely ruled out. A firm conclusion waits for the determination of other isotopic abundances from submillimeter observations, as well as for a better understanding of star formation in extreme environments. A comparison of the chemical properties of the oldest stars in massive local ellipticals with the predictions of the models involving different elements would also be extremely useful.
Acknowledgements {#acknowledgements .unnumbered}
================
We are deeply indebted to Padelis P. Papadopoulos for pointing out the need for a thorough investigation of the role of massive fast rotators in the turbulent media of submillimeter galaxies, as well as for many enlightening discussions. DR acknowledges funding from INAF PRIN-SKA *“Empowering SKA as a Probe of galaxy Evolution with HI (ESKAPE-HI)”* program 1.05.01.88.04 (PI L. K. Hunt). This work benefited from the International Space Science Institute (ISSI, Bern, CH) and the International Space Science Institute–Beijing (ISSI-BJ, Beijing, CN) thanks to the funding of the team *“Chemical abundances in the ISM: the litmus test of stellar IMF variations in galaxies across cosmic time”* (PIs D. Romano, Z.-Y. Zhang). The research shown here acknowledges use of the Hypatia Catalog Database, which was supported by NASA’s Nexus for Exoplanet System Science (NExSS) research coordination network and the Vanderbilt Initiative in Data-Intensive Astrophysics (VIDA).
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: In this paper, we adopt standard definitions of elemental abundance ratios. For two elements X and Y, \[X/Y\] $\equiv \log(N_{\rm X}/N_{\rm Y}) - \log(N_{\rm X}/N_{\rm Y})_\odot$ is the ratio of the abundances by number on a logarithmic scale relative to the solar reference value, while X/Y refers to the ratio of the abundances by mass.
[^3]: reached similar conclusions by relying on less refined chemical evolution models and stellar yields.
[^4]: Many authors have argued that a slope steeper than @1955ApJ...121..161S’s has to be preferred for the local field star IMF independently from chemical evolution arguments; see, e.g., .
[^5]: In models MWG-11 and MWG-12, the mass range for full collapse to black holes is reduced to 60–100 M$_\odot$ for \[Fe/H\] $\ge -$1.
[^6]: The O abundances for the stars in the sample of are taken from .
[^7]: This is true only for the yield sets by [@2018ApJS..237...13L]; always report positive $^{17}$O yields for massive stars.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Daniel Gottesman$^1$[^1], and M. B. Hastings$^{2,3}$[^2]\
*$^1$Perimeter Institute for Theoretical Physics, Waterloo, Ontario, Canada*\
*$^2$Center for Nonlinear Studies and Theoretical Division,*\
*Los Alamos National Laboratory, Los Alamos, NM, USA*\
*$^3$ Microsoft Research, Station Q, University of California, Santa Barbara, CA, 93106*
title: 'Entanglement vs. gap for one-dimensional spin systems'
---
Introduction
============
In many local quantum systems, the long-distance behavior of correlation functions is a good diagnostic of the critical properties of a system near a phase transition as the excitation gap $\Delta$ tends to zero. Entanglement provides a different diagnostic, which gives complementary information about critical behavior, and is important for our ability to simulate these systems on classical computers, as described later.
To quantify entanglement, we define the entanglement entropy of a region $A$ with the complementary region $B$ (assuming the global state on $A \cup B$ is pure) as the von Neumann entropy $S(\rho_A)$, with $\rho_A$ the density matrix of region $A$ (i.e., the global state traced over region $B$). A state of a spin system that obeys an “area law” has the property that the entropy of a region (in one or more dimensions) is bounded by the surface area of the region instead of the usual maximum of the volume of the region. (See [@arealaw] for a review of area laws and references to a full cross-section of work in that area.) Area laws are conjectured to hold for the ground state of any gapped Hamiltonian spin system with short-range interactions in finite dimensions, but have not been proved in general. For one-dimensional systems, the area law simply states that the entropy of an interval along the line is bounded by a constant, no matter how large the interval.
At a critical point, it is possible that the entanglement, like the correlation length, can diverge. However, if we are merely *near* a critical point (either because of additional Hamiltonian terms or because we are considering a finite-size system), the gap $\Delta$ becomes finite, and in one dimension, an area law must hold [@Hastings-area-law]. The entropy can thus provide another way to identify a critical point. Note that diverging entanglement and diverging correlation length do not necessarily occur together; on the one hand, there exist states with all correlation functions short-range but with arbitrarily large entanglement [@expanders; @expanders2], while on the other hand there are critical points in two dimensions where the correlation functions become long-range but the area law is still obeyed [@2dboson].
In one dimension, the entanglement entropy has important implications for our ability to simulate quantum systems on a classical computer. Intuitively, so long as the entanglement entropy remains small, the system is “classical" in some sense. More precisely, we know that whenever the system obeys an area law for the Rényi entropy, $S_{\alpha}\equiv (1-\alpha)^{-1}\ln({\rm tr}(\rho_A^{\alpha}))$, for some $\alpha<1$ (which implies an area law for the von Neumann entropy), there exists a matrix product state approximation to the ground state with an error inverse polynomial in the bond dimension [@renyiapprox]. This approximation lies behind the powerful density-matrix renormalization group algorithm [@dmrg], as well as behind Vidal’s algorithm for time evolution [@tebd].
Hastings [@Hastings-area-law] proved that a general one-dimensional gapped Hamiltonian spin system satisfies an area law. However, the bound is probably not tight: $$S({\rm region}) \leq S_{\rm max} = c_0 \xi' \ln \xi' \ln D \, 2^{\xi' \ln D},
\label{eq:oneDbound}$$ where $c_0$ is a constant of order $1$, $D$ is the Hilbert space dimension of a single spin in the system, and $$\xi' = 6 \max (2v/\Delta, \xi_c),$$ where $v$ and $\xi_c$ are parameters from the Lieb-Robinson bound [@LRbound; @locality; @lr2; @lr3], and $\Delta$ is the spectral gap of the Hamiltonian. The quantity $v$ is the Lieb-Robinson velocity, and is roughly equal to the size of the energy scale of the largest individual terms in the Hamiltonian times the interaction range. The Lieb-Robinson velocity sets an upper bound for the group velocity of excitations, so that for a local operator $O$, the operator $O(t)=\exp(iHt)O\exp(-iHt)$ can be approximately described by an operator supported on the set of sites within distance $vt$ of the support of $O$. The quantity $\xi_c$ is a length of order the interaction range which determines the distance which excitations can “leak" outside the light cone of the Lieb-Robinson bound. In particular, all of these constants can be taken to be order $1$, except for $\Delta$, which could be very small even with all other parameters constant in size; for example, by tuning a system through a quantum critical point the quantity $\Delta$ can be made small. Thus, in the bound (\[eq:oneDbound\]), the entropy is limited by $\exp [O(1/\Delta)]$.
In contrast, physical intuition suggests that entanglement should be able to reach a distance roughly equal to the correlation length, i.e., about $1/\Delta$. However, most one-dimensional systems considered by physicists in fact have an even smaller amount of entanglement. Near a conformal critical point, the bound in the area law diverges as only $\log(1/\Delta)$ [@cftent; @cftent2]. Indeed, as far as we aware, for *all* one-dimensional systems where the area law has previously been calculated (including those near non-conformal critical points), the entropy of an interval is also bounded by about $\log (1/\Delta)$, which is below the bound in eq. (\[eq:oneDbound\]) by not one but [*two*]{} exponentials!
Therefore, we raise the following as an open question: given a one-dimensional system with interaction range and interaction strength both $O(1)$, how big can the entanglement be for a given spectral gap? Someone familiar with the previous work on the subject might well conjecture a bound of $\log(1/\Delta)$. In this paper we falsify that conjecture and get closer to the naive intuitive bound of $1/\Delta$ by demonstrating a family of systems where the entropy of some regions is ${\rm poly} (1/\Delta)$. In particular, we construct systems on a pair of rings coupled along a junction, each with $N$ $3$-state spins. In the ground state, the two rings are highly entangled with each other, and the entropy for one ring is roughly $N$. The spectral gap $\Delta$ is roughly $1/N^4 \log N$, so these systems have entropy for some regions equal to $(-\Delta \log \Delta)^{-1/4}$. The Hamiltonians we present depend on $N$ in two ways: First, they are associated to the ring structure, and the size of the rings changes with $N$, and second, we have some terms in the Hamiltonian which are polynomially small in $N$. Since there are also Hamiltonian terms of size $O(1)$, the presence of some small terms in the Hamiltonian does not affect the bound (\[eq:oneDbound\]). In sec. \[sec:infinite\], we discuss how to let the number of the particles in the system be much larger than $N$, so we can consider $N$ to be a separate parameter from the size of the system.
We have not succeeded in obtaining the naive estimate of $1/\Delta$ and we are very far from $\exp[O(1/\Delta)]$, so at present neither of the authors has any firm intuition for a tight upper bound. Given progress in recent years in understanding the relation between correlations and spectral gap [@locality], hopefully this question will be answered too. Intuitively, an area law is fairly simple: in a gapped system, the entanglement spreads only over a finite length and therefore only the spins near the boundary of $A$ are entangled with those outside of $A$. However, in general any spin within $A$ will be entangled with other spins within $A$ as well as spins in the complement of $A$. Thus, to turn this intuition into practice seems quite tricky as it leads us into the difficult question of multi-party entanglement.
The intuition for the construction used in this paper is somewhat similar to that used for proofs of QMA-completeness and universality of adiabatic quantum computation [@QMA1D], although the details are rather different. In particular, we imagine a computational process that generates entanglement among the spins and write down a Hamiltonian whose ground state is a superposition of time steps of the process.
To be more precise, we imagine the spins have the three states ${|{0}\rangle}$, ${|{1}\rangle}$, and ${|{x}\rangle}$. ${|{0}\rangle}$ and ${|{1}\rangle}$ are two states of a qubit, and most of the spins in each ring will be in this qubit subspace, but one spin in each ring will be in the ${|{x}\rangle}$ state. We can imagine the ${|{x}\rangle}$ as a “hole” in a ring otherwise filled with spin-$1/2$ particles. The rings will meet in two sites, which we number $1$ and $N$ on each ring, with sites $2, \ldots, N-1$ on each ring being connected only to points in order on the same ring. When there are two qubit states at the two number-$1$ sites, we project onto a maximally entangled state such as the singlet ${|{0}\rangle}_L {|{1}\rangle}_R - {|{1}\rangle}_L {|{0}\rangle}_R$. We then move each ${|{x}\rangle}$ state around its ring, shifting the qubit states by one place, allowing us to entangle another pair. The steady-state behavior of this system is thus $N-1$ maximally entangled pairs, aligned roughly between corresponding sites on the two rings, perhaps offset by one place, depending on the locations of the ${|{x}\rangle}$ states.
We then write down a nearest-neighbor Hamiltonian whose ground state is the superposition of such states over all possible locations of the two ${|{x}\rangle}$ states. In the ground state, the entropy of one ring taken alone will be slightly over $N-1$, since we broken all $N-1$ EPR pairs, but we will show that the spectral gap of the Hamiltonian is $\Delta = \Omega(1/N^4 \log N)$. Thus, the entropy of just the left ring, say, is $\Omega((-\Delta \log \Delta)^{-1/4})$, a polynomial in $1/\Delta$.
The ground state differs from those used in the previous QMA-completeness constructions in that there is an ${|{x}\rangle}$ state in each ring, and we take the superposition over possible positions of the [*pair*]{}, whereas in the previous constructions, the ground state was a superposition over time slices. Essentially, the previous Hamiltonians were quantum walks on a line, whereas ours is a walk on a two-dimensional graph.
While the Hamiltonian we write requires fine-tuning, it is not too unnatural from the physics point of view. Regarding the state ${|{0}\rangle}$ as representing an electron with spin up, the state ${|{1}\rangle}$ as an electron with spin down, and ${|{x}\rangle}$ as a hole, within the rings our Hamiltonian is simply the $t-J$ model Hamiltonian of condensed matter physics with $t\neq 0$ and $J=0$. Only where the rings meet do we have a more complicated interaction.
Sandy Irani has shown a similar result independently [@irani]. The two constructions have slightly different motivations and each has different advantages. We were primarily interested in the dependence of entanglement entropy on gap, and we have a more rapid divergence of entanglement entropy with inverse gap and use fewer states per site. Irani’s primary motivation was to study the effect of translational invariance on entanglement entropy. Her construction shows that it is possible to achieve a polynomial dependence of entanglement entropy on gap with a translation-invariant Hamiltonian, whereas our construction explicitly breaks translation invariance.
From Two Rings To An Infinite Line {#sec:infinite}
==================================
As mentioned above, our main concern in this paper is to construct a local Hamiltonian $H$ for a system of $2N$ $3$-state spins. It will consist of two rings containing $N$ sites each which are connected at two adjacent points, and we will show that for the ground state of $H$, the entanglement of the left ring with the right ring is $\Omega((-\Delta \log \Delta)^{-1/4})$ (where $\Delta$ is the spectral gap of $H$). It may not be immediately obvious how this relates to the area law of an infinite one-dimensional system, so before we discuss the technical details of the construction, let us first explain how the result immediately carries over to a system of spins on a line with the property that the entanglement of some regions is polynomial in the inverse gap.
First, note that we can squash the two rings down into a one-dimensional “brick” composed of a total of $N$ $9$-state spins, as illustrated in figure \[fig:brick\].
(300,60)
(10,45)[(20,10)[a)]{}]{} (50,30) (100,30)
(69,36) (62,46) (50,50) (38,46) (31,36) (31,24) (38,14) (50,10) (62,14) (69,24)
(62,34)[(0,0)[1]{}]{} (55,43)[(0,0)[2]{}]{} (45,40)[(0,0)[…]{}]{} (62,24)[(0,0)[N]{}]{}
(50,30)[(12,12)\[tr\]]{} (50,30)[(12,12)\[tl\]]{} (50,30)[(12,12)\[bl\]]{} (50,24)[(1,0)[5]{}]{}
(119,36) (112,46) (100,50) (88,46) (81,36) (81,24) (88,14) (100,10) (112,14) (119,24)
(88,34)[(0,0)[1]{}]{} (95,43)[(0,0)[2]{}]{} (105,40)[(0,0)[…]{}]{} (88,24)[(0,0)[N]{}]{}
(100,30)[(12,12)\[tl\]]{} (100,30)[(12,12)\[tr\]]{} (100,30)[(12,12)\[br\]]{} (100,24)[(-1,0)[5]{}]{}
(160,45)[(20,10)[b)]{}]{} (180,24)[(48,12)]{} (238,24)[(48,12)]{}
(228,36) (216,36) (204,36) (192,36) (180,36) (180,24) (192,24) (204,24) (216,24) (228,24)
(228,46)[(0,0)[1]{}]{} (216,46)[(0,0)[2]{}]{} (204,43)[(0,0)[…]{}]{} (228,14)[(0,0)[N]{}]{}
(286,36) (274,36) (262,36) (250,36) (238,36) (238,24) (250,24) (262,24) (274,24) (286,24)
(238,46)[(0,0)[1]{}]{} (250,46)[(0,0)[2]{}]{} (262,43)[(0,0)[…]{}]{} (238,14)[(0,0)[N]{}]{}
(174,18)[(12,24)]{}
Note that $3$-state spins which are neighboring in the rings get mapped to neighboring $9$-state spins in the brick, so the resulting Hamiltonian is still a nearest-neighbor Hamiltonian. While at the place where the two rings meet, the ring Hamiltonian has a four-spin term (see the following section), all four of those spins are contained in just two $9$-state spins, so the brick Hamiltonian contains only two-spin terms. Note that it is important that we squash the rings this way and not try to connect the rings along their length, as we are interested in studying the area law based on the entropy of a single ring by itself, and we therefore want the surface area of a single ring to be small. In the brick, the surface area of what used to be a ring is simply $2$ spins: the $9$-state spin consisting of the original sites $1$ and $N$, and the $9$-state spin consisting of sites $N/2$ and $N/2+1$. If we had joined the rings length-wise, the surface area would be $N$, the size of a ring, which would be no good.
Now we have a finite truly one-dimensional system whose ground state has the property that the left half of the brick has high entanglement with the right half. We can take an infinite line of $9$-state particles and chop it up into identical bricks of this sort (figure \[fig:line\]).
(400,60)
(30,24)(58,0)[6]{}[(48,12)]{}
(78,36)(58,0)[6]{} (66,36)(58,0)[6]{} (54,36)(58,0)[6]{} (42,36)(58,0)[6]{} (30,36)(58,0)[6]{} (30,24)(58,0)[6]{} (42,24)(58,0)[6]{} (54,24)(58,0)[6]{} (66,24)(58,0)[6]{} (78,24)(58,0)[6]{}
(26,20)(116,0)[3]{}[(114,20)]{}
(6,24)[(20,12)[$\cdots$]{}]{} (372,24)[(20,12)[$\cdots$]{}]{}
If we take a Hamiltonian which is simply the sum of the brick Hamiltonians on each piece, with no terms interacting distinct pieces, the ground state of the resulting Hamiltonian is simply the tensor product of the ground states of the individual bricks. There are many excited states, but the lowest-lying excited states are superpositions of states for which all but one brick is in the ground state, and the remaining brick is in its first excited state. The gap of the line Hamiltonian is thus just equal to $\Delta$, the gap of the original ring Hamiltonian (i.e., $\Omega(1/N^4 \log N)$). Let us examine an interval $A$ which begins at the center of one brick and ends at the center of another brick. Bricks which are completely contained in $A$ do not contribute to the entropy, but the two bricks which are cut in half each contribute about $N$ units of entropy, so the entropy of $A$ in the ground state is roughly $2N$. In particular, the entropy of $A$ is $\Omega((\Delta \log \Delta)^{-1/4})$.
Of course, some regions have very low entropy. Indeed, an interval containing only complete bricks will have entropy $0$. This does not affect the area law for the system, which simply provides an upper bound on the entanglement of any region. Some regions have high entropy whereas others do not. Nevertheless, if one finds this result unsatisfying, one may always consider a larger system constructed by two adjacent lines of $9$-state particles, each constructed as above, but offset by half a brick (as in figure \[fig:overlapping\]).
(400,60)
(30,10)(58,0)[6]{}[(48,12)]{}
(78,22)(58,0)[6]{} (66,22)(58,0)[6]{} (54,22)(58,0)[6]{} (42,22)(58,0)[6]{} (30,22)(58,0)[6]{} (30,10)(58,0)[6]{} (42,10)(58,0)[6]{} (54,10)(58,0)[6]{} (66,10)(58,0)[6]{} (78,10)(58,0)[6]{}
(26,6)(116,0)[3]{}[(114,20)]{}
(6,10)[(20,12)[$\cdots$]{}]{} (372,10)[(20,12)[$\cdots$]{}]{}
(30,35)(58,0)[6]{}[(48,12)]{}
(78,47)(58,0)[6]{} (66,47)(58,0)[6]{} (54,47)(58,0)[6]{} (42,47)(58,0)[6]{} (30,47)(58,0)[6]{} (30,35)(58,0)[6]{} (42,35)(58,0)[6]{} (54,35)(58,0)[6]{} (66,35)(58,0)[6]{} (78,35)(58,0)[6]{}
(84,31)(116,0)[2]{}[(114,20)]{} (21,31)[(1,0)[61]{}]{} (82,31)[(0,1)[20]{}]{} (82,51)[(-1,0)[61]{}]{} (377,31)[(-1,0)[61]{}]{} (316,31)[(0,1)[20]{}]{} (316,51)[(1,0)[61]{}]{}
(6,35)[(20,12)[$\cdots$]{}]{} (372,35)[(20,12)[$\cdots$]{}]{}
Now any interval will cut off on each end at least one-quarter of a brick on one of the two lines. We can combine the two lines into a single line of $81$-state particles, and the resulting Hamiltonian has the property that for its ground state, any interval of at least $N$ spins will have entropy $\Omega((- \Delta \log \Delta)^{-1/4})$.
The Two-Ring Hamiltonian
========================
We will write our Hamiltonian $H$ as a sum of various terms: $$H = H_L + H_R + H_B + H_P + H_V.$$ The terms $H_L$ and $H_R$ cause the spins (containing a qubit state) to hop around the left and right rings, respectively. $H_B$ handles the hopping of the ${|{x}\rangle}$ “holes” between the sites numbered $1$ and $N$ on each ring. In order to keep the two rings in sync, we require that the holes hop in unison when they hop between sites $1$ and $N$, and $H_B$ ensures that this happens. $H_P$ is the projection Hamiltonian, which is responsible for ensuring that any pair of qubits at site $1$ on the left and right rings are in fact in the singlet state. $H_x$ is a self-energy term for the ${|{x}\rangle}$ hole states, ensuring that the ground state has exactly one ${|{x}\rangle}$ in each ring.
More specifically, let $$S_i^{L/R} = \sum_{a,b} {|{ab}\rangle}_{i,i+1}^{L/R} {\langle{ba}|}.$$ We work modulo $N$, so site $N+1$ should be considered the same as site $1$. $S_i^{L/R}$ swaps the particles in sites $i$ and $i+1$ for the left (L) or right (R) ring, regardless of the values of the spins at those sites. Thus we can define $$F_i^{L/R} = S_i^{L/R} {|{x}\rangle}_i^{L/R} {\langle{x}|},$$ which swaps an ${|{x}\rangle}$ at site $i$ with a state at site $i+1$, and annihilates any state which does not have an ${|{x}\rangle}$ at site $i$.
We then let $$H_{L/R} = \sum_{i=1}^{N-1} \left( {|{x}\rangle}_i^{L/R} {\langle{x}|} + {|{x}\rangle}_{i+1}^{L/R} {\langle{x}|} - F_i^{L/R} - (F_i^{L/R})^\dagger \right).$$ $H_L$ thus produces an energy penalty for having the ${|{x}\rangle}$ in the $i$th site in the left ring, but cancels that energy penalty if the ${|{x}\rangle}$ hops to the $(i+1)$th site, swapping with the state there; similarly, there is a penalty for having the ${|{x}\rangle}$ in the $(i+1)$th site of the left ring which is cancelled by letting it hop to the $i$th site. $H_L$ thus tends to produce a superposition of ${|{x}\rangle}$ over sites $1$ through $N$ in the left ring, with the qubit states shifting appropriately to make room, and $H_R$ favors a similar superposition over sites in the right ring. Note, however, that $H_L$ and $H_R$ do not have terms allowing ${|{x}\rangle}$ to hop between sites $1$ and $N$.
Instead, we let $H_B$ provide those terms: $$H_B = {|{xx}\rangle}_N^{LR} {\langle{xx}|} + {|{xx}\rangle}_1^{LR} {\langle{xx}|} - F_N^L F_N^R - (F_N^L)^\dagger (F_N^R)^\dagger.$$ $H_B$ has a structure similar to $H_L$ and $H_R$, but it instead provides a penalty only if the ${|{x}\rangle}$ states for the two rings are both at site $1$ or both at site $N$, and then gives a cancelling energy bonus if they hop together to the other pair (i.e., both hop from site $N$ to site $1$ or from site $1$ to site $N$). In conjunction with $H_L$ and $H_R$, this means that the ${|{x}\rangle}$ states in the left and right rings can wander separately in the two rings, but cannot complete a full circle around a ring except by hopping together.
$H_P$ is defined as a projector: $$H_P = (I - {|{x}\rangle}_1^L {\langle{x}|}) (I - {|{x}\rangle}_1^R {\langle{x}|}) - {|{\Psi^-}\rangle}_1^{LR} {\langle{\Psi^-}|},$$ which produces an energy penalty for any state which has a qubit state in site number $1$ for both rings, unless the two qubit states form a singlet state ${|{\Psi^-}\rangle} = ({|{01}\rangle}-{|{10}\rangle})/\sqrt{2}$.
Finally, let $$H_V^{L/R} = V_1-V_1 \sum_{i=1}^N {|{x}\rangle}_i^{L/R} {\langle{x}|} + V_2 \sum_{i=1}^N {|{xx}\rangle}_{i,i+1}^{L/R} {\langle{xx}|},$$ and $$H_V = H_V^L + H_V^R.$$ We choose $V_1$ and $V_2$ to be positive constants, so $H_V$ will produce an energy bonus proportional to the total number of ${|{x}\rangle}$ states in the two rings, but an energy penalty if any two ${|{x}\rangle}$ states are adjacent in a single ring. Since $H_L$, $H_R$, and $H_B$ will tend to produce superpositions of different locations for the ${|{x}\rangle}$ states in each ring, if there is more than one ${|{x}\rangle}$ within a given ring, $H_L$, $H_R$, and $H_B$ will favor states where some ${|{x}\rangle}$ states are adjacent, which will then get an energy penalty due to the $V_2$ term. By tuning $V_1$ and $V_2$ appropriately, we will ensure that the ground state has exactly one ${|{x}\rangle}$ state in each ring, as desired. The constant term $2V_1$ then gives that ground state $0$ energy.
Note that the terms $H_L$, $H_R$, $H_B$, and $H_P$ are all non-negative operators. However, $H_V^L$ and $H_V^R$ are not, which means we will need to exercise some caution in bounding the energy gap.
The Ground State and the Gap
============================
We now turn to determining the structure of the ground state of $H$ and its spectral gap. We shall prove the following:
Suppose $V_1 = 1/N^4$ and $V_2 = 1$. Then
1. $H$ has a nondegenerate ground state ${|{g}\rangle}$,
2. The spectral gap $\Delta$ of $H$ is $\Omega(1/N^4 \log N)$, and
3. $S(\rho_L) = S(\rho_R) \geq N-1$, where $\rho_{L/R} = \operatorname{Tr}_{R/L} {|{g}\rangle} {\langle{g}|}$.
The first thing to notice is that no term of $H$ changes the number of ${|{x}\rangle}$ states in a ring. Therefore, we can treat the Hilbert space as a direct sum of subspaces with different numbers of ${|{x}\rangle}$ states in the two rings and analyze them separately. Let the subspace with $a$ ${|{x}\rangle}$ states in the left ring and $b$ ${|{x}\rangle}$ states in the right ring be ${\mathcal{H}}_{a,b}$. We start with ${\mathcal{H}}_{1,1}$, which we claim contains the global ground state. We therefore analyze the ground state (which will have energy $0$) and gap within this subspace. Then we show that all states in the other subspaces have energy at least $O(1/N^4)$.
One ${|{x}\rangle}$ in Each Ring
--------------------------------
Within the subspace ${\mathcal{H}}_{1,1}$, the Hamiltonian term $H_V$ becomes uniformly $0$, so we can ignore it. We are thus left with $H = H_L + H_R + H_B + H_P$, and we wish to find the ground state and gap of this Hamiltonian on ${\mathcal{H}}_{1,1}$.
First, let us define some convenient basis states. We start with states which have both ${|{x}\rangle}$ states in site $N$ in their respective rings, and define $${|{\{\alpha_i\}}\rangle} = {|{\alpha_1}\rangle}_1^{LR} {|{\alpha_2}\rangle}_2^{LR} \cdots {|{\alpha_{N-1}}\rangle}_{N-1}^{LR} {|{xx}\rangle}_N^{LR},$$ where ${|{\alpha_i}\rangle}$ is a Bell state, $i = 1, \ldots, N-1$. The state ${|{\{\alpha_i\}}\rangle}$ thus has the two qubit states at the $j$th position in each ring sharing the particular entangled state ${|{\alpha_j}\rangle}$.
Now define $M_{a,b}$ to be an operator that shifts ${|{x}\rangle}$ to site $a$ in the left ring and site $b$ in the right ring: $$M_{a,b} = \left( S_a^{L} S_{a+1}^{L} \cdots S_{N-1}^{L} \right) \left( S_b^{R} S_{b+1}^{R} \cdots S_{N-1}^{R} \right).$$ (Make the convention that if $a$ or $b$ is equal to $N$, do not include the L or R operators, respectively. Thus $M_{N,N} = I$.) Then $M_{a,b} {|{\{\alpha_i\}}\rangle}$ has the ${|{x}\rangle}$ states at sites $a$ and $b$ in the left and right rings, respectively, but some of the entangled states ${|{\alpha_i}\rangle}$ are now misaligned: If $j < \min\{a,b\}$ then the two sites $j$ share ${|{\alpha_j}\rangle}$, and if $j > \max\{a,b\}$, then the two sites $j+1$ share ${|{\alpha_j}\rangle}$. However, if $a < j < b$, then the state ${|{\alpha_j}\rangle}$ is shared between site $j+1$ in the left ring and site $j$ in the right ring, whereas if $b < j < a$, then ${|{\alpha_j}\rangle}$ is shared between site $j$ in the left ring and $j+1$ in the right ring. Note that for $M_{N,1} {|{\{\alpha_i\}}\rangle}$, all pairs are misaligned by one. Thus, even though $S_N^R M_{N,1} {|{\{\alpha_i\}}\rangle}$ has both ${|{x}\rangle}$ sites back at site $N$, the entangled pairs are misaligned, so it is not a state of the form ${|{\{\beta_i\}}\rangle}$ for any $\{\beta_i\}$.
Let $R = M_{1,1}^\dagger S_N^L S_N^R$. Then $R\, {|{\{\alpha_i\}}\rangle}$ also has both ${|{x}\rangle}$ states back at site $N$, but since the two ${|{x}\rangle}$ states have [*both*]{} circled the ring once, the pairs remain in sync. They have all been shifted by one place: $$R\, {|{\{\alpha_i\}}\rangle} = {|{\{\beta_i\}}\rangle},$$ with ${|{\beta_i}\rangle} = {|{\alpha_{i+1}}\rangle}$ ($i = 1, \ldots, N-2$) and ${|{\beta_{N-1}}\rangle} = {|{\alpha_{1}}\rangle}$.
We now look at the action of the Hamiltonian on these states. The states $M_{a,b} {|{\{\alpha_i\}}\rangle}$ form a basis for ${\mathcal{H}}_{1,1}$, and $H_P$ is diagonal in this basis, with eigenvalue $0$ for any state of the form $M_{1,b} {|{\{\alpha_i\}}\rangle}$, $M_{a,1} {|{\{\alpha_i\}}\rangle}$ (for any $a$, $b$), or for $M_{a,b} {|{\{\alpha_i\}}\rangle}$, with $a,b > 1$ and ${|{\alpha_1}\rangle} = {|{\Psi^-}\rangle}$. $H_L$ and $H_R$ will interact different states $M_{a,b} {|{\{\alpha_i\}}\rangle}$ with the same list $\{\alpha_i\}$ of Bell states. $H_B$ will couple $M_{1,1} {|{\{\alpha_i\}}\rangle}$ with $R\, {|{\{\alpha_i\}}\rangle}$, so the combination $H_L + H_R + H_B$ couples all states of the form $M_{a,b} {|{\{\alpha_i\}}\rangle}$ with the list of Bell states $\{\alpha_i\}$ fixed up to cyclic shifts. Suppose the list $\{\alpha_i\}$ has period $p$ under cyclic shifts (i.e., $R^p {|{\{\alpha_i\}}\rangle} = {|{\{\alpha_i\}}\rangle}$). Then it follows that $${|{S(\{\alpha_i\})}\rangle} \equiv \frac{1}{\sqrt{pN^2}} \sum_{r=0}^{p-1} \sum_{a,b} M_{a,b} R^r {|{\{\alpha_i\}}\rangle}$$ is an eigenvector of $H_L + H_R + H_B$ with eigenvalue $0$, and all states annihilated by $H_L + H_R + H_B$ are of this form. We note that the Hilbert subspace ${\mathcal{H}}_{1,1}$ further decomposes into invariant subspaces of $H$, each of which can be labeled by a list of Bell states $\{\alpha_i\}$, with two lists equivalent if they differ only by a cyclic shift on $i$. We will refer to the subspace associated to the list $\{\alpha_i\}$ as ${\mathcal{H}}(\{\alpha_i\})$; we choose some convenient representative list $\{\alpha_i\}$ from the set of equivalent lists to label each ${\mathcal{H}}(\{\alpha_i\})$.
The state ${|{g}\rangle} = {|{S(\{\alpha_i\})}\rangle}$ with ${|{\alpha_i}\rangle} = {|{\Psi^-}\rangle}$ for all $i$ also has eigenvalue $0$ for $H_P$, as every term in it has either ${|{x}\rangle}$ in site $1$ for at least one ring or has a singlet shared between the two $1$ sites. All other states ${|{S(\{\alpha_i\})}\rangle}$ do not have this property: Even if ${|{\alpha_1}\rangle} = {|{\Psi^-}\rangle}$, some ${|{\alpha_i}\rangle} \neq {|{\Psi^-}\rangle}$, and therefore $H_P R^{i-1} {|{\{\alpha_i\}}\rangle} = 1$. ${|{g}\rangle}$ is thus the ground state within ${\mathcal{H}}_{1,1}$ (and, we will show, within the whole Hilbert space), whereas we claim any other state has substantially higher energy.
\[thm:onexgap\] If ${|{\psi}\rangle} \in {\mathcal{H}}(\{\alpha_i\})$ with ${|{\alpha_i}\rangle} \neq {|{\Psi^-}\rangle}$ for at least one value of $i$, then $${\langle{\psi}|} H_L + H_R + H_B + H_P {|{\psi}\rangle} = \Omega(1/p^2 N^2 \log N),$$ where $p$ is the periodicity of $\{\alpha_i\}$ under cyclic permutations.
Since $p \leq N$, the theorem implies that the gap of the full Hamiltonian within ${\mathcal{H}}_{1,1}$ is $\Omega(1/N^4 \log N)$.
As a warmup to proving this, we first consider $H_L + H_R + H_B$ and prove that it has a polynomial gap. While we will not use precisely this proposition in the proof of Theorem \[thm:onexgap\], the proof uses a very similar technique, and the main lemma for Prop. \[prop:xmovegap\] will also be central to proving Theorem \[thm:onexgap\].
The Hamiltonian $H_L + H_R + H_B$ can be thought of as a quantum walk on a graph. In particular, let us restrict to the invariant subspace ${\mathcal{H}}(\{\alpha_i\})$, where $\{\alpha_i\}$ has period $p$ under cyclic shifts. We then get a graph $G_p$ of $pN^2$ nodes. The states ${|{a,b,r}\rangle} \equiv M_{a,b} R^r {|{\{\alpha_i\}}\rangle}$ can be considered as the nodes of the graph, which consists of $p$ $N \times N$ grids, with the $(N,N)$ site of one grid connected to the $(1,1)$ site of the next one, and the $(N,N)$ site of the $p$th grid connected to the $(1,1)$ site of the $1$st grid. (See figure \[fig:grids\].)
(240,240)
(0,0)[(1,1)[15]{}]{}
(15,15)(0,15)[5]{}[(1,0)[60]{}]{} (15,15)(15,0)[5]{}[(0,1)[60]{}]{}
(13,12)[(0,0)\[tl\][${|{1,1,0}\rangle}$]{}]{} (75,12)[(0,0)\[t\][${|{N,1,0}\rangle}$]{}]{} (15,78)[(0,0)\[b\][${|{1,N,0}\rangle}$]{}]{} (78,77)[(0,0)\[tl\][${|{N,N,0}\rangle}$]{}]{}
(75,75)[(1,1)[15]{}]{}
(90,90)(0,15)[5]{}[(1,0)[60]{}]{} (90,90)(15,0)[5]{}[(0,1)[60]{}]{}
(87,88)[(0,0)\[br\][${|{1,1,1}\rangle}$]{}]{} (153,152)[(0,0)\[tl\][${|{N,N,1}\rangle}$]{}]{}
(150,150)[(1,1)[15]{}]{}
(165,165)(0,15)[5]{}[(1,0)[60]{}]{} (165,165)(15,0)[5]{}[(0,1)[60]{}]{}
(162,163)[(0,0)\[br\][${|{1,1,2}\rangle}$]{}]{} (228,227)[(0,0)\[tl\][${|{N,N,2}\rangle}$]{}]{}
(225,225)[(1,1)[15]{}]{}
Edges of the graph correspond to states which are coupled by a single off-diagonal term of the Hamiltonian.
\[prop:xmovegap\] The gap of $H_L + H_R + H_B$ within ${\mathcal{H}}(\{\alpha_i\})$ is $\Omega(1/p^2 N^2 \log N)$.
$H_L$ is a simple quantum walk on a line of length $N$. It can be explicitly diagonalized, with eigenstates $${|{\psi_{k},b,r}\rangle} = \sqrt{\frac{2}{N}} \sum_{a=1}^N \cos \left(\frac{\pi k}{N} (a-1/2)\right) {|{a,b,r}\rangle}$$ and eigenvalues $$\lambda_k = 2(1 - \cos \frac{\pi k}{N})$$ for $k = 1, \ldots, N-1$. $k=0$ is also possible (although with a different normalization constant), and gives the ground state ${|{\psi_0,b,r}\rangle} = (1/\sqrt{N}) \sum_a {|{a,b,r}\rangle}$, with energy $0$. Similarly for $H_R$. Since a single grid has the two coordinates completely uncoupled, the eigenstates ${|{\psi_{kl},r}\rangle}$ of $H_L + H_R$ are similarly simple, with eigenvalues $\lambda_{kl} = \lambda_k + \lambda_l$. The ground state of $H_L + H_R$ (for fixed $r$) is ${|{\psi_{00},r}\rangle} = (1/N) \sum_{a,b} {|{a,b,r}\rangle}$. For general $k, l$, let us write ${|{\psi_{kl},r}\rangle} = \sum_{a,b} f_{kl}(a,b) {|{a,b,r}\rangle}$. Then $$f_{kl}(a,b) = C_{kl} \cos\left(\frac{\pi k}{N} (a-1/2)\right) \cos\left(\frac{\pi l}{N} (b-1/2)\right),$$ where $C_{kl}$ is a normalization constant: $C_{kl} = 1/N$ for $(k,l) = (0,0)$, $C_{kl} = \sqrt{2}/N$ if exactly one of $k$, $l$ is $0$, and $C_{kl} = 2/N$ otherwise.
The ground state of $H_L + H_R + H_B$ is again simple: $${|{{\rm ground}}\rangle} = \frac{1}{\sqrt{p}} \sum_{r=0}^{p-1} {|{\psi_{00},r}\rangle} = \frac{1}{\sqrt{p} N} \sum_{r=0}^{p-1} \sum_{a,b = 1}^{N} {|{a,b,r}\rangle}.$$ However, the higher eigenstates are more complicated. We can simplify slightly by noting that the graph has translation invariance under the action of $R$, so the eigenstates of $H_L + H_R + H_B$ will also be eigenstates of $R$. $R$ has eigenvalues $\exp(2\pi i m/p)$ (since $R^p = I$ on the subspace ${\mathcal{H}}(\{\alpha_i\})$). Any eigenstate ${|{\psi}\rangle}$ of $R$ can be written as $${|{\psi}\rangle} = \frac{1}{\sqrt{p}} \sum_{r=0}^{p-1} \sum_{k,l=0}^{N-1} c_{kl} e^{2\pi i mr/p} {|{\psi_{kl},r}\rangle},
\label{eigenstate}$$ with $m = 0, \ldots, p-1$. Let us consider the expected energy ${\langle{\psi}|} H_L + H_R + H_B {|{\psi}\rangle}$ of any state ${|{\psi}\rangle}$ which is orthogonal to the ground state ${|{{\rm ground}}\rangle}$ and is an eigenstate of $R$. If $m=0$, $c_{00} = 0$ (because the state is orthogonal to the ground state), so $${\langle{\psi}|} H_L + H_R {|{\psi}\rangle} \geq \lambda_1 = \Theta(1/N^2).$$
Therefore, we should search for the first excited state among states with nonzero $m$. In particular, we will show that for any state with $m>0$, either the state is dominated by the ground state of $H_L + H_R$ (i.e., $c_{00}$ is very near $1$), in which case it does not satisfy $H_B$ very well, or there is a significant mixture of higher eigenstates of $H_L + H_R$.
For the proof of Theorem \[thm:onexgap\], we will need a stronger version of the result using unnormalized states. The necessary result is contained in the following lemma:
\[lemma:highm\] Let $${|{\psi}\rangle} = \frac{1}{\sqrt{p}} \sum_{r=0}^{p-1} \sum_{k,l=0}^{N-1} c_{kl} e^{2\pi i mr/p} {|{\psi_{kl},r}\rangle}.$$ Then $${\langle{\psi}|} H_L + H_R + H_B {|{\psi}\rangle} = \Omega(m'^2 |c_{00}|^2/p^2 N^2 \log N),$$ where $m' = \min(m, p-m)$.
When ${|{\psi}\rangle}$ is normalized, it is easy to see that ${\langle{\psi}|} H_L + H_R {|{\psi}\rangle}$ is large ($\Omega(1/N^2)$) unless $|c_{00}|$ is order $1$, so the proposition follows immediately from the lemma.
Let us consider the expectation value of $H_B$ for $m > 0$: $${\langle{\psi}|} H_B {|{\psi}\rangle} = \frac{1}{p} \sum_{r,r' = 0}^{p-1} \sum_{k,l,k',l' = 0}^{N-1} c_{kl} c^*_{k'l'} e^{2\pi i m (r-r')/p} {\langle{\psi_{k'l'},r'}|} H_B {|{\psi_{kl},r}\rangle}.$$ Since $$\begin{gathered}
{\langle{\psi_{k'l'},r'}|} H_B {|{\psi_{kl},r}\rangle} = \delta_{r,r'} \left[ f_{kl}(1,1) f^*_{k'l'}(1,1) + f_{kl}(N,N) f^*_{k'l'}(N,N) \right] \\
- \delta_{r,r'-1} f_{kl}(N,N) f^*_{k'l'}(1,1) - \delta_{r,r'+1} f_{kl}(1,1) f^*_{k'l'}(N,N),\end{gathered}$$ we have, focusing on the $(k,l) = (0,0)$ terms, $$\begin{gathered}
{\langle{\psi}|} H_B {|{\psi}\rangle} =
\frac{1}{N^2} |c_{00}|^2 \left[ 2 - 2 \cos (2 \pi i m /p) \right] + {\langle{\psi'}|} H_B {|{\psi'}\rangle}
\\ +
\frac{1}{N} \sum_{(k,l) \neq (0,0)} c_{kl} c^*_{00} \left[ f_{kl}(1,1)(1 - e^{2\pi i m/p}) + f_{kl}(N,N) (1 - e^{-2\pi i m/p}) \right]
\\ +
\frac{1}{N} \sum_{(k',l') \neq (0,0)} c_{00} c^*_{k'l'} \left[ f_{k'l'}^*(1,1)(1 - e^{-2\pi i m/p}) + f_{k'l'}^*(N,N) (1 - e^{2\pi i m/p}) \right],\end{gathered}$$ where $${|{\psi'}\rangle} = \frac{1}{\sqrt{p}} \sum_{r=0}^{p-1} \sum_{(k,l) \neq (0,0)} c_{kl} e^{2\pi i m r/p} {|{\psi_{kl},r}\rangle}.$$ Since $H_B$ is positive semidefinite, ${\langle{\psi'}|} H_B {|{\psi'}\rangle} \geq 0$, and $$\begin{aligned}
{\langle{\psi}|} H_B {|{\psi}\rangle} \geq &
\ \frac{1}{N^2} |c_{00}|^2 \left[ 2 - 2 \cos (2 \pi m /p) \right] \\
& \quad +
2 {\rm Re} \left\{ \frac{1}{N} \sum_{k,l \neq (0,0)} c_{kl} c^*_{00} \left[ f_{kl}(1,1)(1 - e^{2\pi i m/p}) + f_{kl}(N,N) (1 - e^{-2\pi i m/p}) \right] \right\} \nonumber \\
= &
\ \frac{2}{N^2} |c_{00}|^2 \left[ 1 - \cos (2 \pi m /p) \right] \\
& \quad +
\frac{2}{N} \sum_{k,l \neq (0,0)} {\rm Re} (c_{kl} c^*_{00}) C_{kl} \cos(\pi k/2N) \cos(\pi l/2N) [1 - \cos (2\pi m/p)] [1 + (-1)^{k+l}]
\nonumber \\ & \quad +
\frac{2}{N} \sum_{k,l \neq (0,0)} {\rm Im} (c_{kl} c^*_{00}) C_{kl} \cos(\pi k/2N) \cos(\pi l/2N) \sin (2\pi m/p) [1 - (-1)^{k+l}]
\nonumber \\
\geq &
\ \frac{2 |c_{00}|^2 }{N^2} \left[ 1 - \cos (2 \pi m /p) \right]
\\ & \quad -
\frac{8}{N^2} \sum_{k+l > 0 {\rm\ even}} |c_{kl}| |c_{00}| \left[1 - \cos (2\pi m/p)\right]
-
\frac{8}{N^2} \sum_{k+l {\rm\ odd}} |c_{kl}| |c_{00}| |\sin (2\pi m/p)|
\nonumber \\
= &
\ \frac{2|c_{00}| }{N^2} \left[ 1 - \cos (2 \pi m /p) \right] \left( |c_{00}| - 4 \sum_{k+l > 0 {\rm\ even}} |c_{kl}|
-
4 |\cot (\pi m /p)| \sum_{k+l {\rm\ odd}} |c_{kl}| \right).\end{aligned}$$ For $0 < m/p \leq 1/2$, $|\cot (\pi m /p)| \leq p/(m \pi)$, so $${\langle{\psi}|} H_B {|{\psi}\rangle} \geq
\frac{2 |c_{00}|}{N^2} \left[ 1 - \cos (2 \pi m /p) \right] \left[ |c_{00}| - 4 \sum_{k+l > 0 {\rm\ even}} |c_{kl}|
- 4 \frac{p}{m\pi} \sum_{k+l {\rm\ odd}} |c_{kl}| \right].
\label{eq:hbbound}$$ For $1/2 \leq m/p < 1$, we can just consider $m' = p-m$, in which case $|\cot (\pi m /p)| = |\cot (\pi m'/p)|$ and $\cos (2 \pi m / p) = \cos (2 \pi m'/p)$, and we can just substitute $m'$ for $m$ in equation (\[eq:hbbound\]). Otherwise, we let $m' = m$.
Therefore, one of two things must be true. The first possibility is that the sum is dominated by $c_{00}$: $$|c_{00}| - 4 \sum_{k+l > 0 {\rm\ even}} |c_{kl}| - 4 \frac{p}{m'\pi} \sum_{k+l {\rm\ odd}} |c_{kl}| \geq |c_{00}|/2,$$ in which case $${\langle{\psi}|} H_B {|{\psi}\rangle} \geq \frac{|c_{00}|^2}{N^2} \left[ 1 - \cos (2 \pi m' /p) \right] = \Omega\left(\frac{m'^2 |c_{00}|^2}{p^2 N^2}\right).$$ The second possibility is that there is a non-negligible contribution from the higher $k,l$ eigenvalues: $$|c_{00}| - 4 \sum_{k+l > 0 {\rm\ even}} |c_{kl}| - 4 \frac{p}{m'\pi} \sum_{k+l {\rm\ odd}} |c_{kl}| < |c_{00}|/2,$$ which implies that $$\sum_{k,l \neq (0,0)} |c_{kl}| = \Omega(m' |c_{00}|/p).
\label{eq:csum}$$
However, $${\langle{\psi}|} H_L + H_R {|{\psi}\rangle} = \sum_{k,l = 0}^{N-1} (\lambda_k + \lambda_l) |c_{kl}|^2.$$ Note that $$\lambda_k = 2(1 - \cos (\frac{\pi k}{N})) \geq \frac{4 k^2}{N^2}$$ for $0 \leq k \leq N$, so $${\langle{\psi}|} H_L + H_R {|{\psi}\rangle} \geq \frac{4}{N^2} \sum_{k,l \neq (0,0)} (k^2 + l^2) |c_{kl}|^2.
\label{eq:gridenergy}$$ Under the constraint that $\sum_{k,l \neq (0,0)} |c_{kl}|$ is fixed (as suggested by (\[eq:csum\])), the RHS of (\[eq:gridenergy\]) is minimized when $|c_{kl}| = A/(k^2 + l^2)$ (this minimization can be done using Lagrange multipliers, for instance), with $$\sum_{k,l \neq (0,0)} \frac{A}{k^2 + l^2} = \Omega(m' |c_{00}|/p).$$ For large $N$, $\sum 1/(k^2 + l^2) = \Theta(\log N)$, so $A = \Omega(m' |c_{00}|/p \log N)$. Thus, $${\langle{\psi}|} H_L + H_R {|{\psi}\rangle} \geq \Omega(m'^2 |c_{00}|^2/p^2 N^2 \log^2 N) \sum_{k,l \neq (0,0)} \frac{1}{k^2 + l^2} = \Omega(m'^2 |c_{00}|^2/p^2 N^2 \log N).$$ That is, regardless of the size of $|c_{00}|$, we have that ${\langle{\psi}|} H_L + H_R + H_B {|{\psi}\rangle} = \Omega(m'^2 |c_{00}|^2/p^2 N^2 \log N)$, proving the lemma.
We next turn to proving Theorem \[thm:onexgap\]. The proof is somewhat similar: The ground state ${|{{\rm ground}}\rangle}$ of $H_L + H_R + H_B$ is completely delocalized in the ${|{a,b,r}\rangle}$ basis, so in particular has substantial energy for $H_P$. In order to cancel the bad ${|{a,b,r}\rangle}$ sites, we will need in to add a nontrivial contribution of higher $k$, $l$, and $m$ values, which will increase the energy of $H_L + H_R + H_B$.
Let $${|{\psi_{klm}}\rangle} = \frac{1}{\sqrt{p}} \sum_{r=0}^{p-1} e^{2\pi i mr/p} {|{\psi_{kl},r}\rangle},$$ and $${|{\psi}\rangle} = \sum_{k,l,m = 0}^{N-1} a_{klm} {|{\psi_{klm}}\rangle}.$$
We start by considering $|{\langle{\psi_{klm}}|} H_P {|{\psi_{000}}\rangle}|$. To calculate this, we expand ${|{\psi_{klm}}\rangle}$ and ${|{\psi_{000}}\rangle}$ in the ${|{a,b,r}\rangle}$ basis, since $H_P$ is diagonal in that basis. Then ${\langle{a',b',r'}|} H_P {|{a,b,r}\rangle}$ is $0$ unless $a=a'$, $b=b'$, and $r=r'$. It is also $0$ if $a=1$ or $b=1$. In all other cases, ${\langle{a,b,r}|} H_P {|{a,b,r}\rangle} = 0$ if ${|{\alpha_r}\rangle} = {|{\Psi^-}\rangle}$ and $1$ otherwise. Let $$R = \{r \mbox{ s.t. } {|{\alpha_r}\rangle} \neq {|{\Psi^-}\rangle}\}.$$ That is, $R$ is the set of “bad” $r$, that contribute to a nonzero energy for $H_P$.
Then $$\begin{aligned}
|{\langle{\psi_{klm}}|} H_P {|{\psi_{000}}\rangle}| = & \frac{1}{p} \left| \sum_{r,r'} e^{-2\pi i mr'/p} {\langle{\psi_{kl},r'}|} H_P {|{\psi_{00},r}\rangle} \right| \\
= & \frac{1}{p} \left| \sum_{r \in R} e^{-2\pi i r m/p} {\langle{\psi_{kl},r}|} H_P {|{\psi_{00},r}\rangle} \right| \\
\leq & \frac{|R|}{p} |{\langle{\psi_{kl},0}|} H_P {|{\psi_{00},0}\rangle}|.\end{aligned}$$ In the last line we assume, without loss of generality, that $0 \in R$. It is also worth noting that for $m=0$, we have equality in the last line. Furthermore, $$|{\langle{\psi_{00},0}|} H_P {|{\psi_{00},0}\rangle}| = \frac{1}{N^2} \sum_{a',a,b',b = 1}^{N} {\langle{a',b',0}|} H_P {|{a,b,0}\rangle} = \frac{(N-1)^2}{N^2}.$$ When $l \neq 0$, $$\begin{aligned}
|{\langle{\psi_{0l},0}|} H_P {|{\psi_{00},0}\rangle}| = & \frac{\sqrt{2}}{N^2} \left|\sum_{a',a,b',b = 1}^{N} \cos\left(\frac{\pi l}{N} (b'-1/2)\right) {\langle{a',b',0}|} H_P {|{a,b,0}\rangle} \right| \\
= & \frac{\sqrt{2}(N-1)}{N^2} \left| \sum_{b=2}^{N} \cos\left(\frac{\pi l}{N} (b-1/2)\right) \right|.\end{aligned}$$ But note that we have an orthogonality relationship for cosines, so that $$\sum_{b=1}^{N} \cos\left(\frac{\pi l}{N} (b-1/2)\right) = 0,
\label{eq:orthogonal}$$ so $$|{\langle{\psi_{0l},0}|} H_P {|{\psi_{00},0}\rangle}| = \frac{\sqrt{2}(N-1)}{N^2} \left| \cos\left(\pi l/2N\right) \right|$$ when $l \neq 0$. Similarly, when $k \neq 0$, $$|{\langle{\psi_{k0},0}|} H_P {|{\psi_{00},0}\rangle}| = \frac{\sqrt{2}(N-1)}{N^2} \left| \cos\left(\pi k/2N\right) \right|.$$
Finally, when $k, l \neq 0$, $$|{\langle{\psi_{kl},0}|} H_P {|{\psi_{00},0}\rangle}| = \frac{2}{N^2} \left| \sum_{a,b=2}^{N} \cos\left(\frac{\pi k}{N} (a-1/2)\right) \cos\left(\frac{\pi l}{N} (b-1/2)\right) \right|.$$ We use another orthogonality relationship: $$\begin{aligned}
0 = & \sum_{a,b=1}^{N} \cos\left(\frac{\pi k}{N} (a-1/2)\right) \cos\left(\frac{\pi l}{N} (b-1/2)\right) \\
= & \sum_{a,b=2}^{N} \cos\left(\frac{\pi k}{N} (a-1/2)\right) \cos\left(\frac{\pi l}{N} (b-1/2)\right)
+ \sum_{a=2}^{N} \cos\left(\frac{\pi k}{N} (a-1/2)\right) \cos\left(\frac{\pi l}{2N}\right) \nonumber \\
& \quad
+ \sum_{b=2}^{N} \cos\left(\frac{\pi l}{N} (b-1/2)\right) \cos\left(\frac{\pi k}{2N}\right)
+ \cos\left(\frac{\pi k}{2N}\right) \cos\left(\frac{\pi l}{2N}\right) \\
= & \sum_{a,b=2}^{N} \cos\left(\frac{\pi k}{N} (a-1/2)\right) \cos\left(\frac{\pi l}{N} (b-1/2)\right)
- \cos\left(\frac{\pi k}{2N}\right) \cos\left(\frac{\pi l}{2N}\right),\end{aligned}$$ where in the last line, we have used the orthogonality relationship (\[eq:orthogonal\]) for each of the two single sums. We then get $$|{\langle{\psi_{kl},0}|} H_P {|{\psi_{00},0}\rangle}| = \frac{2}{N^2} \left| \cos\left(\frac{\pi k}{2N}\right) \cos\left(\frac{\pi l}{2N}\right) \right|.$$
We now turn to considering ${\langle{\psi}|} H_P {|{\psi}\rangle}$. Let ${|{\psi'}\rangle} = \sum_{(k,l,m) \neq (0,0,0)} a_{klm} {|{\psi_{klm}}\rangle}$, so ${|{\psi}\rangle} = a_{000} {|{\psi_{000}}\rangle} + {|{\psi'}\rangle}$. Since $H_P$ is positive definite, $$\begin{aligned}
|{\langle{\psi}|} H_P {|{\psi}\rangle}| \geq & |a_{000}|^2 {\langle{\psi_{000}}|} H_P {|{\psi_{000}}\rangle} + 2\, {\rm Re} (a_{000} {\langle{\psi'}|} H_P {|{\psi_{000}}\rangle}) \\
= & |a_{000}|^2 \frac{|R| (N-1)^2}{pN^2} + 2\, {\rm Re} \sum_{(k,l,m) \neq (0,0,0)} a_{000}\, a^*_{klm} {\langle{\psi_{klm}}|} H_P {|{\psi_{000}}\rangle} \\
\geq & |a_{000}|^2 \frac{|R| (N-1)^2}{pN^2} - 2 \sum_{(k,l,m) \neq (0,0,0)} |a_{000}| |a_{klm}| |{\langle{\psi_{klm}}|} H_P {|{\psi_{000}}\rangle}| \\
\geq & |a_{000}| \frac{|R| (N-1)^2}{pN^2} \left[|a_{000}| - 2 \sum_{m>0} |a_{00m}| - \frac{2 \sqrt{2}}{N-1} \sum_{k>0, m \geq 0} |a_{k0m}| \right. \nonumber \\
& \qquad \left. - \frac{2 \sqrt{2}}{N-1} \sum_{l>0, m \geq 0} |a_{0lm}| - \frac{4}{(N-1)^2} \sum_{k,l > 0, m \geq 0} |a_{klm}| \right] \\
= & |a_{000}| \frac{|R| (N-1)^2}{pN^2} \left[|a_{000}| - X - Y - Z \right],\end{aligned}$$ where $$\begin{aligned}
X = & \frac{2 \sqrt{2}}{N-1} \sum_{k>0} |a_{k00}| + \frac{2 \sqrt{2}}{N-1} \sum_{l>0} |a_{0l0}| + \frac{4}{(N-1)^2} \sum_{k,l > 0} |a_{kl0}| \\
Y = & \sum_{m>0} \left[\frac{2 \sqrt{2}}{N-1} \sum_{k>0} |a_{k0m}| + \frac{2 \sqrt{2}}{N-1} \sum_{l>0} |a_{0lm}| + \frac{4}{(N-1)^2} \sum_{k,l > 0} |a_{klm}| \right]\\
Z = & 2 \sum_{m>0} |a_{00m}|.\end{aligned}$$
There are four cases:
- [**If $\mathbf{|a_{000}| - X - Y - Z \geq 1/2}$:**]{} Then $|a_{000}| > 1/2$ as well, so $|{\langle{\psi}|} H_P {|{\psi}\rangle}| \geq \frac{|R| (N-1)^2}{4pN^2} = \Omega(1/p)$.
- [**If $\mathbf{X = \Theta(1)}$:**]{} At least one of the sums $\frac{1}{N-1} \sum_{k>0} |a_{k00}|$, $\frac{1}{N-1} \sum_{l>0} |a_{0l0}|$, and $\frac{1}{(N-1)^2} \sum_{k,l > 0} |a_{kl0}|$ must also be $\Theta(1)$; the others may be smaller. These three sums are all averages, and at least one of the terms in an average must be greater than or equal to the mean value. Thus, there exists $(k_0,l_0) \neq (0,0)$ with $|a_{k_0 l_0 0}| = \Theta(1)$. But $${\langle{\psi}|} H_L + H_R {|{\psi}\rangle} = \sum_{k,l,m} |a_{klm}|^2 \lambda_{kl} \geq |a_{k_0 l_0 0}|^2 \Omega(1/N^2) = \Omega(1/N^2).$$
- [**If $\mathbf{Y = \Theta(1)}$:**]{} At least one of the sums $\frac{1}{N-1} \sum_{m>0} \sum_{k>0} |a_{k0m}|$, $\frac{1}{N-1} \sum_{m>0} \sum_{l>0} |a_{0lm}|$, and $\frac{1}{(N-1)^2} \sum_{m>0} \sum_{k,l > 0} |a_{klm}|$ must be $\Theta(1)$. We have $${\langle{\psi}|} H_L + H_R {|{\psi}\rangle} \geq \sum_{(k,l) \neq (0,0)} \sum_m |a_{klm}|^2 \Omega(1/N^2).$$ When $\sum |a_{klm}|$ is fixed (summed over some subset of $(k,l,m)$), we minimize the sum $\sum |a_{klm}|^2$ (over the same subset of $(k,l,m)$) by taking all $|a_{klm}|$ to be equal (for $(k,l,m)$ in the subset). In this particular case, we should choose $|a_{klm}| = \Theta(1/(p-1))$ for one of the sets $\{k>0, l=0, m>0\}$, $\{k=0, l>0, m>0\}$, or $\{k>0, l>0, m>0\}$ and we find $\sum_{klm} |a_{klm}|^2 = \Omega(N/p)$. (Or $\Omega(N^2/p)$ in the third case.) In any case, it certainly follows that ${\langle{\psi}|} H_L + H_R {|{\psi}\rangle} = \Omega(1/N^2)$.
- [**If $\mathbf{Z = \Theta(1)}$:**]{} This is the most interesting case. We note that $${\langle{\psi}|} H_L + H_R + H_B {|{\psi}\rangle} = \sum_m {\langle{\psi_m}|} H_L + H_R + H_B {|{\psi_m}\rangle},$$ with ${|{\psi_m}\rangle} = \sum_{k,l} a_{klm} {|{\psi_{klm}}\rangle}$. By Lemma \[lemma:highm\], $${\langle{\psi_m}|} H_L + H_R + H_B {|{\psi_m}\rangle} = \Omega(m'^2 |a_{00m}|^2/p^2 N^2 \log N),$$ where $m' = \min(m, p-m)$. Therefore, we wish to minimize $\sum_m m'^2 |a_{00m}|^2$ under the constraint that $\sum_{m>0} |a_{00m}| = \Theta(1)$. The minimum is achieved (as can be shown by Lagrange multipliers, for instance) when $|a_{00m}| = B/m'^2$. Since $\sum_m 1/m'^2 \leq 2 \zeta(2) = \pi^2/3$, it follows that $B = \Theta(1)$ and $\sum_m m'^2 |a_{00m}|^2 = \Omega(1)$. Thus, $${\langle{\psi}|} H_L + H_R + H_B {|{\psi}\rangle} = \Omega(1/p^2 N^2 \log N).$$
As it happens, case 2 cannot actually occur in the large $N$ limit, given the normalization constraint that $\sum |a_{klm}|^2 = 1$, but this is not very important since the gap is set by the fourth case.
Zero or Multiple ${|{x}\rangle}$ States
---------------------------------------
We now turn attention to the subspaces ${\mathcal{H}}_{a,b}$, with either $a$ or $b$ different than $1$. We can treat $H_V^L$ and $H_V^R$ separately, and show that $H_L + H_V^L$ is bounded below when $a \neq 1$, as is $H_R + H_V^R$ when $b \neq 1$. Since the remaining Hamiltonian terms $H_B$ and $H_P$ are non-negative, bounding $H_L + H_V^L$ and $H_R + H_V^R$ is adequate to give a lower bound on $H$. Since the two cases are identical, we focus on the left ring.
If $a=0$, $H_V^L$ has all eigenvalues equal to $V_1$ and $H_L$ is $0$. If $a > 1$, there are some states in ${\mathcal{H}}_{a,b}$ which have a negative eigenvalue for $H_V^L$. However, those states will have a positive eigenvalue for $H_L$, and we must choose $V_1$ and $V_2$ to give such states a positive overall energy.
Let us restrict attention to a particular value of $a$; then $H_V^L + (a-1)V_1$ is a non-negative operator, which allows us to apply the following lemma (see [@Kitaevbook] for a proof):
\[lemma:Kitaev\] Let $A_1$ and $A_2$ be nonnegative linear operators with null spaces $L_1$ and $L_2$. Suppose that $L_1 \cap L_2 = \{0\}$ and that no nonzero eigenvalue of $A_1$ or $A_2$ is less than $v$. Then $$A_1 + A_2 \geq v \sin^2 (\theta/2),$$ where $\theta = \theta(L_1, L_2)$ is the angle between $L_1$ and $L_2$.
For us, $A_1 = H_V^L + (a-1)V_1$ and $A_2 = H_L$. The null space $L_1$ of $A_1$ consists of superpositions of states with $a$ ${|{x}\rangle}$ states in the left ring with no two of those ${|{x}\rangle}$ states are adjacent. The smallest nonzero eigenvalue of $A_1$ is at least $V_2$. (It could potentially be higher than $V_2$, if $a$ is large enough so that all configurations contain multiple adjacent ${|{xx}\rangle}$ pairs in the left ring.) The null space $L_2$ of $A_2$ consists of uniform superpositions over all possible configurations of $a$ ${|{x}\rangle}$ sites on the left ring, with some state for the remaining qubit sites, shuffled around appropriately for different ${|{x}\rangle}$ configurations.
To find the lowest nonzero eigenvalue of $A_2$, we can recognize $H_L$ as the spin-$1/2$ ferromagnetic Heisenberg model on a chain of length $N$ with open boundary conditions by mapping ${|{x}\rangle}$ to spin up and a qubit state to spin down. In this case, we now consider the ${|{x}\rangle}$ states to represent particles hopping around on the chain. Unlike the usual ferromagnetic Heisenberg model, the background states with no particles have some structure, since the background states are actually qubits. However, the qubits are inert and cannot hop or otherwise interact by themselves; they can only switch places with a ${|{x}\rangle}$ particle as it hops. Therefore, the Hilbert space breaks up into invariant subspaces within which the state of the background sites is determined completely by the locations of the ${|{x}\rangle}$ particle states. Within a single such subspace, we have exactly the usual spin-$1/2$ ferromagnetic Heisenberg model, which is known to have gap $1 - \cos(\pi/N) = \Theta(1/N^2)$ [@Heisenberg]. (This gap considers all possible numbers of ${|{x}\rangle}$ states; in fact, the minimum gap is achieved when there is just one ${|{x}\rangle}$.)
The angle $\theta$ between $L_1$ and $L_2$ can be determined by simply noting that there are $\binom{N}{a}$ total possible ${|{x}\rangle}$ configurations, but at most $$\frac{N(N-2)(N-4)(N-6) \cdots (N-2a+2)}{a!} \leq \binom{N}{a} \left(1 - \frac{1}{N-1} \right)$$ configurations with no adjacent ${|{xx}\rangle}$ pairs. This formula comes from choosing the location of $a$ ${|{x}\rangle}$ sites, with each one chosen from possibilities excluding the sites of previously chosen ${|{x}\rangle}$ sites as well as the next higher-numbered site to any previously chosen site; then we divide by $a!$ to eliminate the redundancy due to choosing the sites in order. This of course is an overestimate, as we must also exclude the next lower-numbered site to any previously chosen site, but then the counting becomes complicated, as the site below one previously chosen site can be the same as the site above a previously chosen site. Since an element of $L_2$ consists of a superposition over all $\binom{N}{a}$ configurations and an element of $L_1$ can only contain configurations with no adjacent pairs, it follows that $\sin^2 \theta \geq \frac{1}{N-1}$.
Putting these numbers into lemma \[lemma:Kitaev\], we find that on the subspace with $a>1$, $H_L + H_V^L + (a-1)V_1$ has all eigenvalues $\Omega(1/N^3)$. If we choose $V_1 = \Theta(1/N^4)$ and $V_2 = 1$, then $H_L + H_V^L$ also has all eigenvalues $\Omega(1/N^3)$ on the subspace with $a>1$, while on the sector with $a=0$, $H_L$ will have eigenvalue $\Theta(1/N^4)$. Putting everything together, we find that the ground state is unique, and is the desired ground state ${|{g}\rangle}$ (with energy $0$), while the gap in the whole Hilbert space is $\Omega(1/N^4 \log N)$.
Entanglement of the Ground State
--------------------------------
The ground state of $H$ is thus $${|{g}\rangle} = \frac{1}{N} \sum_{a,b = 1}^{N} M_{a,b} {|{xx}\rangle}_{N}^{LR} \bigotimes_{i=1}^{N-1} {|{\Psi^-}\rangle}_i^{LR},$$ as noted before. Suppose we make a local measurement to each ring that measures the location of the ${|{x}\rangle}$ state but nothing else about the qubit states. The outcomes are uniformly distributed among all possible pairs $(a,b)$, and given a specific outcome, the residual state is $$M_{a,b} {|{xx}\rangle}_{N}^{LR} \bigotimes_{i=1}^{N-1} {|{\Psi^-}\rangle}_i^{LR}.$$ $M_{a,b}$ is just a tensor product of a local unitary operator $M_a$ on the left ring with a local unitary operator $M_b$ on the right ring. By performing $M_a^\dagger$ and $M_b^\dagger$, we can therefore convert the state to ${|{\Psi^-}\rangle}^{\otimes (N-1)}$. Since purely local operations convert ${|{g}\rangle}$ into $N-1$ EPR pairs, the original state has at least $N-1$ ebits of entanglement. That is, $S(\rho_L) = S(\rho_R) \geq N-1$. Since one ring has only $N$ $3$-state particles, the maximum possible entropy would be $N \log_2 3$, so we find that $S(\mbox{one ring}) = \Theta(N)$. If we rephrase this in terms of the energy gap, $\Delta = \Omega(1/N^4 \log N)$, we see that $S(\mbox{one ring}) = \Omega((-\Delta \log \Delta)^{-1/4})$, as claimed.
Discussion
==========
The explicit construction in this paper shows that it is possible to have the entanglement entropy scale polynomially with the inverse gap. On the other hand, it was proven previously that the entanglement entropy cannot scale faster than exponentially in the inverse gap. A tight bound remains open.
One route to a tight bound might be to improve the Lieb-Robinson bound. This bound is a key ingredient in proving upper bounds on entanglement entropy, but our construction suggests that the Lieb-Robinson bound does not fully capture the locality properties of Hamiltonian dynamics. Let us explain by an example: consider a system of free fermions on an infinite line in one dimension, with Hamiltonian $\sum_i \Psi_{i+1}^{\dagger} \Psi_i + h.c.$ Consider the operator equations of motion for the operator $\Psi^{\dagger}_0$. The Lieb-Robinson bound says that $\exp[i H t] \Psi^{\dagger}_0 \exp[-i H t]$ can be well-approximated by an operator supported just on the set of sites within distance $v_{LR} t$ of $0$, where $v_{LR}$ is the Lieb-Robinson velocity. However, at least for this system, we can make a sharper statement — there is no good approximation by an operator supported on too [*few*]{} sites. The exact result for this operator is: $$\exp[i H t]\Psi^{\dagger}_0 \exp[-i Ht]=\sum_x A(x,t) \Psi^{\dagger}_x,$$ where $$A(x,t)\equiv \int \frac{{\rm d}k}{2\pi} \exp[2i\cos(k t)] \exp(i kx).$$ One can check from the asymptotics of the Bessel function that for $x>>2t$ the amplitude $A(x,t)$ is exponentially small, in agreement with the Lieb-Robinson bound. However, the amplitude $A(x,t)$ is symmetric under a change in sign from $x \rightarrow -x$. That is, the operator $ \Psi^{\dagger}_0$ creates both left and right moving excitations. Further, while the amplitude is maximum at $x$ close to $\pm 2t$, there is a non-vanishing spread in the width of $|A(x,t)|^2$: the particle is not perfectly localized at $x=\pm 2t$, but instead has a spread in its position, indicating an uncertainty in the initial velocity of the particle. It is possible to create a wavepacket that moves to the right, but in order to do this we need an operator that creates the particle in a superposition of different sites: $\sum_i a(i) \Psi^{\dagger}_i$ for some function $a(i)$. If $a(i)$ has finite support, then there will always be some spread in the velocity of the wavepacket, and in fact there will be some left-moving component to the wavepacket. Thus, it will not be possible to find, at least in this model, a choice of $a(i)$ which has finite support for which the time evolved operator $\exp[i H t] \Bigl(\sum_i a(i) \Psi^{\dagger}_i \Bigr) \exp[-i H t]$ is approximately localized on a set of sites with a bounded, time-independent diameter. The Lieb-Robinson bound does not capture this effect, but it seems to be a real effect for every Hamiltonian that we can think of: operators with finite support create both left- and right-moving excitations.
A similar phenomenon afflicts our construction. We had to go to some effort to keep the two rings synchronized because of the spread in the velocity of the $|x\rangle$, which, if left unchecked, could wash out the entanglement between the two rings. To do this, we only allowed the $|x\rangle$s to hop from site $1$ to site $N$ in pairs. The fact that the term $H_B$ only has effect when both $|x\rangle$s are in the correct position is one of the reasons that the gap is so small. However, if we had a way of just creating an $|x\rangle$ excitation that moved to the right around the ring at a constant velocity equal to $v_{LR}$, it might be easier to synchronize the rings, and thus to increase the gap for the same entanglement. Hopefully, then, this work will lead to a better Lieb-Robinson bound that captures these effects, and this in turn will lead to stronger bounds on entanglement in Hamiltonian systems.
Acknowledgements {#acknowledgements .unnumbered}
----------------
We thank D. Aharonov for many useful discussions throughout this work. DG was supported by CIFAR, by the Government of Canada through NSERC, and by the Province of Ontario through MRI. MBH was supported by U. S. DOE Contract No. DE-AC52-06NA25396.
[00]{}
J. Eisert, M. Cramer, and M. B. Plenio, “Area laws for the entanglement entropy - a review,” arXiv:0808.3773v1 \[quant-ph\].
M. B. Hastings, “An Area Law for One Dimensional Quantum Systems,” JSTAT, P08024 (2007), arXiv:0705.2024v2 \[quant-ph\].
M. B. Hastings, Phys. Rev. B [**76**]{}, 035114 (2007).
A. Ben-Aroya and A. Ta-Shma, arXiv:quant-ph/0702129.
M. B. Plenio, J. Eisert, J. Dreissig, and M. Cramer, Phys. Rev. Lett. [**94**]{}, 060503 (2005).
F. Verstraete and J. I. Cirac, Phys. Rev. B [**73**]{}, 094423 (2006).
S. R. White, Phys. Rev. Lett. [**69**]{}, 2863 (1992).
G. Vidal, Phys. Rev. Lett. [**93**]{}, 040502 (2004).
E. H. Lieb and D. W. Robinson, Commun. Math. Phys. [**28**]{}, 251 (1972).
M. B. Hastings, Phys. Rev. B [**69**]{}, 104431 (2004).
M. B. Hastings and T. Koma, Commun. Math. Phys. [**265**]{}, 781 (2006).
B. Nachtergaele and R. Sims, Commun. Math. Phys. [**265**]{}, 119 (2006).
P. Calabrese and J. Cardy, J. Stat. Mech., P06002 (2004).
H. Casini and M. Huerta, J. Stat. Mech., P12012 (2005).
D. Aharonov, D. Gottesman, S. Irani, and J. Kempe, “The power of quantum systems on a line,” Proc. 48th IEEE Symposium on the Foundations of Computer Science (FOCS), 373–383 (2007), arXiv:0705.4077 \[quant-ph\].
S. Irani, “Ground States Entanglement in One Dimensional Translationally Invariant Quantum Systems,” arXiv:0901.1107 \[quant-ph\].
A.Y. Kitaev, A.H. Shen, and M.N. Vyalyi, [*Classical and Quantum Computation*]{}, AMS, Providence, RI (2002).
T. Koma and B. Nachtergaele, “The spectral gap of the ferromagnetic $XXZ$ chain,” Lett. Math. Phys. [**40**]{}, no. 1, 1–16 (1997).
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the low energy phenomenology of superstrings. In particular we analyse supersymmetry breaking via gaugino condensate and we compare the phenomenology of the two different approaches to stabilize the dilaton field. We study the cosmological constant problem and we show that it is possible to have supersymmetry broken and zero cosmological constant. Finally, we discuss the possibility of having an inflationary potential. Requiring that the potential does not destabilize the dilaton field imposes an upper limit to the density fluctuations which can be consistent with the COBE data.'
---
-1cm Ø
hep-ph/9501322\
IFUNAM-FT-95-71\
Sept. 1994
[**Phenomenology of Superstrings[^1]$^,$[^2]** ]{}\
\
\
\[8mm\]
[**INTRODUCTION**]{}
Superstrings offers the exciting possibility of predicting all the parameters of the standard model in terms of a single parameter, the string tension. However in order to realize the full predictive power of the superstring it is necessary to determine the origin and effects of supersymmetry breaking. Only after SUSY is broken are the vacuum expectation values (vevs) of moduli determined and these determine the couplings of the effective low energy theory. Also SUSY breaking must be responsible for the splitting of supermultiplets allowing for the superpartners to be heavier than their standard model partners.
The dilaton field $S$ plays a crucial role since it interacts with all scalar fields and has a generic interaction. In the context of gaugino condensate it is the dilaton field that sets the mass hierarchy. Its auxiliary field may be responsible for breaking SUSY in which case the soft supersymmetric breaking terms are universal. Furthermore, the dynamics of the dilaton field does not allow for the scalar potential $V$ to inflate , and therefore $S$ must be at its minimum before the universe expands rapidly. Clearly, a potential must be positive to inflate. Is it then possible to have $S$ stable and $V > 0$ ?
Due to lack of space we will just give a short presentation of the different possibilities to stabilize the dilaton field and a discussion of some phenomenological consequences, vanishing of the cosmological constant and inflation. Unfortunately, we will not be able to talk about many interesting topics like $S$ duality, fermion masses, the strong CP problem, discrete and accidental symmetries and the phenomenology of light scalars and axions.
[**DILATON FIELD AND SUSY BREAKING** ]{}
In the absence of non-perturbative effects, the dilaton field interacts with all scalar fields with an $1/S$ interaction, and the scalar potential does not have a stable solution. There are several possibilities to stabilize the dilaton. Firstly, one can impose an $S$-duality (analogous to the $T$ dual symmetry) invariance to the potential. Another possibility is to consider gaugino condensation. Gaugino condensation offers a very plausible origin for SUSY breaking for it is very reasonable to expect such a condensate to form at a scale between the Planck scale and the electroweak breaking scale if the hidden sector gauge group has a (running) coupling which becomes large somewhere in this domain. Non-perturbative studies in effective supergravity theories resulting from orbifold compactification schemes suggest the dynamics of the strongly coupled gauge sector is such that the gaugino condensate will form and trigger supersymmetry breaking.
Using symmetry and anomaly cancelation arguments one derives an effective superpotential for the gaugino condensate $<\ov{\lam}_L\lam_R>$ in terms of $S$ $$W_0= d(T) \,e^{-3\,S/2\,b_0} \simeq \Lambda_c^{3}$$ where $\Lambda_{c}$ is the condensation scale. The scalar potential is given by $V_0= e^{K}|W_0|^2\,\,[(1+\frac{3\,S_r}{2b_0})^2 -3]=
|<\ov{\lam}_L\lam_R>|^2\,\,\frac{b_0^2}{36}\,\,
[(1+\frac{3\,S_r}{2b_0})^2 -3]
$ and it does not have a stable solution. There are two different approaches to stabilize the potential:
\(I) Consider two gaugino condensates and chiral matter fields with non-vanishing v.e.v. and slightly different one-loop beta function coefficients $b_0^1\simeq b_0^2$ with a superpotential $$W_0=d_1\,e^{-3\,S/2\,b_0^1}-d_2\,e^{-3\,S/2\,b_0^2}.$$ A stable solution is found for vanishing auxiliary field of the dilaton $G_S=W_S-W/S_r \simeq \frac{\partial W_0}{\partial S}=0$. SUSY will then be broken by the auxiliary field of the moduli field $G_T$.
\(II) Consider loop corrections of the 4-Gaugino interaction “à la N-J-L” using the Coleman-Weinberg one-loop potential $V_1$. A stable solution is found for $V=V_0+V_1$ with a single gaugino condensate . The leading contribution to $V_1$ is given by the gaugino mass $m_g$ and since $m_g^2/\Lambda_c^2 << 1$ one has $V_1\simeq
-\frac{n_g}{32\pi^2}\Lambda_{c}^{2}m_g^{2} $ where $n_g$ is the dimension of the hidden gauge group. The scalar potential $V=V_{0}+V_{1}$ can then be written as V&&e\^[K]{}\
V&&e\^[K]{}with $A\equiv h_S \bar h_T - 3\bar W (F_S+F_T) + h.c.,
\; h=S_r G_S= S_r W_S-W=F_S W,\; h_T=\sqrt{3} \,T_r G_T=F_T W,
\;
F_S=-(1+\frac{3S_r}{2b_0}) \gg 1,\, F_T=\sqrt{\frac{3
T_r^2}{4\pi^2}}\, \hat
G_2(T) $ and $ \delta \equiv\frac{n_g b_0^2}{144 \pi^2} \ll 1$. We recover the tree level potential by setting $\delta=0$.
[ **Results**]{}
Let us now compare the results obtained by minimizing the scalar potential in the case of two gaugino condensates (I) and for the case of one gaugino condensate (II). In both cases a large hierarchy can be obtained.
\(I) 2 gaugino condensates (II) 1 gaugino condensate
[|ll|ll|]{} $<S> \simeq $&$ 0.17 \frac{N_2M_1-N_1M_2}{(3N_2-M_2)(3N_1-M_1)}
$&$<S> \simeq
$&$\frac{4\pi}{\sqrt{n_g}}$\
$<T> \simeq $&$1.2 $&$ <T> \simeq
$&$\frac{3\,S_r}{2\,\pi\,b_o\,(1-\alpha_0) }\simeq O(10-20) $\
$m_{3/2}=$ & $ O(1 )TeV $&$ m_{3/2}= $ & $
O(1) TeV$\
$G_S=0, $ & $ G_T\neq 0 $ & $ G_S \gg$ & $G_T
$\
where $b_i=\frac{1}{16\pi^2}(3N_i-M_i)$, $\alpha_0$ is related to the number and weight of the hidden sector fields (for an orbifold with untwisted fields only $\alpha_0=-1/3$) and $G_S, G_T$ are the auxiliary fields of the dilaton and moduli fields respectively. All the parameters are related to the normalization and number of fields of the hidden sector and are fixed for a given compactification scheme. Note that the v.e.v. of the moduli in case (II) are much larger than in case (I).
The phenomenology depends strongly on which auxiliary field breaks SUSY and in case (I) SUSY is broken due to the auxiliary field of the moduli $G_T$ while in case (II) it is mainly due to the auxiliary field of the dilaton $G_S
\gg G_T$. The soft supersymmetric breaking terms are universal if SUSY is broken via $G_S$ while they differ if SUSY is broken via $G_T$ and they have been calculated in ,. Experimental evidence on the neutron dipole momenta show that the scalar masses must be almost degenerated ($(m^2_1-m^2_2)/m^2 < 10^{-2}-10^{-3}$).
[ **UNIFICATION SCALE AND COUPLING**]{}
We will, now, discuss the unification scale and coupling. The fine structure constant at the unification scale is $
\alpha_X^{-1}\simeq \frac{4\pi}{ g^2_{gut}}\simeq 4\pi Re\,S
$ and using the solutions of minimization for case (II) we have \_X\^[-1]{}. Consistency with MSSM unification requires then $33 < n_g < 44$ and this is satisfied only for the gauge groups $SU(6)$ or $SO(9)$[^3]. In case (I) there are more possibilities to obtain a fine structure constant required by MSSM unification and the gauge group is therefore not constraint. However, MSSM unification also imposes constraint on the value of the unification scale. The unification scale $M_X$ is a moduli dependent function with the property to be close to the string scale for $T\simeq 1$. On the other hand if $T$ is larger then there is the possibility of having $M_X\simeq 10^{16}$ as required . As an example we can take an $SU(6)$ with $b_0=15/16\,\pi^2$ for which $T=22$, the unification fine structure constant and scale are $\alpha_X^{-1}= 26.1, M_X= 2.8
\times10^{16}\,GeV$.
[**COSMOLOGICAL CONSTANT**]{}
The vanishing of the cosmological constant is an important and still open problem. Experimental evidence shows that the cosmological constant is very small and it is not clear how to implement it a natural scheme. Another approach, is to study the possibility of having a potential with vanishing cosmological constant by introducing new terms and fine tuning them. In non-supersymmetric models this represents no problem. However, in SUSY potentials the possible terms are constraint. In fact, for global supersymmetry it is not possible, if one requires SUSY to be broken (spontaneously or explicitly). On the other hand, in sugra models one has, in principle, the possibility of having V=0 and SUSY spontaneously broken (SB). The breaking of SUSY is a necessary condition but for the simplest potentials if a symmetry is SB the vacuum energy will then be proportional to the symmetry breaking scale ($\Lambda$), $V=-O(\Lambda^4)$. For realistic hierarchy solution $V\simeq -(10^{-12})^{4}$ which is many orders of magnitude larger than the observational upper limit $|V|
<
10^{-120}$. In the context of supergravity models, the canceling of the cosmological constant must come trough a non-vanishing value of an auxiliary field $G_i \neq 0$.
The condition of zero cosmological constant, considering the tree level potential only, is $
G_{a} (K^{-1})^{a}_{b} G^{b} =3|W|^{2}$ but it is hard to satisfy dynamically. Imposing $T$-duality symmetry and assuming, for simplicity, that the $T$ dependent part of the superpotential can be factorized we have $W=\eta(T)^{-6} \O (S,\pp)$ with $\O=\O_0(S)+\O_{ch}(\pp)$ and $\O_0$ the contribution from the gaugino condensates while $\O_{ch}$ the contribution from the chiral matter fields. The scalar potential becomes V\_0= e\^[K]{}||\^[-12]{} where $\hat{G}_2$ is the Eisenstein function of modular weight 2, $h=S_r \O_S
-\O$ and $ k\equiv K_i\O+\O_i.$
To find the vacuum state with zero cosmological constant one needs to solve the eqs. $V|=V_S|=V_T|=V_i|=0$ where “$|$” denotes that the quantities should be evaluated at the minimum. $V|=V_T|=0$ is satisfied for $T$ at the dual invariant points ($T=1,e^{-\pi/6}$) where $\hat{G}_2=0$. This implies that the auxiliary field of the moduli is zero, $G_T=0$, and it does not break SUSY contrary to case (I) where the condition $V|=0$ was not imposed. The cancelation of the cosmological constant must then be due to $h$ or $k$. In the absence of $k$, for the two gaugino condensates case, the solution to $V_S=0$ is $h=0$ and therefore the condition $V|=0$ must be due to $k$. However, if all superpotential terms $\O_{ch}$ are at least quadratic in $\pp_i$ then $k=0$ for $\pp_i=0$. The only possibility to have $k\neq 0$ is with a linear superpotential $\O_{ch} =c \phi$, where $c$ is an arbitrary constant to be fine tuned to give $V|=0$. Let us take the example $N_1=6, M_1=0,
N_2=7, M_2=6$ . For this example one obtains a large hierarchy and $S=2.16$ if $k=0$ . The numerical solution to $V|=V_S|=V_\pp|=0$ is $S=2.15, \;c=1.2 \times 10^{-15},\;\pp=-0.5 $ corresponding to a stable solution. We note that the variation of $S$ is quite small.
We have thus seen that it is possible to cancel the cosmological constant using the tree level sugra scalar potential. SUSY is also broken but mainly due to the auxiliary field $k=G_\pp$ since $G_T=0$ and $G_S \approx 0$. Unfortunately, most phenomenological terms depend on how SUSY is broken and in this case it is broken via the term which we now least and was introduced with the only motivation of rendering $V|=0$.
If SUSY is broken via a single gaugino condensate, i.e. case (II), one can use the same linear superpotential and the cosmological constant may be arranged to vanish at the minimum. The welcome difference in this case is that SUSY is mainly broken by the auxiliary field of the dilaton $G_S$.
[**INFLATION** ]{}
String models are valid below the Planck scale and it should therefore describe the evolution of the universe. The standard big bang theory has some shortcomings like the horizon and flatness problems. An inflationary epoch, where the universe expanded in an accelerated way, may solve this problems. For arbitrary values of the different fields one expects V to be positive and to evolve to its minimum. In this evolution one would hope for an inflationary period. However, it is difficult to obtain an inflationary potential in string models due to the dynamics of the dilaton field $S$ .
The interaction of the dilaton field is very much constraint and the superpotential $W$ is independent of $S$ perturbatively but it may acquire a non-trivial superpotential non-perturbatively like when gauginos condense. Even in the presence of the non-perturbative superpotential when the scalar potential $V$ evolves to the minimum of the dilaton field, the universe, keeping all other fields fixed, does not go trough an inflationary period. At the minimum, SUSY is broken and for vanishing v.e.v. of the chiral fields, the vacuum energy is negative and of the order of $\Lambda^4$ but as we have seen in the previous section it is possible to have SUSY broken with vanishing cosmological constant. However, in string theory there are many chiral matter fields and its potential may drive an inflationary potential . The condition that these potential terms do not destabilize the dilaton field yields some strong constraint on the magnitude of these terms. Nevertheless, it is still possible to have a potential that inflates enough to solve the horizon and flatness problem. The constraint on the magnitude of these potential terms sets un upper limit on the density fluctuations which is of the order of magnitude as the observed by COBE .
A possible picture is that of a universe that starts with random values of the different fields (dilaton, moduli, chiral matter fields). The universe cools down and it evolves in a standard non-inflationary way until $S$ and $T$ are stabilized. Below this scale, other fields, like the chiral matter fields, could drive an exponentially fast expansion of the universe as long as its potential does not destabilize $S$ and $T$. So, we expect that the universe arrives at an inflationary period naturally when the fields roll down to their minimum and the inflationary conditions are first met.
[**ACKNOWLEDGMENTS** ]{}
I would like thank G.G. Ross and S. Lola for many useful discussions and comments.
[99]{}
For a review see D. Amati, K. Konishi, Y. Meurice, G. Rossi and G. Veneziano, Phys. Rep 162 (1988) 169; J.P.Derendinger, L.E.Ibanez and H.P.Nilles, Phys. Lett. B155 (1985) 65; M.Dine, R.Rohm, N.Seiberg and E.Witten, Phys. Lett. B156 (1985) 55; A. Font, L. Ibanez, D. Lust and F. Quevedo, Phys. Lett. B245 (1990) 401.
R. Brustein and P. J. Steinhardt, Phys. Lett. B302 (1993) 196; P. Binetruy and M. K. Gaillard, Phys. Rev. D34 (1986) 3069.
A. de la Macorra and S. Lola, hep-ph/9411443, IFUNAM -FT-94-63 Mexico preprint, HD-THEP-94-45 Heidelberg preprint.
A. Font, L. E. Ibanez, D. Lüst and F. Quevedo, Phys. Lett. B249 (1990) 35.
B. de Carlos, J. A. Casas and C. Munoz, Phys. Lett. B263 (1991) 248 and ref. therein.
A. De La Macorra and G. G. Ross, Nucl. Phys. B404 (1993) 321; Phys. Lett. B325 (1994) 85.
B. de Carlos, J. A. Casas and C. Munoz, Phys. Lett. B 299 (1993) 234.
A. de la Macorra, [*Unification Scale in String Theory*]{}, Oxford preprint OUTP-93-33P, hep-ph/9401239, to appear in Phys. Lett B.
J. Ellis, S. Kelley and D.V. Nanopoulos, Phys. Lett. B249 (1990) 441; Phys. Lett. B260 (1991) 131; U. Amaldi, W. de Boer and H. Fürstenau, Phys. Lett. B260 (1991) 447; P. Langacker and M. Luo, Phys.Rev.D44 (1991) 817.
A. de la Macorra [*“Vanishing of the Cosmological Constant, Stability of the Dilaton and Inflation”*]{} Mexico preprint IFUNAM-FT-94-64 hep-ph/9501250; A. de la Macorra and G.G. Ross, “[*Supersymmetry Breaking in 4D String Theory*]{}”, (preprint OUTP-31P);
[^1]: Invited talk given at the general meeting of the Canadian, American and Mexican Physics Society “CAM 94”, Cancun, Mexico.
[^2]: To be published by [*AIP Press*]{}
[^3]: Considering only $SU(N)$ and $SO(N)$ gauge groups.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the collapse of an attractive atomic Bose-Einstein condensate prepared in the uniform potential of an optical-box trap. We characterise the critical point for collapse and the collapse dynamics, observing universal behaviour in agreement with theoretical expectations. Most importantly, we observe a clear experimental signature of the counterintuitive weak collapse, namely that making the system more unstable can result in a smaller particle loss. We experimentally determine the scaling laws that govern the weak-collapse atom loss, providing a benchmark for the general theories of nonlinear wave phenomena.'
author:
- 'Christoph Eigen, Alexander L. Gaunt, Aziza Suleymanzade, Nir Navon, Zoran Hadzibabic, and Robert P. Smith'
title: 'Observation of Weak Collapse in a Bose-Einstein Condensate'
---
Introduction
============
Wave collapse occurs in a wide range of physical contexts, including optics, atomic and condensed-matter physics. Generally, collapse occurs if an attractive nonlinearity exceeds a critical value. If the collapse is triggered at time $t=0$, the wave amplitude asymptotically diverges at some point in space as the collapse time $t_{\rm c}$ is approached. In practice, the amplitude divergence results in dissipation of wave energy (or particle loss).
The unifying theoretical framework for understanding different collapse phenomena is provided by the nonlinear Schr[ö]{}dinger equation, which has been extensively studied for various forms of nonlinearity [@Sulem:1999; @Fibich:2015]. This general formalism is applied to self-focusing of light [@Askaryan:1962; @Chiao:1964; @Kelly:1965; @Pilipetskii:1965; @Hercher:1964], collapse of Langmuir waves [@Zakharov:1972; @Wong:1984] and Bose-Einstein condensates (BECs) [@Ruprecht:1995; @Kagan:1996a; @Kagan:1997; @Kagan:1998; @Eleftheriou:2000], and even surface water waves [@Davey:1974; @Papanicolaou:1994].
In this framework, wave collapse is classified as either [*strong*]{} or [*weak*]{} (see Fig. \[fig:Cartoon\]). In a strong collapse, a finite fraction of the wave collapses towards the singularity. On the other hand, in a weak collapse [@Zakharov:1975; @Zakharov:1985; @Zakharov:1986] the fraction of the wave that (in absence of dissipation) ultimately reaches the singularity vanishes. This has the counterintuitive practical implication that making the system [*more unstable*]{}, by quenching the nonlinearity further beyond the critical point, can result in [*less*]{} dissipation [@Zakharov:1986; @Berge:2002]. Qualitatively, once the collapse is triggered, for stronger attractive interactions it happens faster and progresses further before dissipative processes halt it; consequently the wave amplitude is larger at the point in time when dissipation occurs, and for weak collapse this means that a smaller fraction of the wave is actually dissipated. To our knowledge, weak collapse has not been experimentally observed in any physical system.
An atomic BEC with $s$-wave two-body interactions is modelled by the Gross-Pitaevskii (GP) equation, with a cubic nonlinearity proportional to the scattering length $a$, which can be dynamically tuned via a Feshbach resonance [@Chin:2008]. The BEC is prone to collapse for any $a<0$, but a kinetic-energy barrier makes it metastable up to a critical interaction strength [@Ruprecht:1995; @Kagan:1996a; @Kagan:1997; @Kagan:1998; @Eleftheriou:2000]. If the BEC becomes unstable and collapses, dissipation occurs through three-body recombination that results in particle loss. Importantly, the three-dimensional GP equation is expected to provide an example of weak collapse.
Previous collapse experiments with atomic BECs [@Gerton:2000; @Roberts:2001; @Donley:2001; @Cornish:2006; @Altin:2011; @Compton:2012] (see also [@CollapseFootnote1]) were performed in the traditional setting of a harmonic trap. The critical point [@Roberts:2001] and collapse times [@Donley:2001; @Altin:2011] were in general agreement with theoretical expectations [@Ruprecht:1995; @Santos:2002; @Saito:2002; @Adhikari:2002; @Savage:2003; @Metens:2003; @Ueda:2003; @Carr:2004; @Wuster:2005], but no evidence of weak collapse was observed; the atom loss was only seen to grow with $|a|$ [@Cornish:2006].
\[btp\] ![Strong versus weak collapse (cartoon). In strong collapse, a finite portion of the wave (here 100%, for simplicity) collapses towards the singularity. In weak collapse, as time progresses, a diminishing fraction of the wave approaches the singularity, with long tails left behind. []{data-label="fig:Cartoon"}](Figure1.pdf "fig:"){width="1\columnwidth"}
In this article we study BEC collapse in a new experimental setting, using a $^{39}$K condensate [@Roati:2007; @Campbell:2010] prepared in the uniform potential of an optical-box trap [@Gaunt:2013]; for details of our setup see Appendix A. The combination of large system size (up to $41\,\mu$m) and fine tuning of the scattering length (with a resolution of $0.03\,a_0$, where $a_0$ is the Bohr radius) gives us a very large dynamic range: we observe metastable attractive BECs with up to $2\times 10^5$ atoms, and collapse times that vary between $3$ and $400$ ms. We demonstrate the expected scaling of the critical scattering length $a_{\rm c}$ with the BEC atom number $N$ and the system size $L$, and show that the collapse time can be expressed as a universal function of the dimensionless interaction strength $aN/L$. Most importantly, we observe conclusive evidence for weak collapse, namely the counterintuitive decrease of the atom loss with increasing $|a|$, and experimentally determine the scaling laws that govern the weak-collapse atom loss. The weak nature of the collapse is directly revealed only by resolving single collapse events, and is obscured in the multiple collapse regime, which has been seen in previous cold atom experiments.
In Sections \[sec:CriticalPoint\]-\[sec:WeakCollapse\] we address in turn: (i) the [*critical point*]{} for the collapse, (ii) the [*collapse dynamics*]{} in a system that is suddenly made unstable by an interaction quench, and (iii) the [*aftermath*]{} of the collapse, which reveals its weak nature.
Critical Point {#sec:CriticalPoint}
==============
The starting point for our discussion is the GP equation for a homogeneous box potential, with a heuristically added three-body loss term [@Kagan:1998]: $$i \hbar \frac{\partial \psi}{\partial t}=-\frac{\hbar^2}{2m}\nabla^2\psi+ \frac{4\pi \hbar^2 a}{m} |\psi|^2\psi - i \frac{\hbar K_3 }{2}|\psi|^4\psi \, ,
\label{eq:gp}$$ where $m$ is the atom mass, $K_3$ is the three-body loss coefficient, $\psi$ is normalised to the atom number $N$, and the boundary condition is $\psi = 0$ at the trap walls.
We use a cylindrical box trap [@Gaunt:2013] of variable length $L$ and radius $R$, and always set $R=L/2$, so $L$ is the only lengthscale characterising the system size. We may thus rewrite Eq. (\[eq:gp\]) in a dimensionless form, defining $\mathbf{\tilde{r}} = \mathbf{r}/L$ and $\tilde{t} = t / \tau_0$, with characteristic time $\tau_0=2m L^2/\hbar$: $$i \frac{\partial \tilde{\psi}}{\partial \tilde{t}}=-\nabla^2\tilde{\psi}+\alpha|\tilde{\psi}|^2\tilde{\psi}-i\eta |\tilde{\psi}|^4\tilde{\psi} \, ,
\label{eq:rgp}$$ where $$\alpha= \frac{8\pi a N}{L}\quad \textrm{and} \quad \eta= \frac{N^2 mK_3}{\hbar L^4} \, ,
\label{eq:dimcoef}$$ and $\tilde{\psi}$ is initially normalised to unity. For the range of scattering lengths that we study, we assume that $K_3$ is constant [@Shotan:2014; @Lepoutre:2016], with the value $1.3 (5) \times 10^{-41}$m$^{6}$s$^{-1}$ [@Fattori:2008]. The corresponding value of $\eta$, for all our $L$ and $N$ values, is very small ($<6 \times 10^{-4}$) and thus three-body loss is negligible in our (meta)stable condensates. However, if the BEC collapses, significant loss occurs, providing the primary experimental signature of the collapse.
Neglecting the atom loss in a metastable BEC, based on Eq. (\[eq:rgp\]) the critical interaction strength, $\alpha_{\rm c}$, can depend only on the boundary conditions, [*i.e.*]{} the box shape. For a family of self-similar boxes ($R/L=$ const.) it should be a universal constant, so $a_{\rm c} \propto L/N$.
To experimentally study the critical point for collapse, we prepare a stable BEC at $4\,a_0$, then over $1$ s ramp the scattering length to a variable $a<0$, and wait for $2$ s before turning off the trap and imaging the atoms after $100$ ms of time-of-flight (ToF) expansion. We image the cloud along the axial direction of our cylindrical trap, and for ToF we jump the scattering length to $20\,a_0$.
In Fig. \[fig:ac\](a) we show how, for a given initial $N$, the final atom number depends on the negative $a$. A well defined $a_{\rm c}$ is signaled by a sharp drop in the atom number. As shown in Fig. \[fig:ac\](b), the atom loss is accompanied by a qualitative change in the appearance of the cloud in ToF.
\[tbp\] ![Critical point for collapse. (a) Final atom number, $N_{\rm f}$, after ramping to a variable negative $a$, for $L=30(1)~\mu$m and initial $N = 18.7(5) \times 10^4$. The critical scattering length $a_{\rm c}$ is seen in the sharp drop in $N_{\rm f}$. (b) ToF images taken on either side of the critical point. (c) Variation of $a_{\rm c}$ with $N$, for $L=30~\mu$m. The linear fit confirms the expected scaling $a_{\rm c} \propto 1/N$. (d) Variation of $Na_{\rm c}$ with $L$. The linear fit confirms the scaling $N a_{\rm c} \propto L$. []{data-label="fig:ac"}](Figure2.pdf "fig:"){width="1\columnwidth"}
In Fig. \[fig:ac\](c) we plot $a_{\rm c}$ for $ L=30$$\mu$m and a wide range of $N$ values, from $10^4$ to $2\times 10^5$. We clearly observe the expected scaling $a_{\rm c} \propto 1/N$ (see also Appendix B). In Fig. \[fig:ac\](d) we plot the measured $N a_{\rm c}$ versus box size and confirm the scaling $Na_{\rm c} \propto L$. We find that the dimensionless critical interaction strength is $\alpha_{\rm c}= - 4(1)$, where the error includes the systematic uncertainties in box size and absolute atom number calibration. For comparison, numerical simulations of the GP equation for our box geometry give $\alpha_{\rm c}= - 4.3$.
Collapse Dynamics {#sec:Dynamics}
=================
To study the collapse dynamics, we perform interaction-quench experiments [@Donley:2001]. We prepare a BEC just above $a_{\rm c}$ and then quench the scattering length to a variable $a < a_{\rm c}$ to initiate the collapse (see Appendix B for more details). After a variable hold time $t$ we jump the scattering length to $20\,a_0$, switch off the trap, and observe the cloud in ToF.
As shown in the left panel of Fig. \[fig:jump\](a), for quenches close to the critical point (small $|a-a_{\rm c}|$), at $t_{\rm c}$ the atom number suddenly drops to a stable lower value. We understand this as a single collapse event. On the other hand, for large quenches \[right panel of Fig. \[fig:jump\](a)\], the atom number appears to gradually decay until it stabilises. Such behaviour, also seen in [@Donley:2001; @Altin:2011], is understood as arising from a series of multiple (experimentally unresolved) collapses [@Malkin:1988; @Vlasov:1989; @Zakharov:1989; @Kagan:1998; @Saito:2001a; @Saito:2001b; @Saito:2002; @Santos:2002; @Berge:2002], and we accordingly associate $t_{\rm c}$ with the onset of the atom loss [@Definetc]. (For further evidence for the occurrence of single and multiple discrete collapse events see Appendix C.)
![Collapse dynamics. (a) Atom number versus time after quenches to $a=-0.86\,a_0$ (left) and $-2.19\,a_0$ (right); here $L=30\,\mu$m and $N=11.4 \times10^4$, corresponding to $a_{\rm c} = -0.79\,a_0$. Green bands indicate $t_{\rm c}$ and its uncertainty. (b) Typical ToF images at various stages after the quench (here for $a=-1.02\,a_0$). (c) Collapse time versus $a$ for six data sets taken for various $L$ and $N$; see legend in (d). The shaded bands indicate $a_{\rm c}$ values. (d) Universal collapse dynamics. We plot the dimensionless collapse time versus the reduced distance from the critical point, for all six data sets. The solid line shows the results of lossless GP simulations without any free parameters. []{data-label="fig:jump"}](Figure3.pdf){width="1\columnwidth"}
In Fig. \[fig:jump\](b) we show typical ToF images for different times after the quench. At $t < t_{\rm c}$, before any change in the atom number occurs, the swelling of the cloud in ToF reveals the shrinking of the wavefunction in-trap. Right after $t_{\rm c}$, within the first ${\approx}\,10$ ms, we observe that the remnant cloud consists of a lower-energy central part and a higher-energy shell, reminiscent of the atom bursts generated during collapse in [@Donley:2001]. At longer times we observe more irregular patterns. We see a similar shell structure in images taken along a perpendicular direction, which implies that the outgoing atom shell is spherical. Based on its size in ToF, the shell expands at a rate of ${\approx}\,2$ mm/s, which is consistent with it reflecting off the trap walls and interfering with the central part of the cloud after ${\approx}\,10$ ms.
In Fig. \[fig:jump\](c) we plot $t_{\rm c}$ versus $a$ for six data sets taken with different $L$ and $N$ values. We observe $t_{\rm c}$ values that vary between 3 and $400\,$ms. In Fig. \[fig:jump\](d) we show that all the data points fall onto a single universal curve if we plot the dimensionless collapse time, $t_{\rm c}/\tau_0$, versus the reduced distance from the critical point, $(a-a_{\rm c})/a_{\rm c} \equiv (\alpha-\alpha_{\rm c})/\alpha_{\rm c}$. In general, $t_{\rm c}$ could also depend on $\eta$, but the universal behaviour seen in Fig. \[fig:jump\](d) shows that this effect is negligible for our range of $\eta$, between $4 \times 10^{-5}$ and $4 \times 10^{-4}$. The solid line in Fig. \[fig:jump\](d) shows results of lossless GP simulations, without any free parameters; we reproduce a very similar dependence of $t_{\rm c}$ on $\alpha$, although the numerical values are systematically slightly below the experimental ones.
Weak Collapse {#sec:WeakCollapse}
=============
\[btp\] ![Observation of weak collapse. Here $L=30\,\mu$m and $N=20.3 \times10^4$, corresponding to $a_{\rm c} = -0.44\,a_0$. (a) Atom loss versus time after quenches to various $a$ values. For small $|a- a_{\rm c}|$ (left panel) $\Delta N/N$ decreases with increasing $|a|$, as expected for a weak collapse, while for large $| a- a_{\rm c}|$ (right panel) the opposite trend is seen. (b) Summary of atom loss for all $t$ and $a<a_{\rm c}$. The coloured points are the data shown in (a), and the points clustered around $\Delta N =0$ correspond to $t< t_{\rm c}$. Single-collapse loss monotonically decreases with increasing $|a|$ and extrapolates to zero for $a_{\rm c}/|a| \rightarrow 0$ (dot-dashed and solid black lines, see text), confirming the prediction of the weak-collapse theory. For $a_{\rm c}/|a| > - 0.6$ single collapse does not re-stabilise the system and multiple collapse occurs. The dashed purple line shows the equilibrium stability criterion, $\Delta N/N = 1 - a_{\rm c}/a$ (see text). []{data-label="fig:Nlost"}](Figure4.pdf "fig:"){width="1\columnwidth"}
We now turn to the aftermath of the collapse. Since $\tilde{\psi}$ is initially normalised to unity, the [*fractional*]{} atom loss, $\Delta N/N$, should be some universal function of $\alpha$ and $\eta$; here $\Delta N = N-N_{\rm f}$ is the difference between the initial (pre-collapse) and the final (time-dependent) atom number. The counterintuitive implication of the weak-collapse theory is that $\Delta N/N$ [*decreases*]{} if the BEC is made [*more unstable*]{}, by quenching $a$ to a more negative value.
In Fig. \[fig:Nlost\] we focus on one data set, for fixed $L=30\,\mu$m and $N=20.3 \times10^4$. As we illustrate in the left panel of Fig. \[fig:Nlost\](a), close to the critical point, where we observe only single-collapse events, the atom loss indeed decreases with increasing $|a|$, indicating weak collapse. On the other hand, as shown in the right panel of Fig. \[fig:Nlost\](a), in the regime of large quenches and multiple collapse, the atom loss in the long-time limit shows the opposite trend; only this type of behaviour was seen in harmonic-trap experiments [@Donley:2001; @Cornish:2006].
In Fig. \[fig:Nlost\](b) we present a consistent picture of the atom-loss trends for all $a< a_{\rm c}$, from $a/a_{\rm c} \approx 1$ to $a/a_{\rm c} \rightarrow \infty$. Here we plot $\Delta N/N$ versus $a_{\rm c}/|a|$, and for each $a$ show $\Delta N/N$ values observed for all $t$; the points clustered around $\Delta N=0$ correspond to $t<t_{\rm c}$.
The single-collapse regime, $a_{\rm c}/|a| < - 0.6$, is clearly identified by the small spread of the non-zero $\Delta N$ values. The single-collapse atom loss clearly decreases with increasing $|a|$, and extrapolates to zero for $a_{\rm c}/|a| \rightarrow 0$. This is the unambiguous signature of a weak collapse. The dot-dashed black line shows a linear extrapolation, which gives $\Delta N/N = -0.02(2)$ for $a_{\rm c}/|a| = 0$, while the (almost indistinguishable) solid black line shows a power-law fit, $\Delta N/N \propto |a|^{-1.05(7)}$.
For $a_{\rm c}/|a| > - 0.6$, multiple collapse occurs, because the diminishing single-collapse atom loss does not re-stabilise the system. However, we see that even in this regime the minimal loss we observe at each $a$ still follows the weak-collapse trend (solid black line). It is also instructive to plot the function $\Delta N/N = 1 - a_{\rm c}/a$ (dashed purple line); this is atom loss such that, after a quench to a given $a < a_{\rm c}$, the atom number drops to the new critical value $N_{\rm c} (a) = \alpha_{\rm c} L /(8\pi a) = N a_{\rm c}/a$ \[see Eq. (\[eq:dimcoef\])\]. This equilibrium stability criterion is not obviously applicable in the non-equilibrium situation after the first collapse [@Cornish:2006]. Still, it provides a good estimate of both the point, $a_{\rm c}/|a| \approx - 0.6$, beyond which the single-collapse loss is insufficient to re-stabilise the system (see also Appendix C), and the long-time loss at large $a/a_{\rm c}$.
We now extend the study of the weak-collapse atom loss to other $L$ and $N$ values (see Fig. \[fig:weakcollapse\]). In this analysis we include all $a$ values for which only single collapse occurs, and also those where clearly resolved single and double collapses occur (see Appendix C).
\[tbp\] ![Weak-collapse scaling laws. (a) Writing $ \Delta N/N \propto |a|^{-\gamma}$ for fixed $L$ and $N$, for our six data sets we get an average $\bar{\gamma}=1.02(2)$. (b) Assuming $\Delta N/N = C/|a|$, we find $C \propto N^{-0.51(2)}$ (solid black line), with no dependence on $L$. (c) Universal behaviour of the weak-collapse atom loss. We plot all the single-collapse data for different $a$, $N$ and $L$ versus $\eta^{1/4}/|\alpha|$, which vanishes in the limit of infinitely strong attractive interactions, $|\alpha| \rightarrow \infty$. The linear fit gives $\Delta N/N \approx 13\, \eta^{1/4}/|\alpha|$. []{data-label="fig:weakcollapse"}](Figure5.pdf "fig:"){width="1\columnwidth"}
Writing $\Delta N/N \propto |a|^{-\gamma}$ for each data set with fixed $L$ and $N$, as in Fig. \[fig:Nlost\](b), we always get $\gamma$ consistent with unity \[see Fig. \[fig:weakcollapse\](a)\]; averaging over all data sets gives $\bar{\gamma}=1.02(2)$. We then assume the form $ \Delta N/N = C/|a|$ and study the dependence of $C$ on $L$ and $N$. As shown in Fig. \[fig:weakcollapse\](b), on a log-log plot, we find $C \propto N^{-0.51(2)}$, with no clear dependence on $L$; the two points taken with $L=16\,\mu$m and $41\,\mu$m fall onto the same line as the four points taken with $L=30\,\mu$m.
We thus experimentally find that weak-collapse atom loss is described very well by $\Delta N / N \propto 1/(\sqrt{N}|a|)$. From Eq. (\[eq:dimcoef\]), this corresponds to $\Delta N/N \propto \eta^{1/4}/|\alpha|$, which is indeed independent of $L$, and vanishes in the limit of infinitely strong attraction, $|\alpha| \rightarrow \infty$. We note that while the weak collapse atom loss does not depend on $L$ (the overall size of the box) it may depend on the box shape; this is an interesting question for future research.
In Fig. \[fig:weakcollapse\](c) we plot all our single-collapse data versus $\eta^{1/4}/|\alpha|$ and confirm that it falls onto a single universal curve [@CollapseFootnote2]. These experimentally obtained scaling laws should provide useful input for further theoretical work.
Conclusions and outlook {#sec:conclusion}
=======================
In conclusion, we have performed a comprehensive study of the collapse of an attractive BEC confined in the homogeneous potential of a 3D box trap. We have fully characterised the critical point for collapse, and the collapse dynamics of an interaction-quenched BEC, finding universal behaviour in agreement with the theoretical expectations. Most importantly, we have provided conclusive experimental evidence for the counterintuitive weak collapse, and have experimentally determined weak-collapse scaling laws that should provide a useful reference point for the general theories of nonlinear wave phenomena.
Our work also points to many avenues for further research. It would be very interesting to explore quenches from a large positive $a$, where the BEC is initially deep in the Thomas-Fermi regime, and in the case of a box potential the density is uniform. In this case it is not obvious how the condensate would spontaneously ‘choose’ the position at which to collapse, or whether many local collapses would occur instead of a global one. Additionally, since the fractional atom loss cannot exceed 100%, the linear trend seen in Fig. \[fig:weakcollapse\](c) cannot extend to the regime of strong dissipation (large $\eta$). It would be interesting to explore that regime using a different geometry, a different Feshbach resonance, or a different atomic species. Finally, a major extension would be to perform similar experiments with 2D gases, for which a strong collapse and hence fundamentally different behaviour is expected.
We thank Sarah Thomas, Yago del Valle-Incl[á]{}n Redondo and Cornelius Roemer for experimental assistance, and Richard Fletcher, Raphael Lopes and Andreas Nunnenkamp for a critical reading of the manuscript. The GeForce GTX TITAN X used for the numerical simulations was donated by the NVIDIA Corporation. This work was supported by the Royal Society, EPSRC \[Grant No. EP/N011759/1\], ERC (QBox), AFOSR and ARO. A.L.G. and N.N. acknowledge support from Trinity College, Cambridge.
Appendix A - Experimental Setup {#appendix-a---experimental-setup .unnumbered}
================================
Our setup is the first 3D BEC box experiment with tuneable interactions. The setup for producing harmonically trapped $^{39}$K condensates is similar to our previous apparatus [@Campbell:2010]. The main difference is that here we employ the gray molasses technique [@Boiron:1996; @Salomon:2013; @Nath:2013] and directly cool $^{39}$K without the need for sympathetic cooling with rubidium atoms (see also [@Landini:2012; @Salomon:2014]). We load the laser-cooled atoms directly into a crossed optical dipole trap (using a $20$-W $1070$-nm laser) and achieve efficient evaporative cooling using the Feshbach resonance in the $|F,m_F\rangle = |1,1\rangle$ state at 402.70(3) G [@Fletcher:2016]. This results in a quasi-pure BEC of $\approx 2\times 10^5$ atoms. We then load the atoms into a cylindrical optical box formed by blue-detuned ($532$ nm) laser light, and cancel out gravity with a magnetic field gradient, as in [@Gaunt:2013]. The loading procedure is essentially 100% efficient and results in a quasi-pure box-trapped BEC of $\approx 2\times 10^5$ atoms.
The Feshbach resonance in the $|1,1\rangle$ state has a width of $\Delta B =52$ G and the background scattering length is $a_{\rm bg}=-29~a_0$ [@DErrico:2007]. Hence, near the zero-crossing of $a$, at $\approx 350$ G, the variation of the scattering length with the magnetic field is $-a_{\rm bg}/\Delta B \approx 0.6~a_0/$G. We tune $B$ in steps of 50 mG, corresponding to a scattering length resolution of 0.03 $a_0$.
Appendix B - Scattering Length Calibration {#appendix-b---scattering-length-calibration .unnumbered}
==========================================
The exact magnetic field at which the scattering length in the $|1,1\rangle$ state vanishes was independently measured in Ref. [@Fattori:2008b] to be . For Fig. \[fig:ac\](a-c) we calculate our $a$ values assuming . Fitting the data in Fig. \[fig:ac\](c) with a free intercept gives an intercept , which is consistent with zero within the systematic $0.06\,a_0$ error due to the uncertainty in $B_{a=0}$. We take this to be an unbiased confirmation of the zero intercept and the expected scaling $a_{\rm c} \propto 1/N$, and use this scaling to slightly refine the value of the zero-crossing field, to $B_{a=0} = 350.45(3)$ G. The remaining 30 mG uncertainty in $B_{a=0}$ corresponds to a systematic uncertainty in our $a$ values of $\approx 0.02\,a_0$.
For our interaction quenches we have determined, using radio-frequency spectroscopy, that the magnetic field takes 4 ms to change (from 20 to 80 % of the jump). We account for this delay in our determination of the collapse time, and also include an additional 2 ms uncertainty in all the reported $t_{\rm c}$ values.
Appendix C - From Single to Double collapse {#appendix-c---from-single-to-double-collapse .unnumbered}
===========================================
\[b!\]  and the shading shows its uncertainty. The solid black line shows the (single-event) weak-collapse scaling, $\Delta N / N \propto 1/|a|$. []{data-label="fig:Bif"}](Figure6.pdf "fig:"){width="1\columnwidth"}
In Fig. \[fig:Bif\] we present evidence for a gradual transition between single- and double-collapse events, which strongly supports the interpretation that an increasing number of discrete collapse events occur as $|a|$ is increased. This data was taken with $L=30~\mu$m and initial $N=11.4 \times 10^4$.
In Fig. \[fig:Bif\](a) we show the evolution of $N_\text{f}$ after a quench to various $a<a_\text{c}$. A fine scan of $a$ resolves a striking bifurcation of the collapse outcome. We interpret the upper and lower branch as the result of, respectively, one and two collapse events. As $|a|$ is increased, the probability of a double collapse gradually increases. This crossover is highlighted in the histograms shown on the right.
In Fig. \[fig:Bif\](b) we show the fractional atom loss versus $|a|$ on a log-log plot. In the regime where a double collapse occurs, the single-collapse branch still clearly follows the weak-collapse scaling $\Delta N / N \propto 1/|a|$. Note that in this data set the double collapse occurs slightly closer to $a_\text{c}$ than expected from the simple equilibrium stability criterion (purple band).
[61]{}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 [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.13.479) [****, ()](\doibase 10.1103/PhysRevLett.15.1005) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.52.1222) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.79.2604) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.61.043601) @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1016/0167-2789(94)90167-8) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1016/S0375-9601(02)01385-3) @noop [****, ()]{} [****, ()](\doibase 10.1038/408692a) [****, ()](\doibase
10.1103/PhysRevLett.86.4211) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.96.170401) [****, ()](\doibase 10.1103/PhysRevA.84.033632) [****, ()](\doibase 10.1103/PhysRevA.86.063601) @noop [****, ()](\doibase 10.1103/PhysRevLett.90.160405) [****, ()](\doibase
10.1103/PhysRevLett.101.080401) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.108.210401) [****, ()](\doibase 10.1038/nature16485) @noop [ ()]{} @noop [ ()]{} [****, ()](\doibase
10.1103/PhysRevA.66.011602) [****, ()](\doibase 10.1103/PhysRevA.65.033624) [****, ()](\doibase http://dx.doi.org/10.1016/S0375-9601(02)00246-3) [****, ()](\doibase
10.1103/PhysRevA.67.014304) [****, ()](\doibase
10.1103/PhysRevA.68.045601) [****, ()](\doibase
10.1143/JPSJS.72SC.127) [****, ()](\doibase 10.1103/PhysRevLett.92.040401) [****, ()](\doibase 10.1103/PhysRevA.71.033604) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.110.200406) [****, ()](\doibase
10.1103/PhysRevLett.113.053202) @noop [ ()]{} [****, ()](\doibase 10.1103/PhysRevLett.100.080405) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.86.1406) [****, ()](\doibase 10.1103/PhysRevA.63.043601) @noop @noop [****, ()](\doibase 10.1103/PhysRevA.53.R3734) [****, ()](http://stacks.iop.org/0295-5075/104/i=6/a=63002) [****, ()](\doibase 10.1103/PhysRevA.88.053407) [****, ()](\doibase
10.1103/PhysRevA.86.033421) [****, ()](\doibase
10.1103/PhysRevA.90.033405) @noop [ ()]{} [****, ()](\doibase 10.1103/PhysRevLett.101.190405) [****, ()](http://stacks.iop.org/1367-2630/9/i=7/a=223)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Chaotic whispering-gallery modes have significance both for optical applications and for our understanding of the interplay between wave phenomena and the classical ray limit in the presence of chaotic dynamics and openness. In strongly non-convex geometries, a theorem by Mather rules out the existence of invariant curves in phase space corresponding to rays circulating in whispering-gallery patterns, so that no corresponding modes of this type are expected. Here we discuss numerical computations of the electromagnetic fields in planar dielectric cavities that are strongly non-convex because they are coupled to waveguides. We find a family of special states which retains many features of the chaotic whispering-gallery modes known from convex shapes: an intensity pattern corresponding to near-grazing incidence along extended parts of the boundary, and comparatively high cavity Q factors. The modes are folded into a figure-eight pattern, so overlap with the boundary is reduced in the region of self-intersection. The modes combine the phenomenology of chaotic whispering-gallery modes with an important technological advantage: the ability to directly attach waveguides without spoiling the Q factor of the folded mode. Using both a boundary-integral method and the finite-difference time domain technique, we explore the dependence of the phenomenon on wavelength in relation to cavity size, refractive-index contrast to the surrounding medium, and the degree of shape deformation. A novel feature that distinguishes folded from regular whispering gallery modes is that a given shape will support high-Q folded chaotic whispering gallery modes only in certain wavelength windows.'
author:
- 'Kahli Burke and Jens U. N[ö]{}ckel'
title: 'Folded chaotic whispering-gallery modes in non-convex, waveguide-coupled planar optical microresonators'
---
Introduction
============
Wave-chaotic optical cavities are of interest in applications because of the degrees of design freedom that are opened up when the geometric constraints of separability and symmetry are removed [@CaoWiersigReview]: spectral density, lifetimes and emission directionality can all be tailored in sometimes dramatic ways that go far beyond perturbative effects around separable special cases. Among the first examples are planar dielectric ovals that support *chaotic whispering-gallery modes* (WGMs) [@Nockel:1997lr], where the anisotropic shape gives rise to highly universal directional emission patterns that can be explained based on the classical phase-space structure of the ray dynamics. From a theoretical point of view, this connection to classical ray physics is intriguing because the transition to chaos that occurs in generic nonseparable systems must then be studied in the presence of openness. The wave equation in this scenario exhibits quasibound states at complex wavenumber $k=k_{\mathrm{r}}-i\,\kappa$; here $\kappa$ measures the escape rate which is related to the quality factor via $$Q=\frac{k}{2\kappa}.\label{eq:QfactorDef}$$ WGMs in oval resonators are special states that preserve long lifetimes comparable to those found in circular dielectrics [@ISI:000320621600008], due to the fact that they are predominantly localized in the phase-space region corresponding to total internal reflection at the dielectric boundary. When the invariant curves (tori) that foliate this phase-space region in the limit of circular symmetry are gradually broken up following the KAM theorem as applied to convex billiards [@LazutkinBook], regular WGMs become *chaotic* WGMs in which rays explore regions of phase space in which the rays cease to be confined by total internal reflection. The critical angle for total internal reflection defines an escape window in phase space rather than real space, because this angle is directly related to the tangential momentum component at the surface of the resonator. In Ref. [@Nockel:1997lr], an approximate semiclassical quantization of these unconventional modes was proposed based on a separation of time scales between the fast whispering-gallery circulation and a slower spiraling-in toward the escape window, making it possible to identify *adiabatic invariant curves.* Their location then served as an initial condition for ray simulations of the escape directionality, caused by the mixed phase space in the vicinity of the escape window.
It has long been understood that no true WGMs can be sustained in circular domains if *waveguides* are attached, because this creates openings in the boundary that interrupt the whispering-gallery circulation. In phase space, the escape window then depends not only on tangential momentum but on position. For a detailed study of this breakdown in the context of electronic microstructures with otherwise impenetrable walls, see Ref. [@PhysRevB.54.10652]. This can be viewed as a special case of a theorem by Mather [@MatherInvariant], who proved the non-existence of whispering-gallery invariant curves in planar billiards when the boundary is not everywhere convex. An example for Mather’s theorem is the Bunimovich stadium (two semicircles joined by straight, parallel sides) [@BunimovichStadium], whose ray dynamics does not permit whispering-gallery orbits even if $\ell$ is arbitrarily short, even though the weaker condition of non-concavity still holds. Strict convexity is therefore required in order to sustain WG circulation, but any waveguide openings will necessarily introduce corners where this condition breaks down.
By exploiting the preferential emission directions of chaotic WGMs in the near or far field, light can be coupled into and out of the resonator without contacting (and thereby perturbing) the boundary geometry directly. However, monolithically attached waveguides offer some distinct advantages from an engineering point of view. In laser applications, they allow both efficient optical pumping and more complete collection of the emission. Although directional emission patterns from planar cavities can be made highly directional *in the plane* [@Nockel:94], losses by out-of-plane diffraction can be significant [@backes:VerticalAndNotches; @VerticalLossWithWG; @VerticalSpreadingDisk; @LimaconDirectionality]. With attached waveguides, such losses are minimized. Moreover, waveguides afford precise control over the number of input/output channels.
This has motivated several recent proposals to integrate chaotic cavities with waveguides [@ChaoticChannelingPRL; @EndFireInjection; @MicrospiralWG; @SquareMicrolasers; @TriangleAndSquareWG; @octagonWithLead]. Because of Mather’s theorem, however, almost all such proposals involve modes that are *not* of whispering-gallery type, achieving good confinement instead with states localized on periodic orbits: Stable and unstable periodic ray orbits whose reflections occur far enough away from any waveguide openings will be insensitive to the boundary shape at the waveguide apertures. These periodic orbits exist with or without the presence of attached waveguides, and they make lasing possible even in the free-standing Bunimovich stadium [@Fukushima:07].
The appearance of wavefunction scarring due to unstable periodic orbits in [@Fukushima:07] illustrates that wave solutions of the Helmholtz equation can defy ray-optics predictions. This is also true for the predictions of Mather’s theorem, and again the stadium billiard is a case in point: Numerical solutions for closed (hard-wall) cavities [@Prange2001] show that a form of chaotic WGM exists if the straight sides are sufficiently short.
![\[fig:Q-factor-versusEps\]Spatial structure of a folded chaotic whispering-gallery mode (insets), and Q factor of the mode versus deformation parameter $\epsilon$ at refractive index $n=2.4$. The resonator deformation is defined in Eq. (\[eq:ShapeCurve\]); attached waveguides of finite lengths are shown horizontal; their lengths are slightly unequal to remove reflection symmetry. False color represents field intensity. The real part of $k$ (not shown) decreases approximately linearly from $14$ to $12.5$ as $\epsilon$ increases.](Figure1){width="0.9\columnwidth"}
We will describe whispering-gallery type behavior with a different topology, which can be described as *folded* chaotic whispering-gallery modes, cf. Fig. \[fig:Q-factor-versusEps\]. This refers to the fact that the direction of whispering-gallery ray circulation underlying the modes reverses due to a self-intersection near the center of the cavity. The bowtie orbit of Ref. [@Gmachl05061998] also exhibits a self-intersection, but it – like all other such orbits we are aware of – gives rise to modes that explore the boundary only at isolated points corresponding to the discrete vertices where ray reflections occur. In Ref. [@Gmachl05061998], $Q>1000$ was obtained at relatively high refractive index contrast of $n\gtrapprox3$, corresponding to a critical angle for total internal reflection of $\chi_{\mathrm{c}}\approx\arcsin\nicefrac{1}{n}\approx0.34$ (where $\chi$ denotes the angle of incidence with respect to the surface normal). For such high index contrast, even Fabry-Perot cavities with near-normal angle of incidence provide similar Q-factors without the need for additional mirrors.
By contrast, all the high-Q modes in our design are localized predominantly in the whispering-gallery region of phase space, much further from the critical angle, even when the index contrast is below $n=2$. The light is then well-confined by total internal reflection. Reminiscent of a figure-eight shape, folded chaotic WGMs have a waist that allows them to avoid two isolated sections of the boundary; but unlike stable periodic orbits such as the bowtie, folded chaotic WGMs explore the remainder of the perimeter in the same way as would be expected for conventional WGMs. In this sense, these types of modes come as close as possible to the circulating ray patterns that are strictly ruled out by Mather’s theorem.
The self-intersecting topology makes folded chaotic WGMs amenable to the incorporation of waveguides because the openings have only small overlap with the waist of the mode. Although the field looks similar to Gaussian beam in the waist region, the latter are fundamentally different because they are always built upon stable periodic orbits, which in the simplest case requires a configuration corresponding to two focusing mirrors [@BabicBook1972]. In our design, the reflections are near-grazing along the entire convex part of the boundary, and there is no focusing-mirror configuration. Therefore, folded chaotic WGMs cannot be obtained in paraxial optics.
As an important correction to the ray limit in the presence of chaotic dynamics, [*dynamical localization*]{} has been invoked [@Nockel:1997lr] to explain the high Q factors of chaotic WGMs in convex cavities, and circular dielectrics with corrugated surface perturbations [@FangCao2005]. A hallmark of dynamical localization is an exponential decay of wave intensity as a function of a variable which in the classical ray picture exhibits diffusive time evolution. In corrugated perturbed circles, this variable is the angular momentum; and it remains a good quantity to characterize dynamical localization in chaotic WGMs of smooth but convex cavities [@Noeckel:T90-2001]. At each reflection, the angular momentum is proportional to $\sin\chi$. However, an important common feature on which previous work relies is that the underlying classical dynamics exhibits a separation of time scales, without which diffusive behavior cannot be identified. Before $\sin\chi$ changes significantly, a whispering-gallery ray will have completed many round trips, allowing the azimuthal positions of boundary reflections to be averaged out [@Noeckel:T90-2001].
To identify wave localization in the folded WGMs to be discussed here, we have to follow a different route because angular momentum does not undergo the required slow diffusion. It jumps discontinuously because a figure-eight pattern entails periodic reversals of the sense of rotation. Therefore, we begin with a description of the ray dynamics in phase space before presenting the detailed results of our wave calculations in section \[sec:wavecalc\]. These two descriptions are then synthesized in section \[sec:Discussion\] to identify the classical structures on which the modes are localized, using Husimi projections of the numerical wave functions onto the ray phase space.
Cavity shape and ray phase space
================================
![(a-c): Three self-intersecting unstable periodic orbits with angles of incidence $\sin\chi\ge0.6$ in the billiard shape given by Eq.(\[eq:ShapeCurve\]) with $\epsilon=0.43$. The lengths $L_{1}$, $L_{2}$ of the orbits are slightly different. The same topology of orbits persists for a wide range of deformations. (c) is obtained from (b) by adjusting the initial conditions to generate eight instead of three reflections in the top half. For the rectangular unstable periodic orbit in (d), the increasing proximity to the corners of the waveguide aperture is illustrated for increasing $\epsilon$. (e) A ray trajectory launched in the cavity (green dot) and escaping through the waveguide opening (red dot). It retains the folded character of the orbits in (a), (b) for a short time. \[fig:Two-self-intersecting-periodicOrbits\]](Figure3){width="0.85\columnwidth"}
Figure \[fig:Two-self-intersecting-periodicOrbits\] illustrates the non-convex cavity shape in which the long-lived modes are confined. The openings in the boundary are attached to waveguides, but for the purposes of the ray dynamics they constitute escape windows, in addition to the refractive escape mechanism that sets in when the condition for total internal reflection at the dielectric interface is violated. The three self-intersecting periodic orbits shown in Fig. \[fig:Two-self-intersecting-periodicOrbits\] (a - c) have incident angles satisfying $\sin\chi\ge0.6$, which means they are confined by total internal reflection for refractive indices $n>1.7$. Half of the reflections occur with opposite sense of circulation, corresponding to the reversal of angular momentum (with respect to the center) taking place at the self-intersection.
The geometry also permits a rectangular periodic orbit whose angle of incidence has the fixed value $\sin\chi\approx0.707$, but Fig. \[fig:Two-self-intersecting-periodicOrbits\] (d) shows an important distinction to the self-intersecting orbits: as $\epsilon$ increases, the corners of the waveguide aperture encroach on the orbit and eventually touch it. The resulting corner diffraction [@ISI:000323333000016] will degrade the lifetime of any modes based on this orbit. The numerical computations to be described in section \[sec:wavecalc\] have revealed high-Q modes in cavities of the shape (\[eq:ShapeCurve\]), that do show enhanced intensity near the rectangle orbit, but never exclusively on that orbit. Instead, the folded chaotic states consistently show high intensity overlapping with the folded orbits over a wide range of deformations $\epsilon$ and attached waveguide widths.
![\[fig:Poincare-surface-of-section\]Poincar[é]{} surface of section with the curve parameter $\phi$ as position and $\sin\chi$ as momentum variable. Here, $\chi$ is the angle between incident rays and the surface normal, so that $\sin\chi=1$ corresponds to grazing incidence. Ray trajectories show up as point clouds in the chaotic sea, or as one-dimensional lines for regular motion. Shown as insets are a stable bouncing-ball orbit (left) and an unstable periodic orbit (right). Arrows point to the corresponding phase space locations; in particular the 6 black dots indicate the location of the 6 bounce points for the periodic orbit. Shaded boxes indicate the intervals of $\phi$ which describe the attached waveguides. Rays entering these regions will escape the resonator. The thin horizontal lines mark the critical angle for total internal reflection, $\left|\sin\chi_{c}\right|=\nicefrac{1}{n}$ ($n=2$ here). Here and in all subsequent results, $n$ can be viewed as the interior refractive index while the exterior refractive index is unity.](Figure2){width="0.8\columnwidth"}
Figure \[fig:Two-self-intersecting-periodicOrbits\] (e) illustrates a ray in the chaotic interior region that escapes into the waveguide opening after several reflections. By folding a WGM over itself and attaching waveguides in the avoided regions of the boundary, two distinct escape windows in phase space are created: one bounded by the critical angle for refractive escape, and the other by the real-space locations of the waveguide openings. This is illustrated in the Poincar[é]{} surface of section of Fig. \[fig:Poincare-surface-of-section\], depicting the phase space of a billiard parametrized in polar coordinates by $$r(\phi)\propto1+\epsilon\,\cos\!\left(4\phi\right)\label{eq:ShapeCurve}$$ The curve parameter $\phi$ is recorded on the horizontal axis, and $\sin\chi$ is a measure of the conjugate momentum variable.
For the numerical wave calculations of Fig. \[fig:Q-factor-versusEps\], waveguide segments were connected to the open sections of the boundaries depicted in Fig. \[fig:Poincare-surface-of-section\]. The waveguide segments were modeled as a “squircle” [@Squircle]. This curve has a polar-coordinate representation $$r_{\text{WG}}(\phi)=\sqrt{a^{2}\left|\cos\phi\right|+b^{2}\left|\sin\phi\right|}$$ but was also shifted horizontally in order to connect smoothly with the unshifted curve of Eq. (\[eq:ShapeCurve\]).
The waveguide portions of the boundary are shaded in the Poincar[é]{} section because they are irrelevant to the interior ray dynamics that makes the folded orbits possible. There are no stable islands associated with any periodic orbits other than the bouncing-ball trajectories (leftmost inset) whose perpendicular angle of incidence precludes long-lived stable modes from forming. There are no invariant curves in the whispering-gallery region near $\sin\chi\to1$.
Wave calculations {#sec:wavecalc}
=================
The wave patterns in Fig. \[fig:Q-factor-versusEps\] indicate that the mode is not supported by any one of the self-intersecting periodic orbits in Fig. \[fig:Two-self-intersecting-periodicOrbits\] alone, but by a more extended phase-space region in their vicinity. The main panel of Fig. \[fig:Q-factor-versusEps\] follows a single mode over a range of deformations $\epsilon$, revealing an optimal deformation of $\epsilon\approx0.44$ at which this particular mode reaches $Q\approx2500$.
To characterize the parameter-dependence of the Q factor further, we carry out numerical simulations of the wave equation using two different methods. The first approach is a version of the boundary-integral method described by Heider [@HeiderBIM] in which we directly search for the quasibound states satisfying outgoing-wave boundary conditions at infinity [@NockelMcBook1]. The second approach isolates the high-Q modes by harmonic inversion [@Mandelshtam:2001fk] of temporal field variation collected in a finite-difference time domain simulation with perfectly matched layer boundary condition (MEEP with post-processing by Harminv) [@MEEP-citation]. In the direct quasibound-state calculation, we obtain complex wavenumbers $k_{\text{QB}}=k-i\,\kappa$ to determine Q using Eq. (\[eq:QfactorDef\]).
Quasibound states\[subsec:Quasibound-states\]
---------------------------------------------
In the boundary-integral approach, we leverage the fact that the dielectric defining the billiard is uniform so that a Green-function description of the interior and exterior fields purely in terms of the boundary is possible. At the dielectric interface, the electric field is assumed to satisfy the boundary conditions for TM polarization (electric field perpendicular to the plane). After discretization along the interface, this leads to a nonlinear eigenvalue problem $A(k_{\text{QB}})\,\boldsymbol{u}=\boldsymbol{0}$ where $\boldsymbol{u}$ contains the source values of the electric field and its normal derivative, and $A$ is a matrix obtained from the field matching equations. A non-trivial solution requires searching for $k_{\text{QB}}$ in the complex plane, which we do using a predictor-corrector method [@HeiderBIM]. The method requires discretization of the boundary curve, and we generally found good convergence up to wavenumbers of $k\approx60$ by choosing $580$ points. For additional analysis of the exponential convergence of the method with discretization density, see Ref. [@HeiderBIM].
![Small-cavity limit of the folded chaotic WGM, showing only low Q factors as the transverse width of the mode is comparable to the cavity size. Below $k\approx7$, the wavelength is too long to observe the WGM like concentration of intensity near the top and bottom of the cavity. \[fig:Small-cavity-limit\] The horizontal waveguide stubs are attached in a way that smoothly matches the shape given by Eq. (\[eq:ShapeCurve\]) for the vertical lobes. The refractive index is $n=2.4$ inside and $n=1$ outside, and the vertical lobes are described by Eq. (\[eq:ShapeCurve\]) with $\epsilon=0.444$. In this and the following plots, one can discern even and odd parity with respect to reflections at the horizontal axis. Although we intentionally break reflection symmetry across the vertical axis by making the horizontal waveguide lengths unequal, the modes still show approximate antinodes (left) or nodal lines (right) along the vertical axis. This is because the cavity supporting most of the intensity is still left-right symmetric.](Figure4){width="0.6\columnwidth"}
![Wavenumber sweeps of the Q factor for different refractive indices $n$, discarding low-Q modes with $\kappa\ge0.05$ (for $n=1.6$, only modes with $\kappa\ge0.2$ are shown). The deformation is $\epsilon=0.444$. The appearance of distinct Q-factor peaks as a function of quasi-bound state wavenumber is most pronounced at the largest refractive index, $n=2.4$, and becomes nearly unobservable at $n=1.6$ (therefore, data for $n=1.6$ were not collected beyond $nk\approx106.5$). The peaks that do remain observable are approximately at the same values of $nk$ for all $n$.\[fig:Wavenumber-sweeps\]](Figure5){width="0.75\columnwidth"}
By definition, the boundary-integral method assumes that regions of uniform dielectric constant are bounded by closed curves, so we model the waveguides as finite-length attachments, see Fig. \[fig:Q-factor-versusEps\]. By varying the lengths of these waveguides (either together or independently of each other) we verified that the finiteness of the stubs has no significant effect on the mode structure: neither real parts nor imaginary parts of $k_{\text{QB}}$ for the modes with figure-eight topology were affected unless their Q factor was below approximately $200$ to begin with. Because the modes are built on classical ray orbits, they will not be found (with appreciable lifetimes) in small cavities. This is illustrated in Fig. \[fig:Small-cavity-limit\] where the self-intersecting topology is barely discernible in the wave intensities. This raises the question whether the Q factors of such modes will increase *monotonically* with increasing cavity size, or equivalently with shorter wave length. To address this dependence, we fix the cavity geometry at $\epsilon=0.444$ to identify all the modes with lifetimes above a threshold of $\left|\mathrm{Im}(k_{\text{QB}})\right|=\kappa<0.05$ with wavenumber $k\lesssim60$. The particular choice of deformation corresponds to the maximum in Fig. \[fig:Q-factor-versusEps\] at $n=2.4$, but the same phenomena are observed at other values of $\epsilon$.
Figure \[fig:Wavenumber-sweeps\] shows that the Q factors display a *non-monotonic* peak structure. The data are plotted for four different resonators with refractive indices of $n=1.6$, $1.8$, $2.0$ and $2.4$. The peaks are more pronounced for larger refractive index because the critical angle for total internal reflection decreases with $\sin\chi_{\mathrm{c}}=1/n$. By choosing the horizontal axis to display $nk$ instead of the free-space wavenumber $k$, the peak positions for different $n$ moreover line up to a good approximation. Since $nk$ is the wave number inside the cavity, this indicates that the high-Q peak *locations* are determined by the interior wave patterns, not the monotonically $n$-dependent coupling to the surrounding free space. The absolute length scale of the cavity drops out of $Q$ as per Eq. (\[eq:QfactorDef\]).
The folded chaotic whispering-gallery modes are surprisingly resilient even to large waveguide openings. For practical applications, much narrower waveguides will typically be desirable because the input and output should be single mode. We show results for large openings because that is the regime in which the coexistence of whispering-gallery modes with an apparent violation of Mather’s theorem is clearest. In particular, the openings are much wider than the wavelength, so they cannot be treated as small perturbations. We also observe the non-monotonic distribution of $Q$-factors shown in Fig. \[fig:Wavenumber-sweeps\] at $\epsilon=0.444$ for other deformations and waveguide widths. Figure \[fig:Q-factor-scan-e52\] shows this for a higher deformation of $\epsilon=0.52$.
![Q-factor scan versus interior wave number at deformation $\epsilon=0.52$ and refractive index $n=2.4$, showing modes with decay rate $\kappa<0.05$. \[fig:Q-factor-scan-e52\]](Figure6){width="0.8\columnwidth"}
Figure \[fig:Field-intensities-forPeaks\] depicts the wave intensities of the highest-Q folded chaotic WGMs in Fig. \[fig:Wavenumber-sweeps\], with wavenumbers ranging from $k\approx13.6\ldots44$ at $n=2.4$. Each image represents a single quasibound state, not a superposition. The nodal structure that decorates the underlying figure-eight patterns becomes more complex with increasing $k$, in much the same way that transverse nodes appear in higher-order Gaussian beams or radial nodes appear in whispering-gallery modes. The difference between those examples and the folded chaotic WGMs is that the nodal lines show wave dislocations typical of non-separable wave equations. Non-separability goes along with the chaotic ray dynamics of Fig. \[fig:Poincare-surface-of-section\]. Despite the increasingly complex nodal structure, all modes share the whispering-gallery like wave propagation along the convex parts of the boundary.
![Field intensities for the modes corresponding to the first four Q-factor peaks in Fig. \[fig:Wavenumber-sweeps\] for $n=2.4$, $\epsilon=0.444$. The Q factors are (a) $2500$, (b) $2100$, (c) $2400$ and (d) $4500$. \[fig:Field-intensities-forPeaks\]](Figure7){width="0.6\columnwidth"}
The Q-factor of any given mode may depend non-monotonically on the deformation parameter $\epsilon$ of Eq. (\[eq:ShapeCurve\]), as shown in Fig. \[fig:Q-factor-versusEps\]. However, we find other high-Q modes over the entire range of deformations in Fig. \[fig:Q-factor-versusEps\], as illustrated by the examples in Fig. \[fig:Field-intensitiesVaryEps\]. The robustness of the folded morphology against deformation is another interesting feature that these modes share with conventional whispering-gallery modes in convex resonators.
![Field intensities for folded WGMs at different deformations: (a) $\epsilon=0.52$, (b) $\epsilon=0.6$, (c) $\epsilon=0.7$. The refractive index is $n=2.4$. In (c), a smaller wavenumber is chosen, whereas (a) and (b) have comparable wavenumbers. The narrowing of the horizontal stubs is a result of the requirement that its tangents must match the curve described by Eq. (\[eq:ShapeCurve\]) at the corners. \[fig:Field-intensitiesVaryEps\]](Figure8){width="0.9\columnwidth"}
Time-domain results
-------------------
To further validate the results of the boundary-integral method, we repeated the calculations for some of the modes using MEEP, which in particular allows for a different modeling of the waveguide attachments. Instead of assuming them to be finite stubs, we allowed them to extend to the boundary of the simulation domain which includes a perfectly matched layer that prevents back-reflections of a wavepacket which was launched in one arm of the structure. After settling into a late-time decaying state, the wave field of a typical high-Q mode displays the same pattern found in the previous section, cf. Fig. \[fig:MEEPmode1\] (a). Shown in Fig. \[fig:MEEPmode1\] (b) is the same resonator at a roughly tenfold shorter wavelength. The comparison illustrates why we make a distinction between folded chaotic WGMs as in Fig. \[fig:MEEPmode1\] (a) and scarred states. The concentration of intensity on the six-bounce unstable periodic orbit of Fig. \[fig:Two-self-intersecting-periodicOrbits\] (a) identifies the mode of Fig. \[fig:MEEPmode1\] (b) as a scarred state, whereas there is no single periodic orbit that describes the intensity of a folded chaotic WGM.
![(a) Folded chaotic whispering-gallery mode as obtained in a finite-difference time domain computation for the same deformation $\epsilon=0.444$ and refractive index $n=2.4$ shown in Fig. \[fig:Wavenumber-sweeps\]. The wave number corresponds to the top of the first Q-factor peak, $k\approx13.6$ ($Q=2612$). Waveguide and resonator structure are underlaid as green shading, and the false-color scale represents the electric field. (b) Same deformation and refractive index, but approximately ten times shorter wavelength, $k\approx116.0$ and $Q=9898$. (b) is discussed in section \[sec:Discussion\].\[fig:MEEPmode1\]](Figure9){width="0.8\columnwidth"}
Having convinced ourselves that the same results can be obtained with both numerical approaches (boundary-integral and finite-difference time domain), we proceed with MEEP to look for folded chaotic WGMs in resonators with thinner waveguides, cf. Fig. \[fig:Two-different-modesThinWires\]. The size of the opening in Fig. \[fig:Two-different-modesThinWires\] (a) is comparable to the wavelength, whereas it is approximately twice the wavelength in (b). The whispering-gallery circulation along the boundary, characteristic of the earlier results in Fig. \[fig:Q-factor-versusEps\], shows that the phenomenon is robust not only to variations in the refractive index, but also to changes in waveguide width.
![Two different modes with thinner waveguides, obtained by introducing a vertical offset between the two lobes of Eq. (\[eq:ShapeCurve\]) and adjusting$\epsilon$ to match them smoothly to the horizontal waveguides. The refractive index is $n=2.4$. In (a) $k\approx13.3$ and $Q\approx1580$, in (b) $k\approx22.4$ and $Q\approx2550$. \[fig:Two-different-modesThinWires\]](Figure10){width="0.8\columnwidth"}
Discussion\[sec:Discussion\]
============================
Folded chaotic whispering-gallery modes in phase space
------------------------------------------------------
Individual modes of the folded chaotic WGM type do not shift or broaden significantly when the attached waveguides are shortened into stubs. From this we conclude that leakage into the waveguides is not the limiting factor for the lifetimes of individual folded chaotic WGMs. However, this is not straightforward to reconcile with the ray-tracing results depicted in Figs. \[fig:Poincare-surface-of-section\] and \[fig:Two-self-intersecting-periodicOrbits\] (e):
A comparison between the numerically observed peak Q factors in Fig. \[fig:Wavenumber-sweeps\] and the classical ray escape into the waveguides provides a first indication that dynamical wave localization is essential in the formation of folded chaotic WGMs. We find that up to a set of measure zero (the trapped unstable periodic orbits), no matter where in the classical phase space we launch a ray, it escapes into the waveguide openings much too fast to explain the $Q$ factor found in the wave calculations.
The rapid ray escape can be viewed as a result of Mather’s theorem, because the non-convex billiard shape implies that the bounce dynamics does not constitute a twist map, and consequently neither the Lazutkin nor the Poincar[é]{}-Birkhoff theorems apply [@LazutkinBook]. What remains is a largely chaotic phase space as in Fig. \[fig:Poincare-surface-of-section\], and no stable ray orbits around which high-Q modes can be form by the mechanism of paraxial optics [@BabicBook1972].
However, as suggested by Fig. \[fig:Two-self-intersecting-periodicOrbits\] (e), even in a chaotic region of phase space, the motion is nevertheless organized by the periodic orbits: Each unstable periodic orbit is a periodic point of the Poincar[é]{} map which has stable and one unstable manifolds. In the linear regime near a periodic point, they correspond to trajectories that either converge on, or recede from, that point. By launching a large number of rays from a small neighborhood of the periodic point and iterating the billiard map forward and backward in time, these manifolds are traced out, revealing a characteristic web of intersections [@LichtenbergBook1992] – the homoclinic tangle. In the main panel of Fig. \[fig:HusimiPhaseSpace\] (a), this has been done for the six-bounce orbit corresponding to the thick solid dots in Fig. \[fig:Poincare-surface-of-section\]. Embedded in the intersections between the manifolds are higher-order periodic points corresponding to orbits such as the one shown in Fig. \[fig:Two-self-intersecting-periodicOrbits\] (c). Guided by the homoclinic tangle, the phase space flow is far from random on intermediate time scales and in fact helps explain the directional emission from free-standing asymmetric cavities [@NoeckelChapter2002] by predicting at what positions the condition for total internal reflection is first violated for a chaotically diffusing trajectory.
![(a) Stable and unstable manifolds around an unstable periodic orbit. The zoomed-in region shown in (a) corresponds to the top left quadrant of the Husimi projection in (b). The real-space wave intensity of the mode in (b) is shown in the inset to (a). It is the same mode as shown in Fig. \[fig:Field-intensities-forPeaks\] (b). Shaded regions in (b) mark the waveguides, which were treated as escape windows in (a). The thin horizontal lines mark $\left|\sin\chi_{c}\right|=\nicefrac{1}{2.4}$. Refractive escape is not considered in the ray simulation for (a), to get a more complete picture of the manifolds. Note the correspondence in shape between the phase space structure and the areas of large Husimi weight.\[fig:HusimiPhaseSpace\]](Figure11){width="\columnwidth"}
To expose the relationship between the observed modes and the underlying ray phase space structure exemplified in Fig. \[fig:Poincare-surface-of-section\], it is useful to project the numerical wave results onto the phase space by means of the Husimi function. Choosing $\phi$ and $\sin\chi$ as the coordinates for this projection, only the boundary fields are needed, and this is just what the boundary-integral approach provides [@PhysRevE.90.022903]. Therefore, Fig. \[fig:HusimiPhaseSpace\] shows one of the high-Q modes as obtained with this method. In Fig. \[fig:Wavenumber-sweeps\] it corresponds to the top of the peak structure located at $n k \approx 58$ as well as the mode depicted in Fig. \[fig:Field-intensities-forPeaks\] (b). The wave intensity shown in the inset illustrates the appearance of nodal lines which in a conventional WGM would be the radial zeros. As seen in Fig. \[fig:Field-intensities-forPeaks\], successive Q-factor peaks at higher $nk$ in Fig. \[fig:Wavenumber-sweeps\] show additional such “radial” nodes, but the chaotic nature of the underlying phase space makes a rigorous classification in terms of radial nodal lines ambiguous.
The Husimi plot in Fig. \[fig:HusimiPhaseSpace\] (b) instead classifies the mode according to the region of phase space by which it is supported – a procedure that is especially useful in this case because there are no stable ray orbits in the regions of the Poincar[é]{} section bounded by the escape conditions. The six-bounce periodic orbit whose manifolds are explored in Fig. \[fig:HusimiPhaseSpace\] (b) is also shown in Fig. \[fig:HusimiPhaseSpace\] (b), colored yellow and red overlaying the Husimi intensity of the wave solution (the two colors belonging to the two opposite senses in which the figure-eight is traversed). Shown in green and cyan are the self-intersecting period-five orbits. We have identified other periodic orbits with four to six bounces that are confined by total internal reflection in the same area of phase space, and it is not possible to uniquely assign regions of high Husimi intensity to a single orbit. However, Fig. \[fig:HusimiPhaseSpace\] (b) does indicate unambiguously that the wave solution is in fact extended over a region of phase space bounded away from the critical angle $\chi_{c}$ by the V-shaped tangle of manifolds of Fig. \[fig:HusimiPhaseSpace\] (a).
![Comparison of Husimi plots for the same states shown in Fig. \[fig:Field-intensitiesVaryEps\]. Only the positive-$\sin\chi$ half of the phase space is displayed (shaded regions and horizontal lines as in Fig. \[fig:HusimiPhaseSpace\]). The low wavenumber ($k\approx17.2$) in (c) leads to lower phase-space resolution than in (a) and (b) where $k\approx31.4$ and $k\approx31.2$.\[fig:Comparison-of-Husimi\]](Figure12){width="0.5\columnwidth"}
Figure \[fig:Comparison-of-Husimi\] illustrates the common phase-space region shared between all high-Q modes. Just like the real-space wave functions, the Husimi intensity shows some variability between different modes. In particular, the areas of highest Husimi intensity are not centered on a single orbit of the many figure-eight type periodic orbits that exist in the cavity.
When interpreting the Husimi projection, the question arises whether our inability to assign a given mode to a unique unstable periodic orbit is merely a consequence of phase-space resolution. If this were true, it would mean the all the high-Q modes in our geometry are really of the scarred kind shown in Fig. \[fig:MEEPmode1\] (b). However, this is not the conclusion we draw. With the length scale used here, $k$ ranges from $k\approx7$ in Fig. \[fig:Small-cavity-limit\] to $k\approx116$ in Fig. \[fig:MEEPmode1\] (b). Within this approximate window, scarred modes belonging to a single unstable periodic orbit were not found below $k\approx70$. The common feature of those lower-$k$ modes is instead the appearance of a caustic-like concentration of intensity following the curvature of the surface at $\phi\approx\pi/2,\,3\pi/2$ (the top and bottom in the wave plots).
In a convex billiard, whispering-gallery circulation leads to true caustics, but there is also a dense set of periodic orbits in the shape of inscribed polygons with lengths that accumulate at the circumference of the billiard. This particular property of inscribed polygonal orbits in the whispering-gallery region of phase space finds its counterpart in our non-convex geometry: there is a similar accumulation of periodic orbits with figure-eight topology, characterized by an increasing number of bounces along *either* the top *or* the bottom portion of the billiard. The orbit displayed in Fig. \[fig:Two-self-intersecting-periodicOrbits\] (c) is an example.
Two other possible scenarios are known to produce wave solutions localized on phase-space regions where the classical billiard shows no stable structure: marginally stable periodic orbits [@WiersigAnticrossingModes; @CaoWiersigReview], and ray orbits that become stabilized only due to the “softness” of the dielectric billiard boundary as a function of incident angle [@Foster:07; @Unterinninghofen2008]. In these cases, the wave intensity is “scar-like” in the sense that it again coalesces onto identifiable periodic orbits, which is not the case for the folded chaotic WGMs.
Given that purely ray based escape rates as shown in Fig. \[fig:Two-self-intersecting-periodicOrbits\] (e) are too fast to explain the high Q factors, we therefore conclude that the folded chaotic WGMs are dynamically localized on the homoclinic tangle of the self-intersecting periodic orbits in this non-convex cavity.
Q-factor oscillations\[subsec:Q-factor-oscillations\]
-----------------------------------------------------
The results shown in Figs. \[fig:Wavenumber-sweeps\] and \[fig:Q-factor-scan-e52\] are of particular practical importance: in designing resonators that exhibit modes with a folded topology, the geometric shape alone doesn’t guarantee that $Q$ factors will increase monotonically with $k$, so the operating wavelength is constrained to match one of the $nk$ ranges where modes with high $Q$ can be found. Whispering-gallery modes of the perfect circle do *not* show any comparable grouping into peaks of higher and lower $Q$; instead, maximum attainable $Q$ factors grow monotonically with $nk$ because the modes with the lowest radial node number grow more and more concentrated at the surface, corresponding to grazing angle of incidence. Therefore, Figs. \[fig:Wavenumber-sweeps\] and \[fig:Q-factor-scan-e52\] show an aspect of the folded modes that distinguishes them from the conventional WGMs in convex resonators.
By varying the widths and lengths of the attached waveguides, we ruled out that the periodicity in $nk$ is due to threshold effects associated with the opening of new propagation channels. In an attempt to understand the $Q$-factor oscillations at the level of individual modes, we have singled out some high-$Q$ examples for closer investigation. In Fig. \[fig:Wavenumber-sweeps\], the highest-$Q$ mode occurs at $nk\approx105.6$. It is also shown in Fig. \[fig:Field-intensities-forPeaks\] (d) at $k\approx44$. By following this mode over a neighborhood of deformation parameters $\epsilon$ around the shown values, we found that no crossings or anticrossings with other modes occur. Such anticrossings can in principle lead to non-monotonic $Q$-factors as a function of a system parameter, in particular when whispering-gallery and chaotic modes coexist because they often depend differently on said parameters [@Hackenbroich97].
Husimi plots provide phase-space information about the classical structures supporting the folded chaotic WGMs, but their resolution is limited by the Fourier uncertainty relation between position $\phi$ and angular momentum as measured by $\sin\chi$ [@NoeckelChapter2002]. Individual modes also vary in the relative weighting between the neighborhoods of the five- and six-bounce orbits, suggesting that an analysis of the Q-factor oscillations in Fig. \[fig:Wavenumber-sweeps\] is best done not at the level of individual modes, but by adopting a more global point of view.
Such a global perspective can often be obtained by investigating the connection between the classical periodic orbits and the density of states [@BrackBook1997]. In open systems, the density of states is continuous but can exhibit resonant structure because it is proportional to the Wigner-Smith delay time [@NoeckelChapter2002]. Oscillations in the number of high-$Q$ modes could be interpreted as oscillations in a density of states in which short-lived resonances are not counted. In order to relate these oscillations to periodic orbits, we explored the approach of reference [@PhysRevLett.70.568], where beats between a small number of trajectories with similar actions in the periodic orbit sum were successfully used to interpret structure in the density of states for a H[é]{}non-Heiles Hamiltonian. Using the period-orbit lengths $L_{1}$ and $L_{2}$ from Fig. \[fig:Two-self-intersecting-periodicOrbits\] (a, b) to obtain the actions $S_{1,2}=nkL_{1,2}$, we arrive at an estimate for the beating period in Fig. \[fig:Wavenumber-sweeps\] of $$\Delta(nk)\approx\frac{2\pi}{\left|L_{1}-L_{2}\right|}\approx30\label{eq:DeltaKnperiod}$$ at refractive index $n=2.4$. Although this is in reasonable agreement with the spacing between the first two peaks in Fig. \[fig:Wavenumber-sweeps\], we have not yet been able to reproduce the correct dependence of the numerically observed peak positions on deformation parameter $\epsilon$. In particular, we have found additional periodic orbits with larger values of $\left|L_{1}-L_{2}\right|$ that should appear with comparable amplitudes in the periodic-orbit sum for the density of states, so it remains to be seen if a refined version of Eq. (\[eq:DeltaKnperiod\]) will be able to preserve the appealing simplicity of this interpretation in terms of periodic-orbit beats.
Conclusions
===========
We have numerically calculated the quasi-bound state wave functions of a series of planar dielectric cavities with attached waveguides that support the formation of folded chaotic whispering-gallery modes. We considered a large range of wavelengths and refractive indices, using shape deformations that combine two convex halves in a non-focussing configuration. It is possible to sustain useful quality factors ($10^{3}$ or larger) despite the fact that the ray dynamics permits long-lived, trapped trajectories only for a discrete set of unstable periodic orbits. The figure-eight topology of these orbits imprints itself on the wave solutions even though there is no one-to-one correspondence between individual modes and single periodic orbits.
As a result of this folding, the modes are confined by total internal reflection in the uninterrupted convex parts of the resonator boundary, while at the same time exhibiting a waist that reduces their overlap with the opening to the attached waveguides. Because these openings are necessarily in violation of Mather’s theorem, the existence of folded chaotic WGMs with Q-factors larger than $10^{3}$ is a wave localization effect. This localization is visualized with the help of Husimi projections which show that the long-lived modes are supported by a region of the ray phase space that coincides with the heteroclinic/homoclinic tangles of the unstable periodic orbits whose intensity pattern is also discernible in the real-space wave plots. In contrast to previous studies of dynamical localization in chaotic whispering-gallery like cavities, the angular momentum of the rays is not a slowly diffusing variable for our folded modes; it is the chaotic manifold structure that allows the wave modes to remain localized away from all escape windows in phase space.
As seen in the real-space plots, there are varying degrees of emission into the free space surrounding the structures. In this paper, the focus has been on the fact that high Q can be preserved in the presence of waveguide openings, because the latter are more intrusive than the curvature-induced violation of the total-internal reflection condition at the dielectric-air interface.
Having characterized the long-lived states for a range of deformations based on the resonator shape given in Eq. (\[eq:ShapeCurve\]), explorations of modified shapes are currently in progress, with the additional goal of further elucidating the Q-factor peaks discussed in subsection \[subsec:Q-factor-oscillations\]. By breaking the reflection symmetry of the cavity and going to higher wavenumbers in the wave simulations, we expect to improve upon the diagnostic value of the Husimi projections. We have already performed simulations for structures without spatial symmetries in order to make contact with previous work on unidirectional coupling [@MicrospiralWG], but additional work is needed to optimize the Q factors of those shapes.
A complementary approach to the quasibound-state analysis presented here is to investigate the resonances in transmission or reflection with the waveguides as input and output. In the stadium as the prototypical chaotic billiard, the transmission through attached leads has been studied in [@berggrenStadium] in the context of electronic transport, i.e., with impenetrable walls. Sharp resonances are found to be associated with all regions of the chaotic phase space, leading to spectral statistics governed by level repulsion. In our system, the only long-lived modes are of the folded chaotic WGM type, and their groupings shown in Fig. \[fig:Wavenumber-sweeps\] are not described by a universal random-matrix distribution. This goes hand in hand with the observation that no anti-crossings occur when varying the deformation parameter $\epsilon$ of Eq. (\[eq:ShapeCurve\]), indicating that all modes respond to such variations in unison. A study of the transmission statistics for waveguide-coupled resonators of this type will provide insight into the interplay of the two types of escape windows in the phase space depicted in Fig. \[fig:Poincare-surface-of-section\], where chaotic ray dynamics dominates similarly to the stadium, while at the same time the mode structure appears to be much simpler provided that low-Q states are discarded.
Acknowledgment
==============
This work benefited from access to the University of Oregon high performance computer, Talapas.
Boundary Integral Implementation
================================
Outline of the method
---------------------
To discover and investigate the resonant modes described, we developed software based on the boundary-integral approach as described in detail by Heider in Ref. [@HeiderBIM]. The method is similar to that described in Ref. [@Wiersig2003], which we have also applied to our system. The latter invariably produces spurious modes that make a large-scale analysis of the spectrum difficult, whereas the former does not produce any spurious modes at all (to our knowledge). This significant advantage can be traced back to the fact that Heider solves a set of simultaneous disretized of which one half differs from Ref. [@Wiersig2003] in that it is obtained by taking an additional normal derivative of the integral equations for the fields in terms of their values derivatives on the dielectric interface. The price to be paid for this advantage is that great care must be taken in in numerically dealing with the singularities not just in the Green functions but also their derivatives. The following subsections address these details.
Our software is written using the Julia programming language, which was chosen for its numeric performance, access to the necessary special-function libraries with implementations for complex arguments, and ease of integration with other numerical computing software. Our implementation is publicly available on GitHub [@GitHub]. It has been written for reusability and allows the user to specify their own boundary parameterizations. The implementation is written to utilize multiple processors on a single machine to parallelize operation where possible and has been tested on OS X and Linux.
To characterize the spectral patterns for folded chaotic WGMs, it was necessary to scan a wide range of $k$-space. Although the direct sweep method has the ability to find multiple resonances near an initial starting guess $k_{0}$, we observed that the ability to discern and identify relevant starting $k$ values for the secondary RII step decreases rapidly as $\left\Vert k_{0}-k\right\Vert $ increases. This is especially true for high-Q modes, as the resonance width for these modes are very narrow and we empirically observed that the direct sweep method discovers wider resonances in the vicinity of narrow resonances more easily which may mask the presence of the narrow resonances. Therefore, in order to reliably find the high-Q modes it was necessary to adopt a scanning procedure in which small regions of $k$-space were examined piecewise and candidate high-Q modes were identified within those small regions. We used the following algorithm:
1. Given a large $k$ range of interest from $k_{\text{min}}$ to $k_{\text{max}}$, choose a small step size ($k_{\text{step}}=.02$) as a discretization.
2. Choose a target $\kappa$ for the imaginary part of $k$. We chose a small value ($\kappa=.0001$) to target high-Q modes.
3. Starting at $k_{0}=k_{\text{min}}-i\thinspace\kappa$, execute the direct sweep procedure to generate candidate resonances.
4. Filter the candidate resonances to keep only those where $k_{0}-k_{\text{step}}\leq k_{\text{cand}}\leq k_{0}+k_{\text{step}}$, as they will be found again by a closer $k_{0}$ if outside that range and are less likely to converge to a valid resonance further away from $k_{0}$.
5. Increment $k_{0}$ by $k_{\text{step}}$ and repeat this procedure until the desired range has been covered.
Because candidate $k$ values were retained both above and below the center point $k_{0}$, there was overlap between adjacent center points and most of the resonances which would eventually be discovered by the RII procedure would appear twice in the list of candidates. However, since different starting $k_{0}$ values were used in the direct sweep procedure, the exact values of these “duplicate” $k_{\text{cand}}$ were numerically slightly different as the candidates were discovered by a linear approximation to the eigenvalue problem around the vicinity of the initial guess $k_{0}$. In practice, when resolving the more accurate $k_{f}$ and the boundary field $x_{f}$ using the RII procedure, these slightly different starting points would converge to the same resonance within the numerical tolerances specified for the RII algorithm as well as display an identical spatial field pattern so we concluded that they were in fact the same mode. Therefore in this system it would have been possible to be more computationally efficient by combining or averaging very close candidate $k$ values found in adjacent regions of $k$-space spanned by the above algorithm, or in eliminating the overlap and keeping candidates at each step in the range $k_{0}-k_{\text{step}}/2\leq k_{\text{cand}}\leq k_{0}+k_{\text{step}}/2$ or $k_{0}\leq k_{\text{cand}}\leq k_{0}+k_{\text{step}}$. We chose to err on the side of the extra computation in the interest of not missing any important resonances in the scan. It should be noted that there are other boundary geometries (such as the Reuleaux billiard, [@PhysRevE.90.022903]) which exhibit very closely spaced doublets for which a careful approach may be necessary to avoid missing resonances.
Once the list of candidate $k_{0}$ values had been generated, the final $k_{f}$ and $x_{f}$ values were determined by first filtering the candidates to exclude very low-Q modes with $\kappa>1.0$ and then running the direct sweep procedure with starting guess $k_{0}=k_{\text{cand}}$ for each remaining candidate. The reason for the repeated direct sweep procedure was that since a large number of candidates were initially generated, we did not store the boundary fields $x$ associated with them. Since the input to the RII algorithm requires both the starting $k_{0}$ and the boundary field $x_{0}$, we needed to regenerate the boundary field. Since the initial candidates $k_{0}$ were already close to their final $k$, the subsequent direct sweep would in general produce inputs to the RII procedure that were even closer to the final convergent values and thus few iterations of the RII loop were required.
Some corrections to previous work
---------------------------------
We discovered two errors in the equations listed in the Appendix of Ref. [@HeiderBIM]. The first appears in the description of $\tilde{N}$ below equation (A.6). There is a sign error in the last term. The corrected equation is:
$$\begin{aligned}
\tilde{N}(t,\tau) & = & \frac{i}{2}\bar{N}(t,\tau)\left\{ (k\,n_{e})^{2}H_{0}^{(1)}(k\,n_{e}\left|x(t)-x(\tau)\right|)-\frac{2k\,n_{e}H_{1}^{(1)}(k\,n_{e}\left|x(t)-x(\tau)\right|)}{\left|x(t)-x(\tau)\right|}\right\} \\
& & +\frac{i}{2}\frac{k\,n_{e}x'(t)x'(\tau)}{\left|x(t)-x(\tau)\right|}H_{1}^{(1)}(k\,n_{e}\left|x(t)-x(\tau)\right|)\\
& & -\frac{i}{2}\bar{N}(t,\tau)\left\{ (k\,n_{i})^{2}H_{0}^{(1)}(k\,n_{i}\left|x(t)-x(\tau)\right|)-\frac{2k\,n_{i}H_{1}^{(1)}(k\,n_{i}\left|x(t)-x(\tau)\right|)}{\left|x(t)-x(\tau)\right|}\right\} \\
& & -\frac{i}{2}\frac{k\,n_{i}x'(t)x'(\tau)}{2\left|x(t)-x(\tau)\right|}H_{1}^{(1)}(k\,n_{i}\left|x(t)-x(\tau)\right|)\end{aligned}$$
The other error is in the equation for the diagonal terms $\tilde{N}_{2}(t,t)$, where the log terms should not use the norm of $x'(t)$ squared, but simply the norm.The corrected equation is: $$\begin{aligned}
\tilde{N_{2}}(t,t) & =\frac{\left|x'(t)\right|^{2}}{4\pi}\left[((k\,n_{e})^{2}-(k\,n_{i})^{2})(\pi i-1-2C)\right]\\
& +\frac{\left|x'(t)\right|^{2}}{4\pi}\left[-2(k\,n_{e})^{2}\text{{ln}}\left(\frac{k\,n_{e}\left|x'(t)\right|}{2}\right)+2(k\,n_{i})^{2}\text{{ln}}\left(\frac{k\,n_{i}\left|x'(t)\right|}{2}\right)\right]\end{aligned}$$
Derivatives of the integral kernels
-----------------------------------
Heider’s boundary integral approach describes the $A$ matrix and its use in iteratively finding a resonant solution to the boundary integral equations. This iterative procedure requires the use of $A'(k)$, both in the direct sweep procedure (Algorithm 3) and the Newton’s iteration step in the RII procedure (Algorithm 1). The derivation of the derivative matrix is straightforward and the results are outlined below, in terms of the constituent equations provided in the paper’s appendix.
The $A'(k)$ matrix is defined as
$$A'(k)=\begin{bmatrix}-(K_{e}'(k)-K_{i}'(k)) & -(S_{e}'(k)-S_{i}'(k))\\
T_{e}'(k)-T_{i}'(k) & K_{e}^{*}{}'(k)-K_{i}^{*}{}'(k)
\end{bmatrix}$$
with the $K,S,T,$ and $K^{*}$ integral operators being expressed after singularity subtraction in terms of equations $$\tilde{H}_{1}(k),\tilde{H}_{2}(k),\tilde{H}(k),\tilde{M}_{1}(k),\tilde{M}_{2}(k),\tilde{M}(k),\tilde{N}_{1}(k),\tilde{N}_{2}(k),\tilde{N}(k),\tilde{L}_{1}(k),\tilde{L}_{2}(k),\tilde{L}(k).$$
All other equations given in the appendix are independent of the wavenumber $k$.
First the $\tilde{H}$ equations:
$$\begin{aligned}
\partial_{k}\tilde{H}_{1}(t,\tau,k) & = & -\frac{1}{2\pi}n(\tau)(x(t)-x(\tau))\left[k\,n_{e}^{2}J_{0}(k\,n_{e}\left|x(t)-x(\tau)\right|)-k\,n_{i}^{2}J_{0}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right].\\
\partial_{k}\tilde{H}_{1}(t,t,k) & = & 0\\
\partial_{k}\tilde{H}_{2}(t,\tau,k) & = & \partial_{k}\tilde{H}(t,\tau,k)-\partial_{k}\tilde{H}_{1}(t,\tau,k)\ \text{{ln}}\left(4\,\text{sin}^{2}\frac{t-\tau}{2}\right)\\
\partial_{k}\tilde{H}_{2}(t,\tau,k) & = & \partial\tilde{H}(t,\tau,k)-\partial_{k}\tilde{H}_{1}(t,\tau,k)\ \text{{ln}}\left(4\,\text{sin}^{2}\frac{t-\tau}{2}\right)\\
\partial_{k}\tilde{H}_{2}(t,t,k) & = & 0\\
\partial_{k}\tilde{H}(t,\tau,k) & = & \frac{i}{2}n(\tau)\left(x(t)-x(\tau)\right)\left[k\,n_{e}^{2}H_{0}^{(1)}(k\,n_{e}\left|x(t)-x(\tau)\right|)-k\,n_{i}^{2}H_{0}^{(1)}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right]\end{aligned}$$
The $\tilde{H}_{1}^{*}$,$\tilde{H}_{2}^{*}$, and $\tilde{H}^{*}$, equations used in the starred operators $K^{*}$ are defined in terms of the unstarred equations with no additional $k$ dependence so their derivatives can use the results above.
The $\tilde{M}$ equations are:
$$\begin{aligned}
\partial_{k}\tilde{M}_{1}(t,\tau,k) & = & \frac{\left|x(t)-x(\tau)\right|}{2\pi}\left[n_{e}J_{1}(k\,n_{e}\left|x(t)-x(\tau)\right|)-n_{i}J_{1}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right]\\
\partial_{k}\tilde{M}_{1}(t,t,k) & = & 0\\
\partial_{k}\tilde{M}_{2}(t,\tau,k) & = & \partial_{k}\tilde{M}(t,\tau,k)-\partial_{k}\tilde{M}_{1}(t,\tau,k)\ \text{{ln}}\left(4\,\text{sin}^{2}\frac{t-\tau}{2}\right)\\
\partial_{k}\tilde{M}_{2}(t,t,k) & = & 0\\
\partial_{k}\tilde{M}(t,\tau,k) & = & -\frac{i\left|x(t)-x(\tau)\right|}{2}\left[n_{e}H_{1}^{(1)}(k\,n_{e}\left|x(t)-x(\tau)\right|)-n_{i}H_{1}^{(1)}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right]\end{aligned}$$
The $\tilde{N}$ equations are:
$$\begin{aligned}
\partial_{k}\tilde{N}(t,\tau,k) & = & -\frac{i}{2}\left|x(t)-x(\tau)\right|\bar{N}(t,\tau)\left[k\,n_{e}^{3}H_{1}^{(1)}(k\,n_{e}\left|x(t)-x(\tau)\right|)-k\,n_{i}^{3}H_{1}^{(1)}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right]\\
\qquad\qquad & & +\frac{i}{2}x'(t)x'(\tau)\left[k\,n_{e}^{2}H_{1}^{(1)}(k\,n_{e}\left|x(t)-x(\tau)\right|)-k\,n_{i}^{2}H_{1}^{(1)}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right]\\
\partial_{k}\tilde{N}_{1}(t,\tau,k) & = & \frac{1}{2\pi}\left|x(t)-x(\tau)\right|\bar{N}(t,\tau)\left[k^{2}n_{e}^{3}J_{1}(k\,n_{e}\left|x(t)-x(\tau)\right|)-k^{2}n_{i}^{3}J_{1}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right]\\
& & -\frac{1}{2\pi}x'(t)x'(\tau)\left[k\,n_{e}^{2}J_{0}(k\,n_{e}\left|x(t)-x(\tau)\right|)-k\,n_{i}^{2}J_{0}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right]\\
\partial_{k}\tilde{N}_{1}(t,t,k) & = & \frac{\left|x'(t)\right|^{2}k\left(n_{i}^{2}-n_{e}^{2}\right)}{2\pi}\\
\partial_{k}\tilde{N_{2}}(t,\tau,k) & = & \partial_{k}\tilde{N}(t,\tau,k)-\partial_{k}\tilde{N_{1}}(t,\tau,k)\ \text{{ln}}\left(4\,\text{sin}^{2}\frac{t-\tau}{2}\right)\\
\partial_{k}\tilde{N}_{2}(t,t,k) & = & \frac{\left|x'(t)\right|^{2}}{2\pi}\left[-2k\,n_{e}^{2}\text{\,{ln}\ensuremath{\left(\frac{k\,n_{e}\left|x'(t)\right|}{2}\right)}}+2k\,n_{i}^{2}\text{\,{ln}\ensuremath{\left(\frac{k\,n_{i}\left|x'(t)\right|}{2}\right)}}-k(n_{e}^{2}-n_{i}^{2})(2+2C-\pi i)\right]\end{aligned}$$
As in the reference, $C\approx0.5772156649$ refers to the Euler–Mascheroni constant.
Finally, the $\tilde{L}$ equations: $$\begin{aligned}
\partial_{k}\tilde{L}_{1}(t,\tau,k) & = & \frac{\left|x(t)-x(\tau)\right|}{2\pi}\left[(k^{2}n_{e}^{3}J_{1}(k\,n_{e}\left|x(t)-x(\tau)\right|)-k^{2}n_{i}^{3}J_{1}(k\,n_{i}\left|x(t)-x(\tau)\right|))\right]\\
& & -\frac{1}{\pi}\left[k\,n_{e}^{2}J_{0}(k\,n_{e}\left|x(t)-x(\tau)\right|))-k\,n_{i}^{2}J_{0}(k\,n_{i}\left|x(t)-x(\tau)\right|))\right]\\
\partial_{k}\tilde{L}_{1}(t,t,k) & = & \frac{k(n_{i}^{2}-n_{e}^{2})}{\pi}\\
\partial_{k}\tilde{L}_{2}(t,\tau,k) & = & \partial_{k}\tilde{L}(t,\tau,k)-\partial_{k}\tilde{L}_{1}(t,\tau,k)\ \text{{ln}}\left(4\,\text{sin}^{2}\frac{t-\tau}{2}\right)\\
\partial_{k}\tilde{L}_{2}(t,t,k) & = & -\frac{1}{\pi}\left[2k\,n_{e}^{2}\text{\,{ln}}(k\,n_{e})-2k\,n_{i}^{2}\text{\,{ln}}(k\,n_{i})+k\left(n_{e}^{2}-n_{i}^{2}\right)\left(2\text{\,{ln}}(\frac{\left|x'(t)\right|}{2})-\pi i+2C+1\right)\right]\\
\partial_{k}\tilde{L}(t,\tau,k) & = & -\frac{i}{2}\left|x(t)-x(\tau)\right|\left[k^{2}n_{e}^{3}H_{1}^{(1)}(k\,n_{e}\left|x(t)-x(\tau)\right|)-k^{2}n_{i}^{3}H_{1}^{(1)}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right]\\
& & +i\left[k\,n_{e}^{2}H_{0}^{(1)}(k\,n_{e}\left|x(t)-x(\tau)\right|)-k\,n_{i}^{2}H_{0}^{(1)}(k\,n_{i}\left|x(t)-x(\tau)\right|)\right]\end{aligned}$$ It is hoped that the explicit formulas given here provide a complete picture of the computational method we employed in finding the quasibound states.
[38]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, [ ****, ()](https://link.aps.org/doi/10.1103/RevModPhys.87.61).
, [ ****, ()](http://dx.doi.org/10.1038/385045a0).
, , , , , , , , , [ ****, ()](http://dx.doi.org/10.1063/1.4809724).
, ** (, , ).
, , , [ ****, ()](https://link.aps.org/doi/10.1103/PhysRevB.54.10652).
, [ ****, ()](http://dx.doi.org/10.1017/S0143385700002455).
, [ ****, ()](https://doi.org/10.1007/BF01197884).
, , , [ ****, ()](http://ol.osa.org/abstract.cfm?URI=ol-19-21-1693).
, , , , [ ****, ()](https://avs.scitation.org/doi/abs/10.1116/1.590415).
, , , , , , , , , [ ****, ()](http://www.opticsexpress.org/abstract.cfm?URI=oe-21-13-16069).
, [ ****, ()](http://dx.doi.org/10.1109/3.622627).
, , , , , [ ****, ()](http://doi.org/101016/j.optlastec.2013.08.009).
, , , , [ ****, ()](https://doi.org/10.1103/PhysRevLett.108.243902).
, , , , , , , , [ ****, ()](http://www.osapublishing.org/optica/abstract.cfm?URI=optica-5-5-612).
, , , , [ ****, ()](https://doi.org/10.1364/OE.22.000824).
, , , , , , [ ****, ()](http://doi.org/10.1109/LPT.2010.2057417).
, , , , , , [ ****, ()](https://doi.org/10.1007/s11431-009-0306-y).
, , , , , , , , [ ****, ()](http://doi.org/10.1109/JSTQE.2013.2244566).
, , , [ ****, ()]({http://ol.osa.org/abstract.cfm?URI=ol-32-23-3397}).
, , , [ ****, ()]({http://stacks.iop.org/1402-4896/T90/134}).
, , , , , , , , [ ****, ()](http://www.sciencemag.org/content/280/5369/1556.abstract).
, ** (, , ).
, , , , [ ****, ()](http://www.opticsexpress.org/abstract.cfm?URI=oe-13-15-5641).
, [ ****, ()](http://stacks.iop.org/1402-4896/T90/263).
, , , , [ ****, ()](https://doi.org/10.1103/PhysRevE.88.022916).
, [ ****, ()](https://doi.org/10.1080/0020739920230607).
, [ ****, ()](https://doi.org/10.1016/j.camwa.2010.06.044).
, in [ **, edited by (, )](https://arxiv.org/abs/physics/0203063).
, [ ****, ()](http://www.sciencedirect.com/science/article/pii/S0079656500000327).
, , , , , , [ ****, ()](http://www.sciencedirect.com/science/article/pii/S001046550900383X).
, ** (, , ).
, in [ **, edited by (, )](https://arxiv.org/abs/physics/0406134), vol. of **, pp. .
, , , , , [ ****, ()](https://link.aps.org/doi/10.1103/PhysRevE.90.022903).
, [ ****, ()](https://link.aps.org/doi/10.1103/PhysRevLett.97.253901).
, , , [ ****, ()](https://doi.org/10.1364/OL.32.001764).
, , , [ ****, ()](https://link.aps.org/doi/10.1103/PhysRevE.78.016201).
, [ ****, ()](http://stacks.iop.org/0295-5075/39/371).
, ** (, , ).
, , , , [ ****, ()](https://link.aps.org/doi/10.1103/PhysRevLett.70.568).
, , , , [ ****, ()](http://www.sciencedirect.com/science/article/pii/S0010465501003204).
, [ ****, ()](http://stacks.iop.org/1464-4258/5/53).
<https://github.com/kahliburke/BoundaryIntegralMethod.jl>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper we study the boundary controllability of the Gear-Grimshaw system posed on a finite domain $(0,L)$, with Neumann boundary conditions: $$\label{abs}
\begin{cases}
u_t + uu_x+u_{xxx} + a v_{xxx} + a_1vv_x+a_2 (uv)_x =0, & \text{in} \,\, (0,L)\times (0,T),\\
c v_t +rv_x +vv_x+abu_{xxx} +v_{xxx}+a_2buu_x+a_1b(uv)_x =0, & \text{in} \,\, (0,L)\times (0,T), \\
u_{xx}(0,t)=h_0(t),\,\,u_x(L,t)=h_1(t),\,\,u_{xx}(L,t)=h_2(t), & \text{in} \,\, (0,T),\\
v_{xx}(0,t)=g_0(t),\,\,v_x(L,t)=g_1(t),\,\,v_{xx}(L,t)=g_2(t), & \text{in} \,\, (0,T),\\
u(x,0)= u^0(x), \quad v(x,0)= v^0(x), & \text{in} \,\, (0,L).\nonumber
\end{cases}$$ We first prove that the corresponding linearized system around the origin is exactly controllable in $(L^2(0,L))^2$ when $h_2(t)=g_2(t)=0$. In this case, the exact controllability property is derived for any $L>0$ with control functions $h_0, g_0\in H^{-\frac{1}{3}}(0,T)$ and $h_1, g_1\in L^2(0,T)$. If we change the position of the controls and consider $h_0(t)=h_2(t)=0$ (resp. $g_0(t)=g_2(t)=0)$ we obtain the result with control functions $g_0, g_2\in H^{-\frac{1}{3}}(0,T)$ and $h_1, g_1\in L^2(0,T)$ if and only if the length $L$ of the spatial domain $(0,L)$ belongs to a countable set. In all cases the regularity of the controls are sharp in time. If only one control act in the boundary condition, $h_0(t)=g_0(t)=h_2(t)=g_2(t)=0$ and $g_1(t)=0$ (resp. $h_1(t)=0$), the linearized system is proved to be exactly controllable for small values of the length $L$ and large time of control $T$. Finally, the nonlinear system is shown to be locally exactly controllable *via* the contraction mapping principle, if the associated linearized systems are exactly controllable.
address:
- |
Department of Mathematics\
Federal University of Pernambuco, UFPE,\
CEP 50740-545, Recife, PE\
Brazil.
- |
Institute of Mathematics\
Federal University of Rio de Janeiro, UFRJ\
P.O. Box 68530, CEP 21945-970, Rio de Janeiro, RJ\
Brazil.
- |
Institute of Mathematics\
Federal University of Rio de Janeiro, UFRJ\
P.O. Box 68530, CEP 21945-970, Rio de Janeiro, RJ\
Brazil.
author:
- 'Roberto A. Capistrano–Filho'
- 'Fernando A. Gallego'
- 'Ademir F. Pazoto'
title: 'Neumann Boundary Controllability of the Gear–Grimshaw System With Critical Size Restrictions on the Spacial Domain'
---
Introduction
============
Setting of the Problem
----------------------
The goal of this paper is to investigate the boundary controllability properties of the nonlinear dispersive system $$\label{gg1}
\begin{cases}
u_t + uu_x+u_{xxx} + a v_{xxx} + a_1vv_x+a_2 (uv)_x =0, & \text{in} \,\, (0,L)\times (0,T),\\
c v_t +rv_x +vv_x+abu_{xxx} +v_{xxx}+a_2buu_x+a_1b(uv)_x =0, & \text{in} \,\, (0,L)\times (0,T), \\
u(x,0)= u^0(x), \quad v(x,0)= v^0(x), & \text{in} \,\, (0,L),
\end{cases}$$ with the following boundary conditions $$\label{gg2}\begin{cases}
u_{xx}(0,t)=h_0(t),\,\,u_x(L,t)=h_1(t),\,\,u_{xx}(L,t)=h_2(t),\\
v_{xx}(0,t)=g_0(t),\,\,v_x(L,t)=g_1(t),\,\,v_{xx}(L,t)=g_2(t).
\end{cases}$$ In , $a_1, a_2, a, b, c$ and $r$ are real constants, $u=u(x,t)$ and $v=v(x,t)$ are real-valued functions of the two variables $x$ and $t$ and subscripts indicate partial differentiation. The boundary functions $h_i$ and $g_i$, for $i=0,1,2$, are considered as control inputs acting on the boundary conditions. Our purpose is to see weather we can force the solutions of the system to have certain properties by choosing appropriate control inputs. More precisely, we are mainly concerned with the following exact control problem: 0.2 cm [*Given $T>0$ and $u^0,v^0,u^1,v^1\in L^2(0,L)$, can one find appropriate control inputs $h_i$, $g_i$, for $i=0,1,2$, such that the corresponding solution $(u,v)$ of - satisfies $$\label{exactcontrol'a}
(u(x,T),v(x,T))=(u^1(x),v^1(x))?$$*]{}
In order to provide the tools to handle with this problem, we assume that the coefficients $a, b, c$ and $r$ satisfy $$\label{coef}
b, c \mbox{ and } r \mbox{ are positive and } 1-a^2b >0.$$
System was derived by Gear and Grimshaw in [@geargrimshaw1984] as a model to describe strong interactions of two long internal gravity waves in a stratified fluid, where the two waves are assumed to correspond to different modes of the linearized equations of motion (we also refer to [@BoPoSaTo; @SaTz] for an extensive discussion on the physical relevance of the system). This somewhat complicated model has the structure of a pair of Korteweg-de Vries (KdV) equations coupled through both dispersive and nonlinear effects and has been object of intensive research in recent year. It is a special case of a broad class of nonlinear evolution equations for which the well-posedness theory associated to the pure initial-value problem posed on the whole real line $\mathbb{R}$, or on a finite interval with periodic boundary conditions, has been intensively investigated. By contrast, the mathematical theory pertaining to the study of the boundary value problem is considerably less advanced, specially in what concerns the study of the controllability properties. As far as we know, the controllability results for system was first obtained in [@MiOr], when the model is posed on a periodic domain and $r=0$. In this case, a diagonalization of the main terms allows to the decouple the corresponding linear system into two scalar KdV equations and use the previous results available in the literature. Later on, assuming that holds, Micu *et al.* [@micuortegapazoto2009] proved the local exact boundary controllability property for the nonlinear system, posed on a bounded interval, considering the following boundary conditions: $$\label{gg3}\begin{cases}
u(0,t)=0,\,\,u(L,t)=f_1(t),\,\,u_x(L,t)=f_2(t),\\
v(0,t)=0,\,\,v(L,t)=k_1(t),\,\,v_x(L,t)=k_2(t).
\end{cases}$$ The analysis developed in [@micuortegapazoto2009] was inspired by the results obtained by Rosier in [@rosier] for the scalar KdV equation. It combines the analysis of the linearized system and the Banach’s fixed point theorem. Following the classical duality approach [@dolecki; @lions], the exact controllability of system linearized system is equivalent to an observability for the adjoint system. Then, the problem is reduced to prove a nonstandard unique continuation property of the eigenfunctions of the corresponding differential operator. Their main result reads as follows: 0.2 cm [**Theorem A** ]{}(Micu *et al.* [@micuortegapazoto2009]) [*Let $L>0$ and $T >0$. Then there exists a constant $\delta>0$, such that, for any initial and final data $u^0, v^0, u^1, v^1\in L^2(0, L)$ verifying $$||(u^0, v^0)||_{(L^2(0,L))^2}\leq\delta \quad\text{and}\quad ||(u^1, v^1)||_{(L^2(0,L))^2}\leq\delta,$$ there exist four control functions $f_1, k_1\in H^1_0(0, T)$ and $f_2, k_2 \in L^2(0,T)$, such that the solution $$(u, v)\in C([0,T];(L^2(0,L))^2)\cap L^2(0,T;(H^1(0, L))^2)\cap H^1(0,T;(H^{-2}(0, L))^2)$$ of - verifies .*]{}
Later on, the same problem was addressed by Cerpa and Pazoto [@cerpapazoto2011] when only two controls act on the Neumann boundary conditions, i.e., assuming that $f_1=k_1=0$. In this case, the analysis of the linearized system is much more complicated, therefore the authors used a direct approach based on the multiplier technique that gives the observability inequality for small values of the length $L$ and large time of control $T$. The fixed point argument, as well as, the existence and regularity results needed in order to consider the nonlinear system run exactly in the same way as in [@micuortegapazoto2009].
The program of this work was carried out for a particular choice of boundary conditions and aims to establish as a fact that such a model predicts the interesting controllability properties initially observed for the KdV equation. Therefore, to introduce the reader to the theory developed for KdV with the boundary conditions of types and , we present below a summary of the results achieved in [@rosier] and [@caicedo_caspistrano_zhang_2015], respectively.
Rosier, in [@rosier], studied the following boundary control problem for the KdV equation posed on the finite domain $(0,L)$ $$\left\{
\begin{array}
[c]{lll}u_t+u_x+uu_x+u_{xxx}=0 & & \text{ in } (0,L)\times(0,T),\\
u(0,t)=0,\text{ }u(L,t)=0,\text{ }u_x(L,t)=g(t) & & \text{ in }(0,T),\\
u(x,0)=u^0(x) & & \text{ in }(0,L),
\end{array}
\right. \label{2}$$ where the boundary value function $g(t)$ is considered as a control input. First, the author studies the associated linear system $$\left\{
\begin{array}
[c]{lll}u_t+u_x+u_{xxx}=0 & & \text{ in } (0,L)\times(0,T),\\
u(0,t)=0,\text{ }u(L,t)=0,\text{ }u_x(L,t)=g(t) & & \text{ in }(0,T),\\
u(x,0)=u^0(x) & & \text{ in }(0,L)
\end{array}
\right. \label{2a}$$ and discovered the so-called [*critical length*]{} phenomena, i.e., whether the system (\[2a\]) is exactly controllable depends on the length $L$ of the spatial domain $(0,L)$. More precise, the following result was proved:
[**Theorem B**]{} (Rosier [@rosier]) [*The linear system is exactly controllable in the space $L^2(0,L)$ if and only if the length $L$ of the spatial domain $(0,L)$ does not belong to the set*]{} $$\mathcal{N}:=\left\{ \frac{2\pi}{\sqrt{3}}\sqrt{k^{2}+kl+l^{2}}
\,:k,\,l\,\in\mathbb{N}^{\ast}\right\} . \label{critical}$$
Then, by using a fixed point argument, the controllability result was extended to the nonlinear system when $L\notin\mathcal{N}$. 0.2 cm [**Theorem C**]{} (Rosier [@rosier])*Let $T>0$ be given. If $L\notin\mathcal{N}$, there exists $\delta>0$, such that, for any $u^0,u^T\in L^2(0,L)$ with $$||u^0||_{L^2(0,L)}+||u^T||_{L^2(0,L)}\leq\delta,$$ one can find a control input $g\in L^2(0,T)$, such that the nonlinear system admits a unique solution $$u\in C([0,T];L^2(0,L))\cap L^2(0,T;H^1(0,L))$$ satisfying $$u(x,T)=u^T(x).$$*
More recently, in [@caicedo_caspistrano_zhang_2015], Caicedo *et al.* investigated the boundary control problem of the KdV equation with new boundary conditions, namely, the Neumann boundary conditions: $$\left\{
\begin{array}
[c]{lll}u_t+ (1+\beta )u_x+u_{xxx}=0 & & \text{ in } (0,L)\times(0,T),\\
u_{xx}(0,t)=0,\text{ }u_x(L,t)=h(t),\text{ }u_{xx}(L,t)=0 & & \text{ in }(0,T),\\
u(x,0)=u^0(x) & & \text{ in }(0,L).
\end{array}
\right. \label{gg5}$$ In , $\beta $ is a given real constant and $g$ a control input. For any $\beta \ne -1$, the authors obtained the following set of [*critical lengths*]{} $$\mathcal{R}_{\beta} :=\left\{ \frac{2\pi}{\sqrt{3(1+\beta)}}\sqrt{k^{2}+kl+l^{2}}
\,:k,\,l\,\in\mathbb{N}^{\ast}\right\}\cup\left\{\frac{k\pi}{\sqrt{\beta +1}}:k\in\mathbb{N}^{\ast}\right\},
\label{critical_new}$$ and proved that the following result holds: 0.2 cm [**Theorem D**]{} (Caicedo *et al.* [@caicedo_caspistrano_zhang_2015])
**
- If $\beta \ne -1$, the linear system (\[gg5\]) is exactly controllable in the space $L^2 (0,L)$ if and only if the length L of the spatial domain $(0, L)$ does not belong to the set $\mathcal{R}_{\beta}$.
- If $\beta =-1$, then the system (\[gg5\]) is not exact controllable in the space $L^2 (0,L)$ for any $L>0$.
In addition, for the nonlinear system $$\left\{
\begin{array}
[c]{lll}u_t+u_x+uu_x+u_{xxx}=0 & & \text{ in } (0,L)\times(0,T),\\
u_{xx}(0,t)=0,\text{ }u_x(L,t)=h(t),\text{ }u_{xx}(L,t)=0 & & \text{ in }(0,T),\\
u(x,0)=u_0(x) & & \text{ in }(0,L),
\end{array}
\right. \label{gg6}$$ the result below was proved by using a fixed point argument: 0.2 cm [**Theorem E**]{} (Caicedo *et al.* [@caicedo_caspistrano_zhang_2015]) [*Let $T>0$, $\beta \ne -1$ and $L\notin\mathcal{R}_{\beta} $ be given. There exists $\delta>0$, such that, for any $u^0,u^T\in L^2(0,L)$ with $$||u^0-\beta ||_{L^2(0,L)}+||u^T-\beta ||_{L^2(0,L)}\leq\delta,$$ one can find a control input $h\in L^2(0,T)$, such that the system admits unique solution $$u\in C([0,T];L^2(0,L))\cap L^2(0,T;H^1(0,L))$$ satisfying $$u(x,T)=u^T(x).$$*]{}
Both theorems, Theorems B and D, were proved following the classical duality approach [@dolecki; @lions] which reduces the problem to prove an observability inequality for the solutions of the corresponding adjoint system. Then, the controllability is obtained with the aid of a compactness argument that leads the issue to a nonstandard unique continuation principle for the eigenfunctions of the differential operator associated to the model. The critical lengths in and are such that there are eigenfunctions of the linear scalar problem for which the observability inequality associated to the adjoint system fails[^1]. However, in [@caicedo_caspistrano_zhang_2015], the authors encountered some difficulties that require special attention. For instance, the adjoint system of the linear system is given by $$\left\{
\begin{array}
[c]{lll}\psi_t+(1+\beta )\psi_x+\psi_{xxx}=0 & &\text{ in } (0,L)\times(0,T),\\
(1+\beta )\psi(0,t)+\psi_{xx}(0,t)=0& &\text{ in } (0,T),\\
(1+\beta )\psi(L,t)+\psi_{xx}(L,t)=0& &\text{ in } (0,T),\\
\psi_x(0,t)=0& &\text{ in } (0,T),\\
\psi(x,T)=\psi^T(x) & &\text{ in }(0,L).
\end{array}
\right. \label{gg6_a}$$ The exact controllability of system is equivalent to the following observability inequality for the adjoint system : $$||\psi^T||_{L^2(0,L)}\leq C||\psi_x(L,\cdot)||_{L^2(0,T)},$$ for some $C>0$. Nonetheless, the usual multiplier method and compactness arguments used to deal with the system only lead to $$||\psi^T||^2_{L^2(0,L)}\leq C_1||\psi_x(L,\cdot)||^2_{L^2(0,T)}+C_2||\psi(L,\cdot)||^2_{L^2(0,T)},\label{gg7_a}$$ where $C_1$ and $C_2$ are positive constants. In order to absorb the extra term present in , Caicedo *et al.* derived a technical result, which reveals some hidden regularity (sharp trace regularities) for solutions of the adjoint system : 0.2 cm [**Theorem F**]{} (Caicedo *et al.* [@caicedo_caspistrano_zhang_2015]) [*For any $\psi^T\in L^2(0,L)$, the solution $$\psi\in C([0,T];L^2(0,L))\cap L^2(0,T;H^1(0,L))$$ of the problem possesses the following sharp trace properties $$\sup_{x\in(0,L)}||\partial^r_x\psi(x,\cdot)||_{H^{\frac{1-r}{3}}(0,T)}\leq C_r||\psi^T||_{L^2(0,L)},
\label{9}$$ for $r=0,1,2$, where $C_r$ are positive constants.*]{}
0.2 cm
Estimate is then combined with compactness argument to remove the extra term in . We remark that the sharp Kato smoothing properties obtained by Kenig, Ponce and Vega [@KePoVe] for the solutions of the KdV equation posed on the line, played an important role in the proof of the previous result. The same strategy has been successfully applied by Cerpa *et al.* [@cerizh] for the study of a similar boundary controllability problem.
Main Result
-----------
We are now in position to return considerations to the control properties of the system . First, we prove that the corresponding linear system with the following boundary conditions $$\begin{cases}
u_{xx}(0,t)=h_0(t),\,\,u_x(L,t)=h_1(t),\,\,u_{xx}(L,t)=0,\\
v_{xx}(0,t)=g_0(t),\,\,v_x(L,t)=g_1(t),\,\,v_{xx}(L,t)=0,
\end{cases}$$ is exactly controllable in $(L^2(0,L))^2$ with controls $h_0$, $g_0\in H^{-\frac{1}{3}}(0,T)$ and $h_1$, $g_1\in L^2(0,T)$. In this case, any restriction on the length $L$ of the spatial domain is required. However, if we change the position of the controls a critical size restriction can appear. This is the case when we consider the following boundary conditions $$\begin{cases}
u_{xx}(0,t)=0,\,\,u_x(L,t)=h_1(t),\,\,u_{xx}(L,t)=0,\\
v_{xx}(0,t)=g_0(t),\,\,v_x(L,t)=g_1(t),\,\,v_{xx}(L,t)=g_2(t).
\end{cases}$$ In this case, the exact controllability result in $(L^2(0,L))^2$ is derived with controls $g_0$, $g_2\in H^{-\frac{1}{3}}(0,T)$ and $h_1$, $g_1\in L^2(0,T)$ if and only if the length $L$ does not belong of the following set
$${\mathcal{F}}_r:= \left\lbrace 2\pi k \sqrt{\frac{1-a^2b}{r}}: k \in {\mathbb{N}}^{*}\right\rbrace\cup \left\lbrace \pi \sqrt{\frac{(1-a^2b)\alpha(k,l,m,n,s)}{3r}}:k, l, m, n, s \in {\mathbb{N}}^{*} \right\rbrace,
\label{critical_f}$$
where $$\begin{aligned}
\alpha:=\alpha(k,l,m,n,s)=&5k^2+8l^2+9m^2+8n^2+5s^2+8kl+6km\\
&+4kn+2ks+12ml+8ln+3ls+12mn+6ms+8ns.\end{aligned}$$ As in [@caicedo_caspistrano_zhang_2015], the hidden regularity for the corresponding adjoint system was required. Here, the result is given in Proposition \[hiddenregularities\], which is the key point to prove the controllability result.
Finally, for small values of the length $L$ and large time of control $T$ we derive a exact controllability result in $(L^2(0,L))^2$ by assuming that the controls $g_1(t)=0$ (resp. $h_1(t)=0$) and $g_0(t)=g_2(t)=0$. In this case, the analysis of the linearized system is much more complicated, therefore we use a direct approach based on the multipliers technique, as in [@cerpapazoto2011]. In all cases, the result obtained for the linear system allows to prove the local controllability property of the nonlinear system by means of a fixed point argument.
The analysis describe above are summarized in the main result of the paper, Theorem \[main\_theo\]. However, in order to make the reading easier, throughout the paper we use the following notation for the boundary functions:
- $\vec{h}_1=(0,h_1,0),\,\, \vec{g}_1=(g_0,g_1,g_2)$ and $\vec{h}_2=(h_0,h_1,h_2),\,\, \vec{g}_2=(0,g_1,0)$,
- $\vec{h}_3=(h_0,h_1,0),\,\,\vec{g}_3=(g_0,g_1,0)$ and $\vec{h}_4=(0,h_1,h_2),\,\,\vec{g}_4=(0,g_1,g_2)$,
- $\vec{h}_5=(0,h_1,0),\,\,\vec{g}_5=(0,0,0)$ and $\vec{h}_6=(0,0,0),\,\,\vec{g}_6=(0,g_1,0)$.
We also introduce the space $\mathcal{X}:=(L^2(0,L))^2$ endowed with the inner product $$\left\langle (u,v) , (\varphi
,\psi)\right\rangle := \int_0^L u(x)\varphi(x) dx + \frac{b}{c}\int_0^L v(x)\psi(x) dx,\qquad \forall (u,v), (\varphi,\psi) \in {\mathcal{X}},$$ and the spaces $${\mathcal{H}}_T:=H^{-\frac{1}{3}}(0,T)\times L^2(0,T)\times H^{-\frac{1}{3}}(0,T) \mbox{ and } {\mathcal{Z}}_T:= C([0,T];(L^2(0,L))^2)\cap L^2(0,T,(H^1(0,L))^2)$$ endowed with their natural inner products.
Thus, our main result reads as follows:
\[main\_theo\] Let $T>0$. Then, there exists $\delta>0$, such that, for any $(u^0,v^0), (u^1,v^1) \in {\mathcal{X}}:=(L^2(0,L))^2$ verifying $$\|(u^0,v^0)\|_{{\mathcal{X}}} + \|(u^1,v^1)\|_{{\mathcal{X}}} \leq \delta,$$ the following holds:
1. If $L \in (0,\infty) \setminus {\mathcal{F}}_r$, one can find $\vec{h}_i, \vec{g}_i \in {\mathcal{H}}_T$, for $i=1,2$, such that the system - admits a unique solution $(u,v) \in {\mathcal{Z}}_T$ satisfying .
2. For any $L>0$, one can find $\vec{h}_i, \vec{g}_j \in {\mathcal{H}}_T$, for $j=3,4$, such that the system - admits a unique solution $(u,v) \in {\mathcal{Z}}_T$, satisfying .
3. Let $T>0$ and $L>0$ satisfying $$\begin{aligned}
1>\frac{\beta C_T}{T}\left[L +\frac{r}{c} \right],\end{aligned}$$ where $C_T$ is the constant in and $\beta$ is the constant given by the embedding $H^{\frac{1}{3}}(0,T) \subset L ^2(0,T)$. Then, one can find $\vec{h}_k, \vec{g}_k \in {\mathcal{H}}_T$, for $k=5,6$, such that the system - admits a unique solution $(u,v) \in {\mathcal{Z}}_T$, satisfying .
Before close this section, we observe that the exact controllability result given in Theorem A holds without any restriction of the Length $L$. However, we believe that, with another configuration of the controls, it is possible to prove the existence of a critical set for the system .
The article is organized as follows:
—- In Section \[Sec2\], we show that the system - is locally well-posed in ${\mathcal{Z}}_T$, whenever $(u^0, v^0) \in(L^2 (0,L))^2$, $h_0, \ g_0\in H^{-\frac13} (\mathbb{R}^+), \ h_1, \ g_1 \in L^2 (\mathbb{R}^+)$ and $h_2, \ g_2\in H^{-\frac13} (\mathbb{R}^+)$. Various linear estimates, including hidden regularities, are presented for solutions of the corresponding linear system. As we pointed before, such estimates will play important roles in studying the controllability properties.
—- In Section \[Sec3\], the boundary control system is investigated for its controllability. We investigate first the linearized system and its corresponding adjoint system for their controllability and observability. In particular, the hidden regularities for the solutions of the adjoint system presented in the Section \[Sec2\] are used to prove observability inequalities associated to the control problem.
—- Finally, the proof of our main result, Theorem \[main\_theo\], is presented in Section \[Sec4\].
Well-posedness {#Sec2}
==============
Linear System
-------------
In this section, we establish the well-posedness of the linear system associated to (\[gg1\])-(\[gg2\]): $$\label{gglin}
\left\lbrace \begin{tabular}{l l}
$u_t + u_{xxx} + av_{xxx} =0$, & in $(0,L)\times (0,T)$,\\
$v_t +\frac{r}{c}v_x+\frac{ab}{c}u_{xxx} + \frac{1}{c}v_{xxx} =0$, & in $(0,L)\times (0,T)$,\\
$u_{xx}(0,t) = h_0(t),\,\,u_x(L,t) = h_1(t),\,\,u_{xx}(L,t) = h_2(t)$,& in $(0,T)$,\\
$v_{xx}(0,t) =g_0(t),\,\,v_x(L,t) = g_1(t),\,\, v_{xx}(L,t) = g_2(t)$,& in $(0,T)$,\\
$u(x,0)=u^0(x), \quad v(x,0) = v^0(x)$, & in $(0,L)$.
\end{tabular}\right.$$ We begin by considering the following linear non-homogeneous boundary value problem $$\label{gglin1}
\left\lbrace \begin{tabular}{l l}
$u_t + u_{xxx} + av_{xxx} =f$, & in $(0,L)\times (0,T)$,\\
$v_t +\frac{ab}{c}u_{xxx} + \frac{1}{c}v_{xxx} =s$, & in $(0,L)\times (0,T)$,\\
$u_{xx}(0,t) = h_0(t),\,\,u_x(L,t) = h_1(t),\,\,u_{xx}(L,t) = h_2(t)$,& in $(0,T)$,\\
$v_{xx}(0,t) =g_0(t),\,\,v_x(L,t) = g_1(t),\,\, v_{xx}(L,t) = g_2(t)$,& in $(0,T)$,\\
$u(x,0)=u^0(x), \quad v(x,0) = v^0(x)$, & in $(0,L)$,
\end{tabular}\right.$$ with the notation introduced in Section 1. Then next proposition shows that the, problem (\[gglin1\]) is well-posed in the space ${\mathcal{X}}$.
\[prop1\] Let $T>0$ be given. Then, for any $(u^0,v^0)$ in ${\mathcal{X}}$, $f,s$ in $L^1(0,T;L^2(0,L))$ and $\overrightarrow{h},\overrightarrow{g}\in {\mathcal{H}}_T$, problem (\[gglin1\]) admits a unique solution $(u,v) \in {\mathcal{Z}}_T$, with $$\label{hr1}
\partial_x^k u,\partial_x^k v \in L^{\infty}_x(0,L;H^{\frac{1-k}{3}}(0,T)), \quad k=0,1,2.$$ Moreover, there exist $C>0$, such that $$\begin{gathered}
\|(u,v)\|_{{\mathcal{Z}}_T}+\sum_{k=0}^{2}\|(u,v)\|_{L^{\infty}_x(0,L;(H^{\frac{1-k}{3}}(0,T))^2)}\leq C\left\lbrace \|(u^0,v^0)\|_{{\mathcal{X}}} \right. \\
\left.+\|(\overrightarrow{h},\overrightarrow{g})\|_{{\mathcal{H}}_T}+\|(f,s)\|_{L^1(0,T;(L^2(0,L))^2)} \right\rbrace.\end{gathered}$$
We diagonalize the main term in and consider the change of variable $$\left\lbrace\begin{tabular}{l}
$u = 2a \widetilde{u} +2a \widetilde{v}$, \\
$v = \left(\left(\frac{1}{c}-1\right)+\lambda\right) \widetilde{u}+ \left(\left(\frac{1}{c}-1\right)-\lambda\right) \widetilde{v}$,
\end{tabular}\right.$$ where $\lambda=\sqrt{\left(\frac{1}{c}-1\right)^2+\frac{4a^2b}{c}}$. Thus, we can transform the linear system (\[gglin1\]) into $$\label{kdvln1}
\left\{
\begin{array}{ll}\vspace{2mm}
\widetilde{u}_t + \alpha_{-}\widetilde{u}_{xxx} =\widetilde{f},\\
\widetilde{v}_t +\alpha_{+}\widetilde{v}_{xxx}=\widetilde{s}, \vspace{2mm} \\
\widetilde{u}_{xx}(0,t) = \widetilde{h}_0(t), \ \widetilde{u}_x(L,t) = \widetilde{h}_1(t), \ \widetilde{u}_{xx}(L,t) = \bar{h}_2(t) ,\vspace{2mm} \\
\widetilde{v}_{xx}(0,t) = \widetilde{g}_0(t), \ \widetilde{v}_x(L,t) = \widetilde{g}_1(t), \ \widetilde{v}_{xx}(L,t) = \widetilde{g}_2(t) ,\vspace{2mm} \\
\widetilde{u}(x,0)= \widetilde{u}^0(x), \quad \widetilde{v}(x,0) = \widetilde{v}^0(x),
\end{array}
\right.$$ where $\alpha_{\pm} = -\frac{1}{2}\left(\left(\frac{1}{c}-1\right)\pm \lambda\right)$ and $$\left\lbrace\begin{tabular}{l l l l}
$\widetilde{f}=-\frac{1}{2}\left(\frac{\alpha_+}{a\lambda}f+\frac{1}{\lambda}s\right)$, & $\widetilde{u}_0=-\frac{1}{2}\left(\frac{\alpha_{-}}{a\lambda}u^0-\frac{1}{\lambda}v^0\right)$, & $\widetilde{h}_i=-\frac{1}{2}\left(\frac{\alpha_{-}}{a\lambda}h_i-\frac{1}{\lambda}g_i\right),$ & $i=0,1,2,$ \\
\\
$\widetilde{s}=-\frac{1}{2}\left(\frac{\alpha_{-}}{a\lambda}f-\frac{1}{\lambda}s\right),$ & $\widetilde{v}_0=\frac{1}{2}\left(\frac{\alpha_{+}}{a\lambda}u^0-\frac{1}{\lambda}v^0\right)$, & $\widetilde{g}_i=\frac{1}{2}\left(\frac{\alpha_{+}}{a\lambda}h_i-\frac{1}{\lambda}g_i\right),$ & $i=0,1,2.$
\end{tabular}\right.$$ Note that condition guarantees that $\alpha_{\pm}$ are nonzero. Therefore, system can be decoupled into two single KdV equations as follows: $$\left\{
\begin{array}{ll}
\widetilde{u}_t + \alpha_{-}\widetilde{u}_{xxx} =\widetilde{f},\vspace{2mm}\\
\widetilde{u}_{xx}(0,t) = \widetilde{h}_0(t), \ \widetilde{u}_x(L,t) = \widetilde{h}_1(t), \ \widetilde{u}_{xx}(L,t) = \widetilde{h}_2(t),\vspace{2mm}\\
\widetilde{u}(0,x)= \widetilde{u}^0(x)
\end{array}
\right.
\label{kdv_ne_1}$$ and $$\left\{
\begin{array}{ll}\widetilde{v}_t +\alpha_{+}\widetilde{v}_{xxx}=\widetilde{s},\vspace{2mm} \\
\widetilde{v}_{xx}(0,t) = \widetilde{g}_0(t), \ \widetilde{v}_x(L,t) = \widetilde{g}_1(t), \ \widetilde{v}_{xx}(L,t) = \widetilde{g}_2(t), \vspace{2mm}\\
\widetilde{v}(x,0) = \widetilde{v}^0(x).
\end{array}
\right.
\label{kdv_ne_2}$$ Here, we consider the solutions written on the form $\{W_{bdr}^{\pm}(t)\}_{t\geq 0}$ that will be called *the boundary integral operator*. For this purpose we use a Lemma, which can be found in [@CaZhaSun Lemma 2.4], for solutions of (or ):
\[l4-d\] The solution $u$ of the IBVP (or ) can be written in the form $$u(x,t)=[{W}^{+}_{bdr}\vec{\widetilde{h}}](x,t):=[{W}^{+}_{bdr}\vec{h}](x,t):= \sum_{j,m=1}^3
[W^{+}_{j,m}h_{m}](x,t),$$ where $$[W^{+}_{j,m}h](x,t) \equiv [U_{j,m} h](x,t) +\overline{[U_{j,m}h](x,t)}\label{2.2-4}$$ with $$[U_{j,m} h](x,t)\equiv \frac{1}{2\pi } \int ^{+\infty }_{0} e^{i\rho ^3
t} e^{\lambda ^+_j (\rho ) x} 3\rho ^2
[Q_{j,m}^{+}h](\rho )d\rho \label{2.3-4}$$ for $j=1,3, \ m=1,2,3 $ and $$[U_{2,m} h](x,t) \equiv \frac{1}{2\pi } \int ^{+\infty }_{0} e^{i \rho ^3
t} e^{-\lambda ^+_2 (\rho ) (1-x)} 3\rho ^2
[Q_{2,m}^{+}h](\rho )d\rho \label{2.4-4}$$ for $m=1,2,3$. Here $$\label{2.5-4} [Q_{j,m}
^{+}h] (\rho ):=\frac{\Delta ^{+}_{j,m} (\rho ) } {\Delta
^{+} (\rho )} \hat{h} ^+ (\rho ), \qquad [Q_{2,m}^{+}h]
(\rho )=\frac{\Delta ^{+}_{2,m} (\rho ) } {\Delta ^{+} (\rho
)} e^{\lambda ^+_2 (\rho )} \hat{h} ^+ (\rho )$$ for $j=1,3 $ and $m=1,2,3$. Here $\hat{h}^+ (\rho) = \hat{h} (i\rho
^3)$, $\Delta
^{+}(\rho)$ and $\Delta ^{+}_{j,m} (\rho)$ are obtained from $\Delta
(s)$ and $\Delta _{j,m} (s)$ by replacing $s$ with $i \rho
^3 $ and $\lambda _j ^+ (\rho)=\lambda _j (i \rho
^3 )$ where $$\Delta = \lambda _1\lambda _2 \lambda _3\left (\lambda _1(\lambda _3-\lambda _2) e^{-\lambda _1}
+\lambda _2(\lambda _1-\lambda _3) e^{-\lambda _2}+\lambda _3(\lambda _2-\lambda _1) e^{-\lambda
_3}\right );$$ $$\Delta _{1,1} = e^{-\lambda _1}\lambda _2 \lambda _3(\lambda _3-\lambda
_2), \ \Delta _{2,1} = e^{-\lambda _2}\lambda _1 \lambda _3(\lambda
_1-\lambda _3), \ \Delta _{3,1} = e^{-\lambda _3}\lambda _1 \lambda _2(\lambda
_2-\lambda _1);$$ $$\Delta _{1,2} = \lambda ^2_2 \lambda _3^2 (e^{\lambda _2} -
e^{\lambda _3}), \ \Delta _{2,2} = \lambda _1^2\lambda ^2_3( e^{\lambda _3} -
e^{\lambda _1}), \ \Delta _{3,2} = \lambda_1 ^2\lambda _2^2 ( e^{\lambda _1} -
e^{\lambda _2});$$ $$\Delta _{1,3} = \lambda _2\lambda _3 (\lambda _2e^{\lambda _3}-\lambda
_3
e^{\lambda _2}), \ \Delta _{2,3} = \lambda _1 \lambda _3 (\lambda_3e^{\lambda _1}-\lambda
_1
e^{\lambda _3}), \ \Delta _{3,3} = \lambda _1 \lambda _2 (\lambda _1e^{\lambda _2}-\lambda
_2
e^{\lambda _1}).$$
Since $$(\widetilde{u}^0, \widetilde{v}^0) \in {\mathcal{X}},\quad (\widetilde{f},\widetilde{s}) \in L^1(0,T;(L^2(0,L))^2)\, \mbox{ and }\, \overrightarrow{\widetilde{h}}, \overrightarrow{\widetilde{g}} \in {\mathcal{H}}_T,$$ by [@caicedo_caspistrano_zhang_2015 Proposition 2.5], we obtain the existence of $(\widetilde{u},\widetilde{v})\in {\mathcal{Z}}_T$, solution of the system , such that $$\partial_x^k \widetilde{u},\partial_x^k \widetilde{v} \in L^{\infty}_x(0,L;H^{\frac{1-k}{3}}(0,T)), \quad k=0,1,2,$$ and $$\begin{gathered}
\|(\widetilde{u},\widetilde{v})\|_{{\mathcal{Z}}_T}+\sum_{k=0}^{2}\|(\widetilde{u},\widetilde{v})\|_{L^{\infty}_x(0,L;(H^{\frac{1-k}{3}}(0,T))^2)}\leq C\left\lbrace \|(\widetilde{u}^0,\widetilde{v}^0)\|_{{\mathcal{X}}}+\|(\overrightarrow{\widetilde{h}},\overrightarrow{\widetilde{g}})\|_{{\mathcal{H}}_T}\right. \\
\left.+\|(\widetilde{f},\widetilde{s})\|_{L^1(0,T;(L^2(0,L))^2)} \right\rbrace,\end{gathered}$$ for some constant $C>0$. Furthermore, we can write $\widetilde{u}$ and $\widetilde{v}$ in its integral form as follows $$\widetilde{u}(t)=W_0^{-}(t)\widetilde{u}^0+W_{bdr}^{-}(t)\overrightarrow{\widetilde{h}} + \int_0^tW_0^{-}(t-\tau)\widetilde{f}(\tau)d\tau,$$ $$\widetilde{v}(t)=W_0^{+}(t)\widetilde{v}^0+W_{bdr}^{+}(t)\overrightarrow{\widetilde{g}} + \int_0^tW_0^{+}(t-\tau)\widetilde{s}(\tau)d\tau,$$ where $\{W_0^{\pm}(t)\}_{t\geq 0}$ are the $C_0$-semigroup in the space $L^2(0,L)$ generated by the linear operators $$A^{\pm}=-\alpha_{\pm}g''',$$ with domain $$D(A^{\pm})=\{g \in H^3(0,L): g''(0)=g'(L)=g''(L)=0\},$$ and $\{W_{bdr}^{\pm}(t)\}_{t\geq 0}$ are the operator given in Lemma \[l4-d\] (see also [@caicedo_caspistrano_zhang_2015 Lemma 2.1] for more details). Then, by change of variable we can easily verify that $$\begin{cases}
u(t)=W_0^{-}(t)u^0+W_{bdr}^{-}(t)\overrightarrow{h} + \displaystyle\int_0^tW_0^{-}(t-\tau)f(\tau)d\tau, \\
v(t)=W_0^{+}(t)v^0+W_{bdr}^{+}(t)\overrightarrow{g} + \displaystyle\int_0^tW_0^{+}(t-\tau)s(\tau)d\tau
\end{cases}$$ and the result follows.
The global well-posedness of the system (\[gglin\]) is obtained using a fixed point argument.
\[prop2\] Let $T>0$ be given. Then, for any $(u^0,v^0) \in {\mathcal{X}}$ and $\overrightarrow{h}, \overrightarrow{g}\in {\mathcal{H}}_T$, problem (\[gglin\]) admits a unique solution $(u,v) \in {\mathcal{Z}}_T$ with $$\partial_x^k u,\partial_x^k v \in L^{\infty}_x(0,L;H^{\frac{1-k}{3}}(0,T)), \quad k=0,1,2.$$ Moreover, there exist $C>0$, such that $$\begin{gathered}
\|(u,v)\|_{{\mathcal{Z}}_T}+\sum_{k=0}^{2}\|(u,v)\|_{L^{\infty}_x(0,L;(H^{\frac{1-k}{3}}(0,T))^2)}\leq C\left\lbrace \|(u^0,v^0)\|_{{\mathcal{X}}}+\|(\overrightarrow{h},\overrightarrow{g})\|_{{\mathcal{H}}_T} \right.\\
\left.+\|(f,s)\|_{L^1(0,T;(L^2(0,L))^2)} \right\rbrace.\end{gathered}$$
Let ${\mathcal{F}}_T:= \left\lbrace (u,v) \in {\mathcal{Z}}_T: (u,v) \in L^{\infty}_x(0,L;(H^{\frac{1-k}{3}}(0,T))^2), k=0,1,2\right\rbrace$ equipped with the norm $$\|(u,v)\|_{{\mathcal{F}}_T} = \|(u,v)\|_{{\mathcal{Z}}_T}+\sum_{k=0}^{2}\|(u,v)\|_{L^{\infty}_x(0,L;(H^{\frac{1-k}{3}}(0,T))^2)}.$$ Let $0< \beta \leq T$ to be determined later. For each $u,v \in {\mathcal{F}}_{\beta}$, consider the problem $$\label{e1}
\left\lbrace \begin{tabular}{l l}
$\omega_t + \omega_{xxx} + a\eta_{xxx} =0$, & in $(0,L)\times (0,\beta)$,\\
$\eta_t +\frac{ab}{c}\omega_{xxx} + \frac{1}{c}\eta_{xxx} =-\frac{r}{c}v_x$, & in $(0,L)\times (0,\beta)$,\\
$\omega_{xx}(0,t) = h_0(t),\,\,\omega_x(L,t) = h_1(t),\,\,\omega_{xx}(L,t) = h_2(t)$,& in $(0,\beta)$,\\
$\eta_{xx}(0,t) =g_0(t),\,\,\eta_x(L,t) = g_1(t),\,\, \eta_{xx}(L,t) = g_2(t)$,& in $(0,\beta)$,\\
$\omega(x,0)=u^0(x), \quad v(x,0) = v^0(x)$, & in $(0,L)$.
\end{tabular}\right.$$ According to Proposition \[prop1\], we can define the operator $$\Gamma: {\mathcal{F}}_{\beta} \rightarrow {\mathcal{F}}_{\beta}, \quad \text{given by} \quad \Gamma(u,v)=(\omega,\eta),$$ where $(\omega,\eta)$ is the solution of (\[e1\]). Moreover, $$\label{hr3}
\|\Gamma(u,v)\|_{{\mathcal{F}}_\beta}\leq C\left\lbrace \|(u^0,v^0)\|_{{\mathcal{X}}}+\|(\overrightarrow{h},\overrightarrow{g})\|_{{\mathcal{H}}_\beta}+\|(0,v_x)\|_{L^1(0,\beta;(L^2(0,L))^2)} \right\rbrace,$$ where the positive constant $C$ depends only on $T$. Since $$\|(0,v_x)\|_{L^1(0,\beta;L^2(0,L))}\leq \beta^{\frac{1}{2}}\|(u,v)\|_{{\mathcal{F}}_\beta},$$ we obtain a positive constant $C>0$, such that $$\label{e2}
\|\Gamma(u,v)\|_{{\mathcal{F}}_\beta}\leq C\left\lbrace \|(u^0,v^0)\|_{{\mathcal{X}}}+\|(\overrightarrow{h},\overrightarrow{g})\|_{{\mathcal{H}}_\beta}\right\rbrace+C\beta^{\frac{1}{2}}\|(u,v)\|_{{\mathcal{F}}_\beta}.$$ Let $(u,v) \in B_{r}(0):=\left\lbrace (u,v) \in {\mathcal{F}}_{\beta}: \|(u,v)\|_{{\mathcal{F}}_{\beta}}\leq r\right\rbrace$, with $r=2C\left\lbrace \|(u^0,v^0)\|_{{\mathcal{X}}}+\|(\overrightarrow{h},\overrightarrow{g})\|_{{\mathcal{H}}_\beta}\right\rbrace$. Choosing $\beta>0$, satisfying $$\label{beta}
C\beta^{\frac{1}{2}}\leq \frac{1}{2},$$ from (\[e2\]) we obtain $$\|\Gamma(u,v)\|_{{\mathcal{F}}_\beta}\leq r.$$ The above estimate allows us to conclude that $$\Gamma: B_r(0)\subset{\mathcal{F}}_{\beta} \rightarrow B_r(0).$$ On the other hand, note that $\Gamma(u_1,v_1)-\Gamma(u_2,v_2)$ solves the following system $$\left\lbrace \begin{tabular}{l l}
$\omega_t + \omega_{xxx} + a\eta_{xxx} =0$, & in $(0,L)\times (0,\beta)$,\\
$\eta_t +\frac{ab}{c}\omega_{xxx} + \frac{1}{c}\eta_{xxx} =-\frac{r}{c}(v_{1x}-v_{2x}) $, & in $(0,L)\times (0,\beta)$,\\
$\omega_{xx}(0,t) = \omega_x(L,t) = \omega_{xx}(L,t) = 0$,& in $(0,\beta)$,\\
$\eta_{xx}(0,t)=\eta_x(L,t) = \eta_{xx}(L,t) = 0$,& in $(0,\beta)$,\\
$\omega(x,0)=0, \quad v(x,0) = 0$, & in $(0,L)$.
\end{tabular}\right.$$ Again, from Proposition \[prop1\] and , we have $$\begin{array}{l}
\|\Gamma(u_1,v_1)-\Gamma(u_2,v_2)\|_{{\mathcal{F}}_\beta}\leq C\|(0,v_{1x}-v_{2x})\|_{L^1(0,\beta;(L^2(0,L))^2)}\leq C\beta^{\frac{1}{2}}\|(u_1,v_1)-(u_2,v_2)\|_{{\mathcal{F}}_{\beta}}\\
\qquad\qquad\qquad\qquad\qquad\quad\leq \displaystyle\frac{1}{2}\|(u_1,v_1)-(u_2,v_2)\|_{{\mathcal{F}}_{\beta}}.\nonumber
\end{array}$$ Hence, $\Gamma: B_r(0) \rightarrow B_r(0)$ is a contraction and, by Banach fixed point theorem, we obtain a unique $(u,v) \in B_r(0)$, such that $$\Gamma(u,v) = (u,v) \in {\mathcal{F}}_{\beta},$$ and holds, for all $t \in (0,\beta)$. Since the choice of $\beta$ is independent of $(u^0,v^0)$, the standard continuation extension argument yields that the solution $(u,v)$ belongs to ${\mathcal{F}}_T$. The proof is complete.
### Adjoint System
Consider the following homogeneous initial-value problem associated to (\[gg1\])-(\[gg2\]): $$\label{gglin3}
\left\lbrace \begin{tabular}{l l}
$u_t + u_{xxx} + av_{xxx} =0$, & in $(0,L)\times (0,T)$,\\
$v_t +\frac{r}{c}v_x+\frac{ab}{c}u_{xxx} + \frac{1}{c}v_{xxx} =0$, & in $(0,L)\times (0,T)$,\\
$u_{xx}(0,t)=u_x(L,t)=u_{xx}(L,t) = 0$,& in $(0,T)$,\\
$v_{xx}(0,t) =v_x(L,t) = v_{xx}(L,t) = 0$,& in $(0,T)$,\\
$u(x,0)=u^0(x), \quad v(x,0) = v^0(x)$, & in $(0,L)$.
\end{tabular}\right.$$ In order to introduce the backward system associated to , we multiply the first equation of by $\varphi$, the second one by $\psi$ and integrate over $(0,L)\times (0,T)$. Assuming that the functions $u, v, \varphi$ and $\psi$ are regular enough to justify all the computations, we obtain, after integration by parts, the following identity: $$\begin{aligned}
\label{ad1}
\int_0^L&\left( u(x,T)\varphi(x,T)+v(x,T)\psi(x,T)\right)dx-\int_0^L\left(u^0(x)\varphi(x,0)+v^0(x)\psi(x,0)\right) dx = \notag \\
&\int_0^T\int_0^L u(x,t)\left(\varphi(x,t) + \varphi_{xxx}(x,t) + \frac{ab}{c}\psi_{xxx}(x,t) \right)dxdt \notag \\
&+\int_0^T\int_0^Lv(x,t)\left( \psi(x,t) +\frac{r}{c}\psi(x,t)+a\varphi_{xxx}(x,t) + \frac{1}{c}\psi_{xxx}(x,t)\right)dxdt \notag\\
&-\int_0^T u_{xx}(L,t)\left(\varphi(L,t)+\frac{ab}{c}\psi(L,t)\right)dt+\int_0^T u_{xx}(0,t)\left(\varphi(0,t)+\frac{ab}{c}\psi(0,t)\right)dt \notag\\
&+\int_0^T u_x(L,t)\left(\varphi_x(L,t)+\frac{ab}{c}\psi_x(L,t)\right)dt-\int_0^T u_x(0,t)\left(\varphi_x(0,t)+\frac{ab}{c}\psi_x(0,t)\right)dt \notag \\
&-\int_0^T u(L,t)\left(\varphi_{xx}(L,t)+\frac{ab}{c}\psi_{xx}(L,t)\right)dt+\int_0^T u(0,t)\left(\varphi_{xx}(0,t)+\frac{ab}{c}\psi_{xx}(0,t)\right)dt \\
&-\int_0^T v_{xx}(L,t)\left(a\varphi(L,t)+\frac{1}{c}\psi(L,t)\right)dt+\int_0^T v_{xx}(0,t)\left(a\varphi(0,t)+\frac{1}{c}\psi(0,t)\right)dt \notag \\
&+\int_0^T v_{x}(L,t)\left(a\varphi_x(L,t)+\frac{1}{c}\psi_x(L,t)\right)dt-\int_0^T v_{x}(0,t)\left(a\varphi_x(0,t)+\frac{1}{c}\psi_x(0,t)\right)dt \notag \\
&-\int_0^T v(L,t)\left(a\varphi_{xx}(L,t)+\frac{1}{c}\psi_{xx}(L,t)+\frac{r}{c}\psi(L,t)\right)dt\notag \\
&+\int_0^T v(0,t)\left(a\varphi_{xx}(0,t)+\frac{1}{c}\psi_{xx}(0,t)+\frac{1}{c}\psi(0,t)\right)dt. \notag\end{aligned}$$ Having the previous equality in hands, we consider backward system as follows $$\label{linadj}
\begin{cases}
\varphi_t + \varphi_{xxx} + \frac{ab}{c}\psi_{xxx}=0, & \text{in}\,\, (0,L)\times (0,T), \\
\psi_t +\frac{r}{c}\psi_x+a\varphi_{xxx} +\frac{1}{c}\psi_{xxx} =0, & \text{in}\,\, (0,L)\times (0,T)
\end{cases}$$ satisfying the boundary conditions, $$\label{bc1}
\begin{cases}
a\varphi_x(0,t)+\frac{1}{c}\psi_x(0,t) =0, & \text{in}\,\, (0,T),\\
\varphi_x (0,t) +\frac{ab}{c}\psi_x (0,t) =0, & \text{in}\,\, (0,T), \\
\varphi_{xx} (L,t) +\frac{ab}{c}\psi_{xx} (L,t) =0, & \text{in}\,\, (0,T), \\
\varphi_{xx} (0,t) +\frac{ab}{c}\psi_{xx} (0,t) =0, & \text{in}\,\, (0,T), \\
a\varphi_{xx}(L,t)+\frac{1}{c}\psi_{xx}(L,t)+\frac{r}{c}\psi(L,t) =0, & \text{in}\,\, (0,T), \\
a\varphi_{xx}(0,t)+\frac{1}{c}\psi_{xx}(0,t)+\frac{r}{c}\psi(0,t) =0, & \text{in}\,\, (0,T)
\end{cases}$$ and the final conditions $$\label{finaladj}
\varphi(x,T)= \varphi^1(x), \qquad \psi(x,T)= \psi^1(x), \qquad \text{in}\,\, (0,L).$$ Since the coefficients satisfy $1-a^2b >0$, we can deduce from the first and second equations of that the above boundary conditions can be written as $$\label{linadjbound}
\begin{cases}
\varphi_x(0,t)=\psi_x(0,t) =0, & \text{in}\,\, (0,T),\\
\varphi_{xx} (L,t) +\frac{ab}{c}\psi_{xx} (L,t) =0, & \text{in}\,\, (0,T),\\
\varphi_{xx} (0,t) +\frac{ab}{c}\psi_{xx} (0,t) =0, & \text{in}\,\, (0,T),\\
a\varphi_{xx}(L,t)+\frac{1}{c}\psi_{xx}(L,t)+\frac{r}{c}\psi(L,t) =0, & \text{in}\,\, (0,T), \\
a\varphi_{xx}(0,t)+\frac{1}{c}\psi_{xx}(0,t)+\frac{r}{c}\psi(0,t) =0, & \text{in}\,\, (0,T).
\end{cases}$$
The following proposition is the key to prove the controllability of the linear system . The result ensures the hidden regularity for the solution of the adjoint system -.
\[hiddenregularities\] For any $(\varphi^1, \psi^1) \in {\mathcal{X}}$, the system - admits a unique solution $(\varphi, \psi) \in {\mathcal{Z}}_T$, such that it possess the following sharp trace properties $$\label{hr4}
\begin{cases}
\underset{0< x < L}{\sup} \| \partial^k_x \varphi(x,\cdot)\|_{H^{\frac{1-k}3}(0,T)}\le C_T\|\varphi^1\|_{L^2(0,L)}, \\
\underset{0< x <L}{\sup} \| \partial^k_x \psi(x,\cdot)\|_{H^{\frac{1-k}3}(0,T)}\le C_T \|\psi^1\|_{L^2(0,L)},
\end{cases}$$ for $k=0,1,2$, where $C_r$ is a positive constant.
Proceeding as the proof of Proposition \[prop2\], we obtain the result. Indeed, first we consider the change of variable $t\rightarrow T-t$ and $x\rightarrow L-x$, then for any $(\varphi,\psi)$ in ${\mathcal{Z}}_T$, we consider the system $$\begin{cases}
u_t + u_{xxx} + \frac{ab}{c}v_{xxx}=0, & \text{in}\,\, (0,L)\times (0,T), \\
v_t + au_{xxx} +\frac{1}{c}v_{xxx} =-\frac{r}{c}v_x, & \text{in}\,\, (0,L)\times (0,T), \\
\varphi(x,0)= \varphi^0(x)\text{, } \psi(x,0)= \psi^0(x), & \text{in}\,\, (0,L),
\end{cases}$$ with boundary conditions $$\begin{cases}
u_x(L,t)=v_x(L,t) =0, & \text{in}\,\, (0,T),\\
u_{xx} (L,t)=-\frac{ab}{c}\psi_{xx} (L,t), & \text{in}\,\, (0,T),\\
u_{xx} (0,t)=-\frac{ab}{c}\psi_{xx} (0,t), & \text{in}\,\, (0,T),\\
v_{xx}(L,t)=-ac\varphi_{xx}(L,t)-r\psi(L,t), & \text{in}\,\, (0,T), \\
v_{xx}(0,t)=-ac\varphi_{xx}(0,t)-r\psi(0,t), & \text{in}\,\, (0,T).
\end{cases}$$ By using a fixed point argument the result is archived.
The adjoint system possesses a relevant estimate as described below.
\[prop3\] Any solution $(\varphi,\psi)$ of the adjoint system - satisfies $$\begin{aligned}
\label{e6}
\|(\varphi^1,\psi^1)\|_{{\mathcal{X}}}^2 \leq &\frac{1}{T}\|(\varphi,\psi)\|_{L^2(0,T;{\mathcal{X}})}+\frac{1}{2}\|\varphi_x(L,\cdot)\|_{L^2(0,T)}^2+\frac{b}{2c}\|\psi_x(L,\cdot)\|_{L^2(0,T)}^2 +\frac{br}{c^2}\|\psi(L,\cdot)\|_{L^2(0,T)}^2\notag\\
&+\frac{1}{2}\left\|\varphi_x(L,\cdot)+\frac{ab}{c}\psi_x(L,\cdot)\right\|_{L^2(0,T)}^2+\frac{b}{2c}\left\|a\varphi_x(L,\cdot)+\frac{1}{c}\psi_x(L,\cdot)\right\|_{L^2(0,T)}^2,\end{aligned}$$ with initial data $(\varphi^1,\psi^1) \in {\mathcal{X}}$.
Multiplying the first equation of by $-t\varphi$, the second one by $-\frac{b}{c}t\psi$ and integrating by parts over $(0, T)\times (0,L)$, we obtain $$\begin{gathered}
\frac{T}{2}\int_0^L\varphi^2(x,T)dx=\frac{1}{2}\int_0^T\int_0^L\varphi^2(x,t)dxdt + \frac{ab}{c}\int_0^T\int_0^L t\varphi_{xxx}(x,t)\psi(x,t) dxdt\\
-\int_0^Tt\left[\varphi_{xx}(x,t)\varphi(x,t)-\frac{1}{2}\varphi_x^2(x,t)+\frac{ab}{c}\psi_{xx}(x,t)\varphi(x,t)-\frac{ab}{c}\psi_{x}(x,t)\varphi_x(x,t)\right. \\
\left.+\frac{ab}{c}\psi(x,t)\varphi_{xx}(x,t)\right]_0^Ldt\end{gathered}$$ and $$\begin{gathered}
\frac{Tb}{2c}\int_0^L\psi^2(x,T)dx=\frac{b}{2c}\int_0^T\int_0^L\psi^2(x,t)dxdt - \frac{ab}{c}\int_0^T\int_0^L t\varphi_{xxx}(x,t)\psi(x,t) dxdt \\
-\int_0^Tt\left[\frac{b}{c^2}\psi_{xx}(x,t)\psi(x,t)-\frac{b}{2c^2}\psi_x^2(x,t)+\frac{br}{2c^2}\psi^2(x,t)\right]_0^Ldt.\end{gathered}$$ Adding the above identities, it follows that $$\begin{gathered}
\frac{T}{2}\|(\varphi^1,\psi^1)\|_{{\mathcal{X}}}^2 = \frac{1}{2}\|(\varphi,\psi)\|^2_{L^2(0,T;{\mathcal{X}})} - \int_0^Tt\left[\frac{b}{c}\psi(x,t)\left(a\varphi_{xx}(x,t)+\frac1c\psi_{xx}(x,t)+\frac{r}{c}\psi(x,t)\right) \right]_0^Ldt \\
-\int_0^Tt\left[\frac{b}{2c}\psi_x(x,t)\left(a\varphi_{x}(x,t)+\frac1c\psi_{x}(x,t)\right) -\frac{1}{2}\varphi_x(x,t)\left(\varphi_{x}(x,t)+\frac{ab}{c}\psi_{x}(x,t)\right)\right]_0^Ldt \\
+\int_0^Tt\left[\varphi(x,t)\left(\varphi_{xx}(x,t)+\frac{ab}{c}\psi_{xx}(x,t)\right) -\frac{br}{2c^2}\psi^2(x,t)\right]_0^Ldt.\end{gathered}$$ Then, from , we obtain $$\begin{aligned}
\frac{T}{2}\|(\varphi^1,\psi^1)\|_{{\mathcal{X}}}^2 \leq & \frac{1}{2}\|(\varphi,\psi)\|^2_{L^2(0,T;{\mathcal{X}})} + \frac{b T}{2c}\int_0^T\psi_x(L,t)\left(a\varphi_{x}(L,t)+\frac1c\psi_{x}(L,t)\right)dt \\
&+\frac{T}{2}\int_0^T\varphi_x(L,t)\left(\varphi_{x}(L,t)+\frac{ab}{c}\psi_{x,t}(L,t)\right)dt \\
&+\frac{brT}{2c^2}\int_0^T\psi^2(L,t)dt-\frac{brT}{2c^2}\int_0^T\psi^2(0,t)dt.\end{aligned}$$ Finally, is obtained by applying Young inequality in the right hand side of the above inequality.
Nonlinear System
----------------
In this subsection, attention will be given to the full nonlinear system -. The proof of the lemma below is available in [@bonasunzhang2003 Lemma 3.1] and, therefore, we will omit it.
\[lem3\] There exists a constant $C>0$, such that, for any $T>0$ and $(u,v) \in {\mathcal{Z}}_T$, $$\|uv_x\|_{L^1(0,T;L^2(0,L))}\leq C(T^{\frac{1}{2}}+T^\frac{1}{3})\|u\|_{{\mathcal{Z}}_T}\|v\|_{{\mathcal{Z}}_T}.$$
We first show that system - is locally well-posed in the space ${\mathcal{Z}}_T$.
\[nonlinearteo\] For any $(u^0,v^0) \in {\mathcal{X}}$ and $\overrightarrow{h}=(h_0,h_1,h_2), \overrightarrow{g}=(g_0,g_1,g_2) \in {\mathcal{H}}_T$, there exists $T^*>0$, depending on $\|(u^0,v^0)\|_{{\mathcal{X}}}$, such that the problem - admits a unique solution $(u,v) \in {\mathcal{Z}}_{T^*}$ with $$\partial_x^k u,\partial_x^k v \in L^{\infty}_x(0,L;H^{\frac{1-k}{3}}(0,T^*)), \quad k=0,1,2.$$ Moreover, the corresponding solution map is Lipschitz continuous.
Let ${\mathcal{F}}_T= \left\lbrace (u,v) \in {\mathcal{Z}}_T: (u,v) \in L^{\infty}_x(0,L;(H^{\frac{1-k}{3}}(0,T))^2), k=0,1,2\right\rbrace$ equipped with the norm $$\|(u,v)\|_{{\mathcal{F}}_T} = \|(u,v)\|_{{\mathcal{Z}}_T}+\sum_{k=0}^{2}\|(u,v)\|_{L^{\infty}_x(0,L;(H^{\frac{1-k}{3}}(0,T))^2)}.$$ Let $0< T^* \leq T$ to be determined later. For each $u,v \in {\mathcal{F}}_{T^*}$, consider the problem $$\label{e1'}
\left\lbrace \begin{tabular}{l l}
$\omega_t + \omega_{xxx} + a\eta_{xxx} =f(u,v)$, & in $(0,L)\times (0,T^*)$,\\
$\eta_t +\frac{ab}{c}\omega_{xxx} + \frac{1}{c}\eta_{xxx} =s(u,v)$, & in $(0,L)\times (0,T^*)$,\\
$\omega_{xx}(0,t) = h_0(t),\,\,\omega_x(L,t) = h_1(t),\,\,\omega_{xx}(L,t) = h_2(t)$,& in $(0,T^*)$,\\
$\eta_{xx}(0,t) =g_0(t),\,\,\eta_x(L,t) = g_1(t),\,\, \eta_{xx}(L,t) = g_2(t)$,& in $(0,T^*)$,\\
$\omega(x,0)=u^0(x), \quad v(x,0) = v^0(x)$, & in $(0,L)$,
\end{tabular}\right.$$ where $$f(u,v)=-a_1(vv_x)-a_2(uv)_x$$ and $$s(u,v)=-\frac{r}{c}v_x -\frac{a_2b}{c}(uu_x)-\frac{a_1b}{c}(uv)_x.$$ Since $\|v_x\|_{L^1(0,T^*;L^2(0,L))}\leq \beta^{\frac12}\|v\|_{{\mathcal{Z}}_{T^*}}$, from Lemma \[lem3\] we deduce that $f(u,v)$ and $s(u,v)$ belong to $L^1(0,T^*;L^2(0,L))$ and satisfies $$\label{f,g}
\|(f,s)\|_{L^1(0,T^*;(L^2(0,L))^2)}\leq C_1 ((T^*)^{\frac12}+(T^*)^{\frac13})\left( \|u\|_{{\mathcal{Z}}_{T^*}}^2+(\|u\|_{{\mathcal{Z}}_{T^*}}+1)\|v\|_{{\mathcal{Z}}_{T^*}}+\|v\|_{{\mathcal{Z}}_{T^*}}^2\right),$$ for some positive constant $C_1$. Then, according to Proposition \[prop1\], we can define the operator $$\Gamma: {\mathcal{F}}_{T^*} \rightarrow {\mathcal{F}}_{T^*}, \quad \text{given by} \quad \Gamma(u,v)=(\omega,\eta),$$ where $(\omega,\eta)$ is the solution of (\[e1’\]). Moreover, $$\label{f,g,h}
\|\Gamma(u,v)\|_{{\mathcal{F}}_{T^*}}\leq C\left\lbrace \|(u^0,v^0)\|_{{\mathcal{X}}}+\|(\overrightarrow{h},\overrightarrow{g})\|_{{\mathcal{H}}_{T^*}}+\|(f,s)\|_{L^1(0,T^*;(L^2(0,L))^2)} \right\rbrace,$$ where the positive constant $C$ depends only on $T^*$. Combining and , we obtain $$\begin{aligned}
\|\Gamma(u,v)\|_{{\mathcal{F}}_{T^*}}\leq &C\left\lbrace \|(u^0,v^0)\|_{{\mathcal{X}}}+\|(\overrightarrow{h},\overrightarrow{g})\|_{{\mathcal{H}}_{T^*}}\right\rbrace \\
&+CC_1((T^*)^{\frac12}+(T^*)^{\frac13})\left( \|u\|_{{\mathcal{Z}}_{T^*}}^2+(\|u\|_{{\mathcal{Z}}_{T^*}}+1)\|v\|_{{\mathcal{Z}}_{T^*}}+\|v\|_{{\mathcal{Z}}_{T^*}}^2\right).\end{aligned}$$ Let $(u,v) \in B_{r}(0):=\left\lbrace (u,v) \in {\mathcal{F}}_{T^*}: \|(u,v)\|_{{\mathcal{F}}_{T^*}}\leq r\right\rbrace$, where $r=2C\left\lbrace \|(u^0,v^0)\|_{{\mathcal{X}}}+\|(\overrightarrow{h},\overrightarrow{g})\|_{{\mathcal{H}}_T}\right\rbrace$. From the estimate above, it follows that $$\label{e2'}
\|\Gamma(u,v)\|_{{\mathcal{F}}_{T^*}}\leq \frac{r}{2}+CC_1((T^*)^{\frac12}+(T^*)^{\frac13})\left( 3r+1\right)r.$$ Then, by choosing $T^{*}>0$, such that $$\label{beta1}
CC_1((T^*)^{\frac12}+(T^*)^{\frac13})\left( 3r+1\right)\leq \frac{1}{2},$$ from (\[e2’\]), we have $$\|\Gamma(u,v)\|_{{\mathcal{F}}_{T^*}}\leq r.$$ Thus, we conclude that $$\Gamma: B_r(0)\subset{\mathcal{F}}_{T^*} \rightarrow B_r(0).$$ On the other hand, $\Gamma(u_1,v_1)-\Gamma(u_2,v_2)$ solves the system $$\left\lbrace \begin{tabular}{l l}
$\omega_t + \omega_{xxx} + a\eta_{xxx} =f(u_1,v_1)-f(u_2,v_2)$, & in $(0,L)\times (0,T^*)$,\\
$\eta_t +\frac{ab}{c}\omega_{xxx} + \frac{1}{c}\eta_{xxx} =s(u_1,v_1)-s(u_2,v_2) $, & in $(0,L)\times (0,T^*)$,\\
$\omega_{xx}(0,t) = \omega_x(L,t) = \omega_{xx}(L,t) = 0$,& in $(0,T^*)$,\\
$\eta_{xx}(0,t)=\eta_x(L,t) = \eta_{xx}(L,t) = 0$,& in $(0,T^*)$,\\
$\omega(x,0)=0, \quad v(x,0) = 0$, & in $(0,L)$,
\end{tabular}\right.$$ where, $f(u,v)$ and $s(u,v)$ were defined in . Note that $$\begin{aligned}
|f(u_1,v_1)-f(u_2,v_2)| \leq C_2|\left( (v_2-v_1)v_{2,x}+ v_1(v_2-v_1)_x+(u_2(v_2-v_1))_x +((u_2-u_1)v_1)_x\right)|\end{aligned}$$ and $$\begin{gathered}
|s(u_1,v_1)-s(u_2,v_2)| \leq C_2|\left( (v_2-v_1)_x+ (u_2-u_1)u_{2,x}+ u_1(u_2-u_1)_x \right. \\
\left. +(u_2(v_2-v_1))_x +((u_2-u_1)v_1)_x\right)|,\end{gathered}$$ for some positive constant $C_2$. Then, Proposition \[prop1\] and Lemma \[lem3\], give us the following estimate $$\begin{aligned}
\|\Gamma(u_1,v_1)-\Gamma(u_2,v_2)\|_{{\mathcal{F}}_{T^*}}&\leq C\|(f(u_1,v_1)-f(u_2,v_2), s(u_1,v_1)-s(u_2,v_2))\|_{L^1(0,T^*;(L^2(0,L))^2)} \\
&\leq C_3((T^*)^{\frac{1}{2}}+(T^*)^{\frac{1}{3}})(8r+1)\|(u_1-u_2,v_1-v_2)\|_{{\mathcal{F}}_{T^*}},\end{aligned}$$ for some positive constant $C_3$. Choosing $T^*$, satisfying and such that $$C_3((T^*)^{\frac{1}{2}}+(T^*)^{\frac{1}{3}})(8r+1) \leq \frac{1}{2},$$ we obtain $$\|\Gamma(u_1,v_1)-\Gamma(u_2,v_2)\|_{{\mathcal{F}}_{T^*}}\leq \frac{1}{2}\|(u_1-u_2,v_1-v_2)\|_{{\mathcal{F}}_{T^*}}.$$ Hence, $\Gamma: B_r(0) \rightarrow B_r(0)$ is a contraction and, by Banach fixed point theorem, we obtain a unique $(u,v) \in B_r(0)$, such that $\Gamma(u,v) = (u,v) \in {\mathcal{F}}_{T^*}$ and, therefore, the proof is complete.
\[integralform\] From the proof of Proposition \[prop1\], we deduce that solution of the system (\[gg1\])- can be written as $$\begin{aligned}
\left( \begin{array}{cc}
u(t)\\ v(t)
\end{array} \right)&= W_0(t)\left( \begin{array}{cc}
u^0(x) \\ v^0(x)
\end{array} \right) + W_{bdr}(t)\left( \begin{array}{cc}
\overrightarrow{h} \\ \overrightarrow{g}
\end{array} \right)\\
& -\int_0^t W_0(t-\tau) \left( \begin{array}{cc}
a_1(vv_x)(\tau)+a_2(uv)_x(\tau)\vspace{2mm} \\
\frac{r}{c}v_x(\tau)+\frac{a_2b}{c}(uu_x)(\tau)+\frac{a_1b}{c}(uv)_x(\tau)
\end{array} \right)d\tau,\end{aligned}$$ with $$W_0(t) = \left( \begin{array}{cc}
W_0^{-}(t) & 0 \\
0 & W_0^{+}(t)
\end{array}\right) \quad \text{and} \quad W_{bdr}(t) = \left( \begin{array}{cc}
W_{bdr}^{-}(t) & 0 \\
0 & W_{bdr}^{+}(t)
\end{array}\right),$$ where $\{W_0^{\pm}(t)\}_{t\geq 0}$ are the $C_0$-semigroup in the space $L^2(0,L)$ generated by the linear operators $$A^{\pm}=-\alpha_{\pm}g''',$$ where $$\alpha_{\pm} = -\frac{1}{2}\left(\left(\frac{1}{c}-1\right)\pm \sqrt{\left(\frac{1}{c}-1\right)^2+\frac{4a^2b}{c}}\right),$$ with domain $$D(A^{\pm})=\{g \in H^3(0,L): g''(0)=g'(L)=g''(L)=0\},$$ and $\{W_{bdr}^{\pm}(x)\}_{t\geq 0}$ is the operator defined in Lemma\[l4-d\].
Exact Boundary Controllability for the Linear System {#Sec3}
====================================================
In this section, we study the existence of controls $\overrightarrow{h}:=(h_0,h_1,h_2)$ and $\overrightarrow{g}:=(g_0,g_1,g_2) \in {\mathcal{H}}_T$, such that the solution $(u,v)$ of the system $$\label{ggln4}
\left\lbrace \begin{tabular}{l l}
$u_t + u_{xxx} + av_{xxx} =0$ & in $(0,L)\times (0,T)$,\\
$v_t +\frac{r}{c}v_x+\frac{ab}{c}u_{xxx} + \frac{1}{c}v_{xxx} =0$ & in $(0,L)\times (0,T)$,\\
$u(x,0)=u^0(x), \quad v(x,0) = v^0(x)$, & in $(0,L)$,
\end{tabular}\right.$$ satisfying the boundary conditions $$\label{ggln4b}
\left\lbrace\begin{tabular}{l l}
$u_{xx}(0,t) = h_0(t),\,\,u_x(L,t) = h_1(t),\,\,u_{xx}(L,t) = h_2(t)$ & in $(0,T)$,\\
$v_{xx}(0,t) =g_0(t),\,\,v_x(L,t) = g_1(t),\,\, v_{xx}(L,t) = g_2(t)$ & in $(0,T)$,
\end{tabular}\right.$$ satisfies $$\label{finaldata}
u(\cdot,T)=u^1(\cdot), \qquad \text{and} \qquad v(\cdot,T)=v^1(\cdot).$$
More precisely, we have the following definition:
Let $T > 0$. System - is exact controllable in time $T$ if for any initial and final data $(u^0,v^0)$ and $(u^1,v^1)$ in ${\mathcal{X}}$, there exist control functions $\overrightarrow{h}=(h_0,h_1,h_2)$ and $\overrightarrow{g}=(g_0,g_1,g_2)$ in ${\mathcal{H}}_T$, such that the solution of - satisfies .
Without any loss of generality, we shall consider only the case $u^0 = v^0 = 0$. Indeed, let $(u^0,v^0)$, $(u^1,v^1)$ in ${\mathcal{X}}$ and $\overrightarrow{h}$, $\overrightarrow{g}$ in ${\mathcal{H}}_T$ be controls which lead the solution $(\widetilde{u}, \widetilde{v})$ of from the zero initial data to the final state $(u^1,v^1)-(u(T),v(T))$, where $(u,v)$ is the mild solution corresponding to - with initial data $(u^0,v^0)$. It follows immediately that these controls also lead to the solution $(\widetilde{u},\widetilde{v})+(u,v)$ of - from $(u^0,v^0)$ to the final state $(u^1,v^1)$.
In the following pages, we will analyze the exact controllability of the system - for different combinations of four controls and one control.
Four Controls
-------------
### Case 1
Consider the following boundary conditions: $$\label{gglnb1}
\left\lbrace\begin{tabular}{l l}
$u_{xx}(0,t) = h_0(t),\,\,u_x(L,t) = h_1(t),\,\,u_{xx}(L,t) = 0$ & in $(0,T)$,\\
$v_{xx}(0,t) =g_0(t),\,\,v_x(L,t) = g_1(t),\,\, v_{xx}(L,t) = 0$ & in $(0,T)$.
\end{tabular}\right.$$ We first give an equivalent condition for the exact controllability property.
\[equicontrol\] For any $(u^1,v^1)$ in ${\mathcal{X}}$, there exist four controls $\overrightarrow{h}=(h_0,h_1,0)$ and $\overrightarrow{g}=(g_0,g_1,0)$ in ${\mathcal{H}}_T$, such that the solution $(u,v)$ of - satisfies if and only if $$\begin{aligned}
\int_0^L\left(u^1(x)\varphi^1(x)+v^1(x)\psi^1(x)\right) dx =& \int_0^T h_0(t)\left(\varphi(0,t)+\frac{ab}{c}\psi(0,t)\right)dt\nonumber\\
& +\int_0^T h_1(t)\left(\varphi_x(L,t)+\frac{ab}{c}\psi_x(L,t)\right)dt\nonumber \\
&+\int_0^T g_0(t)\left(a\varphi(0,t)+\frac{1}{c}\psi(0,t)\right)dt\label{ad2}\\
&+\int_0^T g_1(t)\left(a\varphi_x(L,t)+\frac{1}{c}\psi_x(L,t)\right)dt,\nonumber\end{aligned}$$ for any $(\varphi^1,\psi^1)$ in ${\mathcal{X}}$, where $(\varphi,\psi)$ is the solution of the backward system - with initial data $(\varphi^1,\psi^1)$.
The relation is obtained by multiplying the equations in by the solution $(\varphi,\psi)$ of -, integrating by parts and using the boundary conditions .
The following observability inequality plays a fundamental role for the study of the controllability properties.
\[prop4\] For $T>0$ and $L> 0$, there exists a constant $C:=C(T,L) > 0$, such that $$\begin{aligned}
\label{obineq1}
\|(\varphi^1, \psi^1)\|_{{\mathcal{X}}}^2\leq C&\left\lbrace \left\|(-\Delta_t)^{\frac16}\left(\varphi(0,\cdot)+\frac{ab}{c}\psi(0,\cdot)\right)\right\|_{L^2(0,T)}^2 + \left \|\varphi_x(L,\cdot)+\frac{ab}{c}\psi_x(L,\cdot)\right \|_{L^2(0,T)}^2 \right. \notag \\
&\left. +\left \|(-\Delta_t)^{\frac16}\left(a\varphi(0,\cdot)+\frac{1}{c}\psi(0,\cdot)\right)\right \|_{L^2(0,T)}^2 + \left \|a\varphi_x(L,\cdot)+\frac{1}{c}\psi_x(L,\cdot)\right \|_{L^2(0,T)}^2\right\rbrace,\end{aligned}$$ for any $(\varphi^1,\psi^1) \in {\mathcal{X}}$, where $(\varphi,\psi)$ is a solution of - with initial data $(\varphi^1,\psi^1)$, where $\Delta_t:=\partial_t^2$.
We argue by contradiction, as in [@rosier Proposition 3.3], and suppose that does not hold. In this case, we obtain a sequence $\{(\varphi^1_n, \psi^1_n)\}_{n \in {\mathbb{N}}}$, satisfying $$\begin{aligned}
\label{e3}
1=\|(\varphi^1_n, \psi^1_n)\|_{{\mathcal{X}}}^2&\geq n\left\lbrace \left \|(-\Delta_t)^{\frac16}\left(\varphi_n(0,\cdot)+\frac{ab}{c}\psi_n(0,\cdot)\right)\right \|_{L^{2}(0,T)}^2 + \left \|\varphi_{n,x}(L,\cdot)+\frac{ab}{c}\psi_{n,x}(L,\cdot)\right \|_{L^2(0,T)}^2 \right. \notag \\
&\left. +\left \|(-\Delta_t)^{\frac16}\left(a\varphi_n(0,\cdot)+\frac{1}{c}\psi_n(0,\cdot)\right)\right \|_{L^2(0,L)}^2 + \left \|a\varphi_{n,x}(L,\cdot)+\frac{1}{c}\psi_{n,x}(L,\cdot)\right \|_{L^2(0,T)}^2\right\rbrace.\end{aligned}$$ Consequently, imply that $$\label{e4}
\begin{cases}
\varphi_n(0,\cdot)+\frac{ab}{c}\psi_n(0,\cdot) \rightarrow 0 & \text{in} \quad H^{\frac{1}{3}}(0,T), \\
a\varphi_n(0,\cdot)+\frac{1}{c}\psi_n(0,\cdot) \rightarrow 0 & \text{in} \quad H^{\frac{1}{3}}(0,T), \\
\varphi_{n,x}(L,\cdot)+\frac{ab}{c}\psi_{n,x}(L,\cdot) \rightarrow 0 & \text{in} \quad L^2(0,T), \\
a\varphi_{n,x}(L,\cdot)+\frac{1}{c}\psi_{n,x}(L,\cdot) \rightarrow 0 & \text{in} \quad L^2(0,T),
\end{cases}$$ as $n\rightarrow\infty$. Since $1-a^2b >0$, guarantees that the following converge hold $$\label{e4'}
\begin{cases}
\varphi_n(0,\cdot) \rightarrow 0, \quad \psi_n(0,\cdot) \rightarrow 0 &\text{in} \quad H^{\frac{1}{3}}(0,T),\\
\varphi_{n,x}(L,\cdot)\rightarrow 0,\quad \psi_{n,x}(L,\cdot) \rightarrow 0 &\text{in} \quad L^2(0,T),\\
\end{cases}$$ as $n\rightarrow\infty$. The next steps are devoted to pass the strong limit in the left hand side of . First, observe that from Proposition \[hiddenregularities\] we deduce that $\{ (\varphi_n,\psi_n)\}_{n \in {\mathbb{N}}}$ is bounded in $L^2(0,T;(H^1(0,L))^2)$. Then, implies that $\{ (\varphi_{t,n},\psi_{t,n})\}_{n \in {\mathbb{N}}}$ is bounded in $L^2(0,T; (H^{-2}(0,L))^2)$ and the compact embedding $$H^1(0,L) \hookrightarrow L^2(0,L) \hookrightarrow H^{-2}(0,L)$$ allows us to conclude that $\{ (\varphi_n,\psi_n)\}_{n \in {\mathbb{N}}}$ is relatively compact in $L^2(0,T;{\mathcal{X}})$. Consequently, we obtain a subsequence, still denoted by the same index $n$, satisfying $$\label{e8}
(\varphi_n,\psi_n) \rightarrow (\varphi,\psi) \mbox{ in } L^2(0,T;{\mathcal{X}}), \mbox{ as } n\rightarrow\infty.$$ On the other hand, and imply that the sequences $$\text{$\{ \varphi_n(0,\cdot)\}_{n\in{\mathbb{N}}}$\mbox{ and } $\{\psi_n(0,\cdot)\}_{n\in{\mathbb{N}}}$ are bounded in $H^{\frac13}(0,T)$.}$$ Then, the following compact embedding $$\label{e15}
H^{\frac13}(0,T) \hookrightarrow L^2(0,T)$$ guarantees that the above sequences are relatively compact in $L^2(0,T)$, that is, we obtain a subsequence, still denoted by the same index $n$, satisfying $$\label{e10}
\begin{cases}
\varphi_n(0,\cdot) \rightarrow \varphi(0,\cdot) \quad \text{in} \quad L^2(0,T), \\
\psi_n(0,\cdot) \rightarrow \psi(0,\cdot) \quad \text{in} \quad L^2(0,T),
\end{cases}$$ as $n\rightarrow\infty$. Then, from and we deduce that $$\varphi(0,\cdot)=\psi(0,\cdot)=0.$$ Moreover, , and imply that $\{\varphi_n(L,t)\}_{n\in{\mathbb{N}}}$ and $\{\psi_n(L,t)\}_{n\in{\mathbb{N}}}$ are relatively compact in $L^2(0,T)$. Hence, we obtain a subsequence, still denoted by the same index, satisfying $$\label{e9}
\begin{cases}
\varphi_n(L,\cdot) \rightarrow \varphi(L,\cdot) \mbox{ in } L^2(0,T), \\
\psi_n(L,\cdot) \rightarrow \psi(L,\cdot) \mbox{ in } L^2(0,T),
\end{cases}$$ as $n\rightarrow\infty$. In addition, according to Proposition \[prop3\], we have $$\begin{aligned}
\|(\varphi^1_n,\psi^1_n)\|_{{\mathcal{X}}}^2 \leq& \frac{1}{T}\|(\varphi_n,\psi_n)\|_{L^2(0,T;{\mathcal{X}})}+\frac{1}{2}\|\varphi_{n,x}(L,\cdot)\|_{L^2(0,T)}^2\\
&+\frac{b}{2c}\|\psi_{n,x}(L,\cdot)\|_{L^2(0,T)}^2 +\frac{br}{c^2}\|\psi_{n}(L,\cdot)\|_{L^2(0,T)}^2\\
&+\frac{1}{2}\left\|\varphi_{n,x}(L,\cdot)+\frac{ab}{c}\psi_{n,x}(L,\cdot)\right\|_{L^2(0,T)}^2+\frac{b}{2c}\left\|a\varphi_{n,x}(L,\cdot)+\frac{1}{c}\psi_{n,x}(L,\cdot)\right\|_{L^2(0,T)}^2.\end{aligned}$$ Then, from , , and we conclude that $\{(\varphi^1_n,\psi^1_n)\}_{n\in{\mathbb{N}}}$ is a Cauchy sequence in ${\mathcal{X}}$ and, therefore, we get $$\label{e16}
(\varphi^1_n,\psi^1_n) \rightarrow (\varphi^1,\psi^1) \mbox{ in } {\mathcal{X}}, \mbox{ as } n\rightarrow\infty.$$ Thus, Proposition \[hiddenregularities\] together with imply that $$\label{eeee1}
\begin{cases}
\varphi_{n,x}(L,\cdot) \rightarrow \varphi_x(L,\cdot) \mbox{ in } L^2(0,T), \\
\psi_{n,x}(L,\cdot) \rightarrow \psi_x(L,\cdot) \mbox{ in } L^2(0,T)
\end{cases}$$ and $$\begin{cases}
\varphi_{n,xx}(L,\cdot)+\frac{ab}{c}\psi_{n,xx}(L,\cdot) \rightarrow \varphi_{xx}(L,\cdot)+\frac{ab}{c}\psi_{xx}(L,\cdot)& \mbox{ in } L^2(0,T), \\
\varphi_{n,xx}(0,\cdot)+\frac{ab}{c}\psi_{n,xx}(0,\cdot) \rightarrow \varphi_{xx}(0,\cdot)+\frac{ab}{c}\psi_{xx}(0,\cdot)& \mbox{ in } L^2(0,T), \\
a\varphi_{n,xx}(L,\cdot)+\frac{1}{c}\psi_{n,xx}(L,\cdot)+\frac{r}{c}\psi_{n}(L,\cdot) \rightarrow a\varphi_{xx}(L,\cdot)+\frac{1}{c}\psi_{xx}(L,\cdot) +\frac{r}{c}\psi(L,\cdot) &\mbox{ in } L^2(0,T),\\
a\varphi_{n,xx}(0,\cdot)+\frac{1}{c}\psi_{n,xx}(0,\cdot)+\frac{r}{c}\psi_{n}(0,\cdot) \rightarrow a\varphi_{xx}(0,\cdot)+\frac{1}{c}\psi_{xx}(0,\cdot) +\frac{r}{c}\psi(0,\cdot) & \mbox{ in } L^2(0,T),
\end{cases}$$ as $n\rightarrow\infty$. Since $(\varphi_n,\psi_n)$ is a solution of the adjoint system, we obtain that $$\begin{cases}
\varphi_{xx}(L,\cdot)+\frac{ab}{c}\psi_{xx}(L,\cdot) =0,\\
\varphi_{xx}(0,\cdot)+\frac{ab}{c}\psi_{xx}(0,\cdot) =0,\\
a\varphi_{xx}(L,\cdot)+\frac{1}{c}\psi_{xx}(L,\cdot) +\frac{r}{c}\psi(L,\cdot)=0, \\
a\varphi_{xx}(0,\cdot)+\frac{1}{c}\psi_{xx}(0,\cdot) +\frac{r}{c}\psi(L,\cdot)=0.
\end{cases}$$ On the other hand, from and , we have $$\varphi_x(L,\cdot)=\psi_x(L,\cdot)=0.$$ Finally, we obtain that $(\varphi,\psi)$ is a solution of $$\label{e11}
\begin{cases}
\varphi_t + \varphi_{xxx} + a \frac{ab}{c}\psi_{xxx}=0, & \text{in} \,\, (0,L)\times(0,T), \\
\psi_t +\frac{r}{c}\psi_x+a\varphi_{xxx} +\frac{1}{c}\psi_{xxx} =0, & \text{in} \,\, (0,L)\times(0,T), \\
a\varphi_{xx}(L,t)+\frac{1}{c}\psi_{xx}(L,t)+\frac{r}{c}\psi(L,t) =0, & \text{in} \,\, (0,T), \\
a\varphi_{xx}(0,t)+\frac{1}{c}\psi_{xx}(0,t)+\frac{r}{c}\psi(0,t) =0, & \text{in} \,\, (0,T), \\
\varphi_{xx} (L,t) +\frac{ab}{c}\psi_{xx} (L,t) =0, & \text{in} \,\, (0,T). \\
\varphi_{xx} (0,t) +\frac{ab}{c}\psi_{xx} (0,t) =0, & \text{in} \,\, (0,T). \\
\varphi_x(0,t)=\psi_x(0,t) =0, & \text{in} \,\, (0,T),\\
\varphi(x,T)= \varphi^1(x), \qquad \psi(x,T)= \psi^1(x), & \text{in} \,\, (0,L),
\end{cases}$$ satisfying the additional boundary conditions $$\label{e12}
\varphi(0,t)=\psi(0,t)=\varphi_x(L,t)=\psi_x(L,t)=0 \,\, \text{ in } \,\, (0,T)$$ and $$\label{e13}
\|(\varphi^1,\psi^1)\|_{{\mathcal{X}}}=1.$$ Observe that implies that the solutions of - can not be identically zero. However, by Lemma \[lem\], one can conclude that $(\varphi,\psi)=(0,0)$, which drive us to a contradiction.
\[lem\] For any $T > 0$, let $N_T$ denote the space of the initial states $(\varphi^1,\psi^1) \in {\mathcal{X}}$, such that the solution of satisfies . Then, $N_T=\{0\}$.
The proof uses the same arguments as those given in [@rosier].
If $N_T\neq\{0\}$, the map $(\varphi^1,\psi^1) \in N_T \rightarrow A(N_T)\subset {\mathbb{C}}N_T$ (where ${\mathbb{C}}N_T$ denote the complexification of $N_T$) has (at least) one eigenvalue. Hence, there exist $\lambda \in {\mathbb{C}}$ and $\varphi_0,\psi_0 \in H^3(0,L)\setminus \{ 0 \}$, such that $$\begin{cases}\lambda\varphi_0+ \varphi'''_0 +\frac{ab}{c}\psi'''_0=0, & \text{in} \,\, (0,L), \\
\lambda\psi_0 +\frac{r}{c}\psi'_0+a\varphi'''_{0} +\frac{1}{c}\psi'''_{0}=0, & \text{in} \,\, (0,L), \\
\varphi'_0(x)=\psi'_0(x)=0, & \text{in} \,\, \{0,L\},\\
a\varphi''_{0}(x)+\frac{1}{c}\psi''_{0}(x)+\frac{r}{c}\psi_0(x) =0, & \text{in} \,\, \{0,L\},\\
\varphi''_{0} (x) +\frac{ab}{c}\psi''_{0} (x) =0, & \text{in} \,\, \{0,L\}, \\
\varphi_0(0)=\psi_0(0)=0. \\\end{cases}$$ The notation $\{0,L\}$, used above, mean that the expression is applied in $0$ and $L$.
Since $1-a^2b >0$, the above system becomes $$\label{e14}
\begin{cases}
\lambda\varphi_0+ \varphi'''_0 + \frac{ab}{c}\psi'''_0=0, & \text{in} \,\, (0,L), \\
\lambda\psi_0 +\frac{r}{c}\psi'_0+a\varphi'''_{0} +\frac{1}{c}\psi'''_{0}=0, & \text{in} \,\, (0,L), \\
\varphi_0(0)=\varphi'_0(0)=\varphi''_0(0)=0, \\
\psi_0(0)=\psi'_0(0) =\psi''_0(0)=0.
\end{cases}$$ By straightforward computations we see that $(\varphi_0,\psi_0)=(0,0)$ is a solution of for all $L > 0$, which concludes the proof of Lemma \[lem\] and Proposition \[prop4\].
The following theorem gives a positive answer for the control problem:
\[teo1\] Let $T > 0$ and $L>0$. Then, the system - is exactly controllable in time T.
Let us denote by $\Gamma$ the linear and bounded map defined by $$\begin{tabular}{r c c c}
$\Gamma :$ & $L^2(0, L) \times L^2(0, L)$ & $\longrightarrow$ & $L^2(0, L) \times L^2(0, L)$ \\
& $(\varphi^1(\cdot), \psi^1(\cdot))$ & $\longmapsto$ & $\Gamma(\varphi^1(\cdot), \psi^1(\cdot))=(u(\cdot,T), v(\cdot,T))$,
\end{tabular}$$ where $(u,v)$ is the solution of -, with $$\label{contr1}
\begin{cases}
h_0(t) = (-\Delta_t)^{\frac13}\left(\varphi(0,t)+\frac{ab}{c}\psi(0,t)\right), & h_1(t)= \varphi_x(L,t)+\frac{ab}{c}\psi_x(L,t), \\
g_0(t)=(-\Delta_t)^{\frac13}\left(a\varphi(0,t)+\frac{1}{c}\psi(0,t)\right), & g_1(t)=a\varphi_x(L,t)+\frac{1}{c}\psi_x(L,t),
\end{cases}$$ and $(\varphi,\psi)$ the solution of the system - with $\Delta_t= \partial^2_t$ and initial data $(\varphi^1,\psi^1)$. According to Lemma \[equicontrol\] and Proposition \[prop4\], we obtain $$\begin{aligned}
\left (\Gamma(\varphi^1, \psi^1), (\varphi^1, \psi^1) \right)_{(L^2(0,L))^2} = &\left\|\varphi_x(L,\cdot)+\frac{ab}{c}\psi_x(L,\cdot)\right\|_{L^2(0,T)}^2 + \left\|a\varphi_x(L,\cdot)+\frac{1}{c}\psi_x(L,\cdot)\right\|_{L^2(0,T)}^2 \\
&+\left((-\Delta_t)^{\frac13}\left(\varphi(0,\cdot)+\frac{ab}{c}\psi(0,\cdot)\right), \varphi(0,\cdot)+\frac{ab}{c}\psi(0,\cdot) \right)_{L^2(0,T)} \\
&+ \left( (-\Delta_t)^{\frac13}\left(a\varphi(0,\cdot)+\frac{1}{c}\psi(0,\cdot)\right), a\varphi(0,\cdot)+\frac{1}{c}\psi(0,\cdot) \right)_{L^2(0,T)} \\
=&\left\|\varphi_x(L,\cdot)+\frac{ab}{c}\psi_x(L,\cdot)\right\|_{L^2(0,T)}^2 + \left\|a\varphi_x(L,\cdot)+\frac{1}{c}\psi_x(L,\cdot)\right\|_{L^2(0,T)}^2 \\
&+\left\|(-\Delta_t)^{\frac16}\left(a\varphi(0,\cdot)+\frac{1}{c}\psi(0,\cdot)\right)\right\|_{L^2(0,T)}^2\\
&+ \left\|(-\Delta_t)^{\frac16}\left(\varphi(0,\cdot)+\frac{ab}{c}\psi(0,\cdot)\right)\right\|_{L^2(0,T)}^2 \\
\geq& C^{-1} \|(\varphi^1, \psi^1)\|_{{\mathcal{X}}}^2.\end{aligned}$$ Thus, by the Lax-Milgram theorem, $\Gamma$ is invertible. Consequently, for given $(u^1,v^1) \in (L^2(0, L))^2$, we can define $(\varphi^1,\psi^1) := \Gamma^{-1}( u^1,v^1)$ to solve the system - and get $(\varphi,\psi) \in {\mathcal{Z}}_T$. Then, if $h_0(t)$, $h_1(t)$, $g_0(t)$ and $g_1(t)$ are given by , the corresponding solution $(u,v)$ of the system -, satisfies $$(u(\cdot,0),v(\cdot,0))=(0,0)\quad \text{and} \quad (u(\cdot,T),v(\cdot,T))=(u^1(\cdot),=v^1(\cdot)).$$
\[remark1\] An important question is whether the exact controllability holds, in time $T>0$, when we consider the boundary condition with another configuration, for example, $$\label{gglnb1'}
\left\lbrace\begin{tabular}{l l l l}
$u_{xx}(0,t) = 0$ & $u_x(L,t) = h_1(t)$ & $u_{xx}(L,t) = h_2(t)$,& in $(0,T)$,\\
$v_{xx}(0,t) =0,$ & $v_x(L,t) = g_1(t),$ & $v_{xx}(L,t) = g_2(t)$,& in $(0,T)$.
\end{tabular}\right.$$ Observe that, in this case it would be necessary to prove that the following observability inequality $$\begin{aligned}
\|(\varphi^1, \psi^1)\|_{{\mathcal{X}}}^2\leq& C\left\lbrace\left \|(-\Delta_t)^{\frac{1}{6}}\left(\varphi(L,\cdot)+\frac{ab}{c}\psi(L,\cdot)\right)\right \|_{L^2(0,T)}^2 +\left \|\varphi_x(L,\cdot)+\frac{ab}{c}\psi_x(L,\cdot)\right \|_{L^2(0,T)}^2 \right. \notag \\
&\left. +\left \|(-\Delta_t)^{\frac{1}{6}}\left(a\varphi(L,\cdot)+\frac{1}{c}\psi(L,\cdot)\right)\right \|_{L^2(0,T)}^2 +\left \|a\varphi_x(L,\cdot)+\frac{1}{c}\psi_x(L,\cdot)\right \|_{L^2(0,T)}^2\right\rbrace,\end{aligned}$$ holds for any $(\varphi^1,\psi^1)$ in ${\mathcal{X}}$, where $(\varphi,\psi)$ is solution of - with initial data $(\varphi^1,\psi^1)$. It can be done using Proposition \[hiddenregularities\] together with the contradiction argument used in the proof of Proposition \[prop4\]. Thus, the next result about the exact controllability of the system - also holds:
\[teo2\] Let $T > 0$ and $L>0$. Then, the system - is exactly controllable in time T.
### Case 2
We consider the following boundary conditions: $$\label{gglnb2}
\left\lbrace\begin{tabular}{l l l l}
$u_{xx}(0,t) = 0$ & $u_x(L,t) = h_1(t)$ & $u_{xx}(L,t) = 0$,& in $(0,T)$,\\
$v_{xx}(0,t) =g_0(t),$ & $v_x(L,t) = g_1(t),$ & $v_{xx}(L,t) = g_2(t)$,& in $(0,T)$.
\end{tabular}\right.$$ First, as in subsection above, we give an equivalent condition for the exact controllability property. It can be done using the same idea of the proof of Lemma \[equicontrol\].
\[equicontrol2\] For any $(u^1,v^1)$ in ${\mathcal{X}}$, there exist four controls $\overrightarrow{h}=(0,h_1,0)$ and $\overrightarrow{g}=(g_0,g_1,g_2)$ in ${\mathcal{H}}_T$, such that the solution $(u,v)$ of - satisfies if and only if $$\begin{aligned}
\int_0^L(u^1(x)\varphi^1(x)+v^1(x)\psi^1(x))dx=&\int_0^T g_0(t)\left(a\varphi(0,t)+\frac{1}{c}\psi(0,t)\right)dt\nonumber\\
&+\int_0^t g_1(t)\left(a\varphi_x(L,t)+\frac{1}{c}\psi_x(L,t)\right)dt\nonumber \\
&-\int_0^Tg_2(t)\left(a\varphi(L,t)+\frac{1}{c}\psi(L,t)\right)dt\label{ad3} \\
&+\int_0^T h_1(t)\left(\varphi_x(L,t)+\frac{ab}{c}\psi_x(L,t)\right)dt\nonumber,\end{aligned}$$ for any $(\varphi^1,\psi^1)$ in ${\mathcal{X}}$, where $(\varphi,\psi)$ is the solution of the backward system -.
To prove the exact controllability property, it suffices to prove the following observability inequality:
\[prop5\] Let $T > 0$ and $L\in (0,\infty)\setminus {\mathcal{F}}_r$, where ${\mathcal{F}}_r$ is given by . Then, there exists a constant $C (T,L) > 0$, such that $$\begin{aligned}
\|(\varphi^1, \psi^1)\|_{{\mathcal{X}}}^2\leq& C\left\lbrace \left \|(-\Delta_t)^{\frac{1}{6}}\left( a\varphi(0,\cdot)+\frac{1}{c}\psi(0,\cdot)\right)\right \|_{L^2(0,T)}^2 +\left \|\varphi_x(L,\cdot)+\frac{ab}{c}\psi_x(L,\cdot)\right \|_{L^2(0,T)}^2 \right. \notag \\
&\left. +\left \|(-\Delta_t)^{\frac{1}{6}}\left(a\varphi(L,\cdot)+\frac{1}{c}\psi(L,\cdot)\right)\right \|_{L^2(0,T)}^2 +\left \|a\varphi_x(L,\cdot)+\frac{1}{c}\psi_x(L,\cdot)\right \|_{L^2(0,T)}^2\right\rbrace,\end{aligned}$$ for any $(\varphi^1,\psi^1)$ in ${\mathcal{X}}$, where $(\varphi,\psi)$ is solution of - with initial data $(\varphi^1,\psi^1)$, where $\Delta_t :=\partial_t^2$.
We proceed as in the proof of Proposition \[prop4\] using the contradiction argument. Therefore, we will summarize it. Firstly, we show that the sequences $\{(\varphi^1_n, \psi^1_n)\}_{n \in {\mathbb{N}}}$, $$\{a\varphi_n(0,\cdot) +\frac{1}{c}\psi_n(0,\cdot)\}_{n\in {\mathbb{N}}},$$ $$\{a\varphi_n(L,\cdot) +\frac{1}{c}\psi_n(L,\cdot)\}_{n\in {\mathbb{N}}},$$ $$\{a\varphi_{n,x}(L,\cdot) +\frac{1}{c}\psi_{n,x}(L,\cdot)\}_{n\in {\mathbb{N}}}$$ and $$\{\varphi_{n,x}(L,\cdot) +\frac{ab}{c}\psi_{n,x}(L,\cdot)\}_{n\in {\mathbb{N}}},$$ are relatively compact in ${\mathcal{X}}$ and $L^2(0,T;{\mathcal{X}})$, respectively. Next, we proceed as in the proof of Proposition \[prop4\] to get that $$a\varphi_n(0,\cdot) +\frac{1}{c}\psi_n(0,\cdot)\rightarrow0,$$ $$a\varphi_n(L,\cdot) +\frac{1}{c}\psi_n(L,\cdot)\rightarrow0,$$ $$\varphi_{n,x}(L,\cdot)\rightarrow0,\,\,\psi_x(L,\cdot)\rightarrow0,$$ as $n\rightarrow\infty$, and $$||(\varphi,\psi)||_{(L^2(0,L))^2}=1.$$ Finally, combining the hidden regularity of the solutions of the adjoint system and the compact embedding $H^{\frac13}(0,T)\hookrightarrow L^2(0,T)$, we conclude that $(\varphi,\psi)$ satisfies $$\label{e17}
\begin{cases}
\varphi_t + \varphi_{xxx} +\frac{ab}{c}\psi_{xxx}=0 & \mbox{ in }(0,L)\times(0,T),\\
\psi_t +\frac{r}{c}\psi_x+a\varphi_{xxx} +\frac{1}{c}\psi_{xxx} =0 & \mbox{ in } (0,L)\times(0,T), \\
\varphi_{xx}(L,t) +\frac{ab}{c}\psi_{xx}(L,t)=0 & \mbox{ in }(0,T),\\
\varphi_{xx}(0,t) +\frac{ab}{c}\psi_{xx}(0,t)=0 & \mbox{ in }(0,T),\\
a\varphi_{xx}(L,t) +\frac{1}{c}\psi_{xx}(L,t) +\frac{r}{c}\psi(L,t)=0 & \mbox{ in }(0,T),\\
a\varphi_{xx}(0,t) +\frac{1}{c}\psi_{xx}(0,t) +\frac{r}{c}\psi(0,t)=0 & \mbox{ in }(0,T),\\
\varphi_x(0,t)=\psi_x(0,t)=0 & \mbox{ in }(0,T),\\
\varphi(x,T)= \varphi^1(x), \,\,
\psi(x,T)= \psi^1(x) & \mbox{ in } (0,L)
\end{cases}$$ and $$\label{e17'}
\begin{cases}
a\varphi(L,t) +\frac{1}{c}\psi(L,t)=0 & \mbox{ in }(0,T), \\
a\varphi(0,t) +\frac{1}{c}\psi(0,t)=0 & \mbox{ in }(0,T), \\
\varphi_x(L,t)=\psi_x(L,t)=0 & \mbox{ in }(0,T), \\
\|(\varphi,\psi)\|_{{\mathcal{X}}} = 1.
\end{cases}$$
Notice that the solutions of - can not be identically zero. Therefore, from Lemma \[lem1\], one can conclude that $(\varphi,\psi)=(0,0)$, which drive us to a contradiction.
\[lem1\] For any $T > 0$, let $N_T$ denote the space of the initial states $(\varphi^1,\psi^1) \in {\mathcal{X}}$, such that the solution of satisfies . Then, for $L \in (0,\infty)\setminus {\mathcal{F}}_r$, $N_T=\{0\}$.
By the same arguments given in [@rosier], if $N_T\neq \{0\}$, the map $(\varphi^1,\psi^1) \in N_T \rightarrow A(N_T)\subset {\mathbb{C}}N_T$ has (at least) one eigenvalue. Hence, there exist $\lambda \in {\mathbb{C}}$ and $\varphi_0,\psi_0 \in H^3(0,L)\setminus \{ 0 \}$, such that $$\label{e18}
\begin{cases}
\lambda\varphi_0+ \varphi'''_0 + \frac{ab}{c}\psi'''_0=0, & \text{in} \,\, (0,L), \\
\lambda\psi_0 +\frac{r}{c}\psi'_0+a\varphi'''_{0} +\frac{1}{c}\psi'''_{0}=0, & \text{in} \,\, (0,L), \\
a\varphi_0(x)+\frac{1}{c}\psi_0(x)=0, & \text{in} \,\, \{0,L\},\\
\varphi'_0(x)=\psi'_0(x)=0, & \text{in} \,\, \{0,L\},\\
\varphi''_{0} (x) +\frac{ab}{c}\psi''_{0} (x) =0, & \text{in} \,\, \{0,L\}, \\
a\varphi''_{0}(x)+\frac{1}{c}\psi''_{0}(x)+\frac{r}{c}\psi_0(x) =0, & \text{in} \,\, \{0,L\}.
\end{cases}$$ To conclude the proof of the Lemma \[lem1\], we prove that this does not hold if $L \in (0,\infty) \setminus {\mathcal{F}}_r$. To simplify the notation, henceforth we denote $(\varphi_0,\psi_0):=(\varphi,\psi)$.
\[lem2\] Let $L>0$. Consider the assertion $$({\mathcal{N}}):\ \ \exists \lambda \in {\mathbb{C}}, \exists \varphi,\psi \in H^3(0,L)\setminus(0,0),\,\, \text{such that}\,\,
\begin{cases}
\lambda\varphi+ \varphi''' + \frac{ab}{c}\psi'''=0, & \text{in} \,\, (0,L), \\
\lambda\psi +\frac{r}{c}\psi'+a\varphi''' +\frac{1}{c}\psi'''=0, & \text{in} \,\, (0,L), \\
a\varphi(x)+\frac{1}{c}\psi(x)=0, & \text{in} \,\, \{0,L\},\\
\varphi'(x)=\psi'(x)=0, & \text{in} \,\, \{0,L\},\\
\varphi'' (x) +\frac{ab}{c}\psi''(x) =0, & \text{in} \,\, \{0,L\}, \\
a\varphi''(x)+\frac{1}{c}\psi''(x)+\frac{r}{c}\psi(x) =0, & \text{in} \,\, \{0,L\}.
\end{cases}$$ Then, $({\mathcal{N}})$ holds if and only if $L \in {\mathcal{F}}_r$.
*Proof.* We use an argument similar to the one used in [@rosier Lemma 3,5]. Let us introduce the notation $\hat{\varphi}(\xi) =\int_0^L e^{-i x \xi}\varphi(x)dx$ and $\hat{\psi}(\xi) =\int_0^L e^{-i x \xi}\psi(x)dx$. Then, multiplying the first and the second equations in $({\mathcal{N}})$ by $e^{-i x \xi}$ and integrating by part in $(0,L)$, it follows that $$\begin{gathered}
\left( (i\xi)^3+\lambda\right) \hat{\varphi}(\xi) +\frac{ab}{c}(i \xi)^3\hat{\psi}(\xi) \\
+ \left[ \left(\left(\varphi''(x)+\frac{ab}{c}\psi''(x)\right) +(i\xi)\left(\varphi'(x)+\frac{ab}{c}\psi'(x)\right)+(i\xi)^2\left(\varphi(x)+\frac{ab}{c}\psi(x)\right)\right) e^{-ix\xi}\right]_0^L = 0\end{gathered}$$ and $$\begin{gathered}
\left( \frac{1}{c}(i\xi)^3+\frac{r}{c}(i\xi)+\lambda\right) \hat{\psi}(\xi) +a(i \xi)^3\hat{\varphi}(\xi) + \left[ \left(\left(a\varphi''(x)+\frac{1}{c}\psi''(x)+\frac{r}{c}\psi(\xi)\right)\right.\right. \\
\left.\left.+(i\xi)\left(a\varphi'(x)+\frac{1}{c}\psi'(x)\right)+(i\xi)^2\left(a\varphi(x)+\frac{1}{c}\psi(x)\right)\right) e^{-ix\xi}\right]_0^L = 0.\end{gathered}$$ The boundary conditions allow us to conclude that $$\label{eqq1}
\begin{cases}
[(i\xi)^3+\lambda]\hat{\varphi}(\xi)+\dfrac{ab}{c}(i\xi)^3\hat{\psi}(\xi) = (i\xi)^2\left(\varphi(0)+\dfrac{ab}{c}\psi(0)- \left(\varphi(L)+\dfrac{ab}{c}\psi(L)\right)e^{-iL\xi}\right), \\
\\
\dfrac{1}{c}[(i\xi)^3+r(i\xi)+c\lambda]\hat{\psi}(\xi) + a(i\xi)^3\hat{\varphi}(\xi)=0.
\end{cases}$$ Then, from the first equation in , we obtain $$\label{eqq2}
\hat{\varphi}(\xi) = \frac{(i\xi)^2\left(\alpha+\beta e^{-iL\xi}\right)}{(i\xi)^3+\lambda} -\frac{ab(i\xi)^3\hat{\psi}(\xi)}{c\left((i\xi)^3+\lambda\right)},$$ where $\alpha= \varphi(0)+\frac{ab}{c}\psi(0)$ and $\beta=-\varphi(L)-\frac{ab}{c}\psi(L)$. Replacing the above expression in the second equation in it follows that $$\frac{1}{c}\left[ (i\xi)^3+r(i\xi)+c\lambda - \frac{a^2b(i\xi)^6}{(i\xi)^3+\lambda}\right]\hat{\psi}(\xi) =-\frac{a(i\xi)^5\left(\alpha+\beta e^{-iL\xi}\right)}{(i\xi)^3+\lambda}.$$ Thus, $$\label{eqq3}
\hat{\psi}(\xi) =-\frac{ac(i\xi)^5\left(\alpha+\beta e^{-iL\xi}\right)}{(1-a^2b)(i\xi)^6+r(i\xi)^4+(c+1)\lambda(i\xi)^3+r\lambda(i\xi)+c\lambda^2}.$$ Replacing in , we obtain $$\begin{aligned}
\hat{\varphi}(\xi)& = \frac{ (i\xi)^2\left( (i\xi)^3+r(i\xi)+c\lambda \right)\left(\alpha+\beta e^{-iL\xi}\right)}{(1-a^2b)(i\xi)^6+r(i\xi)^4+(c+1)\lambda(i\xi)^3+r\lambda(i\xi)+c\lambda^2}.\end{aligned}$$ Setting $\lambda=ip$, $p \in {\mathbb{C}}$, from the previous identities we can write $\hat{\psi}(\xi)=-i\left[ acf(\xi)\right]$ and $\hat{\varphi}(\xi)=-ig(\xi)$, where $$\begin{cases}
f(\xi) = \dfrac{\xi^5\left( \alpha +\beta e^{-iL\xi}\right)}{P(\xi)}, \\
\\
g(\xi)= \dfrac{\xi^2\left(\xi^3-r\xi-cp\right)\left( \alpha +\beta e^{-iL\xi}\right)}{P(\xi)},
\end{cases}$$ with $$P(\xi):=(1-a^2b)\xi^6-r\xi^4-(c+1)p\xi^3+rp\xi+cp^2.$$ Using Paley-Wiener theorem (see [@yosida Section 4, page 161]) and the usual characterization of $H^2({\mathbb{R}})$ functions by means of their Fourier transforms, we see that $({\mathcal{N}})$ is equivalent to the existence of $p \in {\mathbb{C}}$ and $(\alpha,\beta) \in {\mathbb{C}}^2 \setminus (0,0),$ such that
1. $f$ and $g$ are entire functions in ${\mathbb{C}}$,
2. $\displaystyle\int_{{\mathbb{R}}}|f(\xi)|^2(1+|\xi|^2)^2d\xi<\infty$ and $\int_{{\mathbb{R}}}|g(\xi)|^2(1+|\xi|^2)^2d\xi<\infty$,
3. $\forall \xi \in {\mathbb{C}}$, we have that $|f(\xi)|\leq c_1(1+|\xi|)^ke^{L|Im\xi|}$ and $|g(\xi)|\leq c_1(1+|\xi|)^ke^{L|Im\xi|}$, for some positive constants $c_1 $ and $k$.
Notice that if (i) holds true, then (ii) and (iii) are satisfied. Recall that $f$ and $g$ are entire functions if and only if the roots $\xi_0, \xi_1, \xi_2, \xi_3, \xi_4$ and $\xi_5$ of $P(\xi)$ are roots of $\xi^5\left( \alpha +\beta e^{-iL\xi}\right)$ and $\xi^2(\xi^3-r\xi-cp)\left( \alpha +\beta e^{-iL\xi}\right)$. 0.3 cm Let us first assume that $\xi=0$ is not a root of $P(\xi)$. Thus, it is sufficient to consider the case when $\alpha +\beta e^{-iL\xi}$ and $P(\xi)$ share the same roots. Observe that the roots of $\alpha +\beta e^{-iL\xi}$ are simple, unless $\alpha = \beta= 0$ (Indeed, in this case $\varphi(0)+\frac{ab}{c}\psi(0)=0$ and $\varphi(L)+\frac{ab}{c}\psi(L)=0$ and using the system we conclude that $(\varphi,\psi)=(0,0)$, which is a contradiction). Then, (i) holds provided that the roots of $P(\xi)$ are simple. Therefore, it follows that $({\mathcal{N}})$ is equivalent to the existence of complex numbers $p$ and $\xi_0$ and positive integers $k,l,m,n$ and $s$, such that, if we set $$\label{eqq4}
\xi_1=\xi_0+\frac{2\pi}{L}k, \quad \xi_2=\xi_1+\frac{2\pi}{L}l, \quad \xi_3=\xi_2+\frac{2\pi}{L}m, \quad \xi_4=\xi_3+\frac{2\pi}{L}n\quad \text{and} \quad \xi_5=\xi_4+\frac{2\pi}{L}s,$$ $P(\xi)$ can be written as follows $$P(\xi)=(\xi-\xi_0)(\xi-\xi_1)(\xi-\xi_2)(\xi-\xi_3)(\xi-\xi_4)(\xi-\xi_5).$$ In particular, we obtain the following relations: $$\begin{gathered}
\xi_0+\xi_1+\xi_2+\xi_3+\xi_4+\xi_5=0, \label{eqq6}\end{gathered}$$ $$\begin{gathered}
\xi_0(\xi_1+\xi_2+\xi_3+\xi_4+\xi_5)+\xi_1(\xi_2+\xi_3+\xi_4+\xi_5)+\xi_2(\xi_3+\xi_4+\xi_5) \\
+\xi_3(\xi_4+\xi_5) +\xi_4\xi_5=-\frac{r}{1-a^2b}\label{eqq7}\end{gathered}$$ and $$\begin{gathered}
\xi_0\xi_1\xi_2\xi_3\xi_4\xi_5=\left(\frac{c}{1-a^2b}\right) p^2. \label{eqq8}\end{gathered}$$ and imply that $$\begin{gathered}
\xi_0+\left(\xi_0+\frac{2\pi}{L}k\right)+\left(\xi_0+\frac{2\pi}{L}(k+l)\right)+\left(\xi_0+\frac{2\pi}{L}(k+l+m)\right)+\left(\xi_0+\frac{2\pi}{L}(k+l+m+n)\right) \\
+\left(\xi_0+\frac{2\pi}{L}(k+l+m+n+s)\right)=0.\end{gathered}$$ Straightforward computations lead to $$\label{eqq9}
\xi_0 = -\dfrac{\pi}{3L}(5k+4l+3m+2n+s).$$ On the other hand, from , we obtain $$\begin{gathered}
\xi_0\left(5\xi_0+\frac{2\pi}{L}(5k+4l+3m+2n+s)\right) +\left(\xi_0+\frac{2\pi}{L}k\right) \left(4\xi_0+\frac{2\pi}{L}(4k+4l+3m+2n+s)\right) \\
+\left(\xi_0+\frac{2\pi}{L}(k+l)\right) \left(3\xi_0+\frac{2\pi}{L}(3k+3l+3m+2n+s)\right) \\
+\left(\xi_0 +\frac{2\pi}{L}(k+l+m)\right) \left(2\xi_0+\frac{2\pi}{L}(2k+2l+2m+2n+s)\right) \\
+\left(\xi_0+\frac{2\pi}{L}(k+l+m+n)\right) \left(\xi_0+\frac{2\pi}{L}(k+l+m+n+s)\right) = -\frac{r}{1-a^2b}.\end{gathered}$$ Thus, we have $$\label{eqq12}
15\xi_0^2+\frac{2\pi}{L}(25k+20+15m+10n+5s)\xi_0+\frac{4\pi^2}{L^2}\eta = -\frac{r}{1-a^2b},$$ where $$\begin{gathered}
\eta=k(10k+10l+9m+7n+4s) +l(6k+6l+6m+5n+3s) \\
+m(3k+3l+3m+3n+2s)+n(k+l+m+n+s).\end{gathered}$$ Replacing in , we obtain $$\frac{3rL^2}{1-a^2b}= \pi^2\left(5(5k+4l+3m+2n+s)^2-12\eta\right).$$ From the discussion above, we can conclude that $$\label{eqq14}
\begin{cases}
L=\pi \sqrt{\dfrac{(1-a^2b)\alpha(k,l,m,n,s)}{3r}},\\
\\
\xi_0 = -\dfrac{\pi}{3}(5k+4l+3m+2n+s),\\
\\
p= \sqrt{\dfrac{(1-a^2b)\xi_0\xi_1\xi_2\xi_3\xi_4\xi_5}{c}},
\end{cases}$$ where $$\begin{aligned}
\alpha(k,l,m,n,s):=&5k^2+8l^2+9m^2+8n^2+5s^2+8kl+6km+4kn+2ks+12ml \\
&+8ln+3ls+12mn+6ms+8ns.\end{aligned}$$ Now, we assume that $\xi_0= 0$ is a root of $P(\xi)$. Then, it follows that $p=0$ and $$\begin{cases}
f(\xi)= \dfrac{\xi^5\left(\alpha+\beta e^{-iL\xi}\right)}{(1-a^2b)\xi^6-r\xi^4}=\dfrac{\xi\left(\alpha+\beta e^{-iL\xi}\right)}{(1-a^2b)\xi^2-r},\\
\\
g(\xi)= \dfrac{\xi^2\left(\xi^3-r\xi\right)\left( \alpha +\beta e^{-iL\xi}\right)}{(1-a^2b)\xi^6-r\xi^4}=\dfrac{\left(\xi^2-r\right)\left( \alpha +\beta e^{-iL\xi}\right)}{\xi\left((1-a^2b)\xi^2-r\right)}.
\end{cases}$$ In this case, $({\mathcal{N}})$ holds if and only if $f$ and $g$ satisfy (i), (ii) and (iii). Thus, (i) holds provided that $$\xi_0=0, \quad \xi_1=\sqrt{\frac{r}{1-a^2b}} \quad \text{and} \quad \xi_2=-\sqrt{\frac{r}{1-a^2b}}$$ are roots of $\alpha+\beta e^{-iL\xi}$. Therefore, we can write $\xi_1=\xi_0+\frac{2\pi}{L}k$, for $k \in {\mathbb{Z}}$. Consequently, it follows that $$\label{eqq15}
L= 2\pi k \sqrt{\dfrac{1-a^2b}{r}}.$$ Finally, from and , we deduce that $({\mathcal{N}})$ holds if and only if $L \in {\mathcal{F}}_r$, where ${\mathcal{F}}_r$ is given by . This completes the proof of Lemma \[lem2\], Lemma \[lem1\] and, consequently, the proof of Proposition \[prop5\].
The next result gives a positive answer for the control problem, and can be proved using the same ideas presented in Theorem \[teo1\] and, thus, we will omit the proof.
\[teo3\] Let $T > 0$ and $L \in (0,\infty) \setminus {\mathcal{F}}_r$, where ${\mathcal{F}}_r$ is given by . Then, the system - is exactly controllable in time T.
\[remark2\] As in the previous subsection, the question here is weather system - is exactly controllable with another configuration of the boundary condition, for example, $$\label{gglnb2'}
\left\lbrace\begin{tabular}{l l l l}
$u_{xx}(0,t) = h_0(t)$, & $u_x(L,t) = h_1(t)$, & $u_{xx}(L,t) = h_2(t)$& in $(0,T)$,\\
$v_{xx}(0,t) =0,$ & $v_x(L,t) = g_1(t),$ & $v_{xx}(L,t) = 0$& in $(0,T)$.
\end{tabular}\right.$$ The answer for this question is positive if we prove that the following observability inequality $$\begin{aligned}
\|(\varphi^1, \psi^1)\|_{{\mathcal{X}}}^2\leq C&\left\lbrace\left \|(-\Delta)^{\frac{1}{6}}\left(\varphi(0,\cdot)+\frac{ab}{c}\psi(0,\cdot)\right)\right \|_{L^2(0,T)}^2 +\left \|\varphi_x(L,\cdot)+\frac{ab}{c}\psi_x(L,\cdot)\right \|_{L^2(0,T)}^2 \right. \notag \\
&\left. +\left \|(-\Delta)^{\frac{1}{6}}\left(\varphi(L,\cdot)+\frac{ab}{c}\psi(L,\cdot)\right)\right \|_{L^2(0,T)} +\left \|a\varphi_x(L,\cdot)+\frac{1}{c}\psi_x(L,\cdot)\right \|_{L^2(0,T)}^2\right\rbrace,\end{aligned}$$ holds, for any $(\varphi^1,\psi^1)$ in ${\mathcal{X}}$, where $(\varphi,\psi)$ is solution of - with initial data $(\varphi^1,\psi^1)$. Note that it can be proved using Proposition \[hiddenregularities\] together with the contradiction argument as in the proof of Proposition \[prop5\]. Thus, the exact controllability result is also true in this case.
\[teo4\] Let $T > 0$ and $L \in (0,\infty) \setminus {\mathcal{F}}_r$. Then, the system - is exactly controllable in time T.
One Control
-----------
In this subsection, we intend to prove the exact controllability of the system by using only one boundary control $h_1$ or $g_1$ and fixing $h_0=h_2=g_0=g_2=0$, namely, $$\label{gglnb3}
\left\lbrace\begin{tabular}{l l l l}
$u_{xx}(0,t) = 0$ & $u_x(L,t) = h_1(t),$ & $u_{xx}(L,t) = 0$,& in $(0,T)$,\\
$v_{xx}(0,t) =0,$ & $v_x(L,t) = 0,$ & $v_{xx}(L,t) =0$,& in $(0,T)$.
\end{tabular}\right.$$ or $$\label{gglnb3'}
\left\lbrace\begin{tabular}{l l l l}
$u_{xx}(0,t) = 0$ & $u_x(L,t) = 0$ & $u_{xx}(L,t) = 0$,& in $(0,T)$,\\
$v_{xx}(0,t) =0,$ & $v_x(L,t) = g_1(t),$ & $v_{xx}(L,t) =0$,& in $(0,T)$.
\end{tabular}\right.$$ The result below give us an equivalent condition for the exact controllability and the proof is analogous to the proof of the Lemma \[equicontrol\].
\[equicontrol3\] For any $(u^1,v^1)$ in ${\mathcal{X}}$, there exist one control $\overrightarrow{h}=(0,h_1,0)$ and $\overrightarrow{g}=(0,0,0)$ (resp. $\overrightarrow{h}=(0,0,0)$ and $\overrightarrow{g}=(0,g_1,0)$) in ${\mathcal{H}}_T$, such that the solution $(u,v)$ of - (resp. -) satisfies if and only if $$\begin{gathered}
\int_0^L(u^1(x)\varphi^1(x)+v^1(x)\psi^1(x))dx= \int_0^T h_1(t)\left[\varphi_x(L,t)+\frac{ab}{c}\psi_x(L,t)\right]dt \\
\left(\text{resp.}\quad \int_0^L(u^1(x)\varphi^1(x)+v^1(x)\psi^1(x))dx= \int_0^T g_1(t)\left[a\varphi_x(L,t)+\frac{1}{c}\psi_x(L,t)\right]dt \right)\end{gathered}$$ for any $(\varphi^1,\psi^1)$ in ${\mathcal{X}}$, where $(\varphi,\psi)$ is the solution of the backward system -.
Note that using the change of variable $x' = L-x$ and $t' = T-t$, the system - is equivalent to the following back-forward system $$\begin{aligned}
&\begin{cases}\label{linadj1}
\varphi_t +\varphi_{xxx} +\frac{ab}{c}\psi_{xxx}=0, & \text{in} \,\, (0,L)\times (0,T),\\
\psi_t +\frac{r}{c}\psi_x+a\varphi_{xxx} +\frac{1}{c}\psi_{xxx} =0, & \text{in} \,\, (0,L)\times (0,T), \\
\varphi(x,0)= \varphi^0(x), \,\,
\psi(x,0)= \psi^0(x), & \text{in} \,\, (0,L),
\end{cases}\end{aligned}$$ with boundary conditions $$\label{linadjbound1}
\begin{cases}
\varphi_{xx} (x,t) +\frac{ab}{c}\psi_{xx} (x,t) =0, & \text{in} \,\, \{0,L\}\times (0,T),\\
a\varphi_{xx}(x,t)+\frac{1}{c}\psi_{xx}(x,t)+\frac{r}{c}\psi(x,t) =0, & \text{in} \,\, \{0,L\}\times (0,T), \\
\varphi_x(L,t)=\psi_x(L,t) =0, & \text{in} \,\, (0,T).
\end{cases}$$ It is well know (according to the previous sections) that the observability inequality $$\begin{aligned}
\label{obineq3}
\|(\varphi^0, \psi^0)\|_{{\mathcal{X}}}^2\leq C \left \|\varphi_x(0,\cdot)+\frac{ab}{c}\psi_x(0,\cdot)\right \|_{L^2(0,T)}^2\end{aligned}$$ or $$\begin{aligned}
\label{obineq3'}
\|(\varphi^0, \psi^0)\|_{{\mathcal{X}}}^2\leq C \left \|a\varphi_x(0,\cdot)+\frac{1}{c}\psi_x(0,\cdot)\right\|_{L^2(0,T)}^2\end{aligned}$$ plays a fundamental role for the study of the controllability. To prove (resp. ), we use a direct approach based on the multiplier technique that gives us the observability inequality for small values of the length $L$ and large time of control $T$.
\[prop6\] Let us suppose that $T > 0$ and $L>0$ satisfy $$\begin{aligned}
\label{Lsmall}
1>\frac{\beta C_T}{T}\left[L +\frac{r}{c} \right],\end{aligned}$$ where $C_T$ is the constant in and $\beta$ is the constant given by the embedding $H^{\frac{1}{3}}(0,T) \subset L ^2(0,T)$. Then, there exists a constant $C (T,L) > 0$, such that for any $(\varphi^0,\psi^0)$ in ${\mathcal{X}}$ the observability inequality (resp. ) holds, where $(\varphi,\psi)$ is solution of - with initial data $(\varphi^0,\psi^0)$.
We multiply the first equation in by $(T-t)\varphi$, the second one by $\frac{b}{c}(T-t)\psi$ and integrate over $(0,T)\times(0,L)$, to give us:
$$\begin{aligned}
\frac{T}{2}\int_0^L\left( \varphi_0^2(x) +\frac{b}{c}\psi_0^2(x)\right)dx=&\frac{1}{2}\int_0^T\int_0^L \left(\varphi^2(x,t)+\frac{b}{c}\psi^2(x,t)\right)dxdt \\
&+\int_0^T (T-t)\left [ \varphi(L,t)\left(\varphi_{xx}(L,t)+\frac{ab}{c}\psi_{xx}(L,t)\right)\right]dt\\
&-\int_0^T (T-t)\left [ \varphi(0,t)\left(\varphi_{xx}(0,t)+\frac{ab}{c}\psi_{xx}(0,t)\right)\right]dt \\
&+\int_0^T (T-t)\left [\frac{b}{c}\psi(L,t)\left(a\varphi_{xx}(L,t)+\frac{\psi_{xx}(L,t)}{c}+\frac{r}{2c}\psi(L,t)\right)\right]dt \\
&+\int_0^T (T-t)\left [ -\frac{b}{c}\psi(0,t)\left(a\varphi_{xx}(0,t)+\frac{\psi_{xx}(0,t)}{c}+\frac{r}{2c}\psi(0,t)\right) \right]dt \\
&+\frac12\int_0^T(T-t)\left[\varphi_x^2(0,t)+\frac{2ab}{c}\psi_x(0,t)\varphi_x(0,t)+\frac{b}{c^2}\psi_x^2(0,t)\right]dt.\end{aligned}$$
From the boundary conditions , we have that $$\begin{aligned}
\|(\varphi^0,\psi^0)\|_{{\mathcal{X}}}^2\leq& \frac{1}{T}\|(\varphi,\psi)\|_{L^2(0,T;{\mathcal{X}})}^2 +\frac{br}{c^2T}\|\psi(0,\cdot)\|_{L^2(0,T)}^2-\frac{br}{c^2}\int_0^T\frac{T-t}{T}\psi(L,t)^2dt\\
& +\int_0^T\left[\varphi_x^2(0,t)+\frac{2ab}{c}\psi_x(0,t)\varphi_x(0,t)+\frac{b}{c^2}\psi_x^2(0,t)\right]dt,
\\
\leq&\frac{1}{T}\|(\varphi,\psi)\|_{L^2(0,T;{\mathcal{X}})}^2 +\frac{\beta b r}{c^2T}\|\psi(0,\cdot)\|_{H^{\frac{1}{3}}(0,T)}^2 +\frac{1}{a^2b}\left\|\varphi_x(0,\cdot)+\frac{ab}{c}\psi_x(0,\cdot)\right\|_{L^2(0,T)}^2,\end{aligned}$$ $$\left(\text{resp.} \,\, \|(\varphi^0,\psi^0)\|_{{\mathcal{X}}}^2\leq\frac{1}{T}\|(\varphi,\psi)\|_{L^2(0,T;{\mathcal{X}})}^2 +\frac{\beta b r}{c^2T}\|\psi(0,\cdot)\|_{H^{\frac{1}{3}}(0,T)}^2 +\frac{1}{a^2}\left\|a\varphi_x(0,\cdot)+\frac{1}{c}\psi_x(0,\cdot)\right\|_{L^2(0,T)}^2 \right)$$ where $\beta$ is the constant given by the compact embedding $H^{\frac{1}{3}}(0,T) \subset L ^2(0,T)$. On the other hand, note that $L^{\infty}(0,L) \subset L^{2}(0,L)$, thus $$\|\varphi(\cdot,t)\|^2_{L^2(0,L)} \leq L \|\varphi(\cdot,t)\|^2_{L^{\infty}(0,L)}, \quad \text{and} \quad \|\psi(\cdot,t)\|^2_{L^2(0,L)} \leq L \|\psi(\cdot,t)\|^2_{L^{\infty}(0,L)},$$ Hence, $$\begin{aligned}
\|(\varphi,\psi)\|_{L^2(0,T;{\mathcal{X}})}^2 &= \int_0^T \left\lbrace \|\varphi(\cdot,t)\|^2_{L^2(0,L)} + \frac{b}{c}\|\psi(\cdot,t)\|^2_{L^2(0,L)}\right\rbrace dt \\
&\leq L\int_0^T \left\lbrace \|\varphi(\cdot,t)\|^2_{L^{\infty}(0,L)} + \frac{b}{c}\|\psi(\cdot,t)\|^2_{L^{\infty}(0,L)}\right\rbrace dt \\
&\leq L\beta \|\varphi\|_{H^{\frac{1}{3}}(0,T;L^{\infty}(0,L))}^2+\frac{bL\beta}{c}\|\psi\|_{H^{\frac{1}{3}}(0,T;L^{\infty}(0,L))}^2 . \\\end{aligned}$$ Thanks to the Proposition \[hiddenregularities\], we obtain $$\begin{aligned}
\|(\varphi^0,\psi^0)\|_{{\mathcal{X}}}^2 \leq & \frac{L \beta C_T}{T} \|\varphi^0\|_{L^2(0,L)}^2+\frac{b L \beta C_T}{cT}\|\psi^0\|_{L^2(0,L)}^2 +\frac{\beta C_T b r}{c^2T}\|\psi^0\|_{L^2(0,L)}^2 \\
& +\frac{1}{a^2b}\left\|\varphi_x(0,\cdot) +\frac{ab}{c}\psi_x(0,\cdot)\right\|_{L^2(0,T)}^2 \\
\leq &\frac{L \beta C_T}{T} \|(\varphi^0,\psi^0)\|_{{\mathcal{X}}}^2 +\frac{\beta C_T r}{cT}\|(\varphi^0,\psi^0)\|_{{\mathcal{X}}}^2 +\frac{1}{a^2b}\left\|\varphi_x(0,\cdot) +\frac{ab}{c}\psi_x(0,\cdot)\right\|_{L^2(0,T)}^2. \\\end{aligned}$$ Finally, it follows that $$\begin{aligned}
\|(\varphi^0,\psi^0)\|_{{\mathcal{X}}}^2 \leq & K \left\|\varphi_x(0,\cdot)+\frac{ab}{c}\psi_x(0,\cdot)\right\|_{L^2(0,T)}^2\end{aligned}$$ under the condition $$\label{TLcondition}
K= \frac{1}{a^2b}\left(1- \frac{\beta C_T}{T}\left[L +\frac{r}{c} \right]\right)^{-1}>0.$$
From the observability inequality , the following result holds.
\[teo5\] Let $T > 0$ and $L >0$ satisfying . Then, the system - (resp. -) is exactly controllable in time T.
Consider the map $$\begin{tabular}{r c c c}
$\Gamma :$ & $L^2(0, L) \times L^2(0, L)$ & $\longrightarrow$ & $L^2(0, L) \times L^2(0, L)$ \\
& $(\varphi^1(\cdot), \psi^1(\cdot))$ & $\longmapsto$ & $\Gamma(\varphi^1(\cdot), \psi^1(\cdot))=(u(\cdot,T), v(\cdot,T))$
\end{tabular}$$ where $(u,v)$ is the solution of -, with $$\begin{cases}
h_1(t)=\varphi_x(L,t)+\frac{ab}{c}\psi_x(L,t), \\
g_1(t)= a\varphi_x(L,t)+\frac{1}{c}\psi_x(L,t),
\end{cases}$$ and $(\varphi,\psi)$ is the solution of the system - with initial data $(\varphi^1,\psi^1)$. By (resp. ) and the Lax-Milgram theorem, the proof is achieved.
The Nonlinear Control System {#Sec4}
============================
We are now in position to prove our main result considering several configurations of the control in the boundary conditions. Let $T>0$, from Theorems \[teo1\], \[teo2\], \[teo3\], \[teo4\] and \[teo5\], we can define the bounded linear operators $$\Lambda_i : {\mathcal{X}}\times {\mathcal{X}}\longrightarrow {\mathcal{H}}_T \times {\mathcal{H}}_T \qquad (i=1,2,3,4,5),$$ such that, for any $(u^0, v^0) \in {\mathcal{X}}$ and $(u^1, v^1) \in {\mathcal{X}}$, $$\Lambda_i\left( \left( \begin{array}{cc} u^0\\v^0 \end{array}\right), \left( \begin{array}{cc} u^1\\v ^1 \end{array}\right)\right ):= \left( \begin{array}{cc} \vec{h}_i\\ \vec{g}_i \end{array}\right),$$ where $\vec{h}_i$ and $\vec{g}_i$ were defined in the Introduction.
We treat the nonlinear problem - using a classical fixed point argument.
According to Remark \[integralform\], the solution can be written as $$\begin{aligned}
\left( \begin{array}{cc} u(t)\\v(t) \end{array}\right) =& W_0(t) \left( \begin{array}{cc} u_0\\v_0 \end{array}\right) + W_{bdr}(t)\left(\begin{array}{cc}\vec{h}_i \\ \vec{g}_i \end{array}\right)\\
&- \int_0^t W_0 (t-\tau)\left( \begin{array}{cc} a_1 (vv_x)(\tau)+a_2 (uv)_x(\tau) \\ \frac{r}{c}v_x(\tau)+ \frac{a_2b}{c}(uu_x)(\tau)+\frac{a_1b}{c}(uv)_x(\tau) \end{array}\right) d\tau,\end{aligned}$$ for $i=1, 2, 3, 4, 5$, where $\{W_0(t)\}_{t\ge 0}$ and $\{W_{bdr}(t)\}_{t\ge 0}$ are the operators defined in Proposition \[prop1\]. We only analyze the case $i=1$, since the other cases are analogous we will omit them.
For $u,v \in {\mathcal{Z}}_T$, let us define $$\left(\begin{array}{cc} \upsilon \\ \nu (T,u,v)\end{array}\right) := \int_0^T W_0(T-\tau) \left( \begin{array}{cc} a_1 (vv_x)(\tau) + a_2 (uv)_x (\tau) \\
\frac{a_2b}{c}(uu_x)(\tau) + \frac{a_2b}{c} (uv)_x(\tau) \end{array} \right) d\tau$$ and consider the map $$\begin{aligned}
\Gamma\left( \begin{array}{cc} u\\v \end{array}\right) = & W_0(t) \left( \begin{array}{cc} u^0\\v^0 \end{array}\right) + W_{bdr}(x) \Lambda_1\left( \left( \begin{array}{cc} u^0\\v^0 \end{array}\right), \left( \begin{array}{cc} u^1 \\ v^1 \end{array}\right) + \left( \begin{array}{cc} v \\ \nu(T,u,v) \end{array}\right) \right) \\
&- \int_0^t W_0 (t-\tau)
\left( \begin{array}{cc} a_1 (vv_x)(\tau)+a_2 (uv)_x(\tau) \\ \frac{r}{c}v_x(\tau)+ \frac{a_2b}{c}(uu_x)(\tau)+\frac{a_1b}{c}(uv)_x(\tau) \end{array}\right) d\tau.\end{aligned}$$ If we choose $$\label{controli}
\left( \begin{array}{cc} \vec{h}_1 \\ \vec{g}_1 \end{array}\right) = \Lambda_1\left( \left( \begin{array}{cc} u^0\\v^0 \end{array}\right), \left( \begin{array}{cc} u^1 \\ v^1 \end{array}\right) + \left( \begin{array}{cc} v \\ \nu(T,u,v) \end{array}\right) \right),$$ from Theorem \[teo3\], we get $$\Gamma\left( \begin{array}{cc} u\\v \end{array}\right)\Big|_{t=0}= \left( \begin{array}{cc} u^0\\v^0\end{array}\right)$$ and $$\Gamma\left( \begin{array}{cc} u\\v \end{array}\right)\Big|_{t=T}= \left( \begin{array}{cc} u^1\\v^1\end{array}\right)+\left(\begin{array}{cc} v \\ \nu (T,u,v)\end{array}\right)-\left(\begin{array}{cc} v \\ \nu (T,u,v)\end{array}\right)= \left( \begin{array}{cc} u^1\\v^1\end{array}\right).$$ Now we prove that the map $\Gamma$ is a contraction in an appropriate metric space, then its fixed point $(u,v)$ is the solution of - with $\vec{h}_1$ and $\vec{g}_1$ defined by , satisfying . In order to prove the existence of the fixed point we apply the Banach fixed point theorem to the restriction of $\Gamma$ on the closed ball $$B_r=\left\{ ( u, v) \in {\mathcal{Z}}_T : \left\| ( u, v )\right\|_{{\mathcal{Z}}_T} \le r \right\},$$ for some $r>0$.
- *$\Gamma$ maps $B_r$ into itself.*
Using Proposition \[prop2\] there exists a constant $C_1>0$, such that
$$\begin{aligned}
\left\|\Gamma\left( \begin{array}{cc} u\\v \end{array}\right) \right\|_{{\mathcal{Z}}_T} \le& C_1\left\lbrace \left\| \left( \begin{array}{cc} u^0\\v^0 \end{array}\right) \right\|_{{\mathcal{X}}}
+\left\| \Lambda_1\left( \left( \begin{array}{cc} u^0\\v^0 \end{array}\right), \left( \begin{array}{cc} u^1 \\ v^1 \end{array}\right) + \left( \begin{array}{cc} v \\ \nu (T,u,v)\end{array}\right) \right)\right\|_{{\mathcal{H}}_T}
\right\rbrace\\
&+C_1\left\lbrace\int_0^t \left\| \left( \begin{array}{cc} a_1 (vv_x)(\tau)+a_2 (uv)_x(\tau) \\ \frac{r}{c}v_x(\tau)+ \frac{a_2b}{c}(uu_x)(\tau)+\frac{a_1b}{c}(uv)_x(\tau) \end{array}\right)\right\|_{{\mathcal{X}}} d\tau\right\rbrace.\end{aligned}$$
Moreover, since $$\begin{gathered}
\left\| \Lambda_1\left( \left( \begin{array}{cc} u^0\\v^0 \end{array}\right), \left( \begin{array}{cc} u^1 \\ v^1 \end{array}\right) + \left( \begin{array}{cc} v \\ \nu (T,u,v)\end{array}\right) \right)\right\|_{{\mathcal{H}}_T} \leq C_2 \left\lbrace \left\| \left( \begin{array}{cc} u^0\\v^0 \end{array}\right) \right\|_{{\mathcal{X}}}+ \left\|\left( \begin{array}{cc} u^1 \\ v^1 \end{array}\right) \right\|_{{\mathcal{X}}} \right. \\
\left. + \left\| \left( \begin{array}{cc} v \\ \nu (T,u,v)\end{array}\right)\right\|_{{\mathcal{X}}} \right\rbrace,\end{gathered}$$ applying Lemma \[lem3\], we can deduce that $$\left\|\Gamma\left( \begin{array}{cc} u\\v \end{array}\right) \right\|_{{\mathcal{Z}}_T} \le C_3\delta +C_4(r+1)r,$$ where $C_4$ is a constant depending only $T$. Thus, choosing $r$ and $\delta$ such that $$r=2C_3\delta$$ and $$2C_3C_4\delta +C_4\leq \frac12,$$ the operator $\Gamma$ maps $B_r$ into itself for any $(u,v) \in {\mathcal{Z}}_T$.
- *$\Gamma$ is contractive.*
Proceeding as the proof of Theorem \[nonlinearteo\], we obtain $$\left\|\Gamma\left( \begin{array}{cc} u\\v \end{array}\right) - \Gamma\left( \begin{array}{cc} \widetilde{u}\\\widetilde{v} \end{array}\right) \right\|_{{\mathcal{Z}}_T} \le C_5(r+1)r \left\|\left( \begin{array}{cc} u-\widetilde{u}\\v-\widetilde{v} \end{array}\right) \right\|_{{\mathcal{Z}}_T},$$ for any $(u,v), (\widetilde{u},\widetilde{v}) \in B_r$ and a constant $C_5$ depending only on $T$. Thus, taking $\delta >0$, such that $$\gamma= 2C_3C_5\delta +C_5 < 1,$$ we obtain $$\left\|\Gamma\left( \begin{array}{cc} u\\v \end{array}\right) - \Gamma\left( \begin{array}{cc} \widetilde{u}\\\widetilde{v} \end{array}\right) \right\|_{{\mathcal{Z}}_T} \le \gamma \left\|\left( \begin{array}{cc} u-\widetilde{u}\\v-\widetilde{v} \end{array}\right) \right\|_{{\mathcal{Z}}_T}.$$ Therefore, the map $\Gamma$ is a contraction. Thus, from (i), (ii) and the Banach fixed point theorem, $\Gamma$ has a fixed point in $B_r$ and its fixed point is the desired solution. The proof of Theorem \[main\_theo\] is, thus, complete.
**Acknowledgments:** Roberto A. Capistrano–Filho was supported by CNPq (Brazil), Project PDE, grant 229204/2013-9, Fernando A. Gallego was supported by CAPES (Brazil) and Ademir F. Pazoto was partially supported by CNPq (Brazil).
[99]{}
J. L. Bona, G. Ponce, J.-C. Saut and M. M. Tom, A Model System for Strong Interaction Between Internal Solitary Waves, [*Commun. Math. Phys.*]{} 143 (1992), 287–313.
J.J. Bona, S.M. Sun and B.-Y. Zhang, A nonhomogeneous boundary-value problem for the Korteweg-de Vries equation posed on a finite domain, [*Comm. Partial Differential Equations*]{}, 28 (2003), 1391–1438.
M. Caicedo, R. A. Capistrano–Filho and Bingyu Zhang, Neumann boundary controllability of the Korteweg-de Vries equation on a bounded domain, [*arXiv preprint*]{}, (2015), arXiv:1508.07525.
R. A. Capistrano–Filho, S. M. Sun and B.-Y. Zhang, General boundary value problems of the Korteweg-de Vries equation on a bounded domain. [*Preprint*]{}, (2016).
E. Cerpa and A. F. Pazoto, A note on the paper On the controllability of a coupled system of two Korteweg-de Vries equations, [*Commun. Contemp. Math*]{}, 13 (2011), 183–189.
E. Cerpa, I. Rivas and B.-Y. Zhang, Boundary controllability of the Korteweg-de Vries equation on a bounded domain, [*SIAM J. Control Optim.*]{}, 51 (2013), 2976–3010.
S. Dolecki and D. L. Russell, A general theory of observation and control, [*SIAM J. Control Optimization*]{}, 15 (1977), 185–220.
J. A. Gear and R. Grimshaw, Weak and strong interactions between internal solitary waves, [*Studies in Appl. Math*]{}, 70 (1984), 235-258.
J.-L. Lions, [*Contrôlabilité Exacte, Perturbations et Stabilisation de Systèmes Distribués,*]{} Tome 1, Masson, Paris, (1988).
L. Rosier, Exact boundary controllability for the Korteweg-de Vries equation on a bounded domain, [*ESAIM Control Optim. Cal. Var*]{}, 2 (1997), 33-55.
S. Micu and J. H. Ortega, On the controllability of a linear coupled system of Korteweg–de Vries equations, [*in Mathematical and Numerical Aspects of Wave Propagation* ]{}(Santiago de Compostela, 2000) (SIAM, Philadelphia, PA, 2000), 1020–1024.
S. Micu, J. Ortega and A. Pazoto, On the Controllability of a Coupled system of two Korteweg-de Vries equation, [*Commun. Contemp. Math*]{}, 11 (5) (2009), 779–827.
C. E. Kenig, G. Ponce and L. Vega, Well-posedness of the initial value problem for the Korteweg–de Vries equation,[*J. Amer. Math. Soc.*]{}, 4 (1991), 323–347.
J.-C. Saut and N. Tzvetkov, On a model system for the oblique interaction of internal gravity waves, [*M2AN Math. Model. Numer. Anal.*]{}, 34 (2000), 501–523.
K. Yosida, [*Functional Analysis*]{}, Springer-Verlag, Berlin Heidelberg New York, (1978).
[^1]: In the case of $L\in\mathcal{N}$ (resp. $L\in\mathcal{R}_{\beta}$), Rosier (resp. Caicedo *et al.* in [@caicedo_caspistrano_zhang_2015]) proved in [@rosier] that the associated linear system is not controllable; there exists a finite-dimensional subspace of $L^2(0,L)$, denoted by $\mathcal{M}=\mathcal{M}(L)$, which is unreachable from $0$ for the linear system. More precisely, for every nonzero state $\psi\in\mathcal{M}$, $g\in L^2(0,T)$ and $u\in C([0,T];L^2(0,L))\cap L^2(0,T;H^1(0,L))$ satisfying and $u(\cdot,0)=0$, one has $u(\cdot,T)\neq\psi$. A spatial domain $(0,L)$ is called *critical* for the system (\[2a\]) (resp. ) if its domain length $L\in\mathcal{N}$ (resp. $L\in\mathcal{R}_{\beta}$).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study prismatic sets analogously to simplicial sets except that realization involves prisms, i.e., products of simplices rather than just simplices. Particular examples are the prismatic subdivision of a simplicial set $S$ and the prismatic star of $S$. Both have the same homotopy type as $S$ and in particular the latter we use to study lattice gauge theory in the sense of Phillips and Stone. Thus for a Lie group $G$ and a set of parallel transport functions defining the transition over faces of the simplices, we define a classifying map from the prismatic star to a prismatic version of the classifying space of $G$. In turn this defines a $G$-bundle over the prismatic star.'
address:
- |
Department of Mathematics\
Dokuz Eylül University\
TR-35160 Izmir, Turkey
- |
Department of Mathematics\
University of Aarhus\
DK-8000 [Å]{}rhus C, Denmark
author:
- 'B.Akyar'
- 'J. L. Dupont'
title: Lattice Gauge Field Theory and Prismatic Sets
---
**[Introduction]{}\[one\]**
===========================
In the study of global properties of locally trivial fibre bundles it is a fundamental difficulty that the usual combinatorial methods of algebraic topology depends on the use of simplicial complexes which structure behaves badly with respect to local trivializations. By a theorem of Johnson [@J], the base and total space of a locally trivial smooth fibre bundle with projection $\pi:E\to{B}$ can be triangulated in such a way that $\pi$ is a simplicial map. But obviously even in this case a general fibre is not a simplicial complex in any natural way. However such a fibre has a natural decomposition into prisms, i.e., products of simplices, and the whole triangulated bundle gives the basic example of a prismatic set, analogous to the notion of a simplicial set derived from a simplicial complex. Prismatic sets were introduced and used by the second author and R. Ljungmann in [@DLj] (see also Ljungmann’s thesis [@Lj]) in order to construct an explicit fibre integration map in smooth Deligne cohomology, see also [@DK]. But the important special case of the prismatic subdivision of a simplicial set was used in Akyar [@B] in connection with “Lattice Gauge Theory” in the sense of Phillips and Stone [@PS], [@PS1], [@PS5] and similar constructions have been used in other connections, see e.g. [@MS]. One can see [@L] for further information about Lattice Gauge Fields.
In this paper we shall give a more systematic treatment of prismatic sets and their properties but we shall concentrate on the applications to lattice gauge theory extending the work of Phillips and Stone to arbitrary simplicial sets and all dimensions. For an arbitrary simplicial set $S$ and a given Lie group $G$ together with a set of parallel transport functions in their sense, we construct a prismatic set $\bar{P}(S)$ of the same homotopy type as $S$ and a classifying map from $\bar{P}(S)$ to a prismatic version of the standard model for $BG$. This is one of our main results (Theorem 8.1). Geometrically, for $S$ a simplicial complex, $\bar{P}(S)$ is closely related to the nerve of the covering by stars of vertices (Theorem 5.1). In turn this gives a principal $G$-bundle with a connection and thus in principle gives rise via the usual Chern-Weil and Chern-Simons theory to explicit formulas for characteristic classes (Corollary 8.2). We shall return to this elsewhere. One can see [@CrS], [@CnS], [@D], [@F3], [@W] for further information about Chern-Simons Theory.
The paper is organized as follows:
In chapter 2, prismatic sets are defined and their various geometric realizations are studied.
The third chapter introduces the prismatic triangulation of a simplicial map and in particular of a simplicial set. Furthermore, we comment on the calculation of the homology of the geometric realization of a prismatic set.
In chapter 4 we study prismatic sets associated to stars of simplicial complexes. It turns out that the prismatic set $\bar{P}_.(S)$ given in this chapter in the case of a simplicial complex is the nerve of the covering by stars of vertices.
In the fifth chapter, we compare the two star simplicial sets and prove that there is a natural surjective map $\bar{p}:\bar{P}_.(S)\to{P\operatorname{St}{S}}$. It turns out that this map is an isomorphism for $S=K^s$, where $K$ is a simplicial complex.
In chapter 6, we introduce a prismatic version of the classifying space. This is done by replacing the Lie group $G$ by the singular simplicial set of continuous maps $\text{Map}(\Delta^q,G)$.
In chapter 7, we introduce the notion of “compatible transition functions” similar to the “parallel transport functions” of Phillips-Stone [@PS1] for a simplicial complex $K$. We show how a given bundle on the realization of a simplicial set and socalled “admissible trivializations” give rise to a set of compatible transition functions and vice versa. We end the chapter with a remark on the relation between the compatible transition functions and parallel transport along a piecewise linear path.
Finally in the last chapter we construct the classifying map for a given set of compatible transition functions. For this we construct a prismatic map from $\bar{P}(S)$ to the prismatic model for the classifying space constructed in chapter 6.
**Acknowledgements:** We would like to thank Marcel Bökstedt for his interests and comments during the preparation of this paper.
**[Prismatic Sets]{}\[two\]**
=============================
Prismatic sets are similar to simplicial sets but they are realized by using prisms instead of only simplices.
Let $\Delta^p=\{(t_0,\dots,t_p)\in{{\Bbb{R}}^{p+1}}\text{ }|\text{
}\sum_i t_i=1,t_i\leq{1}\}$ be a standard $p$-simplex given with barycentric coordinates. A prism is a product of simplices, that is, a set of the form $\Delta^{q_0...q_p}=\Delta^{q_0}\times{\dots}\times{\Delta^{q_p}}$.
The motivating example is triangulated fibre bundles:
\[2.1\]
Given a smooth fibre bundle $\pi:Y\to{Z}$ with dim$Y=m+n$, dim$Z=m$ and compact fibres possibly with boundary. By a theorem of Johnson [@J], there are smooth triangulations $K$ and $L$ of $Y$ and $Z$, respectively and a simplicial map $\pi
^\prime:K\to{L}$ in the following commutative diagram
$$\xymatrix{ |K| \ar[d]^{|\pi^{\prime}|} \ar[r]^{\approx}
& Y \ar[d]^\pi \\
|L| \ar[r]^{\approx} & Z }$$
and the horizontal maps are homeomorphisms which are smooth on each simplex, here $|K|=\bigcup_{\tau\in{K_k}}\Delta^k\times{\tau}/\sim$, $k=0,...,\text{dim}K$, is the geometric realization.
One can extend a given such triangulation of $\partial Y\to{Z}$ to a triangulation of $Y\to{Z}$.
{width="9cm"}\
A simplex $\tau$ in $K$ has vertices $\tau=(b_0^0,\dots,b_{q_0}^0|...|b_0^p,\dots,b_{q_p}^p)$ with $\sigma=(a_0,\dots,a_p)$ such that $\pi^\prime(b_j^i)=a_i$. Here, we give the set of vertices of the total space, the lexicographical order. So geometrically, for an open simplex $\stackrel{\circ}{\sigma}$ in $L$, we have $$\pi^{-1}(\lvert\stackrel{\circ}{\sigma}\rvert)\approx{\lvert\stackrel{\circ}{\sigma}\rvert}\times{\bigcup_{\tau\in{\pi^{-1}(\sigma)}}}\Delta^{q_0...q_p}\times{\tau}.$$
We collect all these in the formal definition below using simplicial sets. For these we recall the notation but refer otherwise to Mac Lane [@McL], May [@May].
\[2.2\] A simplicial set $S_.=\{S_q\}$ is a sequence of sets with face operators $d_i:S_q\to{S_{q-1}}$ and degeneracy operators $s_i:S_q\to{S_{q+1}}$, $i=0,...,q$, satisfying the following identities: $$\begin{aligned}
d_id_j= \left\{ \begin{array}{r@{\quad: \quad}l}d_{j-1}d_i & i<j \\
d_jd_{i+1} & i\geq{j}, \end{array} \right.\end{aligned}$$ $$\begin{aligned}
s_is_j= \left\{ \begin{array}{r@{\quad: \quad}l}s_{j+1} s_i & i\leq{j} \\
s_j s_{i-1} & i>j, \end{array} \right.\end{aligned}$$ and $$d_is_j= \left\{ \begin{array}{r@{\quad:
\quad}l}s_{j-1}d_i & i<j \\ \text{id} & i=j,i=j+1
\\ s_jd_{i-1} & i>j+1. \end{array} \right.$$
\[2.3\] A simplicial complex $K_.$ gives a simplicial set where $K_p=\{(a_{i_0},...,a_{i_p})\text{ }|\text{
}\text{some nondecreasing sequences for a given linear ordering of}\text{ }
K_0\}$ is the set of $p$-simplices.
\[2.4\] Given an open cover $\mathcal{U}={U_i}$ of $Z$ we have the nerve $N\mathcal{U}=\{N\mathcal{U}(p)\}$ of the covering, where $$N\mathcal{U}(p)=\bigsqcup_{i_0,...,i_p}U_{i_0}\cap{...}\cap{U_{i_p}},$$ and $(i_0,...,i_p)$ is nondecreasing for a given linear order of the index set.
Let us denote $U_{i_0}\cap{...}\cap{U_{i_p}}$ by $U_{i_0,...,i_p}$. $N\mathcal{U}$ is a simplicial manifold, where the face and degeneracy maps come from the followings
$$\begin{aligned}
d_j:U_{i_0,...,i_p}&\to&{U_{i_0,...,{\hat{i}}_j,...,i_p}}\\
s_j:U_{i_0,...,i_p}&\to&{U_{i_0,...,i_j,i_j,...,i_p}}\end{aligned}$$
That is, $N\mathcal{U}(p)$ is a smooth manifold for each $p$ and the face and degeneracy maps are smooth. There is also a corresponding simplicial set $N_d\mathcal{U}=\{N_d\mathcal{U}(p)\}$ called the discrete nerve of the covering. Here $N_d\mathcal{U}(p)$ is simply the set consisting of an element for each non-empty intersection of $p+1$ open sets from $\mathcal{U}$. So there is a natural forgetful map $N\mathcal{U}\to{N_d\mathcal{U}}$.
[**Note:**]{} If $S_.$ has only face operators, then it is called a $\Delta$-set.
\[2.5\]
Given $p\geq{0}$, a $(p+1)$-multi-simplicial set is a sequence $\{S_{q_0,...,q_p}\}$ which is a simplicial set in each variable $q_i$, $i=0,...,p$.
\[2.6\] A prismatic set $P=\{P_{p,.}\}$ is a sequence $P_{p,.}=\{P_{p,q_0,...,q_p}\}$ of $(p+1)$-multi-simplicial sets, i.e., with face and degeneracy operators
$$\begin{aligned}
d_j^i:P_{p,q_0,...,q_p}&\to&{P_{p,q_0,...,q_i-1,...,q_p}}\\
s_j^i:P_{p,q_0,...,q_p}&\to&{P_{p,q_0,...,q_i+1,...,q_p}}\end{aligned}$$
such that $d_j^i$, $s_j^i$ commute with $d_l^k$, $s_l^k$ for $i\neq{k}$, and such that $d_j^i$, $s_j^i$ for fixed $i$ satisfy the identities in Definition \[2.2\].
Furthermore there are face operators $$d_k:P_{p,q_0,...,q_p}\to{P_{p-1,q_0,...,\hat{q}_k,...,q_p}}$$ commuting with $d_j^i$ and $s_j^i$ (interpreting $d_j^k=s_j^k=\text{id}$ on the right) such that $\{P_{p,.}\}$ is a $\Delta$-set.
\[2.7\] If similarly $P$ has degeneracy operators $$s_k:P_{p,q_0,...,q_p}\to{P_{p+1,q_0,...,q_k,q_k,...,q_p}}$$ then $P$ is called a strong prismatic set.
[**Remark:**]{} In this case $(P_p,d_k,s_k)$ is a usual simplicial set.
In general, degeneracy operators given in Definition \[2.7\] do not exist naturally so in this case $(P_{p,.},d_k)$ is only a $\Delta$-set.
Example \[2.1\] continued: A triangulated fibre bundle $$\pi:|K|\to{|L|}$$ gives a strong prismatic set $P_.(K/L)$ by letting $$P_p(K/L)_{q_0...q_p}\subseteq{K_{p+q_0+...+q_p}\times{L_p}}$$ be the subset of pairs of simplices $(\tau,\sigma)$ so that $q_i+1$ of the vertices in $\tau$ lies over the $i$-th vertex in $\sigma$. Then we have face and degeneracy operators defined in the obvious way. It is now straight forward to check that this is a strong prismatic set.
\[ES\] For a given simplicial set $S$, consider $E_pS=\underbrace{S_.\times{...}\times{S_.}}_{p+1-times}$.$\pi_i:E_pS\to{E_{p-1}S}$ is the projection which deletes the $i$-th factor. Similarly, the diagonal map $\delta_i:E_pS\to{E_{p+1}}S$ repeats the $i$-th factor. This is a strong prismatic set.
Prismatic sets have various geometric realizations.
\[2.9\] First, we have for each $p$ the thin (geometric) realization $$\begin{aligned}
\label{2.10}
|P_{p,.}|=\bigsqcup_{q_0,...,q_p}\Delta^{q_0...q_p}\times{P_{p,q_0,...,q_p}}/\sim\end{aligned}$$ with equivalence relation “$\sim$” generated by the face and degeneracy maps $$\begin{aligned}
\varepsilon_j^i&:&\Delta^{q_0...q_i...q_p}\to{\Delta^{q_0...q_i+1...q_p}}\text{ }\text{ }\text{and}\\
\eta_j^i&:&\Delta^{q_0...q_i...q_p}\to{\Delta^{q_0...q_i-1...q_p}},\end{aligned}$$ respectively. $\{|P_{p,.}|\}$ is a $\Delta$-space hence it gives a fat realization $$\begin{aligned}
\label{2.10*} \|\text{ }|P_.|\text{
}\|=\bigsqcup_{p\geq{0}}\Delta^p\times{|P_{p,.}|}/\sim\end{aligned}$$ by only using face operators $d_k$.
The face and degeneracy operators $d_k$, $s_k$ act on $\Delta^{q_0...q_p}$ as the projection and the diagonal, respectively so they induce a structure of a simplicial set on $|P_p|$. In other words, the projection $\pi_i:\Delta^{q_0...q_p}\to{\Delta^{q_0...{\hat{q}}_i...q_p}}$ deletes the $i$-th coordinate and the diagonal map $\Delta_i:\Delta^{q_0...q_p}\to{\Delta^{q_0...q_iq_i...q_p}}$ repeats the $i$-th factor. Then the further equivalence relation on $|P_.|$ given in (\[2.10\*\]) is generated by $$\begin{aligned}
(\varepsilon^it,s,\sigma)\sim(t,\pi_is,d_i\sigma),\text{ }t\in{\Delta^{p-1}},\text{ }s\in{\Delta^{q_0...q_p}},\text{ }\sigma\in{P_{p,q_0,...,q_p}}.\\\end{aligned}$$
If $P_.$ is strong then we also have a thin realization $$|P_.|=\||P_.|\|/\sim$$ given by the above and the further relation $$\begin{aligned}
(\eta^it,s,\sigma)\sim(t,\Delta_is,s_i\sigma),\text{
}t\in{\Delta^{p+1}},\text{ }s\in{\Delta^{q_0,...,q_p}},\text{
}\sigma\in{P_{p,q_0,...,q_p}}.\end{aligned}$$
Similarly, we can define for each $p$, the fat realization $||P_.||$, that is $$\|P_p\|=\bigsqcup_{q_0,...,q_p}\Delta^{q_0...q_p}\times{P_{p,q_0,...,q_p}}/\sim$$ with equivalence relation given by only the face maps $d_j^i$.
Moreover, we have the very fat realization $$\|\text{ }\|P_.\|\text{
}\|=\bigsqcup_{p\geq{0}}\Delta^p\times{\|P_{p,.}\|}/\sim$$ using only face operators.
For a given simplicial set $S$ and $E_pS$ as in Example \[ES\] we have $\|\text{ }|E_.S|\text{ }\|$ as the fat realization of the space which maps $p$-th term to $\underbrace{|S_.|\times{...}\times{|S_.|}}_{p+1-times}$.
Define for a space $X$, $E_pX=\underbrace{X_.\times{...}\times{X_.}}_{p+1-times}$ Let us say $X=|S|$ then $\|\text{ }|E_.S|\text{ }\|$ is contractible.
**[Prismatic Triangulation]{}\[three\]**
========================================
Let us return to the case of a triangulated fibre bundle $|K|\to{|L|}$. In this case the natural map $$P_p(K/L)_{q_0,...q_p}\to{K_{q_0+...+q_p+p}}$$ induces a homeomorphism
$$\xymatrix{ |P_.(K/L)| \ar[d]^{|\pi^{\prime}|} \ar[r]^{\phantom{123}\approx}
&|K| \ar[d]^\pi \\
|L| \ar[r]^{=} & |L| }$$
In this diagram, the top horizontal map we shall call the prismatic triangulation homeomorphism $$\lambda:|P_.(K/L)|\stackrel{\cong}\to{|K|}$$ induced by $$\begin{aligned}
\label{3.1}
\lambda(t,s^0,...,s^p,(\tau,\sigma))=(t_0s^0,...,t_ps^p,\tau)\in{\Delta^{p+q}\times{K_{p+q}}},\end{aligned}$$ where $(t,s,\tau,\sigma)\in{\Delta^p}\times{\Delta^{q_0...q_p}}\times{P_p(K/L)_{q_0...q_p}}$ and $q=q_0+...+q_p$.
[**Note:**]{} If $\stackrel{\circ}{\sigma}$ is an open $p$-simplex in $L$ then $\lambda$ provides a natural trivialization of $|K|_\sigma=\pi^{-1}(\stackrel{\circ}{\sigma})$, that is, a homeomorphism $$\lambda:\stackrel{\circ}{\sigma}\times{|P_p(K/\sigma)}|\stackrel{\approx}\to{|K|_\sigma}.$$
We can generalize this construction to any simplicial map:
\[3.3\] [*Prismatic triangulation of a simplicial map.*]{} Let $f:S_.\to{\bar{S_.}}$ be a simplicial map of simplicial sets and define $P_.(f)$ by $$P_p(f)_{q_0,...,q_p}=\{(\sigma,\bar{\sigma})\in{S_{q_0+...+q_p+p}}\times{{\bar{S}}_p}\text{
}|\text{ }
f(\sigma)=\mu_{q_0,...,q_p}(\bar{\sigma})\}$$ where the corresponding map $$\mu^{q_0,...,q_p}:\Delta^{q_0+...+q_p+p}\to{\Delta^p}$$ is given by $$\{0,...,q_0|...|q_0+...+q_{p-1}+p,...,q_0+...+q_p+p\}\to{\{0,...,p\}}.$$ By this, we mean that the basis vectors $e_0,...,e_{q_0}$ are mapped to $e_0$, and $e_{q_0+1},...,e_{q_0+q_1+1}$ are mapped to $e_1$ and etc. Explicitly $$\mu_{q_0,...,q_p}=\hat{s}_{q+p}\circ{s_{(q_0+...+q_p+p-1)...(q_0+...+q_{p-1}+p)}}\circ{...}\circ{\hat{s}_{q_0}}\circ{s_{(q_0-1)...(0)}},$$ where the $\hat{s}_i$ are left out and $$s_{(q_0+...+q_i+i-1)...(q_0+...+q_{i-1}+i)}=s_{q_0+...+q_i+i-1}\circ{...}\circ{s_{q_0+...+q_{i-1}+i}},$$ $i=0,...,p$. The boundary maps in the fibre direction $$d_j^i:P_p(f)_{q_0,...,q_p}\to{P_p(f)_{q_0,...,q_i-1,...,q_p}}$$ are inherited from the face operators defined on $S_{q+p}$. Thus $$d_j^i(\sigma,\bar{\sigma})=(d_{q_0+...+q_{i-1}+i+j-1}\sigma,\bar{\sigma}).$$ Similarly the degeneracy maps $s^i_j$ on $P_p(f)_{q_0,...,q_p}$ $$s^i_j:P_p(f)_{q_0,...,q_p}\to{P_p(f)_{q_0,...,q_i+1,...,q_p}}$$ are inherited from the ones on $S_{q+p}$. That is, $$s_j^i(\sigma,\bar{\sigma})=(s_{q_0+...+q_{i-1}+i+j-1}\sigma,\bar{\sigma}).$$ The boundary maps $$d^i:P_p(f)_{q_0,...,q_p}\to{P_{p-1}(f)_{q_0,...,{\hat{q}}_i,...,q_p}}$$ are determined by the boundary maps defined on both $S_{q+p}$ and ${\bar{S}}_p$. Thus $$d^i(\sigma,\bar{\sigma})=(d_{q_0+...+q_{i-1}+i-1}\circ{...}\circ{d_{q_0+...+q_i+i-1}}\sigma,d_i\bar{\sigma}),$$ here the composition of the face operators can be shortly written as $$d_{(q_0+...+q_{i-1}+i-1)...({q_0+...+q_i+i-1})}=d_{q_0+...+q_{i-1}+i-1}\circ{...}\circ{d_{q_0+...+q_i+i-1}}.$$
[**Note:**]{} $P_.(f)$ is a prismatic set, but in general not a strong one.
\[3.4\] There is a pullback diagram
$$\xymatrix{ \|\text{ }|P_.(f)|\text{ }\| \ar[d]^{\|f\|}
\ar[r]^{\phantom{12}\lambda}
&|S_.| \ar[d]^{|f|} \\
\|\bar{S}_.\| \ar[r]^q & |\bar{S}_.| }$$
In particular $\lambda$ is a homotopy equivalence.
[**Proof:**]{} The map $\lambda:\Delta^p\times{\Delta^{q_0...q_p}}\times{P_p(f)_{q_0...q_p}}\to{\Delta^{q+p}\times{S_{q+p}}}$ is given by $\lambda(t,s,\sigma,\bar{\sigma})=(t_0s^0,...,t_ps^p,\sigma).$ The commutativity of the diagram follows from the definition of $P_.(f)$ since $$P_p(f)_{q_0,...,q_p}\subseteq{S_{q+p}\times{{\bar{S}}_p}}$$ consists of pairs $(\sigma,\bar{\sigma})\in{S_{q+p}}\times{{\bar{S}}_p}$ such that $f(\sigma)=\mu_{q_0,...,q_p}(\bar{\sigma})\in{{\bar{S}}_p}$.
By the commutativity of the diagram, $\lambda$ factors over the pullback $|S_.|\times_{|\bar{S}_.|}{\|\bar{S}_.\|}$ in the diagram
$$\xymatrix{ |S_.|\times_{|\bar{S}_.|}{\|\bar{S}_.\|} \ar[d]^{\|f\|}
\ar[r]^{\phantom{1234}{\text{pr}}_1}
&|S_.| \ar[d]^{|f|} \\
\|\bar{S}_.\|
\ar[r]^q & |\bar{S}_.| }$$
Here elements in the pullback $|S_.|\times_{|\bar{S}|}{\|\bar{S}_.\|}$ are represented by pairs $((t,\sigma),(\bar{t},\bar{\sigma}))$ such that $f(\sigma)=\mu_{q_0,...,q_p}(\bar{\sigma})$ and $\bar{t}=\mu^{q_0,...,q_p}(t)$, where $\sigma\in{S_{q+p}}$, $\bar{\sigma}\in{{\bar{S}}_q}$. Therefore $\lambda\times{\|f\|}:\|\text{
}|P_.(f)|\text{ }\|\to{|S|\times{\|\bar{S}\|}}$ induces $\Lambda$ in the diagram
$$\xymatrix{ \|\text{ }|P_.(f)|\text{ }\| \ar[r]^{\Lambda} \ar[d]
& |S_.|\times_{|\bar{S}_.|}{\|\bar{S}_.\|}
\ar[r]^{\phantom{1234}\text{pr}_1} \ar[d]^{{\text{pr}}_2}
& |S_.| \ar[d]^{|f|}\\
\|\bar{S}_.\| \ar[r]^{\text{id}} & \|\bar{S}_.\| \ar[r] &
|\bar{S}_.|}$$
Now $\Lambda$ is a homotopy equivalence. Indeed, an argument similar to the note following (\[3.1\]) gives a homeomorphism of the preimage $\|f\|$ of an open simplex in $\|\bar{S}_.\|$. Hence $\Lambda$ is shown to by a homeomorphism by induction over skeleton of $\|\bar{S}_.\|$.
$\hfill \Box$
\[3.5\] [*Prismatic triangulation of a simplicial set*]{}. Let $S_.$ be a simplicial set and $\bar{S_.}=*$ the simplicial set with one element in each degree. Here $P_p(f)=P_pS$ is called the $p$-th prismatic subdivision of $S$ and for each $t\in{\stackrel{\circ}{\Delta}^p}$ the map $\lambda_p(t,-):|P_pS|\to{|S_.|}$ is a homeomorphism. In this case, Theorem \[3.4\] gives a homeomorphism $\Lambda:\|\text{
}|P_.S|\text{ }\|\stackrel{\approx}\to{\|*\|}\times{|S_.|}$, here $\|*\|=\bigcup_n\Delta^n/\partial\Delta^n$. In particular $\lambda:\|\text{ }|P_.S|\text{ }\|\to{|S_.|}$ is a homotopy equivalence. We shall call $P_.S$ the prismatic triangulation of $S$.
For later use, let us give the explicit construction of the $p+1$-prismatic set $P_.S_.$ and its realization:
$$P_pS_{q_0,...,q_p}=S_{q_0+...+q_p+p}.$$
The face operators $${d_j}^{(i)}:P_pS_{q_0,...,q_i,...,q_p}=S_{q+p}\to{P_pS_{q_0,...,q_i-1,...,q_p}}=S_{q+p-1}$$ are defined by $$d_j^{(i)}:=d_{q_0+...+q_{i-1}+i+j},$$ $j=0,...,q_i$. Similarly, the degeneracy operators $${s_j}^{(i)}:P_pS_{q_0,...,q_i,...,q_p}=S_{q+p}\to{P_pS_{q_0,...,q_i+1,...,q_i}}=S_{q+p+1}$$ can be defined by $$s_j^{(i)}:=s_{q_0+...+q_{i-1}+i+j},$$ $j=0,...,q_i$. The face maps $$d_{(i)}:P_pS_{q_0,...,q_p}\to{P_{p-1}S_{q_0,...,{\hat{q}}_i,...q_p}}$$ are the operators corresponding to $${\varepsilon}^{(i)}:{\Delta}^{q_0+...+{\hat{q}}_i+...+q_p+p-1}\to{{\Delta}^{q_0+...+...+q_p+p}}$$ take $(e_0,...,e_{q_0+...+{\hat{q}}_i+...+q_p+p-1})$ to $(e_0,...,e_{q_0+...+q_p+p})$, deleting the elements $q_0+...+q_{i-1}+i,...,q_0+...+q_i+i$. It deletes $(q_i+1)$- elements. In contrary to this, there is no degeneracy operator.
Now we turn to the realizations. For the sequences of spaces $\{|P_.S_.|\}$, we obtain the fat realization: 0.3 cm $$||\text{ }|P_.S_.|\text{
}||=\bigsqcup_{p\geq{0}}{\Delta}^p\times{|P_pS_.|}/_{\sim},$$ where $$|P_pS_.|=\bigsqcup{\Delta}^{q_0...q_p}\times{S_{q_0+...+q_p+p}}/_{\sim}$$ and the face operators $\pi_i:|P_pS_.|\to{|P_{p-1}S_.|}$ are given by $\pi_i={\text{proj}}_i\times{d_{(i)}}$ with ${\text{proj}}_i:\Delta^{q_0...q_p}\to{\Delta^{q_0...{\hat{q}}_i...q_p}}$ beeing the natural projection.
Note that $\lambda_p:\Delta^p\times{|P_pS_.|}\to{|S_.|}$ satisfies $$\lambda_p\circ{(\varepsilon^i\times{\text{id}})}=\lambda_{p-1}\circ{(\text{id}\times{\pi_i})}.$$ Thus $\lambda_p$ induces the map $\lambda$ on the fat realization.
Let $\|\text{ }|P_.S_.|\text{ }\|^p$ respectively $\|\text{
}|S_.|\text{ }\|^p$ denote the subcomplexes generated by $\Delta^p\times{|P_pS|}$ respectively $\Delta^p\times{|S_.|}$. Then the restriction of $\Lambda$ to $\|\text{ }|P_.S_.|\text{ }\|^p$ is given by $$\Lambda_p(t,s,\sigma)=(t,\lambda_p(t,s,\sigma)).$$
The map $\Lambda_p$ induce a homeomorphism $$\Lambda:\|\text{ }|P_.S_.|\text{ }\|\to{\|\text{ }|S_.|\text{
}\|}\approx{\|*\|\times{|S_.|}}.$$
\[composition\] The composition map ${\text{proj}}_2\circ{\Lambda}=\lambda$ $$\|\text{ }|P_.S_.|\text{ }\|\to{\|\text{ }|S_.|\text{
}\|}\to{|S_.|}$$ is a homotopy equivalence.
[**Remark 1:**]{} We can calculate the homology of the geometric realization of a prismatic set as follows:
A prismatic set $P_{.,.}$ has a double complex $(C_{p,n}(PS),\partial_F,\partial_H)$. Here $$C_{p,n}(PS)=\bigoplus_{q_0+...+q_p=n}C_{p,q_0,...,q_p}(PS)$$ is the associated chain complex $C_p(PS)$ generated by $P_{p,q_0,...,q_p}$. The vertical boundary map is defined by using boundary maps in the fibre direction $${\partial_F}^i:PC_{p,{q_0,...,q_p}}\to{PC_{p,{q_0,...,q_i-1,...,q_p}}}$$ defined by ${\partial_F}^i=\sum{(-1)}^jd_j^i$, where, if $q_i=0$ then ${\partial^i}_F=0$. The total vertical boundary map is then $$\begin{aligned}
\label{vertical}
\partial_V={\partial^0}_F+{(-1)}^{q_0+1}{\partial^1}_F+...+{(-1)}^{q_0+...+q_{p-1}+p}{\partial^p}_F.\end{aligned}$$ There is also a horizontal boundary map $$\begin{aligned}
\label{horizontal}
\partial_H=\partial_0+{(-1)}^{q_0+1}\partial_1+...+{(-1)}^{q_0+...+q_{p-1}+p}\partial_p,\end{aligned}$$ where $$\begin{aligned}
\partial_k= \left\{ \begin{array}{r@{\quad: \quad}l}0 & \text{ }\text{if}\text{ }q_k>0 \\
d_k & \text{ }\text{if}\text{ }q_k=0, \end{array} \right.\end{aligned}$$ so that $\partial=\partial_V+\partial_H$ is a boundary map in the total complex $PC_*$ which is the cellular chain complex for the geometric realization. Hence it calculates the homology. In the case of $P_.(f)$ for $f:S\to{\bar{S}}$ a simplicial map, the double complex gives rise to a spectral sequence which for a triangulated fibre bundle is the usual Leray-Serre spectral sequence.
[**Remark 2:**]{} For each $p$ and each $t\in{{\stackrel{\circ}{\Delta}}^p}$, $\lambda_p(t)^{-1}:|S|\to{\{t\}\times|P_pS|}$ induces a map of cellular chain complexes $$aw:C_*(S)\to{C_{*,*}(PS)}$$ given by $$aw(x)=\sum_{q_0+...+q_p=n}s_{q_0+...+q_{p-1}+p-1}\circ{...}\circ{s_{q_0}}(x)_{(q_0,...,q_p)},$$ where $x\in{S_n}$.
**[Prismatic Sets and Stars of Simplicial Complexes]{}\[four\]**
================================================================
For a simplicial set $S$ and the prismatic triangulation $P_.S$ there is another closely related prismatic set ${\bar{P}}_pS_.$ which as we shall see for a simplicial complex is the nerve of the covering by stars of vertices considered as a prismatic set.
\[3.7\]
For $S$ a simplicial set let $\bar{P_.}S$ be the prismatic set given by $${\bar{P}}_pS_{q_0,...,q_p}:=S_{q_0+...+q_p+2p+1}.$$ where face and degeneracy operators on $\bar{P_p}S_{q_0,...,q_p}$ are inherited from the ones of $S_{q+2p+1}$ as follows:
Let $q=q_0+...+q_p$, the face operators $$d_j^{(i)}:S_{q+2p+1}={\bar{P}}_pS_{q_0,...,q_p}\to{S_{q+2p}}={\bar{P}}_pS_{q_0,...,q_i-1,...,q_p}$$ are defined by $$d_j^{(i)}:=d_{q_0+...+q_{i-1}+2i+j},\text{ }j=0,...,q_i\text{ }\text{but}\text{ }j\not={2i+1+\sum_{k=0}^iq_k}.$$ So $\bar{P_p}S_{q_0,...,q_p}$ has only $q+p$-face operators, i.e., we skip the following $p+1$ face operators $$\{d_{q_0+1},d_{q_0+q_1+3},...,d_{q+2p+1}\}.$$
Similarly the degeneracy operators $$s_j^{(i)}:S_{q+2p+1}\to{S_{q+2p+2}}$$ can be defined by $$s_j^{(i)}:=s_{q_0+...+q_{i-1}+2i+j},\text{ }j=0,...,q_i,\text{ }\text{but}\text{ }j\not={2i+1+\sum_k^iq_k}.$$ Furthermore the face operators are $$d_{(i)}:S_{q+2p+1}={\bar{P}}_pS_{q_0,...,q_p}\to{S_{q+2p-q_i-1}}={\bar{P}}_{p-1}S_{q_0,...,{\hat{q}}_i,...,q_p}$$ corresponding to $$\varepsilon^{(i)}:{\Delta}^{q+2p-q_i-1}\to{{\Delta}^{q+2p+1}}$$ which take $(e_0,...,e_{q_0+...+{\hat{q}}_i+...+q_p+2p-1})$ to $(e_0,...,e_{q+2p+1})$, by deleting the vectors with indices $(q_0+...+q_{i-1}+2i,...,q_0+...+q_i+2i+1)$. So it deletes $q_i+2$ elements. That is, $$d_{(i)}=d_{q_0+...+q_{i-1}+2i}\circ{...}\circ{d_{q_0+...+q_i+2i+1}},\text{ }i=0,...,p.$$
[**Remark:**]{} As $P_.S$, ${\bar{P}}_.S$ is a prismatic set but in general not a strong prismatic set.
[**Realization of ${\bar{P}_.S_.}$:**]{}
Notice that $$\|\text{ }|\bar{P_.}S_.|\text{
}\|=\bigsqcup_{p\geq{0}}{\Delta}^p\times{{\Delta}^{q_0...q_p}}
\times{{\bar{P}}_pS_{q_0,...,q_p}}/_{\sim}$$ where the equivalence relation apart from the internal relations in $|{\bar{P}}_pS|$ using $d_j^{(i)}$ and $s_j^{(i)}$, include the relations $$(\varepsilon^it,(s,y))\sim(t,\pi_i(s,y)),$$ with $\pi_i=({\text{proj}}_i)\times{d_{(i)}}$ the face operators on $|\bar{P_p}S_.|$.
The relation of ${\bar{P}}_.S$ with $S_.$ and $P_.S$ is as follows:
\[3.9\]
Let $i:\|S_.\|\hookrightarrow{\|\text{ }|{\bar{P}}_pS_.|\text{ }\|}$ be an inclusion defined for $(t,x)\in{\Delta^p\times{S_p}}$ by $$i(t,x)=(t,1,s_0\circ{...}\circ{s_p}x)\in{\Delta^p\times{{(\Delta^0)}^{p+1}}\times{S_{2p+1}}}\subseteq{\Delta^p\times{|{\bar{P}}_pS_.|}},$$ and $r:\|\text{ }|{\bar{P}}_.S_.|\text{ }\|\to{\|S_.\|}$ be the retraction defined for $(t,s,y)\in{\Delta^p\times{\Delta^{q_0...q_p}}\times{S_{q+2p+1}}}$ $$r(t,s,y)=(t,d_{0...q_0}\circ{{\hat{d}}_{q_0+1}}\circ{...}\circ{d_{(q_0+...+q_{p-1}+2p)...(q+2p)}}\circ{{\hat{d}}_{q+2p+1}}y),$$ where the ${\hat{d}}_i$ are left out and $d_{(q_0+...+q_{i-1}+2i)...(q_0+...+q_i+2i)}=d_{q_0+...+q_{i-1}+2i}\circ{...}\circ{d_{q_0+...+q_i+2i}}$, $i=0,...,p$.
1\) $i$ is a deformation retract with the retraction $r$.
2\) There is a commutative diagram of homotopy equivalences
$$\begin{aligned}
\xymatrix{
||S_.||\ar[r]^{i}\ar[dr] \ar[ddr]
&\|\text{ }|{\bar{P}}_.S_.|\text{ }\| \ar[d]^{f}\\
&\|\text{ }|P_.S_.|\text{ }\| \ar[d]^{\Lambda}\\
&\|\text{ }|S_.|\text{ }\|
}\end{aligned}$$
where $f:\Delta^p\times{\Delta^{q_0...q_p}\times{S_{q+2p+1}}}\to{\Delta^p\times{\Delta^{q_0...q_p}}}\times{S_{q+p}}$ takes $(t,s^0,...,s^p,x)$ to $(t,s^0,...,s^p,d_{q_0+1}\circ{d_{q_0+q_1+3}}\circ{...}\circ{d_{q+2p+1}}x)$, $x\in{S_{q+2p+1}}$.
The proof is straight forward see [@B] for details.
For a simplicial complex $K$ there is another prismatic complex defined using the stars of simplices. That is, let $K_0=\{a_i|i\in{I}\}$, where $I=\{1,...,N\}$, be the set of vertices and let $K_n=\{\sigma=(a_{i_0},...,a_{i_n})|i_0<...<i_n\}$ be the set of $n$-simplices such that if $\sigma\in{K_n}$ then any face $\tau=(a_{i_{j_0}},...,a_{i_{j_k}})$ lies in $K_k$. We shall write $\tau\preccurlyeq\sigma$ in this case. Now $K\times{K}$ is also a simplicial complex with the lexicographical order of the vertices $$(a_i,b_j)<(a_{i^\prime},b_{j^\prime})
\Leftrightarrow\text{either}\text{ }i<{i^\prime} \text{
}\text{or}\text{ }i=i^\prime\text{ }\text{and}\text{ }j<j^\prime,$$ where $\{(a_{i_0},b_{j_0}),...,(a_{i_n},b_{j_n})\}\in{K\times{K}}$.
\[4.1\] Let $K$ be a simplicial complex. The Star of $K$ is defined as $$\text{St}(K)=\{(\sigma,\tau)\in{K\times{K}}\text{
}|\text{ }\exists\text{ }{\sigma^\prime}\text{ }\text{such
that}\text{ }\sigma\cup\tau\preccurlyeq\sigma^\prime\}\subseteq{K\times{K}}.$$ This is equivalent to say that $$\text{St}(K)=\{\text{faces of}\text{
}\sigma^\prime\times{\sigma^\prime}\subseteq{K\times{K}}\}.$$
[**Remark 1:**]{} For each $\sigma\in{K}$, $(\{\sigma\}\times{K})\cap{\text{St}}(K)$ is the closure of the usual open star of $\sigma$, i.e., the union of the open simplices having $\sigma$ as a face. Whence the name $\text{St}(K)$. Note that $\text{St}(K)\subseteq{K\times{K}}$ is a subcomplex.
Let $K^s$ denote the simplicial set associated to the simplicial complex $K$. That is, $${K_n}^s=\{(a_{i_0},...,a_{i_n})\text{ }|\text{
}\{a_{i_0},...,a_{i_n}\}\text{ }\text{a simplex of}\text{ }K\text{ }\text{(with repetitions)}\text{ }i_0\leq{...}\leq{i_n}\}.$$
$\operatorname{St}(K)_n ^s$ are the following:
Let $(\sigma,\tau)\in{K\times{K}}$, where $\sigma=(a_{i_0},...,a_{i_p})$, $\tau=(b_{j_0},...,b_{j_q})$. For $(\sigma,\tau)\in{\text{St}(K)}$, let $\sigma^\prime=\sigma\cup\tau=(c_{k_0},...,c_{k_n})$. By allowing repetitions in Definition \[4.1\], i.e., by taking $\sigma^\prime\in{K^s}$, we can assume $n=p+q$ so that either $c_{k_s}=a_{i_t}$ or $c_{k_s}=b_{j_u}$, where $t=0,...,p$, $u=0,...,q$. Also we can assume $c_{k_n}=a_{i_p}$, and if $a_{i_t}=b_{j_u}$ then $b_{j_u}$ comes before $a_{i_t}$. In other words $(\sigma,\tau)$ is of the form $$\sigma=d_{\nu_1...\nu_q}\sigma^\prime,\text{ }\tau=d_{\mu_1...\mu_p}\sigma^\prime,$$ where $0\leq{\nu_1}<...<\nu_q<n$ and $0\leq{\mu_1}<...<\mu_p\leq{n}$ and $\mu_i\neq{\nu_j}$, $\forall{i,j}$. Therefore we introduce for a general simplicial set $S$ the following.
\[4.2\] Let $\text{St}(S)$ be the simplicial subset of the diagonal $\delta(S\times{S})$ containing all simplices of the form
$$(s_{\nu_q...\nu_1}\circ{d_{\nu_1...\nu_q}}\sigma^\prime,s_{\mu_p...\mu_1}\circ{d_{\mu_1...\mu_p}}\sigma^\prime),$$
where $0\leq{\nu_1}<...<\nu_q<n$ and $0\leq{\mu_1}<...<\mu_p\leq{n}$ with $\mu_i\neq{\nu_j}$, $\forall{i,j}$ as above. Here $s_{\nu_q...\nu_1}=s_{\nu_q}\circ{...}\circ{s_{\nu_1}}$ , $d_{\nu_1...\nu_q}=d_{\nu_1}\circ{...}\circ{d_{\nu_q}}$, $s_{\mu_p...\mu_1}=s_{\mu_p}\circ{...}\circ{s_{\mu_1}}$ and $d_{\mu_1...\mu_p}=d_{\mu_1}\circ{...}\circ{d_{\mu_p}}$.
\[4.3\] For $K$ a simplicial complex, there is a map $$\operatorname{st}:\operatorname{St}(K)^s\to{\operatorname{St}(K^s)}$$ which is an isomorphism.
: By the discussion made before, there is a well-defined map st. Indeed, $\text{St}(K)^s$ is a simplicial set generated by $$\begin{aligned}
\{(\sigma,\tau)\in{K\times{K}}\text{ }|\text{ }\exists \text{ }
\sigma^\prime\in{K^s}\text{ }
\text{such that}\text{ }
\sigma=d_{\nu_1...\nu_q} \sigma ^\prime,
\tau=d_{\mu_1...\mu_p} \sigma ^\prime\\
\text{ }\text{and}\text{ }
(s_{\nu_q...\nu_1}
\sigma,s_{\mu_p...\mu_1} \tau)\in{\delta(K^s\times{K^s})}
\}.\end{aligned}$$
By Definition \[4.2\], we can put $\operatorname{st}(\sigma,\tau)=(s_{\nu_q...\nu_1}\sigma,s_{\mu_p...\mu_1}\tau)\in{\operatorname{St}(K^s)}$. Clearly $\operatorname{st}$ is an isomorphism since for $(\sigma,\tau)\in{\delta(K^s\times{K^s})}$ and $\sigma^\prime$ as in Definition \[4.2\], $(\sigma,\tau)\in{K^s\times{K^s}}$ determines an element in $K\times{K}$ by deleting repetitions and this is unique.
[**Remark 2:**]{} The projection on the first factor $\pi_1:S\times{S}\to{S}$ gives a simplicial map $\pi_1:\text{St}_.(S)\to{S}$. Hence, we obtain a prismatic set $P_.\text{St}(S)=P_.(\pi_1)$ as in Example \[3.3\]. Here with $q=q_0+...+q_p$ and $\sigma=s_{\nu_q...\nu_1}\circ{d_{\nu_1...\nu_q}}\sigma^\prime=\mu_{q_0,...,q_p}\bar{\sigma}$, $\tau=s_{\mu_p...\mu_1}\circ{d_{\mu_1...\mu_p}}\sigma^\prime$, we have $$P_p\text{St}(S)_{q_0,...,q_p}=\{(\sigma,\tau,
\bar{\sigma})\in{\text{St}(S)_{q+p}}\times{S_p}\subset{\delta}(S\times{S})_{q+p}\times{S_p}
\text{ }|\text{ } \text{ }\sigma,\tau\text{
}\text{given above}
\}.$$ That is, $\pi_1(\sigma,\tau)=\mu_{q_0,\dots,q_p}(\bar{\sigma})$, where $\bar{\sigma}=d_{\nu_1...\nu_q}\sigma^\prime\in{S_p}$. So The elements in $P_p\operatorname{St}(S)_{q_0,\dots,q_p}$ are of the form $(\mu_{q_0,\dots,q_p}\bar{\sigma},\tau,\bar{\sigma})$, where $\tau\in{S_q}$. Here explicitly
$$\mu_{q_0,\dots,q_p}={\hat{s}}_{q+p}\circ{s_{({q+p-1})...({q_0+...+q_{p-1}+p}})}...{\hat{s}}_{q_0+q_1+1}
s_{(q_0+q_1)...(q_0+1)}{\hat{s}}_{q_0}s_{(q_0-1)...(0)}.$$
**[Comparison of the two Star Simplicial Sets]{}\[five\]**
==========================================================
We shall now prove that this is closely related to the prismatic set $\bar{P}S$ defined in the previous section.
\[4.4\]
1\) There is a natural **(surjective)** map $$\bar{p}:\bar{P}_.S_.\to{P{\operatorname{St}}_.(S)_.}$$
2\) If $S=K^s$, where $K$ is a simplicial complex, then $\bar{p}$ is an isomorphism.
1\) Take an element $\gamma\in{\bar{P}_pS_{q_0,...,q_p}}=S_{q_0+...+q_p+2p+1}$. Then $\gamma$ and $q_0,...,q_p$ determine an element $\bar{p}(\gamma)$ in $P_p\operatorname{St}(S)_{q_0,...,q_p}$ together with a $(p+1,q+p+1)$-partition $(i_1,...,i_p,i_{p+1},j_1,...,j_{q+p+1})$ of $n=q+2p+1$, where $q=q_0+...+q_p$. Here $$\begin{aligned}
i_1&=&q_0+1\\
i_2&=&q_0+q_1+3\\
&.&\\
&.&\\
&.&\\
i_p&=&q_0+...+q_{p-1}+2p-1\\
i_{p+1}&=&q_0+...+q_p+2p+1\end{aligned}$$ correspond to the $\mu_i$’s defined in Definition \[4.2\] and the $j$’s correspond to the complement, that is, $j_1,...,j_{q_0+1},j_{q_0+2},...,j_{q_0+q_1+2},...,j_{q_0+...+q_{p-1}+p},...,j_{q_0+...+q_p+p+1}$, are $0,...,q_0,q_0+2,...,q_0+q_1+2,q_0+q_1+4,...,q_0+...+q_{p-2}+2p-2,...,q_0+...+q_{p-1}+2p,q_0+...+q_{p-1}+2p,...,
q_0+...+q_p+2p$, respectively. Then, in terms of Remark 2 at the end of Section 4, we define $$\bar{p}(\gamma)=(\sigma,\tau,
\bar{\sigma})\in{P_p\text{St}(S)_{q_0...q_p}}\subseteq{S_{q+p}\times{S_{q+p}}\times{S_p}}$$ where $$\begin{aligned}
\bar{\sigma}&=&d_{0...q_0}\circ{\hat{d}_{q_0+1}}\circ{...}
\circ{d_{(q_0+...+q_{p-1}+2p)...(q_0+...+q_p+2p)}}
\circ{\hat{d}_{q+2p+1}}(\gamma)=d_{j_1...j_{q+p+1}}(\gamma)\\
\tau&=&d_{q_0+1}\circ{d_{q_0+q_1+3}}\circ{...}
\circ{d_{q_0+...+q_p+2p+1}}(\gamma)=d_{i_1...i_{p+1}}(\gamma)\\
\sigma&=&\hat{s}_{q_0+...+q_p+p}\circ{s_{(q_0+...+q_p+p-1)...(q_0+...+q_{p-1}+p)}}\circ
\hat{s}_{q_0+...+q_{p-1}+p-1}\circ{...}\circ\\
& &\hat{s}_{q_0+q_1+1}\circ{s_{(q_0+q_1)...(q_0+1)}}\circ
\hat{s}_{q_0}\circ{s_{(q_0-1)...(0)}}(\bar{\sigma})\\
&=&s_{(q+p-1)...(q+p-q_p)}\circ{...}\circ{s_{(q_0+q_1)...(q_0+1)}}\circ{s_{(q_0-1)...(0)}}(\bar{\sigma})\\
&=&\mu_{q_0,...,q_p}(\bar{\sigma}).\end{aligned}$$ Using the above expression for $\bar{\sigma}$ in terms of $d$’s and $\gamma$, we get $$\sigma=s_{(q+p-1)...(q+p-q_p)}\circ{...}\circ{s_{(q_0+q_1)...(q_0+1)}}\circ{s_{(q_0-1)...0}}\circ{d_{j_1...j_{q+p+1}}}(\gamma).$$ Now $\operatorname{St}(S)_{q_0+...+q_p+2p+1}$ contains the simplex $$\begin{aligned}
(s_{j_{q+p+1}...j_1}\circ{}s_{(q+p-1)...(q+p-q_p)}&\circ{...}&\circ{}s_{(q_0+q_1)
...(q_0+1)}
\circ{s_{(q_0-1)...0}}\circ{d_{j_1...j_{q+p+1}}}(\gamma),\\
s_{i_{p+1}...i_1}d_{i_1...i_{p+1}}(\gamma))&=&(s_{j_{q+p+1}...j_1}\sigma,s_{i_{p+1}...i_1}\tau).\\\end{aligned}$$ It follows that $(\sigma,\tau)\in{\operatorname{St}(S)}$ and hence $\bar{p}(\gamma)=(\sigma,\tau,\bar{\sigma})\in{P_p\operatorname{St}(S)_{q_0,...,q_p}}$.
Now $\bar{p}$ is a surjective map: Suppose $(\sigma,\tau,\bar{\sigma})\in{P_p\operatorname{St}(S)_{q_0,...,q_p}}$ and we shall find $\gamma\in{\bar{P}_p}S_{q_0,...,q_p}$ such that $\bar{p}(\gamma)=(\sigma,\tau,\bar{\sigma})$. Thus $(\sigma,\tau,\bar{\sigma})\in{P_p\operatorname{St}(S)_{q_0,...,q_p}}\subset{\delta(S\times{S})}_{q+p}$ is such that $$\pi_1(\sigma,\tau)\in{\text{Im}\{\mu_{q_0,...,q_p}:S_p\to{S_{q+p}}\}}$$ where $\bar{\sigma}\in{S_p}$.
Again use the partition $(p+1,q+p+1)$ as above, put $\gamma=s_{i_{p+1}...i_1}\sigma\in{S_{q+2p+1}}$. Indeed since $$(s_{j_{q+p+1}}\circ{s_{j_q-q_p+p-1}}\circ{...}\circ{}s_{j_{q_0+q_1+2}}\circ{s_{j_{q_0+1}}}\sigma,s_{i_{p+1}...i_1}\tau)$$ is of the required form as in Definition \[4.2\] and since $$(\sigma,\tau)=(d_{k_1....k_{p+1}}\times{}
d_{i_1...i_{p+1}})(s_{j_{q+p+1}}\circ{s_{j_q-q_p+p-1}}\circ{...}\circ{}s_{j_{q_0+q_1+3}}\circ{s_{j_{q_0+1}}}\sigma,s_{i_{p+1}...i_1}\tau)$$ here $d_{k_1....k_{p+1}}=d_{q_0+1}\circ{}d_{q_0+q_1+3}\circ{...}\circ{}d_{q_0+...+q_{p-1}+2p-1}\circ{}d_{q+2p+1}$. So the $d_K$’s and the $d_I$’s are the same, where $d_I=d_{i_1...i_{p+1}}$.
We have $(\sigma,\tau)\in{\operatorname{St}(S)}$. Hence $\bar{p}(\gamma)\in{P_.\operatorname{St}(S)_.}$.
2\) If $S=K^s$, $K$ simplicial complex then $$\begin{aligned}
P_p\text{St}(K^s)_{q_0,\dots,q_p}=\{(\sigma,\tau)\in{\text{St}}(K^s)_{q+p}\subset{\delta}(K^s\times{K^s})_{q+p}\\
\text{
}|\text{ }
\pi_1(\sigma,\tau)\in{\text{Im}}\{\mu_{q_0,\dots,q_p}:{K_p}^s\to{{K_{q+p}^s}}\}
\}.\end{aligned}$$
The map $\mu_{q_0,\dots,q_p}:{K_p}^s\to{{K_{q+p}^s}}$ takes $(i_0,...,i_p)$ to $(\underbrace{i_0,\dots,i_0}_{q_0+1-\text{times}},\dots,
\underbrace{i_p,\dots,i_p}_{q_p+1-\text{times}})$. Then $$\begin{aligned}
\sigma&=&(a_{i_0},\dots,a_{i_0},\dots,a_{i_p},\dots,a_{i_p})\in{{K^s}_{q+p}},\\
\tau&=&(b_{j_0},\dots,b_{j_{q_0}},\dots,b_{j_{q_0+...+q_{p-1}+p}},\dots,b_{j_{q+p}})\in{{K^s}_{q+p}}.\end{aligned}$$ By the definition $\bar{P}_p(K^s)_{q_0,\dots,q_p}=P_p(K^s)_{q_0+1,\dots,q_p+1}$. Then $\gamma$ in ${K^s}_{q+2p+1}$ given by $\gamma=(c_0,\dots,c_{q_0+1}|...|c_{q_0+...+q_{p-1}+2p},\dots,c_{q+2p+1})\in{K^s_{q+2p+1}}$ is uniquely determined by $\sigma$ and $\tau$.
Explicitly the inverse map ${\bar{p}}^{-1}:P_p\text{St}(K^s)_{q_0,\dots,q_p}\to{\bar{P}_p{K^s}_{q_0,\dots,q_p}}$ is defined by ${\bar{p}}^{-1}(\sigma,\tau)=\gamma$, where
$$\begin{aligned}
\sigma&=&(a_{i_0},\dots,a_{i_p}),\\
\tau&=&(b_{j_0},\dots,b_{j_{q+p}})\text{ }\text{and}\\
\gamma&=&({j_0}^\prime,\dots,j_{q_0}^\prime,i_0|{j_{q_0+1}}^\prime,\dots,j_{q_0+q_1+1}^\prime,i_1|...|j_{q_0+...+q_{p-1}+p}^\prime,\dots,j_{q+p}^\prime,i_p)\end{aligned}$$
such that for $\sum_{i=0}^{k-1}q_i+p\leq{s}\leq{\sum_{i=0}^kq_i+p}$
$$j_s^\prime= \left\{
\begin{array}{r@{\quad: \quad}l}i_{k-1} &
j_s\le{i_{k-1}} \\ j_s & i_{k-1}<j_s<i_k \\
i_k & i_k\le{j_s}, \end{array} \right.$$
$k=1,...,p$. Hence $\gamma\in{\bar{P_p}}{K^s}_{q_0,...,q_p}$ exists and is uniquely determined by ${(\sigma,\tau)}\in{\text{St}}(K^s)_{q+2p+1}$.
Therefore $\bar{p}:\bar{P_.}K^s\to{P_.\text{St}(K^s)}$ is an isomorphism.
[**Remark:**]{} Note that $\bar{p}$ is not injective for a simplicial set in general since for constructing the inverse map $P\text{St}(S)\to{\bar{P}S}$, there is no unique choice for the element $\gamma$ in $\bar{P}_.S_.$. In fact, we do not know which degeneracy operators we will use in order to define $\gamma$, so in general the inverse is not well-defined.
**[The Classifying Space and Lattice Gauge Theory]{}\[six\]**
=============================================================
For the definition of a classifying map we need a prismatic version of the standard construction of the classifying space.
Let $G$ be a topological group and the usual classifying space $BG=EG/G$ which is constructed as a simplicial space $EG_p=\underbrace{G_.\times{...}\times{G_.}}_{p+1-\text{times}}$, $BG_p=(G\times{...}\times{G})/G$.
In order to make this simplicial set discrete we can replace $G$ by the singular simplicial set of continuous maps $S_qG=\text{Map}(\Delta^q,G)$ and $E_.S_.G$ as in Example 2.8. is a prismatic set. However we shall need another model constructed as follows: For a continuous map $a\in{\text{Map}(\Delta^p\times{\Delta^{q_0...q_p}},G^{p+1})}$. Then we define $$a(t,s^0,...,s^p)=(a_0(t,s^0),a_1(t,s^0,s^1),...,a_p(t,s^0,...,s^p)),$$ where $(t,s^0,...,s^p)\in{\Delta^p\times{\Delta^{q_0...q_p}}}$. $S_.G$ acts on this prismatic set and we define $$\begin{aligned}
P_pEG_{q_0,...,q_p}&=&\{a:\Delta^p\times{\Delta^{q_0...q_p}}\to{G^{p+1}}
\text{
}|\text{ }
a_j(\varepsilon^it,s)\text{ }{\text{is independent of}}\\
& &\text{ } s^i\text{ }
\text{for all}\text{ }j\text{ }\text{different from}\text{ }i
\}.\end{aligned}$$ $P_.BG=P_.EG/S_.G$, that is, $$P_pBG_{q_0,...,q_p}=P_pEG_{q_0,...,q_p}/S_pG.$$
\[5.1\] The evaluation maps give horizontal homotopy equivalences in the diagram
$$\xymatrix{ \|\text{ }|P_.EG_.|\text{ }\| \ar[d]^{\|\text{
}|\gamma|\text{ }\|} \ar[r]^{\phantom{123}ev}
&EG \ar[d]^\gamma \\
\|\text{ }|P_.BG_.|\text{ }\| \ar[r]^{\phantom{123}ev} & EG/G }$$
Furthermore the top map is equivariant with respect to the homomorphism $ev:|S_.G|\to{G}$.
First notice that the evaluation map $ev:|S_.G|\to{G}$ is a homotopy equivalence. Also the equivariance is obvious by the commutative diagram $$\xymatrix{\|\text{ }|P_.EG_.|\text{ }\| \times{|S_.G|}
\ar[d]^{{\text{pr}_1}} \ar[r]^{\phantom{1234567}ev\times{ev}}
&EG\times{G} \ar[d]^{{\text{pr}_1}} \\
\|\text{ }|P_.EG_.|\text{ }\| \ar[r]^{\phantom{12}ev}
& EG }$$ Since $\|\text{ }|P_.EG_.|\text{ }\|$ and $EG$ are both contractible, the evaluation map induces a homotopy equivalence on the quotient.
**[Lattice Gauge Theory, Parallel Transport Function]{}\[six\]**
================================================================
In Lattice gauge theory in the sense of Phillips and Stone [@PS1] they construct for a given Lie group $G$ and a simplicial complex $K$ a $G$-bundle with connection on $|K|$ associated to a set of $G$-valued continuous functions defined over the faces of a simplex. These they call “parallel transport functions” since they are determined by parallel transport for the connection. In this section we shall introduce similar “compatible transition functions” for $K$ replaced by a simplicial set $S$ and in the following section we shall use these to construct a classifying map on the star complex $\bar{P_.}S_.$. First we consider $G$-bundles over simplicial sets.
\[6.2\] A bundle over $|S|$ is a sequence of bundles over ${\Delta}^p\times{\sigma}$ for all ${p}$, where $\sigma\in{S_p}$ and with commutative diagrams; $$\begin{aligned}
\xymatrix{
F_{{d}_j\sigma} \ar[d] \ar[r]^-{{\bar{\varepsilon}}^j}
&{F_{\sigma}} \ar[d] \\
{\Delta}^{p-1}\times{{d}_j\sigma} \ar[r]^-{{\varepsilon}^j}
& {\Delta}^p\times{\sigma}
}\end{aligned}$$ and $$\begin{aligned}
\xymatrix{
F_{{s}_j\sigma} \ar[d] \ar[r]^-{{\bar{\eta}}^j}
&{F_{\sigma}} \ar[d] \\
{\Delta}^{p+1}\times{{s}_j\sigma} \ar[r]^-{{\eta}^j}
& {\Delta}^p\times{\sigma}
}\end{aligned}$$ with the compatibility conditions: $$\begin{aligned}
{\bar{\varepsilon}}^j{\bar{\varepsilon}}^i= \left\{ \begin{array}{r@{\quad: \quad}l}{\bar{\varepsilon}}^i {\bar{\varepsilon}}^{j-1} & i<j \\
{\bar{\varepsilon}}^{i+1}
{\bar{\varepsilon}}^j & i\geq{j}, \end{array} \right.\end{aligned}$$
$$\begin{aligned}
{\bar{\eta}}^j{\bar{\eta}}^i= \left\{ \begin{array}{r@{\quad: \quad}l}{\bar{\eta}}^i {\bar{\eta}}^{j+1} & i\leq{j} \\
{\bar{\eta}}^{i-1}
{\bar{\eta}}^j & i>j, \end{array} \right.\end{aligned}$$
and $$\begin{aligned}
{\bar{\eta}}^j{\bar{\varepsilon}}^i= \left\{ \begin{array}{r@{\quad: \quad}l}{\bar{\varepsilon}}^i {\bar{\eta}}^{j-1} & i<j \\
1 &i=j,i=j+1\\
{\bar{\varepsilon}}^{i-1}
{\bar{\eta}}^i & i>{j+1}. \end{array} \right.\end{aligned}$$ 0.3 cm Given a $G$-bundle $F\to{|S|}$, $G$ a Lie group, since ${\Delta}^p$ is contractible, we can choose a trivialization ${\varphi}_{\sigma}:F_{\sigma}\to{{\Delta}^p}\times{\sigma}\times{G}$ for a non-degenerate $\sigma\in{S_p}$. If $\sigma$ is degenerate, that is, there exists $\tau$ such that $\sigma={s}_i\tau$, then the trivialization of $\sigma$ is defined as pullback of the trivialization of $\tau$, that is, ${\varphi}_{\sigma}={{\eta}^i}^*({\varphi}_{\tau})$.
\[6.3\] ( Admissible Trivializations ) A set of trivializations is called admissible, in case ${\varphi}_{\sigma}$ for $\sigma={s}_i\tau$ is given by ${\varphi}_{\sigma}={{s}^i}^*({\varphi}_{\tau})$.
\[6.4\] Admissible trivializations always exist.
Now, let us construct the transition functions for a simplex $\sigma\in{S_p}$ before giving the following proposition:
\[6.5\]
Given a bundle and a set of trivializations, we get for each face ${\tau}$ of say dim$\tau=q<p$ in $\sigma$, a transition function $v_{\sigma,\tau}:{\Delta}^q\to{G}$. E.g., if $\tau={d}_i\sigma$ then the transition function $v_{\sigma,{d}_i\sigma}:{\Delta}^{p-1}\to{G}$ is given by the diagram $$\xymatrix{
{\Delta}^{p-1}\times{({d}_i\sigma)}\times{G} \ar[d]
\ar[r]^-{\Theta}
&{\Delta}^p\times{(\sigma)}\times{G} \ar[d] \\
{\Delta}^{p-1}\times{{d}_i\sigma} \ar[r]^-{{\varepsilon}^i}
& {\Delta}^p\times{\sigma}
}$$ where ${d}_i\sigma=\tau$ and $\Theta={\varphi}_{\sigma}\circ{{\bar{\varepsilon}}^i}\circ{{{\varphi}_{{d}_i\sigma}}^{-1}}$. So $$\{v_{\sigma,\tau}|\sigma\in{S_p}\text{ }\text{and} \text{ }\tau\text{ }\text{is a face of}\text{ }\sigma\}$$ are [**the transition functions**]{} for the bundle over $|S|$.
[**Remark :**]{} The transitions functions are generalized lattice gauge fields. Classically Lattice gauge fields are defined only on 1-skeletons but one can extend them to $p-1$ simplices for all $p$, given rise to transition functions on $\Delta^p$, as above.
We now list a number of propositions stating the properties of these. The proofs are straight forward. For details see Akyar [@B].
\[6.6\]
Given a bundle on a simplicial set and admissible trivializations, the transition function $v_{\sigma,\tau}$, where $\tau$ is a face of $\sigma$, satisfies;
i\) $\sigma$ is nondegenerate: if $\gamma={d}_j\sigma$ and $\tau={d}_i\gamma$ then $$v_{\sigma,\tau}=(v_{\sigma,\gamma}\circ{{\varepsilon}^i}).v_{\gamma,\tau}.$$
This is called the cocycle condition.
ii\) $\sigma$ is degenerate: If $\sigma={s}_j{\sigma}^{\prime}$ and $\tau={d}_i\sigma$ then when $i<j$ for $\tau={s}_{j-1}{\tau}^{\prime}$ one gets ${\tau}^{\prime}={d}_i{ \sigma}^{\prime}$ and when $i>j+1$ for $\tau={s}_j{\tau}^{\prime}$ one gets ${\tau}^{\prime}={d}_{i-1}{ \sigma}^{\prime}$. For the other cases, $i=j$ or $i=j+1$, $\tau={\sigma}^{\prime}$. Then the transition functions satisfy: $$v_{\sigma,\tau}= \left\{ \begin{array}{r@{\quad: \quad}l}v_{{\sigma}^{\prime},{\tau}^{\prime}}\circ{{\eta}^{j-1}} & i<j \\
1 &i=j,i=j+1\\
v_{{\sigma}^{\prime},{\tau}^{\prime}}\circ{{\eta}^j}
& i>{j+1}. \end{array} \right.$$
iii\) If $\tau$ is a composition of face operators of $\sigma$, e.g., $\tau={\tilde{d}}^{p-(i-1)}\sigma$, $i=1,...,p$, where ${\tilde{d}}^{p-(i-1)}=d_i\circ{...}\circ{d_p}$ then $$v_{\sigma,\tau}=(v_{\sigma,{\tilde{d}}^1\sigma}\circ{{({\varepsilon}^i)}^{p-i}}).
(v_{{\tilde{d}}^1\sigma,{\tilde{d}}^2\sigma}\circ{{({\varepsilon}^i)}^{p-i-1}})...
(v_{{\tilde{d}}^{p-(i+1)}\sigma,{\tilde{d}}^{p-i}\sigma}\circ{{\varepsilon}^i}).v_{{\tilde{d}}^{p-i}\sigma,\tau}.$$
\[6.7\]
Assume that we have a bundle over $|S|$. Then
1\) There exists admissible trivializations such that the transition function is given by $$v_{\sigma,{d}_i\sigma}=1\text{ }\text{if}\text{ }i<p.$$
2\) For $\tau={\tilde{d}}^{p-(i-1)}\sigma$, $i=1,...,p$, we get $v_{\sigma,\tau}$ as product of some transition functions: $$v_{\sigma,\tau}=(v_{\sigma}\circ{{({\varepsilon}^i)}^{p-i}}).
(v_{{\tilde{d}}^1\sigma}\circ{{({\varepsilon}^i)}^{p-i-1}}).(v_{{\tilde{d}}^2\sigma}\circ{{({\varepsilon}^i)}^{p-i-2}})...
(v_{{\tilde{d}}^{p-(i+1)}\sigma}\circ{{({\varepsilon}^i)}^1}).(v_{{\tilde{d}}^{p-i}\sigma}).$$
3\) The transition functions $v_{\sigma,\tau}$ satisfy the compatibility conditions: $$v_{\sigma}\circ{{\varepsilon}^i}= \left\{ \begin{array}{r@{\quad: \quad}l}v_{{d}_i\sigma}
& i<p-1 \\ v_{{d}_{p-1}\sigma}.{v_{{d}_p\sigma}}^{-1}
& i={p-1} \end{array} \right.$$
4\) For a degenerate $\sigma$, we have $$v_{\sigma}\circ{{\eta}^j}=v_{{s}_j\sigma}$$ $\forall {j}$.
\[6.8\]
Given a bundle, one can find admissible trivializations such that the transition functions are determined by functions $v_\sigma:{\Delta}^{p-1}\to{G}$ for $\sigma\in{S_p}$ nondegenerate.
\[6.9\]
Suppose given a set of functions $$v_{\sigma}:{\Delta}^{p-1}\to{G}$$ for $\sigma\in{S_p}$ for all ${p}$, satisfying the compatibility conditions $$v_{\sigma}\circ{{\varepsilon}^i}= \left\{ \begin{array}{r@{\quad: \quad}l}v_{{d}_i\sigma}
& i<p-1 \\ v_{{d}_{p-1}\sigma}.{v_{{d}_p\sigma}}^{-1}
& i={p-1} \end{array} \right.$$ and $$v_{s_j\sigma}=v_{\sigma}\circ{\eta^j}.$$
Then one can define for each $\sigma\in{S_p}$ and each lower dimensional face $\tau$ of $\sigma$, a function $v_{\sigma,\tau}$ such that i) and ii) in Proposition \[6.6\] hold and such that $$v_{\sigma,\tau}= \left\{ \begin{array}{r@{\quad: \quad}l}v_{\sigma} & i=p \\
1 & i<p. \end{array} \right.$$
\[6.10\]
Given a set of transition functions $v_{\sigma,\tau}$ satisfying i) and ii) in Proposition \[6.6\], there is a bundle $F$ over $|S|$ and trivializations with transition functions $v_{\sigma,\tau}$.
\[6.11\]
Given a set of functions $v_{\sigma}$ satisfying the compatibility conditions in Proposition \[6.7\], one can construct a bundle $F$ over $|S|$ and the trivializations with the transition functions $v_{\sigma,d_p\sigma}=v_{\sigma}$ and $v_{\sigma,d_i\sigma}=1$ when $i<p$ and $v_{s_i\sigma}=v_{\sigma}\circ{\eta^i}$ for a degenerate $\sigma$.
\[6.12\] A set of functions $\{v_{\sigma}\}_{\sigma\in{S}}$ as in Proposition \[6.9\] are called a set of “compatible transition functions”.
We end this section by comparing these compatible transition functions with the “parallel transport functions” (p.t.f.) of Phillips and Stone [@PS1]. For $S=K^s$ these consist of a set of maps, $V_\sigma:c_\sigma\to{G}$ for each $r$-simplex $\sigma$ of $K$, $r\ge{1}$, $c_\sigma$ is the $(r-1)$-cube given by $0\le{s_{a_1}}\le{1},...,0\le{s_{a_{r-1}}}\le{1}$, where $\sigma=<a_0,...,a_r>\in{K}$ with the compatibility conditions
1\. Cocycle condition $$V_\sigma(s_{a_1},...,
s_{a_p}=1,
...,s_{a_{r-1}})=
V_{<a_0,...,a_p>} (s_{a_1},...,s_{a_{p-1}} ).
V_{<a_p,...,a_r>})(s_{a_{p+1}},...,s_{a_{r-1}}).$$
2\. Compatibility condition
$$V_\sigma(s_{a_1},..., s_{a_p}=0 ,...,s_{a_{r-1}})=
V_{<a_0,...,{\hat{a}}_p,...,a_{r-1}>}
(s_{a_1},...,{\hat{s}}_{a_p},...,s_{a_{r-1}}).$$
Now, suppose we have compatible transition functions $\{v_\sigma\}$ for a principal $G$-bundle $E\to{|K|}$ with triangulated base. Then for $\sigma=<a_0,...,a_r>$, the p.t.f. $V_\sigma:c_\sigma\to{G}$ is given by the parallel transport $E_{a_0}\to{E_{a_r}}$ along paths determined as follows:
Let $\sigma=<a_0,...,a_r>\in{K^s}$ and $s=(s_{a_0},...,s_{a_{r-1}})\in{c_\sigma}$.
We pick $r-1$-points as $P_1,...,P_{r-1}$ so that $P_1$ is on the line segment from $a_0$ to $a_1$, that is, $$P_1=(1-s_{a_1})a_0+s_{a_1}a_1=((1-s_{a_1},s_{a_1}),<a_0,a_1>)\in{|K|}.$$ Similarly, $P_2$ is on the line segment from $P_1$ to $a_2$, $P_2=(1-s_{a_2})P_1+s_{a_2}a_2$. Then $$P_2=((1-s_{a_2})(1-s_{a_1}),(1-s_{a_2})s_{a_1},s_{a_2},<a_0,a_1,a_2>).$$ By continuing in the same way, we get $$P_{r-1}=(1-s_{a_{r-1}})P_{r-2}+s_{a_{r-1}}a_{r-1}.$$ Let $\alpha$ be the piecewise linear path from $a_0$ through $P_1,...,P_{r-1}$ to $a_r$. In other words, $\alpha$ is determined uniquely up to parametrization by $r-1$ numbers $s_{a_1},...,s_{a_{r-1}}$. For $P_{r-1}=(t,d_r\sigma)\in{\Delta^{r-1}}\times{K_{r-1}}$, $d_r\sigma=<a_1,...,a_{r-1}>$, the element$$V_{\sigma}(s_1,...,s_{r-1})=v_{\sigma}(t)\in{G}$$ is to be interpreted as the parallel transport along $\alpha$.
{width="5cm"}\
**[The Classifying Map]{}\[eight\]**
====================================
[**[The construction of The Classifying Map]{}**]{}
For a given set of compatible transition functions (c.t.f.) $\{v_\sigma\}$ satisfying Proposition \[6.9\] we have seen in Proposition \[6.10\] that there is an associated $G$- bundle $F$ over $|S_.|$. Recall that the composite map $\text{proj}\circ{L}:\|\text{
}|\bar{P}_.S|\text{ }\|\to{\|\text{ }|S_.|\text{ }\|}\to{|S|}$ is a homotopy equivalence, where $L=\Lambda\circ{f}$ is given as in Proposition 4.2. In this section, we construct a classifying map for the bundle $(\text{proj}\circ{L})^*F$ over $\|\text{
}|{\bar{P}}_.S_.|\text{ }\|$.
\[7.4\] [**1)**]{} For given c.t.f.’s $\{v_\sigma\}$, there is a canonical prismatic map $m:{\bar{P}}_.S_.\to{P_.\text{BG}}$.
[**2)**]{} The induced map of geometric realizations $$ev\circ{\|\text{ }|m|\text{ }\|}=\bar{m}:\|\text{
}|\bar{P}_.S|\text{ }\|\stackrel{\|\text{ }|m|\text{
}\|}\to{\|\text{ }|P_.BG|\text{ }\|}\stackrel{ev}\to{BG}$$ is a classifying map for the $G$-bundle $(\text{proj}\circ{L})^*F$ over $\|\text{ }|{\bar{P}}_.S_.|\text{ }\|$.
[**Proof :**]{}
[**1)**]{} The map $m:{\bar{P}}_.S_.\to{P_.BG}$ is defined as $$m(\sigma)=[(a_0,a_1,...,a_p)]$$ where $\sigma\in{{\bar{P}}_pS_{q_0...q_p}}=S_{q+2p+1}$, $q=q_0+\dots+q_p$ and $a_i:\Delta^p\times{\Delta^{q_0...q_i}}\to{G}$ are given below. In the following, we use for convenience the interior coordinates $(t_1,\dots,t_p)$ of the standard simplex with barycentric coordinates $(t_0^\prime,\dots,t_p^\prime)$. $$t_1=1-t_0^\prime,\text{ }t_2=1-t_0^\prime-t_1^\prime,\text{ }\dots\text{ }t_{r-1}=1-t_0^\prime-\dots-t_{r-1}^\prime,\text{ }t_r=t_r^\prime$$ such that $0\leq{t_i}\leq{1}$, $i=1,\dots,p$, $1\geq{t_1}\geq{\dots}\geq{t_p}\geq{0}$ and $\sum_{i=0}^pt_i^\prime=1$, $t_i^\prime\leq{1}$, $i=0,\dots,p$.
In these terms the map $\Lambda$ from Section 3 is induced by the maps $\lambda_p:\Delta^p\times{\Delta^{q_0...q_p}}\to{\Delta^{q+2p+1}}$ given by
$$\begin{aligned}
\lambda_p(t,s^0,0,...,0,s^p,0)&=&(s_1^0(1-t_1)+t_1,\dots,s_{q_0}^0(1-t_1)+t_1,t_1,t_1,\\
& &s_1^1(t_1-t_2)+t_2,\dots,s_{q_1}^1(t_1-t_2)+t_2,t_2,t_2,\\
& &...,\\
& &s_1^{p-1}(t_{p-1}-t_p)+t_p,\dots,s_{q_{p-1}}^{p-1}(t_{p-1}-t_p)+t_p,t_p,t_p,\\
& &s_1^pt_p,...,s_{q_p}^pt_p,0).\end{aligned}$$
For convenience, we drop $p$ in $\lambda_p(t)(s)$ and write $\lambda(t)(s)$. Next, let $\rho^{(i)}:\Delta^{q+2p+1}\to{\Delta^{q_0+\dots+q_{i-1}+2i-1}}$ be the degeneracy map for $i=1,\dots,p$ defined by $$\rho^{(i)}:=\eta^{q_0+\dots+q_{i-1}+2i-1}\circ{\dots}\circ{\eta^{q+2p}}$$ deleting the last $q_i+\dots+q_p+2(p-i+1)$ coordinates. So e.g. $$\begin{aligned}
\rho^{(p)}\lambda(t)(s)&=&(s_1^0(1-t_1)+t_1,\dots,s_{q_0}^0(1-t_1)+t_1,t_1,t_1,\\
& &s_1^1(t_1-t_2)+t_2,\dots,s_{q_1}^1(t_1-t_2)+t_2,t_2,t_2,\\
& &...,\\
& &s_1^{p-1}(t_{p-1}-t_p)+t_p,\dots,s_{q_{p-1}}^{p-1}(t_{p-1}-t_p)+t_p,t_p),\end{aligned}$$ where $\rho^{(p)}:=\eta^{q-q_p+2p-1}\circ{...}\circ{\eta^{q+2p}}$ is deleting the last $q_p+2$ coordinates. With this notation, the maps $a_i:\Delta^p\times{\Delta^{q_0...q_i}}\to{G}$ defining the classifying map $m(\sigma)$ are given by $$\begin{aligned}
a_p(t,s^0,0,...,s^p,0)&=&1\\
a_{p-1}(t,s^0,0,...,s^{p-1},0)&=&v_{\sigma,d_{(p)}\sigma}(\rho^{(p)} (\lambda(t)(s))){}^{-1},\\
a_{p-2}(t,s^0,0,...,s^{p-2},0)&=&v_{\sigma,{\tilde{d}}^{(2)}\sigma}(\rho^{(p-1)} (\lambda(t)(s))){}^{-1}\\
&.&\\
&.&\\
&.&\\
a_1(t,s^0,0,s^1,0)&=&v_{\sigma,d_{(2)...(p)}\sigma}(\rho^{(2)} (\lambda(t)(s))){}^{-1}\\
a_0(t,s^0,0)&=&v_{\sigma,d_{(1)...(p)}\sigma}(\rho^{(1)} (\lambda(t)(s))){}^{-1}.\end{aligned}$$ Here the boundary operators used above are given as follows: $$d_{(p)}:S_{q+2p+1}\to{S_{q+2p-q_p-1}}$$ is defined by $d_{(p)}:=d_{q+2p-q_p}\circ{...}\circ{d_{q+2p+1}}$, deleting $q_p+2$ elements. On the other hand, in the formula ${\tilde{d}}^{(1)}={\hat{d}}_{(p)}=d_{(p)}$. Let’s denote $${\tilde{d}}^{(p-i)}={\hat{d}}_{(i+1)}\circ{...}\circ{{\hat{d}}_{(p)}}$$ $i=0,...,p-1$, which deletes the elements $(q_0+...+q_i+2i-1,...,q+2p+1)$. It deletes $q_{i+1}+...+q_p+2(p-i)=q-(q_0+...+q_i)+2(p-i)$ elements. Here $${\hat{d}}_{(i)}:S_{q+2i+1-\sum_0^{p-i-1}q_{p-j}}\to{S_{q+2i-1-\sum_0^{p-i}q_{p-j}}}$$ $i=1,...,p$. By using the equivalence relations on $m$ we can see that $m(d_{(i)}\sigma)$ is independent of $s^i$ for all $j$ different from $i$. Take $t_0^\prime=0$ then $v_{\sigma,{\tilde{d}}^{(p)}\sigma}(1,...,...,1,1,1)$ does not depend on $s^0$ where $j=1\neq{0}=i$.
[**2)**]{} For given c.t.f.’s $v_{\sigma}$, we now have the map of realizations $\|\text{ }|m|\text{ }\|:\|\text{ }|{\bar{P}}_.S_.|\text{ }\|\to{\|\text{ }|P_.BG|\text{ }\|}$ given by $$\begin{aligned}
\|\text{ }|m|\text{ }\|(t,s,\sigma)=(t,s,[(a_0,\dots,a_p)]).\end{aligned}$$
The associated bundle map is given as follows:
We have a bundle $F$ on $|S|$ by Proposition \[6.10\] and $|{\bar{P}}_pS_.|\to{|S_.|}$ is an epimorphism, so by pulling back we get a bundle $\bar{F}\to{|{\bar{P}}_pS_.|}$, i.e., $$\xymatrix{
\bar{F} \ar[d] \ar[r]
&F \ar[d] \\
|{\bar{P}}_.S_.| \ar[r]
& |S_.|
}$$ Transition functions used to define the classifying map $\tilde{m}$ are taken from the bundle $F\to{|S|}$. Let’s take $\sigma\in{S_{q+2p+1}}$ and there is a fibre at $(\lambda(t,s^0,0\dots,s^p,0),\sigma)$, by using the trivialization $\varphi_\sigma:F_\sigma\to{{\Delta}^{q+2p+1}\times{\sigma}\times{G}}$ and the projection on the last factor, we get $F_\sigma\to{G}$. Let’s denote this composition by ${\bar{\varphi}}_\sigma(\tilde{f})$ where $\tilde{f}:=(\lambda(t,s^0,0,\dots,s^p,0),\sigma)$, $\tilde{f}_\sigma\in{F_{(\lambda(t,s^0,0,\dots,s^p,0),\sigma)}}$, $\sigma\in{S_{q+2p+1}}$. On the other hand $${\varphi}_{d_{(p)}\sigma}:F_{d_{(p)}\sigma}\to{{\Delta}^{q+2p-q_p-1}}\times{d_{(p)}\sigma}\times{G}$$ gives us $${\bar{\varphi}}_{d_{(p)}\sigma}:F_{d_{(p)}\sigma}\to{G}.$$
By the definition, $${\bar{\varphi}}_\sigma({\bar{d}}^{(p)}\tilde{f}_\sigma):=v_{\sigma,d_{(p)}\sigma}(\rho^{(p)}\lambda(t,s^0,0,\dots,s^p,0)).{\bar{\varphi}}_{d_{(p)}\sigma}{(\tilde{f})}_{d_{(p)\sigma}},$$ where the compatible transition function is $$v_{\sigma,d_{(p)}\sigma}:{\Delta}^{q+2p-q_p-1}\to{G}.$$
The last component in $\|\text{ }|m|\text{ }\|(t,s,{\tilde{f}}_\sigma)$ is defined via the trivialization $\varphi_\sigma(\tilde{f})$ which is ${\bar{\varphi}}_\sigma(\tilde{f})$. By using the compatible transition function $v_{\sigma,d_{(p)}\sigma}$ we find the $p$-th component as $${v_{\sigma,d_{(p)}\sigma}(\rho^{(p)} \lambda(t,s^0,0,\dots,s^p,0))}^{-1}.{\bar{\varphi}}_\sigma(\tilde{f}).$$ We can apply the same method several times to get the other coordinates in $\|\text{ }|m|\text{ }\|(t,s,{\tilde{f}}_\sigma)$.
By the definition $PEG/SG=PBG$, $PEG=\|N\bar{G}\|$ and $\gamma:N\bar{G}\to{NG}$ we can identify $PBG=\|NG\|$. Then the required map $\bar{m}$ is $$\bar{m}(t,s,\sigma)=[(a_0,...,a_p)].$$
$\hfill \Box$
In particular for a simplicial complex $K$ we get the following (c. f. [@PS1])
(Phillips-Stone) [**1)**]{} A set of compatible transition functions $\{v_\sigma\}$ for $K$ a simplicial complex there is a natural prismatic map $$P_.\operatorname{St}(K^s)\to{PBG}.$$
[**2)**]{} The induced map on geometric realization gives a classifying map for the bundle $F$ pulled back to $|\operatorname{St}(K)|\subseteq{|K|\times{|K|}}$.
[**Proof:**]{} In the second part of Theorem \[4.4\], we have showed that $\bar{p}:\bar{P}K^s\to{P\text{St}K^s}$ is an isomorphism. On the other hand in the previous proposition, we have defined the classifying map $m$. This is also valid when $S=K^s$. So the p.t.f. $v_\sigma$ will determine a natural map $$m:P\text{St}K^s\to{PBG}.$$
Furthermore $\pi_1:P\text{St}(K^s)\to{K}$ is a homotopy equivalence.
[**Remark:**]{} The point of the corollary is that there is a connection in the prismatic universal bundle in the simplicial sense (see [@DLj]) which thus pulls back to a connection in the bundle over the star complex. We shall return to this elsewhere.
[99]{}
B. Akyar, Lattice Gauge Field Theory, [*Phd. thesis*]{}, Aarhus University, 2002.
J. Cheeger and J. Simons, [*Differential characters and geometric invariants*]{}, in [*Geometry and Topology, Proc. Spec. Year, College Park/ Md. 1983/84*]{}, eds. J. Alexander and J. Harer, pp. 50–80, Lecture Notes in Math., **1167**, Springer-Verlag, Berlin-Heidelberg-New York, 1985.
S.-S. Chern, J. Simons, [*Characteristic forms and geometric invariants*]{}, Ann. of Math., **99** (1974), 48–69.
J. L. Dupont, [*Curvature and Characteristic Classes*]{}, Lecture Notes in Math. **640**, Springer–Verlag, Berlin–Heidelberg–New York, 1978.
J. L. Dupont and F. W. Kamber, [*On a generalization of Cheeger-Chern-Simons classes*]{}, Illinois J. Math., **34** (1990), 221–255.
J. L. Dupont, R. Ljungmann, [*Integration along Deligne Cohomology*]{}, Math. Scand., **96**, (2005).
D. Freed, [*Classical Chern-Simons theory*]{}, Houston J. Math., **28** (2002), 293–310.
F. E. A. Johnson, [*On the triangulation of smooth fibre bundles*]{}, Fund. Math., **118** (1), (1985), 39-58.
M. Luscher, [*Topology of Lattice Gauge Fields*]{}, Comm. Math. Phys., **85**, (1982), 39-48.
R. Ljungmann, Secondary invariants for families of bundles, [*Phd. Thesis*]{}, Aarhus University 2006.
S. MacLane, Homology, Grundlehren Math. Wissensch, **114** 1963, Springer-Verlag, Berlin-Göttingen-Heidelberg.
P. May, Simplicial Objects in Algebraic Topology, Chicago Lectures in Math., 1967.
J. E. McClure, J. H. Smith, [*A solution of Deligne’s Hochschild cohomology conjecture*]{}, Contemp. Math., **293**, (2002), 153-193.
A. V. Phillips, D. A. Stone, [*Lattice Gauge Fields and Chern-Weil Theory*]{}, Geometry and Topology: Manifolds, varieties and knots, Athens, 1985.
A.V. Phillips, D.A. Stone, [*The Chern-Simons character of a lattice gauge field*]{}, Quantum Topology (1993), 244-291.
A.V. Phillips, D.A. Stone, [*Lattice gauge fields, Principal Bundles and the calculation of topological charge*]{}, Comm. Math. Phys., **103** (1986), 599-636.
A.V. Phillips, D.A. Stone, [*Topological Chern-Weil Theory*]{}, Memoirs Amer. Math. Soc., **105**, number 504 (1993).
E. Witten, [*Quantum Field Theory and the Jones polynomial*]{}, Comm. Math. Physics, **121** (1989), 351-399.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Active matter exhibits various forms of non-equilibrium states in the absence of external forcing, including macroscopic steady-state currents. Such states are often too complex to be modelled from first principles and our understanding of their physics relies heavily on minimal models. These have mostly been studied in the case of “dry" active matter, where particle dynamics are dominated by friction with their surroundings. Significantly less is known about systems with long-range hydrodynamic interactions that belong to “wet" active matter. Dilute suspensions of motile bacteria, modelled as self-propelled dipolar particles interacting solely through long-ranged hydrodynamic fields, are arguably the most studied example from this class of active systems. Their phenomenology is well-established: at sufficiently high density of bacteria, there appear large-scale vortices and jets comprising many individual organisms, forming a chaotic state commonly known as *bacterial turbulence*. As revealed by computer simulations, below the onset of collective motion, the suspension exhibits very strong correlations between individual microswimmers stemming from the long-ranged nature of dipolar fields. Here we demonstrate that this phenomenology is captured by the minimal model of microswimmers. We develop a kinetic theory that goes beyond the commonly used mean-field assumption, and explicitly takes into account such correlations. Notably, these can be computed exactly within our theory. We calculate the fluid velocity variance, spatial and temporal correlation functions, the fluid velocity spectrum, and the enhanced diffusivity of tracer particles. We find that correlations are suppressed by particle self-propulsion, although the mean-field behaviour is not restored even in the limit of very fast swimming. Our theory is not perturbative and is valid for any value of the micro-swimmer density below the onset of collective motion. This work constitutes a significant methodological advance and allows us to make qualitative and quantitative predictions that can be directly compared to experiments and computer simulations of micro-swimmer suspensions.'
author:
- Viktor Škultéty
- Cesare Nardini
- Joakim Stenhammar
- Davide Marenduzzo
- Alexander Morozov
bibliography:
- 'Refs.bib'
title: Swimming suppresses correlations in dilute suspensions of pusher microorganisms
---
Introduction {#section:introduction}
============
In recent years active systems emerged as a new state of matter with unique properties that are absent from their passive counterparts [@Ramaswamy2010; @Marchetti2013]. Such systems comprise particles that are capable of extracting energy from their environment and using it to exert forces and torques on their surroundings. The resulting self-propulsion and interactions between particles break detailed balance at the microscopic level, often leading to steady states that are not invariant under time reversal and exhibit macroscopic currents [@Cates2012]. Such currents, or collective motion, have been reported in a variety of systems [@Vicsek2012], including Vicsek particles [@Chate2008], mixtures of microtubules and molecular motors [@Sanchez2012], light-activated colloids [@Palacci2013], Quincke rollers [@Bricard2013; @Karani2019], bacterial colonies [@Zhang2010], sperm cells [@Creppy2015], locusts [@Buhl2006], birds, and fish [@Parrish1997]. The omnipresence of collective motion raises the need to classify various active systems according to common features of their phenomenological behaviour. Marchetti *et al.* [@Marchetti2013] recently introduced two broad universality classes for active systems, “dry" and “wet", comprising particles dominated by friction with their surroundings and long-ranged hydrodynamic interactions, respectively. Each class is expected to be defined by a few, relatively simple model systems, and significant effort has been invested into finding such models. For dry active matter, these include Vicsek-like models [@Vicsek2012; @Chate2020], that describe cases where alignment interactions are dominant, and Active Brownian Particles [@Howse2007; @Romanczuk2012] or Run and Tumble particles [@Schnitzer1993], that describe systems dominated by steric forces randomising their self-propulsion direction either smoothly or in a discontinuous manner. In this work, we study dilute suspensions of motile bacteria that, arguably, play the same role for wet active matter [@Koch2011; @Saintillan2013].
Collective motion in bacteria has been extensively studied in dilute [@Soni2003; @Dombrowski2004; @Gachelin2014] and dense [@Mendelson1999; @Wu2006; @Sokolov2007; @Sokolov2009; @Cisneros2011; @Sokolov2012; @Wensink2012a; @Dunkel2013] suspensions. These studies reveal the following sequence of dynamical states. At very low densities, bacterial suspensions appear featureless and disordered [@Wu2006; @Gachelin2014]. At higher, yet still sufficiently low densities, collective motion sets in on the scale of the system. In this state, bacterial motion takes the form of large-scale jets and vortices with typical speeds that are larger than the swimming speeds of individual organisms [@Soni2003; @Dombrowski2004; @Gachelin2014]. At significantly higher densities, there emerges a typical lengthscale of the vortices, which is comparable to about $5-10$ times the bacterial size [@Sokolov2012; @Dunkel2013; @Ryan2013]. Although this sequence of dynamical states has never been simultaneously observed in a single systematic bulk experiment, with the exception of Sokolov *et al.* [@Sokolov2009], the transition scenario is supported by computer simulations of self-propelled particles interacting through various forms of long-ranged hydrodynamic fields and short-ranged steric repulsion [@Hernandez-Ortiz2005; @Saintillan2007; @Wolgemuth2008; @Underhill2008; @Hernandez-Ortiz2009; @Lushi2013; @Lushi2014; @Krishnamurthy2015; @Wioland2016; @Saintillan2012; @Stenhammar2017; @Theillard2017; @Schwarzendahl2018; @Bardfalvy2019; @Theillard2019].
Bulk experiments with *E.coli* [@Gachelin2014] and *B.subtilis* [@Dombrowski2004] show that the transition to collective motion occurs around a volume fraction of bacterial bodies of about $1-2\%$. At such densities, the typical distance between the organisms is about $5-8$ times their body length, collisions are rare, and the far-field hydrodynamic interactions are thought to be dominant [@Koch2011; @Saintillan2013]. The latter are well-described by a “pusher”-like Stokesian dipolar field [@Lauga2009; @Drescher2011], generated when two point forces of equal magnitude and pointing away from each other are applied to a viscous fluid. Self-propelled pusher-like dipolar particles thus form a minimal model for dilute bacterial suspensions.
The transition to collective motion in dilute bacterial suspensions can be understood in terms of a mean-field kinetic theory [@Koch2011; @Saintillan2013] incorporating the minimal ingredients discussed above. Such theory identifies re-orientation of bacteria in the velocity field created by other organisms as the key ingredient leading to a global isotropic-nematic transition. The globally ordered state is, however, linearly unstable through a long-wavelength generic instability [@AditiSimha2002; @Marchetti2013], and there ensue never-settling dynamics as a compromise between the two instabilities. The critical density of bacteria at the onset of collective motion is determined by the strength of their dipolar interactions, their shape, and the way individual organisms change their orientation: either by occasionally re-orienting in a random way (tumble), or by rotational diffusion [@Saintillan2008; @Saintillan2008a; @Subramanian2009; @Hohenegger2010; @Krishnamurthy2015]. Typically, the critical threshold density is significantly lower in the latter case, and going to zero in the absence of a decorrelation mechanism for individual bacterium orientation. The mean-field kinetic theory has also been extended to systems with steric interactions [@Ezhilan2013; @Ryan2013; @Heidenreich2016; @Reinken2018] and to microswimmers suspended in non-Newtonian fluids [@Bozorgi2013; @Bozorgi2014; @Li2016].
Below the onset of collective motion, the mean-field kinetic theory predicts that the suspension is homogeneous and isotropic, as featureless as a suspension of non-interacting microswimmers. These assumptions are widely used when describing rheological properties of very dilute suspensions [@Hatwalne2004; @Chen2007; @Sokolov2009a; @Saintillan2010; @Underhill2011; @Lopez2015; @Alonso-Matilla2016; @Nambiar2017; @Guo2018; @Nambiar2019; @Liu2019; @Saintillan2018] and enhanced diffusivity of tracer particles [@Wu2000; @Kim2004; @Underhill2008; @Leptos2009; @Dunkel2010; @Ishikawa2010; @Childress2010; @Childress2011; @Kurtuldu2011; @Mino2011; @Mino2013; @Jepson2013; @Pushkin2013; @Pushkin2013jfm; @Morozov2014; @Kasyap2014; @Thiffeault2015; @Patteson2016; @Burkholder2017]. However, recent large-scale Lattice-Boltzmann simulations of dipolar swimmers [@Stenhammar2017; @Bardfalvy2019] revealed the presence of very strong correlations below the onset of collective motion. It was shown that various observables deviate from their mean-field values at any density of microswimmers [@Stenhammar2017], with the deviation diverging in the vicinity of the onset. The origin of such strong correlations can be readily attributed to the slow spatial decay of the dipolar velocity field, implying a simultaneous coupling between all microswimmers in the system. While this argument is intuitive enough, its implementation as a theoretical framework presents major technical challenges [@Stenhammar2017; @Qian2017; @Nambiar2019a], and only simplified cases were studied until now, notably by Stenhammar *et al.* [@Stenhammar2017], who considered a suspension of “shakers" – particles that apply forces to the fluid but do not self-propel.
In this work we develop a kinetic theory that goes beyond the mean-field assumption for the general model of dilute microswimmer suspensions described above. Our theory explicitly includes particle self-propulsion, and is valid at any density of microswimmers below the onset of collective motion. This constitutes simultaneously a significant methodological development compared to the work by Stenhammar *et al.* [@Stenhammar2017], and a major advance in our understanding of one of the key models defining “wet" active matter. Our theory allows us to make explicit predictions for observables that can be directly set against experiments and numerical simulations.
The paper is organised as follows. In Section \[section:theory\] we formulate a kinetic theory for a model suspension of pusher-like dipolar microswimmers. We explicitly find the dynamics of fluctuations around the homogeneous and isotropic state that describe the system below the onset of collective motion. Since our theory differs significantly from the previous work [@Stenhammar2017], we present its derivation in detail. We appreciate, however, that some readers might only be interested in the results of our theory without feeling the need to go through the rather technical Section \[section:theory\]. We, therefore, present our results in a stand-alone Section \[section:results\], which can be read without Section \[section:theory\]. There, we calculate the temporal and spatial correlation functions, fluid velocity variance, energy spectra, and the enhanced diffusivity of tracer particles. We conclude in Section \[section:discussion\], while Appendices contain additional derivations for technically oriented readers.
Kinetic theory of strongly interacting suspensions {#section:theory}
==================================================
Microscopic model {#subsection:microscopic}
-----------------
We consider a collection of $N$ microswimmers contained in a volume $V$ at a finite number density $n=N/V$. The microswimmers are suspended in a Newtonian fluid with the viscosity $\mu$. Each microswimmer is described by its instantaneous position ${\bm x}_i$ and orientation ${\bm p}_i$, that we collectively denote by ${\bm z}_i = \left( {\bm x}_i, {\bm p}_i\right)$, where $i=1\dots N$ enumerates the particles. Within our model, the dynamics of the suspension is governed by the following equations of motion $$\begin{aligned}
\dot{x}_{i}^{\alpha} &= v_{s} p_{i}^{\alpha} + \mathcal{U}^{\alpha}\left( {\bm x}_i \right),
\label{xdot} \\
\dot{p}_{i}^{\alpha} &= \mathbb{P}_{i}^{\alpha\beta} \left(\mathcal{W}^{\beta\gamma}\left( {\bm x}_i \right) + B \mathcal{E}^{\beta\gamma}\left( {\bm x}_i \right) \right) p_{i}^{\gamma},
\label{pdot}\end{aligned}$$ where the dot denotes the time derivative, the superscript indices denote Cartesian components of vectors, and the subscript indices label the particles. Throughout this work, we utilise the Einstein summation convention for the superscript indices, while no summation is assumed over repeated subscript indices.
Equations of motion and incorporate the following physical ingredients. First of all, each swimmer self-propels with the speed $v_s$ in the direction of its orientation. To induce self-propulsion, swimmers generate long-ranged flows in the suspending fluid [@Lauga2009]. The superposition of these flows at the position of the $i$-th swimmer, $\mathcal{U}^{\alpha}\left( {\bm x}_i \right)$, advects that particle in addition to its self-propulsion, see Eq., and re-orients it according to Jeffrey’s equation . The latter describes the dynamics of a passive particle in an external flow [@kimkarrila], with $$\begin{aligned}
&\mathcal{W}^{\beta\gamma}\left( {\bm x}_i \right) = \frac{1}{2} \left( \nabla^{\gamma} \mathcal{U}^{\beta}\left( {\bm x}_i \right) - \nabla^{\beta} \mathcal{U}^{\gamma}\left( {\bm x}_i \right) \right), \\
&\mathcal{E}^{\beta\gamma} \left( {\bm x}_i \right)= \frac{1}{2} \left(\nabla^{\gamma} \mathcal{U}^{\beta}\left( {\bm x}_i \right) + \nabla^{\beta} \mathcal{U}^{\gamma}\left( {\bm x}_i \right) \right),
\label{eq:Strain}\end{aligned}$$ being the Cartesian components of the vorticity and rate-of-strain tensors, respectively. In Eq., $\mathbb{P}_{i}^{\alpha\beta} = \delta^{\alpha\beta} - p^{\alpha}_{i}p_{i}^{\beta}$, is the projection operator, $\delta^{\alpha\beta}$ denotes the Kronecker delta, $\nabla_i^\alpha = \partial/\partial x_i^\alpha$, and $B = \left( a^2 - 1\right)/\left( a^2 + 1\right)$ is the measure of the swimmer’s nonsphericity [@kimkarrila] based on its aspect ratio $a$. For strongly elongated particles, $B\rightarrow 1$, while for spheres, $B=0$. Finally, each swimmer randomly changes its orientation with a rate $\lambda$, thus mimicking the run-and-tumble motion commonly exhibited by bacteria [@Berg1993]. We note here that we neglect the effects of rotational and translational diffusion on the particle’s dynamics, and random tumbling is thus the only source of stochasticity in our model.
The velocity field generated by a self-propelled particle sufficiently far away from its surface is often well-described by the field produced by a point dipole with the same position and orientation [@Lauga2009; @Drescher2011]. In a dilute suspension of microswimmers, where the particles are sufficiently separated from each other, we can approximate $\mathcal{U}^{\alpha}\left( {\bm x}_i \right)$ by a sum of dipolar contributions $$\begin{aligned}
\mathcal{U}^{\alpha}\left( {\bm x}_i \right) = \sum_{j\neq i}^{N} u_{d}^{\alpha}({\bm x}_{i}; {\bm z}_{j}),\end{aligned}$$ where $$\begin{aligned}
{\bm u}_{d}({\bm x}_{i}; {\bm z}_{j}) = \frac{\kappa}{8\pi} \left[ 3
\frac{\left( {\bm p}_j\cdot {\bm x}'\right)^2 {\bm x}' + \epsilon^2 \left( {\bm p}_j\cdot {\bm x}'\right) {\bm p}_j }{\left( x'^2 + \epsilon^2 \right)^{5/2}} \right. \nonumber \\
\left. -\frac{{\bm x}'}{\left( x'^2 + \epsilon^2 \right)^{3/2}} \right]
\label{realspacedipole}\end{aligned}$$ is the velocity field generated at ${\bm x}_{i}$ by a hydrodynamic dipole located at ${\bm x}_{j}$ with the orientation ${\bm p}_{j}$. Here, $\kappa = F l/\mu$ is the reduced dipolar strength, where $F$ is the magnitude of the forces applied to the fluid, $l$ is the dipolar length, and $\mu$ is the viscosity of the fluid; ${\bm x}' = {\bm x}_i -{\bm x}_j$, and $x'$ denotes the length of ${\bm x}'$. The dipole consists of two regularised Stokeslets, that were introduced by Cortez *et al.* [@Cortez2005], with $\epsilon$ being the regularisation length of the order of swimmer size. For pushers, $\kappa > 0$.
The main goal of our work is to calculate spatial and temporal correlations of the fluid velocity in microswimmer suspensions described by the model above. Both quantities can be succinctly expressed through a combined correlation function $$\begin{aligned}
&C(R,T) \nonumber \\
&\qquad = \lim_{t\rightarrow\infty} \frac{1}{V}\int d{\bm x} \,\overline{U^{\alpha}\left( {\bm x}, t \right) U^{\alpha}\left( {\bm x} + {\bm R}, t+T \right)},
\label{CRTgeneral}\end{aligned}$$ where $U^{\alpha}\left( {\bm x}, t \right)$ is the fluid velocity at the position ${\bm x}$ at time $t$, and the large-$t$ limit guarantees independence of the initial conditions. The spatial and temporal correlation functions are trivially recovered by setting $T=0$ and $R=0$, respectively. The bar in Eq. denotes the average over the history of tumble events, and reflect the stochastic nature of our model. To calculate this and similar averages, below we formulate a kinetic theory of microswimmer suspensions based on our macroscopic model. Such theories have been extensively studied at the mean-field level [@Liao2007; @Lau2009; @Zaid2011; @Underhill2011; @Belan2019; @Bardfalvy2019]. Here, we go beyond the mean-field approximation and explicitly take into account strong correlations between the swimmers caused by the long-range nature of their hydrodynamic fields, Eq..
Kinetic theory and BBGKY hierarchy
----------------------------------
The starting point of our theory is the $N$-particle probability distribution function $F_N\left({\bm z}_1, {\bm z}_2, \dots, {\bm z}_N, t \right)$ that gives the geometric probability of the system occupying a particular point in the $6N$-dimensional phase space $\left\{ {\bm z}_1, \dots, {\bm z}_N\right\} $ at time $t$. The $N$-particle probability distribution function is symmetric with respect to swapping particle labels, reflecting their indistinguishability, and is normalised $$\begin{aligned}
\int d{\bm z}_1 \dots d{\bm z}_N F_N\left({\bm z}_1, \dots, {\bm z}_N, t \right) = 1.
\label{normalisation}\end{aligned}$$ Its time dynamics is governed by the Master equation [@Balescu1975] $$\begin{aligned}
\partial_{t} F_{N} & + \sum_{i=1}^{N} \Big[ \nabla_{i}^{\alpha}(\dot{x}_{i}^{\alpha} F_{N}) + \partial_{i}^{\alpha}(\dot{p}_{i}^{\alpha} F_{N}) \Big] \nonumber \\
& = - N \lambda F_{N} + \frac{\lambda}{4\pi} \sum_{i=1}^{N} \int \mathrm{d} {\bm p}_{i} F_{N},
\label{Liouville}\end{aligned}$$ where we introduced $\partial_{i}^{\alpha} = \mathbb{P}_{i}^{\alpha\beta} \partial/\partial p^\beta_i$. The l.h.s. of Eq. describes the probability fluxes to and from a particular point in the phase space due to the deterministic particle dynamics given by Eqs. and , while the r.h.s. gives the changes of the probability due to random tumbling from and into that phase space point [@Subramanian2009; @Koch2011]. Next, we introduce the $s$-particle correlation functions defined as $$\begin{aligned}
&F_{s}\left({\bm z}_1,\dots, {\bm z}_s, t \right) = \frac{N!}{(N-s)!N^{s}} \nonumber \\
&\qquad\qquad \times \int d{\bm z}_{s+1} \dots d{\bm z}_N F_N\left({\bm z}_1, \dots, {\bm z}_N, t \right),
\label{Fs}\end{aligned}$$ Below, we will only be interested in the first partial correlation functions $F_1$, $F_2$, and $F_3$, that we further express as $$\begin{aligned}
F_2\left( {\bm z}_1, {\bm z}_2, t\right) = F_1\left( {\bm z}_1, t\right) F_1\left( {\bm z}_2, t \right) + G\left( {\bm z}_1, {\bm z}_2, t\right),
\label{F2}\end{aligned}$$ and $$\begin{aligned}
F_3\left( {\bm z}_1, {\bm z}_2, {\bm z}_3, t\right) = F_1\left( {\bm z}_1, t\right) F_1\left( {\bm z}_2, t \right) F_1\left( {\bm z}_3, t \right) \nonumber \\
+ G\left( {\bm z}_1, {\bm z}_2, t\right) F_1\left( {\bm z}_3, t \right) + G\left( {\bm z}_1, {\bm z}_3, t\right) F_1\left( {\bm z}_2, t \right) \nonumber \\
+ G\left( {\bm z}_2, {\bm z}_3, t\right) F_1\left( {\bm z}_1, t \right) + H\left( {\bm z}_1, {\bm z}_2, {\bm z}_3, t\right),
\label{F3}\end{aligned}$$ where $G$ and $H$ are the irreducible (connected) correlation functions [@Balescu1975]. The time evolution of $F_s$ can be deduced from the Master equation by integrating it over $\left\{{\bm z}_{s+1},\dots,{\bm z}_N\right\}$. Integrating by parts and using Eqs. and , we obtain the following equations for the one- and two-particle irreducible correlation functions
$$\begin{aligned}
& \partial_t F_1({\bm z}, t) + {\mathcal L}[F_1({\bm z}, t)]({\bm z}) \nonumber \\
& \qquad\qquad = -N \nabla^\alpha \int d{\bm z}' G({\bm z},{\bm z}',t) u^\alpha_d ({\bm x}; {\bm z}')
- N \mathbb{P}^{\alpha\beta} \frac{\partial}{\partial p^\beta} \int d{\bm z}' G({\bm z},{\bm z}',t)
p^\gamma \mathbb{X}^{\alpha\mu\nu\gamma} \nabla^\mu u^\nu_d ({\bm x}; {\bm z}'),
\label{EqF1}
\\
\nonumber \\
& \partial_t G({\bm z}_1,{\bm z}_2, t) + {\mathcal L}[G({\bm z}_1,{\bm z}_2, t)]({\bm z}_1) + {\mathcal L}[G({\bm z}_1,{\bm z}_2, t)]({\bm z}_2) \nonumber \\
& \qquad \qquad + N \nabla_1^\alpha \left[ F_1({\bm z}_1,t) \int d{\bm z}' G({\bm z}_2,{\bm z}',t) u_d^\alpha({\bm x}_1;{\bm z}')\right]
+ N \nabla_2^\alpha \left[ F_1({\bm z}_2,t) \int d{\bm z}' G({\bm z}_1,{\bm z}',t) u_d^\alpha({\bm x}_2;{\bm z}')\right]
\nonumber \\
& \qquad \qquad + N \mathbb{P}_1^{\alpha\beta} \frac{\partial}{\partial p_1^\beta}\left[ F_1({\bm z}_1,t) p_1^\gamma \mathbb{X}_1^{\alpha\mu\nu\gamma} \int d{\bm z}' G({\bm z}_2,{\bm z}',t)\nabla_1^\mu u_d^\nu({\bm x}_1;{\bm z}')\right]
\nonumber \\
& \qquad \qquad + N \mathbb{P}_2^{\alpha\beta} \frac{\partial}{\partial p_2^\beta}\left[ F_1({\bm z}_2,t) p_2^\gamma \mathbb{X}_2^{\alpha\mu\nu\gamma} \int d{\bm z}' G({\bm z}_1,{\bm z}',t)\nabla_2^\mu u_d^\nu({\bm x}_2;{\bm z}')\right] \nonumber \\
& \qquad \qquad = - \mathcal{S}_{1,2}^{F} - \mathcal{S}_{2,1}^{F} - \mathcal{S}_{1,2}^{G} - \mathcal{S}_{2,1}^{G} - \mathcal{S}_{1,2}^{H} - \mathcal{S}_{2,1}^{H},
\label{EqG}\end{aligned}$$
where we have introduced the operator $$\begin{aligned}
&{\mathcal L}[\Phi]({\bm z}) \nonumber \\
& \qquad =
v_s p^\alpha \nabla^\alpha \Phi({\bm z})
+ N \nabla^\alpha \big[ \Phi({\bm z}) \mathcal{U}^\alpha_{\text{MF}}({\bm x})\big]
+ N \mathbb{P}^{\alpha\beta} \frac{\partial}{\partial p^\beta}\big[ \Phi({\bm z}) p^\gamma \mathbb{X}^{\alpha\mu\nu\gamma} \nabla^\mu \mathcal{U}^\nu_{\text{MF}}({\bm x})\big]
+ \lambda \Phi({\bm z})
- \frac{\lambda}{4\pi} \int d{\bm p}\, \Phi({\bm z}),
\label{operatorL}\end{aligned}$$
acting on the variable $\bm z$ of an arbitrary function $\Phi=\Phi({\bm z}_1,\dots,{\bm z}_N)$, and defined the mean-field velocity field as $$\begin{aligned}
\mathcal{U}^\alpha_{\text{MF}}({\bm x}) = \int d{\bm z}' F_1({\bm z}',t) u^\alpha_d({\bm x}; {\bm z}').
\label{UMF}\end{aligned}$$ The rank-4 tensor $$\begin{aligned}
\mathbb{X}_i^{\alpha\mu\nu\gamma} = \mathbb{P}_i^{\alpha\beta} \left[ \frac{B+1}{2}\delta^{\mu\gamma}\delta^{\nu\beta} + \frac{B-1}{2}\delta^{\mu\beta}\delta^{\nu\gamma} \right],\end{aligned}$$ encodes the tensorial structure of Jeffrey’s equation , and the r.h.s. of Eq. is given in terms of $$\begin{aligned}
&\mathcal{S}_{i,j}^{F} =
F_1({\bm z}_j,t)\Big\{ \nabla_i^\alpha \left[ F_1({\bm z}_i,t) u_d^\alpha({\bm x}_i;{\bm z}_j)\right] \nonumber \\
&+ \mathbb{P}_i^{\alpha\beta} \frac{\partial}{\partial p_i^\beta}\left[ F_1({\bm z}_i, t) p_i^\gamma \mathbb{X}_i^{\alpha\mu\nu\gamma} \nabla_i^\mu u_d^\nu({\bm x}_i; {\bm z}_j) \right] \Big\},\end{aligned}$$ $$\begin{aligned}
& \mathcal{S}_{i,j}^{G} =
\nabla_i^\alpha \left[ G({\bm z}_i,{\bm z}_j,t) u_d^\alpha({\bm x}_i;{\bm z}_j)\right] \nonumber \\
&+ \mathbb{P}_i^{\alpha\beta} \frac{\partial}{\partial p_i^\beta}\left[ G({\bm z}_i,{\bm z}_j,t) p_i^\gamma \mathbb{X}_i^{\alpha\mu\nu\gamma} \nabla_i^\mu u_d^\nu({\bm x}_i; {\bm z}_j) \right], \end{aligned}$$ and $$\begin{aligned}
& \mathcal{S}_{i,j}^{H} =
N \int d{\bm z}' \Big\{
\nabla_i^\alpha \left[ H({\bm z}_i,{\bm z}_j,{\bm z}',t) u_d^\alpha({\bm x}_i;{\bm z}')\right] \nonumber \\
&+ \mathbb{P}_i^{\alpha\beta} \frac{\partial}{\partial p_i^\beta}\left[ H({\bm z}_i,{\bm z}_j,{\bm z}',t) p_i^\gamma \mathbb{X}_i^{\alpha\mu\nu\gamma} \nabla_i^\mu u_d^\nu({\bm x}_i; {\bm z}') \right] \Big\}.\end{aligned}$$ Eqs. and are the beginning of a BBGKY hierarchy of equations for partial distribution functions [@Balescu1975]. As such, they do not form a closed system as they also depend on the three-particle irreducible distribution function $H$.
Before discussing our choice of closure for this system of equations, let us briefly review the predictions of the mean-field approximation to Eqs. and , which consists of neglecting all correlation functions beyond $s=1$. The remaining equation determines the mean-field approximation to the one-particle correlation function $$\begin{aligned}
\partial_t F_1^{\text{MF}}({\bm z}, t) + {\mathcal L}[F_1^{\text{MF}}({\bm z}, t)]({\bm z}) = 0,
\label{EqF1MF}\end{aligned}$$ that has been extensively studied before [@Saintillan2008; @Saintillan2008a; @Subramanian2009; @Hohenegger2010; @Koch2011; @Saintillan2013; @Krishnamurthy2015]. One of the solutions of this equation is given by a constant, which is fixed to $F_1^{\text{MF}}({\bm z}, t)=1/(4\pi V)$ by the normalisation condition Eq.. This solution, which is valid at any number density, corresponds to a homogeneous and isotropic suspension of microswimmers. For pushers $(\kappa>0)$, this state loses its stability [@Saintillan2008; @Saintillan2008a; @Subramanian2009; @Hohenegger2010; @Stenhammar2017] at the critical number density of microswimmers $n_{crit} = 5\lambda/(B\kappa)$, while for pullers $(\kappa<0)$, the homogeneous and isotropic state is always linearly stable within the mean-field approximation.
The homogeneous and isotropic mean-field solution implies that $N F_1^{\text{MF}}\sim n \sim O(1)$ is finite in the thermodynamic limit. This, in turn, implies that, to leading order, $G\sim O(N^{-2})$, $H\sim O(N^{-3})$, etc. A more comprehensive discussion of this statement, together with the required rescaling of the correlation functions, system parameters, and time is given elsewhere [@Stenhammar2017].
Building upon these results, here we *assume* that upon approaching the thermodynamic limit, $F_1$ is well-approximated by $F_1^{\text{MF}}$, since the r.h.s. of Eq. is $O(1/N)$ compared its l.h.s. In the homogeneous and isotropic state, the mean-field velocity vanishes $\mathcal{U}^\alpha_{\text{MF}}({\bm x}) = 0$, since the integral in Eq. is then proportional to the total flow rate through a surface surrounding the dipole. The latter is zero due to incompressibility. Fluctuations around the homogeneous and isotropic state are then governed by Eq. with $F_1=1/(4\pi V)$, and $\mathcal{S}_{i,j}^{G}$ and $\mathcal{S}_{i,j}^{H}$ neglected $$\begin{aligned}
& \partial_t G({\bm z}_1,{\bm z}_2, t) + {\mathcal L}_{12} [G] + {\mathcal L}_{21} [G] \nonumber \\
& \qquad\qquad = \frac{3B}{\left(4\pi V\right)^2} \Big\{ p_1^\mu p_1^\nu \nabla_1^\mu u_d^\nu({\bm x}_1;{\bm z}_2) \nonumber \\
& \qquad\qquad\qquad\qquad\qquad+ p_2^\mu p_2^\nu \nabla_2^\mu u_d^\nu({\bm x}_2;{\bm z}_1) \Big\},
\label{EqGfinal}\end{aligned}$$ where $$\begin{aligned}
& {\mathcal L}_{ij} [G] =
v_s p_i^\alpha \nabla_i^\alpha G({\bm z}_1,{\bm z}_2, t) \nonumber \\
&\qquad\qquad - \frac{3 n B}{4\pi} p_i^\mu p_i^\nu \int d{\bm z}' G({\bm z}_j,{\bm z}', t) \nabla_i^\mu u_d^\nu({\bm x}_i;{\bm z}') \nonumber \\
& \qquad\qquad + \lambda G({\bm z}_1,{\bm z}_2, t)
- \frac{\lambda}{4\pi} \int d{\bm p}_i\, G({\bm z}_1,{\bm z}_2, t).
\label{Lij}\end{aligned}$$ This equation has previously been derived and analysed for the case of shakers ($v_s=0$) [@Stenhammar2017]. Here, we solve it in the general case $v_s>0$.
Phase-space density fluctuations
--------------------------------
While the two-point distribution function $G$, given by Eq., contains statistical information about fluctuations in the system, it is not straightforward to relate it to the spatial and temporal correlation function $C(R,T)$, Eq., that we seek to calculate. To establish this connection, we introduce a method based on the phase space density $$\begin{aligned}
\varphi({\bm z},t) = \sum_{i=1}^N \delta({\bm z} - {\bm z}_i(t)),\end{aligned}$$ pioneered by Klimontovich [@klimontovich1967book]. Here, $\delta({\bm z})$ is the three-dimensional Dirac delta function. The average of the phase space density is related to $F_1$ as can be seen from $$\begin{aligned}
&\overline\varphi({\bm z},t) = \int d{\bm z}_1 \dots d{\bm z}_N \sum_{i=1}^N \delta({\bm z} - {\bm z}_i) F_N({\bm z}_1, \dots, {\bm z}_N, t) \nonumber \\
& \qquad\qquad = N F_1({\bm z}, t),\end{aligned}$$ where we used Eq.. Fluctuations of the phase space density can formally be defined as $\delta \varphi = \varphi - \overline\varphi$, and their second moment is given by $$\begin{aligned}
& G_K( {\bm z}', {\bm z}'', t) \equiv \overline{\delta \varphi({\bm z}',t) \delta \varphi({\bm z}'',t)} \nonumber \\
& \qquad = N^2 G ( {\bm z}', {\bm z}'', t) + N F_1({\bm z}',t) \delta({\bm z}' - {\bm z}'').
\label{GK}\end{aligned}$$ Below, we refer to $G_K$ as the Klimontovich correlation function. Its utility is evident if one considers the spatial correlation function $C(R)$, defined in Eq. as $$\begin{aligned}
&C(R) = \lim_{t\rightarrow\infty} \frac{1}{V}\int d{\bm x} \overline{U^{\alpha}\left( {\bm x}, t \right) U^{\alpha}\left( {\bm x} + {\bm R}, t \right)}.\end{aligned}$$ The velocity of the fluid at a position $\bm x$ is given by the superposition of the velocity fields generated by all swimmers $$\begin{aligned}
&U^{\alpha}\left( {\bm x}, t \right) = \sum_{i=1}^N u_d^\alpha({\bm x}; {\bm z}_i(t)) \nonumber \\
&\qquad\qquad = \int d{\bm z}' \varphi({\bm z}',t) u_d^\alpha({\bm x}; {\bm z}'). \end{aligned}$$ Separating the phase space density into its average and fluctuations, $\varphi =\overline\varphi + \delta \varphi$, the spatial correlation function becomes $$\begin{aligned}
&C(R) = \lim_{t\rightarrow\infty} \frac{1}{V}\int d{\bm x} \int d{\bm z}'d{\bm z}'' u_d^\alpha({\bm x}; {\bm z}') u_d^\alpha({\bm x}+{\bm R}; {\bm z}'') \nonumber \\
& \qquad\qquad\qquad \times \Bigg[ \left(\frac{n}{4\pi}\right)^2 + G_K( {\bm z}', {\bm z}'', t)\Bigg].\end{aligned}$$ The integral with the constant term vanishes, demonstrating that $G_K$ fully determines the spatial correlation function.
Time evolution of the Klimontovich correlation function can readily be derived from Eqs. and , yielding $$\begin{aligned}
& \partial_t G_K({\bm z}_1,{\bm z}_2, t) + {\mathcal L}_{12} [G_K] + {\mathcal L}_{21} [G_K] \nonumber \\
& \qquad\qquad =
2 \lambda \frac{n}{4\pi} \delta({\bm x}_1 - {\bm x}_2 ) \left[ \delta({\bm p}_1 - {\bm p}_2 ) - \frac{1}{4\pi} \right],
\label{EqGK}\end{aligned}$$ where $ {\mathcal L}_{ij}$ is defined in Eq., and we used $F_1 = 1/(4\pi V)$ in the homogeneous and isotropic state. To solve Eq., we introduce an auxiliary field $h({\bm z}_1, t)$, that satisfies the following equation $$\begin{aligned}
\partial_t h({\bm z}_1, t) + {\mathcal L}_{11} [h] = \xi({\bm z}_1, t),
\label{Eqh}\end{aligned}$$ where $\xi$ is a noise term with the following properties $$\begin{aligned}
& \langle \xi({\bm z}_1, t) \rangle = 0,
\label{noiseaverage}\\
& \langle \xi({\bm z}_1, t) \xi({\bm z}_2, t')\rangle = 2 \lambda \frac{n}{4\pi} \delta(t-t') \delta({\bm x}_1 - {\bm x}_2 )
\nonumber \\
&\qquad\qquad \times \left[ \delta({\bm p}_1 - {\bm p}_2 ) - \frac{1}{4\pi} \right].
\label{noisevariance}\end{aligned}$$ Here, the angular brackets denote the average over the realisations of the noise $\xi$, and should not be confused with the ensemble averages that we denoted by bars in the equations above. Eq. allows us to factorise the Klimontovich correlation function as $$\begin{aligned}
G_K({\bm z}_1,{\bm z}_2, t) = \langle h({\bm z}_1, t) h({\bm z}_2, t)\rangle,\end{aligned}$$ which replaces the deterministic Eq. by a significantly simpler stochastic Eq. with a fictitious noise $\xi$ with properly chosen spectral properties. Remarkably, the non-equal time correlations of the phase space density can be expressed through the same auxiliary field $$\begin{aligned}
\overline{\delta \varphi({\bm z}',t') \delta \varphi({\bm z}'',t'')} = \langle h({\bm z}', t') h({\bm z}'', t'')\rangle,\end{aligned}$$ as implied by a seminal work of Klimontovich and Silin [@Silin1962]. This, finally, leads to a direct relationship between the field $h$, which encodes the statistical properties of fluctuations in the suspension, and the combined correlation function $$\begin{aligned}
&C(R,T) = \lim_{t\rightarrow\infty} \frac{1}{V}\int d{\bm x} \int d{\bm z}'d{\bm z}'' \nonumber \\
& \times u_d^\alpha({\bm x}; {\bm z}') u_d^\alpha({\bm x}+{\bm R}; {\bm z}'') \langle h({\bm z}', t) h({\bm z}'', t+T) \rangle.
\label{CRTtmp}\end{aligned}$$
Dynamics of the auxiliary field $h$ {#subsection:h}
-----------------------------------
Here, we explicitly find the solution to Eq. together with Eqs. and . Since Eq. is linear in $h$, we introduce the Fourier $$\begin{aligned}
h({\bm z}, t) = \frac{1}{(2\pi)^3} \int d{\bm k} e^{i {\bm k}\cdot{\bm x}} \hat{h}({\bm k}, {\bm p}, t),
\label{FT}\end{aligned}$$ and the Laplace transforms $$\begin{aligned}
\hat{h}({\bm k}, {\bm p}, s) = \int_0^\infty dt e^{-s t} \hat{h}({\bm k}, {\bm p}, t).\end{aligned}$$ We will also require the Fourier transform of the regularised dipolar field, Eq., which is given by $$\begin{aligned}
& u_d^\nu({\bm x}; {\bm z}') = \frac{-i \kappa}{(2\pi)^3} \int d{\bm k} \, e^{i {\bm k}\cdot ({\bm x}-{\bm x}')} \nonumber \\
& \qquad\qquad\qquad \times \frac{A(k \epsilon)}{k} (\hat{\bm k} \cdot {\bm p}') \left( \delta^{\nu\delta} - \hat{k}^\nu \hat{k}^\delta \right) p'^{\delta},
\label{udfourier}\end{aligned}$$ where $\hat{\bm k}={\bm k}/k$, and $k=|{\bm k}|$. The function $A$, defined as $$\begin{aligned}
A(x)=\frac{1}{2} x^2 K_2(x),\end{aligned}$$ with $K_2(x)$ being the modified Bessel function of the second kind, is close to unity for $x<1$, and quickly approaches zero for $x>1$. It will serve as a regularisation of the integrals over $k$, suppressing contributions from lengthscales smaller than the size of individual microswimmers.
Performing the Fourier and Laplace transforms of Eq., we obtain after re-arranging $$\begin{aligned}
&\hat{h}({\bm k}, {\bm p}, s) = \frac{1}{\sigma({\bm k},{\bm p}, s)}
\Bigg[\hat{h}_0({\bm k}, {\bm p}) + \hat{\xi}({\bm k}, {\bm p},s) \nonumber \\
& \qquad + \frac{\lambda}{4\pi} I^{(0)}({\bm k},s)
+ \frac{15 \lambda}{4\pi} \Delta A(k\epsilon) \bigg\{ (\hat{\bm k}\cdot {\bm p}) I^{(1)}({\bm k},{\bm p}, s) \nonumber \\
& \qquad\qquad\qquad\qquad\qquad - (\hat{\bm k}\cdot {\bm p})^2 I^{(2)}({\bm k}, s)
\bigg\}
\Bigg].
\label{hfullsolution}\end{aligned}$$ Here, $\hat{\xi}({\bm k}, {\bm p},s)$ is the Fourier-Laplace transform of the noise, $\sigma({\bm k},{\bm p}, s) = s + \lambda + i v_s ({\bm k}\cdot{\bm p})$, and we defined $$\begin{aligned}
I^{(0)}({\bm k},s) &= \int d{\bm p } \,\hat{h}({\bm k}, {\bm p}, s), \label{I0} \\
I^{(1)}({\bm k},{\bm p}, s) &= \int d{\bm p}' (\hat{\bm k}\cdot {\bm p}')({\bm p}\cdot {\bm p}') \hat{h}({\bm k}, {\bm p}', s), \label{I1} \\
I^{(2)}({\bm k},s) &= \int d{\bm p }\, (\hat{\bm k}\cdot {\bm p})^2 \hat{h}({\bm k}, {\bm p}, s). \label{I2}\end{aligned}$$ In Eq., $\hat{h}({\bm k}, {\bm p}, t=0) = \hat{h}_0({\bm k}, {\bm p})$ denotes some arbitrary initial condition; below we demonstrate that the long-time statistical properties of the suspension are insensitive to $\hat{h}_0({\bm k}, {\bm p})$. In Eq., we have also introduced an important dimensionless parameter $\Delta = n/n_{crit}$, where $n_{crit} = 5\lambda/(B\kappa)$ is the mean-field onset of collective motion in pusher suspensions, $\kappa>0$. For pushers, $\Delta$ measures the dimensionless distance from the onset, with $\Delta=1$ corresponding to the instability. Eq. is a linear integral equation for $\hat{h}({\bm k}, {\bm p}, s)$ and its solution is straightforward. Substituting Eq. into Eqs.-, gives $$\begin{aligned}
&I^{(0)}({\bm k},s) \nonumber \\
&\qquad\qquad= \frac{1}{1-\frac{\lambda}{4\pi}f_0} \int d{\bm p} \frac{\hat{h}_0({\bm k}, {\bm p}) + \hat{\xi}({\bm k}, {\bm p},s)}{\sigma({\bm k},{\bm p}, s)},
\label{I0sol}\end{aligned}$$ $$\begin{aligned}
&I^{(2)}({\bm k},s) = \frac{\lambda}{4\pi} f_1 I^{(0)}({\bm k},s) \nonumber \\
& \qquad\qquad + \int d{\bm p} (\hat{\bm k}\cdot {\bm p})^2\frac{\hat{h}_0({\bm k}, {\bm p}) + \hat{\xi}({\bm k}, {\bm p},s)}{\sigma({\bm k},{\bm p}, s)},
\label{I1sol}\end{aligned}$$ $$\begin{aligned}
&I^{(1)}({\bm k},{\bm p}, s) = \frac{1}{1+\frac{15 \lambda}{8\pi} \Delta A(k\epsilon) (f_2 - f_1)} \Bigg[ \nonumber \\
& \quad \int d{\bm p}' (\hat{\bm k}\cdot {\bm p}')({\bm p}\cdot {\bm p}') \frac{\hat{h}_0({\bm k}, {\bm p}') + \hat{\xi}({\bm k}, {\bm p}',s)}{\sigma({\bm k},{\bm p}', s)} \nonumber \\
& \quad + (\hat{\bm k}\cdot {\bm p}) \bigg\{ \frac{\lambda}{4\pi} f_1 I^{(0)}({\bm k},s) \nonumber \\
& \qquad\qquad\qquad + \frac{15 \lambda}{8\pi} \Delta A(k\epsilon) (f_2 - f_1) I^{(2)}({\bm k},s)
\bigg\} \Bigg],
\label{I2sol}\end{aligned}$$ where $$\begin{aligned}
f_n = 2\pi \int_{-1}^{1} dx \frac{x^{2n}}{s+\lambda + i v_s k x}.\end{aligned}$$
Having found the explicit expression for $\hat{h}({\bm k}, {\bm p},s)$, we proceed to calculate the combined correlation function, Eq.. Below, we show that only a small number of terms from Eqs. and - contribute to $C(R,T)$.
$C(R,T)$ in terms of $\hat{h}({\bm k}, {\bm p},s)$ {#subsection:hhat}
--------------------------------------------------
In what follows, it will be convenient to re-write $C(R,T)$ in terms of the Fourier and Laplace transforms of all quantities. Substituting Eq. into Eq., and using the Fourier representation of the regularised dipolar field, Eq., we obtain $$\begin{aligned}
&C(R,T) = \lim_{t\rightarrow\infty} {\mathcal L}^{-1}_{s_1,t} {\mathcal L}^{-1}_{s_2,t+T} \frac{\kappa^2}{(2\pi)^3 V} \nonumber \\
&\qquad \times \int d{\bm k} e^{-i {\bm k}\cdot{\bm R}} \frac{A^2(k\epsilon)}{k^2}
\int d{\bm p}_1 d{\bm p}_2
(\hat{\bm k} \cdot {\bm p}_1) (\hat{\bm k} \cdot {\bm p}_2) \nonumber \\
& \qquad\qquad\qquad \times \big( \delta^{\alpha\beta} - \hat{k}^\alpha \hat{k}^\beta \big) p_1^{\beta}
\big( \delta^{\alpha\gamma} - \hat{k}^\alpha \hat{k}^\gamma \big) p_2^{\gamma} \nonumber \\
& \qquad\qquad\qquad \times \langle \hat{h}({\bm k}, {\bm p}_1,s_1)\hat{h}(-{\bm k}, {\bm p}_2,s_2) \rangle_{\hat{\xi}},
\label{CRTlaplace}\end{aligned}$$ where ${\mathcal L}^{-1}_{s,t}$ formally denotes the inverse Laplace transform from $s$ to $t$, given by the Bromwich integral [@Doetsch1974]. The angular brackets $\langle\dots\rangle_{\hat{\xi}}$ denote the average with the Fourier-Laplace components of the noise $\xi$, with the following spectral properties $$\begin{aligned}
& \langle \hat{\xi}({\bm k}, {\bm p}, s) \rangle_{\hat{\xi}} = 0,
\label{noiseaverageFL}\\
& \langle \hat{\xi}({\bm k}, {\bm p}_1, s_1) \hat{\xi}(-{\bm k}, {\bm p}_2, s_2)\rangle_{\hat{\xi}} = 2 \lambda V \frac{n}{4\pi}
\nonumber \\
&\qquad\qquad\qquad\qquad \times \frac{1}{s_1+s_2} \left[ \delta({\bm p}_1 - {\bm p}_2 ) - \frac{1}{4\pi} \right],
\label{noisevarianceFL}\end{aligned}$$ obtained by applying the Fourier-Laplace transform to Eqs. and . While the average in Eq. can readily be formed using the solution for $\hat h$ found in Section \[subsection:h\], the result is very cumbersome. Before proceeding, we make two observations that greatly reduce the number of terms contributing to Eq..
First, we observe that $$\begin{aligned}
\int d{\bm p} \, \big( \delta^{\alpha\beta} - \hat{k}^\alpha \hat{k}^\beta \big) p^{\beta} f(\hat{\bm k} \cdot {\bm p}) = 0,\end{aligned}$$ where $f$ is an arbitrary function of $\hat{\bm k} \cdot {\bm p}$. This statement is readily demonstrated by representing $\bm p$ in spherical coordinates with $\hat{\bm k}$ selected along the $z$-axis, and performing the angular integrals component-wise. This result has profound implications for the average $\langle \hat{h}({\bm k}, {\bm p}_1,s_1)\hat{h}(-{\bm k}, {\bm p}_2,s_2) \rangle_{\hat{\xi}}$ in Eq.. Every term in $\hat{h}({\bm k}, {\bm p}_1,s_1)$, Eq., that only depends on ${\bm p}_1$ through its dependence on $(\hat{\bm k} \cdot {\bm p}_1)$ does not contribute to $C(R,T)$, as its integral over ${\bm p}_1$ with the corresponding dipolar field in Eq. vanishes. The same applies to $\hat{h}(-{\bm k}, {\bm p}_2,s_2)$.
The second observation is related to the initial condition. All terms that involve $\hat{h}_0({\bm k}, {\bm p})$ only depend on the Laplace frequency $s$ through $1/\sigma({\bm k},{\bm p}, s)$, and their inverse Laplace transform can be readily performed before any other integration. Since the inverse Laplace transform of $1/(s+a)$ is $e^{-a t}$, where $a$ is a complex number, the dominant long-time behaviour of such terms is given by $e^{-\lambda t}$, where we ignored the subdominant oscillatory dependencies. In Eq., we are interested in the $t\rightarrow$ limit, and these terms also do not contribute to $C(R,T)$.
With these observations in mind, Eq. can be significantly simplified to read $$\begin{aligned}
&\hat{h}({\bm k}, {\bm p}, s) \cong \frac{\hat{\xi}({\bm k}, {\bm p},s)}{\sigma({\bm k},{\bm p}, s)} \nonumber \\
&\qquad\qquad + \frac{ (\hat{\bm k}\cdot {\bm p}) }{\sigma({\bm k},{\bm p}, s)}
\frac{\frac{15 \lambda}{4\pi} \Delta A(k\epsilon) }{1+\frac{15 \lambda}{8\pi} \Delta A(k\epsilon) (f_2 - f_1)} \nonumber \\
& \qquad\qquad\qquad \times \int d{\bm p}' (\hat{\bm k}\cdot {\bm p}')({\bm p}\cdot {\bm p}') \frac{\hat{\xi}({\bm k}, {\bm p}',s)}{\sigma({\bm k},{\bm p}', s)},
\label{hneeded}\end{aligned}$$ where $\cong$ signifies that we only kept the terms that contribute to $C(R,T)$. Now, the average $\langle \hat{h}({\bm k}, {\bm p}_1,s_1)\hat{h}(-{\bm k}, {\bm p}_2,s_2) \rangle_{\hat{\xi}}$ assumes a tractable form that can be used in Eq.. Separating the terms independent of $\Delta$, we obtain $C(R,T) = C_0(R,T) + C_1(R,T)$. Here, $$\begin{aligned}
&C_0(R,T) = \frac{\lambda n \kappa^2}{16\pi^4 } \lim_{t\rightarrow\infty} {\mathcal L}^{-1}_{s_1,t} {\mathcal L}^{-1}_{s_2,t+T} \int d{\bm k} e^{-i {\bm k}\cdot{\bm R}} \frac{A^2(k\epsilon)}{k^2} \nonumber \\
&\quad \times \int d{\bm p}
(\hat{\bm k} \cdot {\bm p})^2 \left[1- (\hat{\bm k} \cdot {\bm p})^2\right] \frac{1}{s_1+s_2}\nonumber \\
& \qquad \times
\frac{1}{\lambda + s_1 + i v_s k (\hat{\bm k}\cdot {\bm p})}
\frac{1}{\lambda + s_2 - i v_s k (\hat{\bm k}\cdot {\bm p})},
\label{C0Laplace}\end{aligned}$$ represents correlations in the fluid created by non-interacting swimmers. The double inverse Laplace transform in the equation above can be performed using the method outlined in Appendix \[appendix:Cesare\]. It yields $$\begin{aligned}
& \lim_{t\rightarrow\infty} {\mathcal L}^{-1}_{s_1,t} {\mathcal L}^{-1}_{s_2,t+T} \frac{1}{s_1+s_2} \frac{1}{\lambda + s_1 + i v_s k (\hat{\bm k}\cdot {\bm p})} \nonumber \\
& \qquad\times \frac{1}{\lambda + s_2 - i v_s k (\hat{\bm k}\cdot {\bm p})} = \frac{e^{-\lambda T + i v_s k T (\hat{\bm k} \cdot {\bm p})}}{2\lambda}.
\label{dilt_example}\end{aligned}$$ Performing the angular integration, we finally obtain $$\begin{aligned}
& C_0(R,T) = \frac{n \kappa^2 e^{-\lambda T}}{\pi^2} \int_{0}^\infty dk \frac{\sin{k R}}{k R} A^2(k\epsilon) \nonumber \\
&\quad \times
\frac{ y(12-y^2) \cos{y} - (12-5y^2)\sin{y}
}{y^5}
\Bigg\vert_{y=v_s k T}.
\label{C0final}\end{aligned}$$
All other terms in Eq. correspond to additional correlations generated by the hydrodynamic interactions among the swimmers, and, as such, they are dependent on the dimensionless microswimmer density $\Delta$. Performing the angular integration over ${\bm p}_1$ and ${\bm p}_2$, gives $$\begin{aligned}
&C_1(R,T) \nonumber \\
& = \frac{2 \lambda n \kappa^2 }{15 \pi^2} \lim_{t\rightarrow\infty} {\mathcal L}^{-1}_{s_1,t} {\mathcal L}^{-1}_{s_2,t+T} \int_{0}^\infty dk \frac{\sin{k R}}{k R} A^2(k\epsilon) \nonumber \\
&\times \frac{1}{\lambda + s_1}\frac{1}{\lambda + s_2}\frac{1}{s_1 + s_2} \frac{z_1 \psi(z_1) + z_2 \psi(z_2)}{z_1+z_2}
\nonumber \\
&\qquad \times\Bigg[ \frac{z_1 \psi(z_1)}{\omega - z_1 \psi(z_1)} + \frac{z_2 \psi(z_2)}{\omega - z_2 \psi(z_2)} \nonumber \\
&\qquad\qquad
+ \frac{z_1 z_2 \psi(z_1)\psi(z_2)}{\left(\omega - z_1 \psi(z_1)\right)\left(\omega - z_2 \psi(z_2)\right)} \Bigg].
\label{C1laplace}\end{aligned}$$ Here, we introduced $\omega = v_s k/(\lambda \Delta A(k\epsilon))$, and the function $\psi(z)$, defined as $$\begin{aligned}
\psi(z) = \frac{5}{2}\frac{3z+2z^3-3(1+z^2)\arctan{z}}{z^5},
\label{psi}\end{aligned}$$ which is related to $f_2-f_1$ used in the previous Section. The variable $z_i=v_s k/(\lambda + s_i)$ allows us to write Eq. in a compact form but hides its complex dependence on the Laplace frequencies $s_1$ and $s_2$. Its inverse Laplace transform is discussed below.
Approximate double inverse Laplace transform {#subsection:approximateDILT}
--------------------------------------------
The integrand of Eq. is not a rational function of $s_1$ and $s_2$, and we were unable to calculate its double inverse Laplace transform exactly. Instead, here we develop a rational approximation to $\psi(z)$ that will allow us to find $C_1(R,T)$ analytically.
First, we observe that if the poles of an analytic function are known, its large-$t$ behaviour is determined by the pole with the smallest negative real part [@Doetsch1974]. Therefore, the presence of the pole at $-\lambda$ in Eq. makes all poles with real parts smaller than $-\lambda$ irrelevant in the large-$t$ limit. This reflects the fact that individual tumbling events are always a source of de-correlation between microswimmers.
Next, we introduce $L=v_s/(\lambda \epsilon)$, which compares the typical runlength of a swimmer to its size. In this work we consider $L=0-25$, ranging from non-swimming (shaker) particles to wild-type *E.coli* bacteria [@Berg1993]. Contributions to the integrand in Eq. with $k \epsilon>1$ are strongly suppressed by the regularising factor $A(k\epsilon)$, and therefore, when approximating $\psi(z)$, the relevant domain is $-\lambda<Re(s)<0$, with $v_s k/\lambda$ not exceeding $L$.
In Appendix \[appendix:psia\] we show that a surprisingly good approximation to $\psi(z)$ on this domain is given by $$\begin{aligned}
\psi_a(z) = \frac{7}{7+3 z^2}.
\label{psia}\end{aligned}$$ The simple structure of this expression allows us to deduce the pole structure of the integrand in Eq.. Indeed, with $\psi(z)$ replaced by $\psi_a(z)$, and factorising $$\begin{aligned}
\frac{1}{\omega - z \psi(z)} = \frac{7+3z^2}{3\omega\left(z-z_{+}\right)\left(z-z_{-}\right)},\end{aligned}$$ where $$\begin{aligned}
z_{\pm} =\frac{7}{6\omega} \left[ 1\pm\sqrt{1-\frac{12}{7}\omega^2} \right],
\label{z012}\end{aligned}$$ the denominators in Eq. can now be written as products of linear polynomials in $s_1$ and $s_2$. It is now straightforward to perform the inverse Laplace transform of this expression using the method outlined in Appendix \[appendix:Cesare\]. Taking the limit of $t\rightarrow\infty$, finally gives $$\begin{aligned}
&C_1(\rho,\tau) \nonumber \\
& = e^{-\tau} \frac{n \kappa^2 }{15 \pi^2 \epsilon} \int_{0}^\infty d\xi \frac{\sin{\xi \rho}}{\xi \rho} A^2(\xi)
\Bigg[
-\cos{\left(\sqrt{\frac{3}{7}}L\xi\tau\right)} \nonumber \\
&+ \frac{e^{\frac{1}{2}A(\xi) \Delta \tau}}{1-A(\xi)\Delta + \frac{3}{7}L^2\xi^2} \Bigg\{ \nonumber \\
& \frac{2-A(\xi)\Delta + \frac{6}{7}L^2\xi^2}{2-A(\xi)\Delta}\cosh{\left(\frac{1}{2}A(\xi) \Delta \tau \sqrt{1-\frac{12 L^2\xi^2}{7 A^2(\xi)\Delta^2}}\right)} \nonumber \\
&\qquad +\frac{\sinh{\left(\frac{1}{2}A(\xi) \Delta \tau \sqrt{1-\frac{12 L^2\xi^2}{7 A^2\left(\xi\right)\Delta^2}}\right)}}{\sqrt{1-\frac{12 L^2\xi^2}{7 A^2\left(\xi\right)\Delta^2}}}
\Bigg\}
\Bigg],
\label{C1final}\end{aligned}$$ where we changed the integration variable to $\xi = k\epsilon$, and introduced the dimensionless parameters $\rho = R/\epsilon$ and $\tau = \lambda T$. In Appendix \[appendix:dilt\], we verify that Eq. provides a good approximation to the long-time behaviour of Eq..
Results {#section:results}
=======
For the benefit of the readers who have skipped Section \[section:theory\], we repeat our main result, which comprises an explicit expression for the combined correlation function $C(R,T)$, defined in Eq.. It describes the steady-state correlations between the fluid velocity at two points in space separated by a distance $R$, and two instances in time separated by a time-interval $T$. Our theory includes full hydrodynamic interactions between microswimmers and is valid at any density up to the onset of collective motion. The result consists of the non-interacting part, $$\begin{aligned}
& C_0(\rho,\tau) = \frac{n \kappa^2 e^{-\tau}}{\pi^2\epsilon} \int_{0}^\infty d\xi \frac{\sin{\xi \rho}}{\xi \rho} A^2(\xi) \nonumber \\
&\quad \times
\frac{ y(12-y^2) \cos{y} - (12-5y^2)\sin{y}
}{y^5}
\Bigg\vert_{y=L\xi\tau},
\label{ResultsC0}\end{aligned}$$ and the interacting correlation function $C_1(\rho,\tau)$, given in Eq.. Here, $\rho = R/\epsilon$, where $\epsilon$ is a lengthscale comparable to the microswimmer size, and $\tau = \lambda T$, where $\lambda$ is the tumbling rate. The parameter $L=v_s/(\lambda \epsilon)$ compares the typical distance covered by a swimming microorganism between two tumble events to its size. Finally, $\Delta = n/n_{crit}$ is the dimensionless number density of the particles, where $n_{crit} = 5\lambda/(B\kappa)$ is the onset of collective motion for pusher-like microswimmers [@Subramanian2009; @Hohenegger2010; @Stenhammar2017]; the parameter $B$, is defined after Eq.. Our theory is valid for $\Delta<1$.
The full expression, $C(\rho,\tau)=C_0(\rho,\tau)+C_1(\rho,\tau)$, given as a definite integral, constitutes the main technical result of our study. We now explicitly work out its predictions for the spatial and temporal correlation functions, and other experimentally accessible observables. When discussing their physical meaning, we are going to vary the dimensionless persistence length $L$, while keeping all the other parameters of the microswimmers fixed. We note that in reality the dipolar strength and shape of a microorganism uniquely determine its swimming speed, and hence $L$. We, however, see varying $L$ as a tool to disentangle the effects of self-propulsion (ability to change one’s position in space) from the strength of the hydrodynamic disturbances it causes. In particular, we will consider two limiting cases: shakers ($L=0$) and fast swimmers ($L\rightarrow\infty$). The former case corresponds to microswimmers that exert dipolar forces on the fluid but do not self-propel, and only change their positions due to being advected by the velocity fields created by other microswimmers [@Stenhammar2017]. The latter case, while obviously non-physical, is a useful tool to assess the effect of fast swimming on various quantities of interest. Finally, we note that the terms representing hydrodynamic interactions in Eq. are proportional to the swimmer’s nonsphericity $B$ that enters Jeffrey equation, Eq.. The limit of non-interacting microswimmers therefore corresponds to setting $B$ to zero, which, in turn, can be achieved by setting $\Delta=0$, while keeping $n$ finite.
Velocity variance
-----------------
![The fluid velocity variance $\langle U^2 \rangle$ normalised by its non-interacting value $\langle U^2 \rangle_0$ for various values of $L$. The dotted line represents the non-interacting case $\langle U^2 \rangle=\langle U^2 \rangle_0$. Note that the $L\rightarrow\infty$ line turns sharply upwards and diverges in the vicinity of $\Delta=1$ in a way that cannot be resolved on the scale of this graph.[]{data-label="Fig_Variance"}](Variance.pdf){width="0.8\linewidth"}
Our first quantity of interest is the fluid velocity variance, $\langle U^2 \rangle \equiv C(\rho=0,\tau=0)$. In the absence of thermal noise, re-arrangements of the microswimmer positions and orientations is the sole source of fluid velocity fluctuations. For this reason, it was used in previous studies as an order parameter to identify the onset of collective motion [@Stenhammar2017; @Bardfalvy2019]. Summing up Eqs. and , and setting $\rho=0$ and $\tau=0$, we obtain $$\begin{aligned}
& \langle U^2 \rangle = \frac{\kappa^2 n}{15 \pi^2 \epsilon} \int_{0}^\infty d\xi A^2(\xi) \nonumber \\
& \qquad\qquad \times \frac{2-A(\xi)\Delta+\frac{6}{7}L^2 \xi^2}{\left(2-A(\xi)\Delta\right)\left(1-A(\xi)\Delta+\frac{3}{7}L^2 \xi^2\right)}.
\label{resultvarience}\end{aligned}$$ We evaluate this integral numerically and plot the fluid velocity variance normalised by its value in the non-interacting case, $\langle U^2 \rangle(\Delta=0)\equiv \langle U^2 \rangle_0$, given by [@Bardfalvy2019] $$\begin{aligned}
\langle U^2 \rangle_0 = \frac{\kappa^2 n}{15 \pi^2 \epsilon} \int_{0}^\infty d\xi A^2(\xi) = \frac{21\kappa^2 n}{2048\epsilon}.\end{aligned}$$ Note that $\langle U^2 \rangle_0$ corresponds to a superposition of uncorrelated fluctuations in the fluid velocity, which, by virtue of the central limit theorem, is proportional to $n$. Any deviations of $\langle U^2 \rangle$ from that value signify the presence of correlations.
As can be seen in Fig.\[Fig\_Variance\], the fluid velocity fluctuations exhibit significant correlations at any density of the microswimmers, as was recognised previously [@Stenhammar2017]. Starting from its non-interacting value at $\Delta=0$, the variance increases with $\Delta$, until it diverges at the onset of collective motion. The strongest correlations are exhibited by suspensions of shakers, while swimming acts to reduce correlations. For large but finite values of $L$, the variance increases mildly from its non-interacting value, until it rises sharply in a small vicinity of $\Delta=1$, with the size of this region shrinking with $L$. Interestingly, the rise of $\langle U^2 \rangle_0$ for $\Delta<1$ remains finite even in the $L\rightarrow\infty$ limit. In other words, while swimming clearly reduces correlations, it does not remove them entirely, and the suspension is never described by the mean-field theory.
To determine the scaling of the fluid velocity variance as $\Delta\rightarrow 1$, we observe that in that limit the integrand in Eq. is dominated by small values of $\xi$, where $A(\xi) \approx 1 - \xi^2/4$. Using this approximation in Eq. and replacing the upper integration limit by unity, we obtain $$\begin{aligned}
& \langle U^2 \rangle \sim \frac{\kappa^2 n}{15 \pi \epsilon} \frac{1}{\sqrt{1+\frac{12}{7}L^2}\sqrt{1-\Delta}}, \quad \Delta \rightarrow 1.
\label{varianceasymptotic}\end{aligned}$$ Therefore, our theory predicts that the fluid velocity variance diverges as $(1-\Delta)^{-1/2}$ in the vicinity of the transition to collective motion, irrespectively of $L$.
Spatial correlations
--------------------
{width="1.0\linewidth"}
Our next quantity of interest is the equal-time spatial correlation function, $C(\rho, T=0)$, given by $$\begin{aligned}
&C(\rho) = \frac{\kappa^2 n}{15 \pi^2 \epsilon} \int_{0}^\infty d\xi \frac{\sin{\xi \rho}}{\xi \rho} A^2(\xi) \nonumber \\
& \qquad\qquad \times \frac{2-A(\xi)\Delta+\frac{6}{7}L^2 \xi^2}{\left(2-A(\xi)\Delta\right)\left(1-A(\xi)\Delta+\frac{3}{7}L^2 \xi^2\right)}.
\label{spatialcorrelations}\end{aligned}$$ While this integral cannot be evaluated analytically, a good approximation can be obtained by setting $A(\xi)=1$ in the integrand, yielding $$\begin{aligned}
&C(\rho) \approx \frac{\kappa^2 n}{30 \pi \epsilon \left(1-\Delta\right)\rho} \nonumber \\
&\qquad\qquad \times \left[ 1 - \frac{\Delta}{2-\Delta} \exp{\left\{-\sqrt{\frac{7}{3}}\sqrt{1-\Delta}\frac{\rho}{L} \right\}}\right].
\label{spatialcorrelationsanalytic}\end{aligned}$$ For $\Delta=0$, this equation reproduces the result obtained previously for non-interacting swimmers [@Zaid2011; @Underhill2011; @Belan2019; @Bardfalvy2019].
In Fig.\[Fig\_spatial\] we evaluate Eq. numerically and compare it against the analytic approximation, Eq.; $\kappa^2 n_{crit}/(15 \pi^2 \epsilon)$ is chosen as the normalisation factor. For all values of $L$ and $\Delta$, the approximation works well for all but small spatial separations $\rho$, where the spatial correlation function is, essentially, equal to the fluid velocity variance. As with the fluid velocity variance, the strongest correlations are exhibited by suspensions of shakers, $L=0$. In this case, the spatial correlation function changes very slowly at short distances, and decays as $\rho^{-1}$ at large distances. Close to the onset of collective motion, the typical scale $\rho_0$ at which the crossover occurs can be estimated from Eqs. and , by requiring that $C(\rho_0) = \langle U^2 \rangle$. For $L=0$, this yields $\rho_0 \sim (1-\Delta)^{-1/2}$. This is readily verified by the data in Fig.\[Fig\_spatial\]A: As the system approaches the onset of collective motion, the overall strength of the correlations grows, with the region of strong correlations extending to progressively larger scales.
![The spatial correlation function $C(\rho)$ as a function of the distance $\rho$ for $\Delta=0.9$ and various values of $L$. At sufficiently large distances, $C(\rho)$ recovers the shaker behaviour, while at small distances correlations are suppressed by swimming. Note that the $L\rightarrow\infty$ line, serving as the limit beyond which correlations cannot be suppressed, joins the shaker line at $\rho\rightarrow\infty$.[]{data-label="Fig_spatial_delta09"}](SpatialDelta09.pdf){width="0.8\linewidth"}
The effect of swimming on the behaviour of $C(\rho)$ is demonstrated in Figs.\[Fig\_spatial\]B-\[Fig\_spatial\]C. As $L$ increases, the strongly correlated core at moderate separations shrinks, indicating that the steady growth of orientational correlations is reduced by the mixing introduced by swimming. The overall strength of correlations inside the core also decreases with $L$, reflecting the reduction of the fluid velocity variance by swimming. At large distances, $C(\rho)$ recovers the behaviour seen in shakers, with the crossover distance given by $\rho_1 \sim L (1-\Delta)^{-1/2}$, as can be deduced from the exponential in Eq.. This behaviour is further demonstrated in Fig.\[Fig\_spatial\_delta09\], where we plot $C(\rho)$ for $\Delta=0.9$ and various values of $L$. In the limit of fast swimming, $L\rightarrow\infty$, the correlation function deviates modestly from the non-interacting case for almost all values of $\Delta$, exhibiting a quick rise and the divergence associated with the onset of collective motion only in a very small vicinity of $\Delta=1$.
The data in Fig.\[Fig\_spatial\] and Eq. demonstrate that $C(\rho)$ exhibits an algebraic decay for large distances, and a true correlation length can thus not be defined. A phenomenological correlation length $\xi_{corr}$ can nevertheless be defined as a distance over which $C(\rho)$ decreases by certain amount, as has been employed in [@Gachelin2014; @Bardfalvy2019]. Setting $C(\xi_{corr}) = \alpha \langle U^2 \rangle$, with $\alpha<1$, we obtain $$\begin{aligned}
\xi_{corr}\sim (1-\Delta)^{-1/2}, \quad \Delta\rightarrow1, \end{aligned}$$ similar to any other typical distance discussed above.
Fluid velocity spectrum
-----------------------
{width="1\linewidth"}
Next, we discuss the fluid velocity energy spectrum $E(k)$ that is closely related to the spatial correlation function $C(\rho)$. Defined as $$\begin{aligned}
E(k) = 4\pi k^2 \overline{\hat{U}^{\alpha}\left( {\bm k} \right) \hat{U}^{\alpha}\left( -{\bm k} \right)},
\label{eq:spectrum}\end{aligned}$$ this quantity is often used in turbulence research to study the cascade of the kinetic energy [@Townsend1980]. Although the kinetic energy is not a useful concept for Stokesian flows, $E(k)$ provides an insight into the relative strength of fluid motion at various scales. The energy spectrum is proportional to the Fourier transform of $C(\rho)$, and, up to a prefactor is given by the integrand of Eq. $$\begin{aligned}
&E(\xi) = \frac{8\pi}{15} \kappa^2 n A^2(\xi) \nonumber \\
&\qquad\qquad \times \frac{2-A(\xi)\Delta+\frac{6}{7}L^2 \xi^2}{\left(2-A(\xi)\Delta\right)\left(1-A(\xi)\Delta+\frac{3}{7}L^2 \xi^2\right)},
\label{energyspectrum}\end{aligned}$$ where, again, $\xi=k\epsilon$. This expression is plotted in Fig.\[Fig\_Spectra\] for various values of $\Delta$ and $L$.
First, we observe that $E(\xi)$ has significant energy content at all large scales, $\xi<1$, that quickly decays to zero at the organism-size scales, $\xi\sim1$, due to the regularising factor $A(\xi)$. This is not caused by some form of energy cascade, but is due to the nature of the dipolar field created by the microswimmers. Indeed, the dipolar velocity field decays in space as $r^{-2}$, while its Fourier transform scales as $k^{-1}$. Together with the definition of $E(k)$, Eq., this implies that $E(k)\sim k^0$ even for a single microswimmer, i.e. the dipolar field has a constant energy content at every scale.
In the presence of interactions, the energy spectrum of shakers ($L=0$) preserves the overall structure described above, while its absolute value increases with $\Delta$ and, eventually, diverges at $\Delta=1$. For swimmers, the increase in the energy content is mostly confined to large scales, while in the limit of fast swimming (not shown), the rise in the energy content on the approach to the onset of collective motion is confined to the largest scales available ($k\rightarrow0$) and starts to be visible only in a very close vicinity of $\Delta=1$.
Temporal correlations {#section:results:temporal}
---------------------
{width="1.0\linewidth"}
The temporal correlation function $C(\tau)=C_0(\rho=0,\tau)+C_1(\rho=0,\tau)$ is given by Eqs. and . The corresponding expressions do not simplify significantly in the limit $\rho=0$, and we do not repeat them here. In Fig.\[Fig\_Temporal\] we plot $C(\tau)$ normalised by its value at $\tau=0$, which is given by the fluid velocity variance $\langle U^2\rangle$. As with the other quantities discussed above, the temporal correlation function exhibits a progressively slower decay as $\Delta$ approaches the onset of collective motion, eventually diverging at $\Delta=1$. For swimmers, this is offset by a decay of $C(\tau)$ at short times that becomes more pronounced as $L$ increases. For very large swimming speeds, the temporal correlations differ only marginally from the non-interacting case for most values of $\Delta$, eventually exhibiting a rapid increase and divergence in a very small vicinity of $\Delta=1$.
To understand the behaviour of $C(\tau)$ at long times, we analyse its individual contributions. The integral in the non-interacting part, $C_0(T)$, can be explicitly evaluated giving $$\begin{aligned}
& C_0(\tau) = \frac{n \kappa^2 }{\pi \epsilon} \frac{e^{-\tau}}{8\alpha^4 (4+\alpha^2)^2}
\Bigg[ \nonumber \\
& \qquad\qquad\qquad 4(24+8\alpha^2+\alpha^4) \mathbb{E}\left( -\frac{\alpha^2}{4}\right) \nonumber \\
& \qquad\qquad\qquad-(4+\alpha^2)(24+5\alpha^2) \mathbb{K}\left( -\frac{\alpha^2}{4}\right)
\Bigg],
\label{}\end{aligned}$$ where $\alpha=L\tau$, and $\mathbb{K}(x)$ and $\mathbb{E}(x)$ are the complete elliptic integrals of the first and second order, respectively. In the limits of small and large $\alpha$ this equation predicts $$\begin{aligned}
C_0(\tau) \sim \frac{n \kappa^2 }{\pi \epsilon} e^{-\tau} \times
\begin{cases}
\frac{21\pi}{2048},\quad &L\tau\rightarrow0, \\ \\
\frac{1}{4 (L\tau)^3},\quad &L\tau\rightarrow\infty.
\end{cases}\end{aligned}$$ At short times, tumbling is the leading source of decorrelation, while at large $\tau$ the non-interacting temporal correlation function $C_0$ decays as $\tau^{-3}e^{-\tau}$, as reported previously [@Belan2019; @Bardfalvy2019]. The crossover time is set by $\alpha=L\tau=v_s T/\epsilon\sim1$, and corresponds to the time interval needed for a microswimmer to swim its own size.
To understand the large-$\tau$ asymptotic behaviour of $C_1(\tau)$, we observe that $$\begin{aligned}
e^{-\tau} \int_0^\infty d\xi A(\xi)^2 \left\{
\begin{array}{c}
\sin{\gamma \tau \xi} \\
\cos{\gamma \tau \xi}
\end{array} \right\}
\underset{\tau\rightarrow\infty}{\sim}
e^{-\tau} \left\{
\begin{array}{c}
\tau^{-1}\\
\tau^{-5}
\end{array} \right\},\end{aligned}$$ where $\gamma$ is a real constant. This result implies that a trigonometric function in the integrand of Eq. generates a contribution to $C_1(\tau)$ that decays on the same timescale as the non-interacting part $C_0(\tau)$, and does not contribute to the slow decay in Fig.\[Fig\_Temporal\]. In turn, this restricts the integration domain to $\xi \in [0,\xi_*]$, with $$\begin{aligned}
\xi_* = \sqrt{\frac{7}{12}}\frac{\Delta}{L},\end{aligned}$$ which ensures that the arguments of the hyperbolic functions in Eq. are real. Introducing $\zeta = \xi/\xi_*$, $C_1(\tau)$ can be approximated as $$\begin{aligned}
&C_1(\tau)
\underset{\tau\rightarrow\infty}{\sim}
\frac{n \kappa^2 }{15 \pi^2 \epsilon} e^{-\tau\left(1-\frac{1}{2}\Delta\right)} \xi_* \int_{0}^1 d\zeta \frac{1}{1-\Delta + \frac{1}{4}\Delta^2\zeta^2} \nonumber \\
& \qquad \times \Bigg\{
\frac{2-\Delta +\frac{1}{2}\Delta^2\zeta^2}{2-\Delta}\cosh{\left(\frac{1}{2}\Delta \tau \sqrt{1-\zeta^2}\right)} \nonumber \\
&\qquad\qquad\qquad + \frac{1}{\sqrt{1-\zeta^2}} \sinh{\left(\frac{1}{2}\Delta \tau \sqrt{1-\zeta^2}\right)}
\Bigg\},
$$ where we used $A(\xi<\xi_*)\sim 1$ for not-too-small values of $L$. In the limit of large $\tau$, this can be further approximated by $$\begin{aligned}
&C_1(\tau)
\underset{\tau\rightarrow\infty}{\sim}
\frac{n \kappa^2 }{15 \pi^2 \epsilon} e^{-\tau\left(1-\Delta\right)} \frac{\xi_*}{1-\Delta} \int_{0}^1 d\zeta \,
e^{-\frac{1}{4}\tau \Delta \zeta^2} \nonumber \\
& = \frac{n \kappa^2 }{15 \pi^2 \epsilon} \sqrt{\frac{7\pi}{12}\frac{\Delta}{\tau}} \frac{1}{L(1-\Delta)} e^{-\tau\left(1-\Delta\right)} \mathrm{erf}\left( \frac{1}{2}\tau\Delta\right),
\label{C1asymptotics}\end{aligned}$$ where $\mathrm{erf}(x)$ denotes the error function. Predictions of Eq. are plotted in Fig.\[Fig\_Temporal\]B and C as dashed lines. We find a good agreement between its prediction and the true decay of $C(\tau)$ as $\tau\rightarrow\infty$.
To extract the typical timescale $\tau_{corr}$ of the fluid velocity fluctuations on the approach to collective motion, we combine Eqs. and to obtain $$\begin{aligned}
\frac{C(\tau)}{\langle U^2\rangle}
\underset{\tau\rightarrow\infty}{\sim}
\frac{e^{-\tau\left(1-\Delta\right)}}{\sqrt{\tau \left(1-\Delta\right)}}, \quad \Delta \rightarrow 1,\end{aligned}$$ which implies $$\begin{aligned}
\tau_{corr} \sim (1-\Delta)^{-1}.\end{aligned}$$
Enhanced diffusivity
--------------------
As the final observable, we consider here the enhanced diffusivity of a passive tracer particle embedded in a suspension of motile microorganisms. The tracer is assumed to be neutrally buoyant and move due to advection by the velocity fields created by the microswimmers. Brownian diffusion of the tracer is significantly weaker than its enhanced counterpart, and is neglected for simplicity. This problem has been extensively studied both experimentally [@Wu2000; @Kim2004; @Leptos2009; @Kurtuldu2011; @Mino2011; @Mino2013; @Jepson2013; @Patteson2016] and theoretically [@Underhill2008; @Dunkel2010; @Ishikawa2010; @Childress2010; @Childress2011; @Pushkin2013; @Pushkin2013jfm; @Morozov2014; @Kasyap2014; @Thiffeault2015; @Burkholder2017] in the dilute regime, where $\Delta\ll 1$, and for arbitrary densities of shakers [@Stenhammar2017]. Here, we consider the case of arbitrary density $\Delta < 1$ and $L$.
The position of the tracer ${\bm a}(T)$ obeys the following equation of motion $$\begin{aligned}
\dot{\bm a}(T) = {\bm U}({\bm a}(T),T),
\label{tracer_position}\end{aligned}$$ which implies that the tracer is point-like and follows the velocity of the fluid at its position. The long-time behaviour of such a tracer is diffusive [@Wu2000; @Leptos2009; @Thiffeault2015], and the associated diffusion coefficient can be extracted in the usual way $$\begin{aligned}
D = \lim_{T\rightarrow\infty} \frac{1}{6T} \overline{{\bm a}(T)\cdot {\bm a}(T)}.\end{aligned}$$ Here, the bar denotes the average over the history of tumble events, and has the same meaning as in Eq.. Solving formally Eq., ${\bm a}(T)={\bm a}(0) + \int_0^T dt' {\bm U}({\bm a}(t'),t')$, the diffusion coefficient can be written as [@Kubo1966] $$\begin{aligned}
&D = \frac{1}{3} \lim_{t\rightarrow\infty}\int_{0}^\infty dT\, \overline{{\bm U}({\bm a}(t+T),t+T) \cdot {\bm U}({\bm a}(t),t)}.\end{aligned}$$ Here, $t$ is sufficiently large so that any influence of the initial conditions has died away. To proceed, we observe that ${\bm U}({\bm a}(t+T),t+T)$ can be iteratively calculated by substituting the formal solution for ${\bm a}(T)$ into its spatial argument, i.e. $$\begin{aligned}
&{\bm U}({\bm a}(t+T),t+T) = {\bm U}({\bm a}(t),t+T)\nonumber \\
& +\nabla {\bm U}({\bm a}(t),t+T) \cdot \int_{t}^{t+T} dt' {\bm U}({\bm a}(t'),t') + \cdots.
\label{DUapprox}\end{aligned}$$ As was argued by Pushkin and Yeomans [@Pushkin2013], for very dilute suspensions velocity gradients over the typical distance travelled by the tracer particle during the microswimmer runtime are small compared to the velocity of the fluid at any of these positions, and can be neglected. Therefore, we can approximate the diffusion coefficient as $$\begin{aligned}
& D \approx \frac{1}{3} \int_{0}^\infty dT\, \overline{{\bm U}({\bm a}(t),t+T) \cdot {\bm U}({\bm a}(t),t)}\nonumber \\
&\qquad\qquad = \frac{1}{3} \int_{0}^\infty dT\, C(T).
\label{DvsCT}\end{aligned}$$ As we have seen in Section \[section:results:temporal\], as $\Delta$ increases, the correlation time increases from $\lambda^{-1}$ (corresponding to $\tau_{corr}=1$) in the very dilute regime to progressively larger values, eventually diverging as $\Delta\rightarrow 1$, implying that the second, etc. terms in Eq. grow rapidly in this limit. However, the fluid velocity variance, which sets the magnitude of the leading term in Eq. also diverges as $\Delta\rightarrow 1$. Further work is required to assess the validity of the approximation above for all values of $\Delta$. Here, we proceed by using Eq. with the potential caveat that it might not be accurate in the vicinity of $\Delta=1$.
The integral in Eq. can be evaluated explicitly, leading to $D = D_0 + D_1$, where the non-interacting and interacting contributions are given by $$\begin{aligned}
D_0 = \frac{\kappa^2 n}{45 \pi^2 \lambda \epsilon} \int_{0}^\infty d\xi A^2(\xi) \psi\left( \xi L\right),
$$ and $$\begin{aligned}
&D_1 = \frac{\kappa^2 n \Delta}{45 \pi^2 \lambda \epsilon} \int_{0}^\infty d\xi A^3(\xi) \nonumber\\
&\qquad\qquad \times \frac{ 2-A(\xi)\Delta+\frac{6}{7}L^2 \xi^2 }{\left( 1+\frac{3}{7} L^2\xi^2\right)\left( 1-A(\xi)\Delta+\frac{3}{7}L^2 \xi^2\right)^2},\end{aligned}$$ respectively, and $\psi(x)$ is defined in Eq.. At this point, we would like to comment on the shaker limit of these expressions, when they should reduce to the ones obtained by Stenhammar *et al.* [@Stenhammar2017]. Instead, we observe that the expression for $D_1$ reported there erroneously contained $A^2(\xi)$ instead of $A^3(\xi)$ under the integral. We note, however, that since $A(\xi)$ is a regularised representation of a step function, this has almost no bearing on the numerical evaluation of $D_1$ presented in [@Stenhammar2017].
The integral in the non-interacting part $D_0$ cannot be represented in terms of special functions, but its limiting behaviour can readily be obtained. Combining the asymptotic results for $L=0$ and $L\rightarrow\infty$, results in the following approximation $$\begin{aligned}
D_0 \approx \frac{\kappa^2 n}{\lambda \epsilon} \frac{7}{2048 + 336 \pi L}.
\label{D0approx}\end{aligned}$$
To derive an approximate expression for $D_1$, we set $A(\xi)\approx 1$ under the integral sign, to obtain $$\begin{aligned}
D_1 \approx \frac{\kappa^2 n}{90 \pi \lambda \epsilon \sqrt{1+\frac{12}{7}L^2}} \left\{ \frac{2-\Delta}{\left( 1-\Delta\right)^{3/2}} -2\right\}.
\label{D1approx}\end{aligned}$$
{width="0.9\linewidth"}
In Fig.\[Fig\_Diffusivity\] we compare the numerical evaluation of $D_0$ and $D_1$ against Eq. and . We observe that while the uniform approximation Eq. does not work well for small but finite values of $L\sim1$, all other values of $L$ are well-represented by the approximation. The interacting part of the diffusivity is well-approximated by Eq..
Finally, we remark that Eq. predicts that $$\begin{aligned}
D_1 \sim (1-\Delta)^{-3/2}, \quad \Delta\rightarrow1,\end{aligned}$$ even though this prediction should be treated with caution, as discussed above.
Discussion and Conclusion {#section:discussion}
=========================
In this work, we have presented a kinetic theory for dilute suspensions of pusher-like microswimmers interacting via long-ranged dipolar fields. Our theory goes beyond the mean-field assumption and explicitly includes correlations between microswimmers. We have overcome a significant technical difficulty in including particle self-propulsion that has limited our previous work to the case of shaker microswimmers [@Stenhammar2017]. Our theory makes explicit predictions for various experimentally relevant observables for any density of microswimmers up to the onset of collective motion. All of its parameters can be independently measured in experiments, and its predictions can be directly compared against experimental data.
The results of our theory, presented in Section \[section:results\], reveal that all observables considered deviate from their mean-field values, which can be recovered from our results by setting $\Delta=0$, indicating that the mean-field theory is incorrect at any density below the onset of collective motion. We have also uncovered the following interplay between the strength of correlations between microswimmers and their self-propulsion speed. For all observables considered, the strongest correlations are exhibited by suspensions of shakers, $L=0$. This can be readily seen by observing that, in the absence of self-propulsion, the microswimmer positions only change due to their mutual advection. In dilute suspension, displacements thus accumulated over one correlation time are small compared to the interparticle distances, and, to first approximation, shaker suspensions perform orientational dynamics only. In turn, this implies that they spend maximum amount of time possible adjusting to the orientational fields created by other microswimmers. In contrast, motile microswimmers are aligning in a local velocity field that constantly changes due to their self-propulsion, implying weaker correlations in such suspensions. This effect becomes stronger as $L$ increases.
The degree to which correlations are suppressed by self-propulsion depends on the nature of the observable. Spatial-like observables (the fluid velocity variance, the energy spectrum, and the spatial correlation function) are significantly reduced as $L$ increases, but do not reach their mean-field values even in the limit $L\rightarrow\infty$. For instance, as can be seen from Fig.\[Fig\_Variance\], the fluid velocity variance is significantly larger than its mean-field value at any density $\Delta$, even in the limit of fast swimming. In a similar fashion, as $L\rightarrow\infty$, the spatial correlation function in Fig.\[Fig\_spatial\_delta09\] does not reduce to its mean-field behaviour, which is given by the $\Delta\rightarrow0$ limit in Fig.\[Fig\_spatial\]. Instead, it recovers the strongly correlated shaker-like behaviour at sufficiently large distances. This can be understood by employing the same argument as above. For any value of $L$, there are such separations $\rho$ that the typical distance travelled by a microswimmer during one correlation time of the suspension is small compared to $\rho$. For such separations, the difference between swimmers and shakers vanishes and $C(\rho)$ recovers its shaker-like behaviour.
On the other hand, temporal-like observables (the temporal correlation function and the enhanced diffusivity of tracer particles) are almost completely suppressed as $L\rightarrow\infty$ for $\Delta<1$, though they still diverge in the limit of $\Delta\rightarrow1$. This behaviour mirrors the dependence of their mean-field values on $L$, which vanish in the limit of fast swimming below the onset of collective motion. An intuitive argument for this behaviour has been put forward by Dunkel *et al.* [@Dunkel2010], who demonstrated that the total displacement of a tracer by a single motile particle vanishes as the length of a straight path covered by the swimmer diverges. This is fundamentally related to the time-reversibility of Stokesian flows. The presence of correlations between microswimmers breaks this time-reversibility: although the pathway between two states in phase space is still reversible, the probabilities of finding the suspension in those states are a priori different. Strong swimming introduces effective phase space mixing and recovers equal a priori probabilities for the phase space states. Again, this argument only holds for $\Delta<1$ when $L$ is large, yet finite.
**Observable** **Scaling law for $\Delta\rightarrow1$**
----------------------------- ------------------------------------------
Fluid velocity variance $(1-\Delta)^{-1/2}$
(Pseudo-)correlation length $(1-\Delta)^{-1/2}$
Correlation time $(1-\Delta)^{-1}$
Enhanced diffusivity $(1-\Delta)^{-3/2}$
: Critical exponents[]{data-label="table:exponents"}
To gain further insight into the nature of the transition to collective motion exhibited by our model, we extracted the scaling behaviour of the observables considered in this work upon the approach to the onset, $\Delta=1$. All of these quantities diverge at $\Delta=1$ and the values of the critical exponents predicted by our theory are summarised in Table \[table:exponents\]. We want to stress that these exponents rely on the approximation introduced in Section \[subsection:approximateDILT\], and while we are confident that it semi-quantitatively captures the spatial and temporal behaviour of the generalised correlation function $C(\rho,\tau)$ for $\Delta<1$, its quality in the close vicinity of $\Delta=1$ is untested. The values presented in Table \[table:exponents\] should thus be seen as a starting point, and more work is needed to make a conclusive statement about the critical exponents predicted by our model. This is especially important in the context of defining universality subclasses of “wet" active matter models.
In this work, we have only considered pusher-like microswimmers below the onset of collective motion. Recent simulations suggest [@Stenhammar2017; @Bardfalvy2019] that suspensions of pullers also exhibit strong correlations, although their effect is opposite to what is observed for pushers. The results presented in this work cannot be used to study this effect, i.e. by replacing $\Delta$ with $-\Delta$ in the relevant expressions. Instead, to extend our theory to pullers, one would have to re-evaluate the long-term behaviour of the approximate double inverse Laplace transform in Section \[subsection:approximateDILT\] for negative values of $\Delta$.
Finally, we note that the scenario emerging from our work is that microswimmer motility suppresses correlations and brings the suspension closer to the mean-field state. This observation could prove to be instrumental in understanding the origins of a surprising observation made by Stenhammer *et al.* [@Stenhammar2017] regarding the nature of the transition to collective motion in bacteria. There, we performed large-scale Lattice Boltzmann simulations of dilute suspensions of pusher-like microswimmers, and observed that all quantities of interest increased sharply around the threshold predicted by the mean-field theory ($\Delta=1$ in the notation adopted here). This increase, however, fell short of a divergence that should be associated with a true transition. While it would be natural to assume that this is the result of finite-size effects always present in computer simulations, Stenhammar *et al.* [@Stenhammar2017] did not see any sharpening of the curves around the supposed transition point upon a systematic increase of the system size. This phenomenon is currently not understood. Either the finite-size effects are decaying very slowly and realistic-size computer simulations cannot reach such large system sizes, or the presence of strong correlations below the onset of collective motion changes the stability properties of the homogeneous and isotropic state, replacing the mean-field transition with a correlation-smoothened cross-over. The results of this work provide a way to test the latter hypothesis: Since swimming suppresses correlations, sharpening of the potential transition/cross-over should be observed in suspensions with progressively increased swimming speed.
Double inverse Laplace transform of an archetypal term {#appendix:Cesare}
======================================================
Here, we show how to calculate the double inverse Laplace transform of Eq.. The derivation of Eq. is similar, though lengthy, and we do not present it here.
We start by observing that the double inverse Laplace transform in Eq., given in terms of two Bromwich integrals [@Doetsch1974], can be written as $$\begin{aligned}
\label{app:B-1}
\lim_{t\to\infty}
\int_{\Gamma_1}\frac{ds_1}{2\pi i} \frac{e^{s_1 t}}{\lambda + s_1 + i v_s k (\hat{\bm k}\cdot {\bm p})}J(s_1),\end{aligned}$$ where $$\begin{aligned}
\label{app:B-2}
J(s_1)
=
\int_{\Gamma_2}\frac{ds_2}{2\pi i}
\frac{1}{s_1+s_2}
\frac{e^{s_2 (t+T)}}{{\lambda + s_2 - i v_s k (\hat{\bm k}\cdot {\bm p})}}.\end{aligned}$$ By the definition of the inverse Laplace transform [@Doetsch1974], the contours defining the integrals above have to be chosen such that $\Gamma_2$ passes on the right of $-s_1$ and of $-\lambda +iv_s k(\hat{\bm k}\cdot {\bm p})$, while $\Gamma_1$ should pass on the right of all the poles of $J(s_1)$ and of $-\lambda-i v_s k(\hat{\bm k}\cdot {\bm p})$. Observe that the first condition implies that $\Gamma_2$ should be chosen on the right of $-\Gamma_1$.
Next, we observe that, again from the definition of the inverse Laplace transform, $J(s_1)$ is only defined for $\textrm{Re}(s_1)>0$. To proceed, we follow the method often utilised in plasma physics to describe the Landau damping [@Balescu1975]. We perform the analytic continuation of $J(s_1)$ to purely imaginary values of $s_1$ (recall that the analytic continuation of a complex function defined on an open set is the only function $\hat{J}(s_1)$ that is analytic, defined on a larger set, and equals $J(s_1)$ on the original set), and replace $J(s_1)$ with $\hat{J}(s_1)$ in Eq. . Since $\lambda>0$, the difficulty in performing the analytic continuation of $J(s_1)$ lies in the pole at $s_2=-s_1$ of the integrand from Eq.. We, therefore, define $$\begin{aligned}
& \hat{J}(s_1) = \int^* \frac{ds_2}{2\pi i}
\frac{e^{s_2 (t+T)}}{s_2+\lambda-i v_s k(\hat{\bm k}\cdot {\bm p})} \nonumber \\
& \qquad\quad +\frac{1}{2} \ell(-s_1)
\frac{e^{s_2 (t+T)}}{s_2+\lambda-i v_s k(\hat{\bm k}\cdot {\bm p})}\Bigg\vert_{s_2=-s_1},\end{aligned}$$ where $$\begin{aligned}
\ell(s_1) = \begin{cases}
0, & \textrm{Re}(s_1)>0, \\
1, & \textrm{Re}(s_1)=0, \\
2, & \textrm{Re}(s_1)<0.
\end{cases}\end{aligned}$$ The meaning of the integral denoted by $\int^* ds_1$ above depends on the sign of $\textrm{Re}(s_1)$: If $\textrm{Re}(s_1)>0$, it is just a standard complex integral over a contour passing on the right of $-s_1$ and of $-\lambda +iv_s k(\hat{\bm k}\cdot {\bm p})$; If $\textrm{Re}(s_1)=0$, $\int^* ds_1$ stands for a principal value integral; Finally, if $\textrm{Re}(s_1)<0$, $\int^* ds_1$ stands for a standard complex integral over a contour passing on the left of $-s_1$ but on the right of $-\lambda +iv_s k(\hat{\bm k}\cdot {\bm p})$. With the definitions above, it is easy to show that $\hat{J}(s_1)$ is holomorphic in an infinitesimal stripe around $s_1\in \mathbb{R}$. Hence, it is the analytic continuation of $J(s_1)$.
Replacing $J(s_1)$ by $\hat{J}(s_1)$ in Eq., we obtain two terms. The first term, containing $\int^*ds_2$, vanishes for $t\to\infty$, since we are now free to choose the integration contours $\Gamma_1$ and $\Gamma_2$ such that $\textrm{Re}(s_1+s_2)<0$. The other term reads $$\begin{aligned}
& \lim_{t\to\infty} \frac{1}{2}\int_{\Gamma_1}\frac{ds_1}{2\pi i}
\frac{1}{s_1+\lambda+i v_s k(\hat{\bm k}\cdot {\bm p})} \nonumber \\
&\qquad\qquad\qquad\qquad \times \frac{\ell(-s_1)\,e^{-s_1 T}}{-s_1+\lambda-i v_s k(\hat{\bm k}\cdot {\bm p})}.\end{aligned}$$ Closing the contour at $+\infty$, the only pole contributing to the integral is at $s_1=\lambda-i v_s k(\hat{\bm k}\cdot {\bm p})$, and we obtain $$\begin{aligned}
&\lim_{t\to\infty}
\int_{\Gamma_1}\frac{ds_1}{2\pi i} \frac{e^{s_1 t}}{s_1+\lambda+i v_s k(\hat{\bm k}\cdot {\bm p})}J(s_1) \nonumber \\
&\qquad\qquad\qquad\qquad = \frac{e^{-\lambda T + i v_s k T (\hat{\bm k} \cdot {\bm p})}}{2\lambda}.\nonumber\end{aligned}$$ This completes the proof of the equality in Eq..
Approximating $\psi(z)$ {#appendix:psia}
=======================
Here, we develop an approximation to $\psi(z)$ from Eq.. Our goal is to find a rational function with a pole structure that is similar to the original $\psi(z)$. As discussed in Section \[subsection:approximateDILT\], the relevant domain is set by the values of $z$ given by $z = \beta/(1+s/\lambda)$, with $\beta=v_s k/\lambda$ varying from $0$ to $L=v_s/(\lambda \epsilon)=0-25$, and by the real part of $s$ ranging from $-\lambda$ to $0$.
Our starting point are the observations that as $z\rightarrow0$, $\psi(z) \rightarrow 1-3z^2/7$, while for $z\rightarrow\infty$, $\psi(z) \rightarrow 0$. Both asymptotic behaviours can be combined into $\psi_a(z)= 7/(7+3z^2)$. Now we show that this is a surprisingly good approximation to $\psi(z)$, both reproducing its global shape and having a similar pole structure.
![Comparison between $\psi(z)$ (solid lines) and $\psi_a(z)$ (dotted lines) for $z = \beta/(1+s/\lambda)$ for real values of $s$ and various values of $\beta$.[]{data-label="psi_vs_psia"}](Psia.pdf){width="0.8\linewidth"}
In Fig.\[psi\_vs\_psia\] we compare $\psi(z)$ and $\psi_a(z)$ for real values of $s$. We observe a good agreement between the two functions for various values of $\beta$. Similar, semi-quantitative, degree of agreement is observed for larger values of $\beta$ and also for complex values $s$.
To demonstrate that $\psi_a(z)$ also reproduces the pole structure of $\psi(z)$, we consider a typical term from the analysis in Section \[subsection:hhat\] $$\begin{aligned}
\frac{1}{s+\lambda-\lambda\Delta A(k\epsilon)\psi(z)} = \frac{1}{\lambda \Delta A(k\epsilon)}\frac{z}{\omega - z \psi(z)}.
\label{A_original}\end{aligned}$$ We compute its inverse Laplace transform numerically, using the original function $\psi(z)$, and compare the result with the analytic expression, which we obtain by replacing $\psi(z)$ with $\psi_a(z)$ in the expression above. The latter is straightforward: factorising $\omega \left(7+3z^2\right) - 7 z = 3\omega(z-z_{+})(z-z_{-})$, where $z_{\pm}$ are given in Eq., we obtain $$\begin{aligned}
&\frac{1}{\lambda \Delta A(k\epsilon)}\frac{z}{\omega - z \psi_a(z)} \nonumber \\
&\qquad\qquad = \frac{1}{7}\frac{1}{s+\lambda}\frac{7(s+\lambda)^2 + 3(v_s k)^2}{\left(s+\lambda - \frac{v_s k}{z_{+}}\right)\left(s+\lambda - \frac{v_s k}{z_{-}}\right)}.\end{aligned}$$ Performing the inverse Laplace transform of this expression and introducing the dimensionless time $\tau = \lambda t$ yields $$\begin{aligned}
&e^{-\tau} \Bigg[ 1 +\frac{2}{\sqrt{1-\frac{12\beta^2}{7\Delta^2}}} \exp{\left(\frac{\tau \Delta }{2}\right)} \nonumber \\
& \qquad\qquad\qquad\qquad \times \sinh{\left(\frac{\tau \Delta}{2}\sqrt{1-\frac{12\beta^2}{7\Delta^2}}\right)}\Bigg].
\label{A_test_analytic}\end{aligned}$$ Since both $A(k\epsilon)$ and $\Delta$ take values between $0$ and $1$, for the purpose of comparing to its numerical counterpart, we set $A(k\epsilon)=1$ in the expression above, without loss of generality.
{width="1.0\linewidth"}
The inverse Laplace transform of the original function Eq. written in terms of the same parameters is given by the Bromwich integral $$\begin{aligned}
\frac{1}{2\pi i} \int_{\gamma-i \infty}^{\gamma+i \infty} d{\tilde s} \frac{e^{\tilde s \tau}}{{\tilde s}+1-\Delta \psi(\beta/(\tilde s+1))},
\label{A_test_numerics}\end{aligned}$$ where $\gamma$ is a real number, chosen to be greater than the real part of any singularity of the integrand [@Doetsch1974]. We perform this integral numerically, using the Gaver–Wynn–Rho algorithm as presented by Valko and Abate [@Valko2005]. Valko and Abate provide an explicit Mathematica function GWR [@math1], which we use here. A Mathematica notebook with the details of this calculation can be found here [@SI].
In Fig.\[ILTcombined\]A we compare Eq. against the numerical Laplace transform of Eq. for $\Delta=0.1$ and $\beta=0.1$, $1$, and $10$. We observe a very good agreement, which is not surprising: At small microswimmer densities, the hydrodynamic interactions between particles affect their dynamics only weakly, and correlations decay as $e^{-\tau}$. This regime does not test the quality of our approximation. A more stringent test is provided, on the other hand, in Fig.\[ILTcombined\]B, were we compare the two Laplace transforms for $\Delta=0.9$. For $\beta<1$, we observe a very good agreement even at such high values of $\Delta$ (close to the mean-field transition). This is the most interesting regime, corresponding to large-scale motion in the suspension, and it is encouraging that our approximation shows quantitative agreement with the numerical data. Note that the black line and the black circles, corresponding to $\beta=0.1$, do not follow $e^{-\tau}$, i.e. our approximation is capable of capturing a non-trivial decay rate. At higher values of $\beta$, corresponding to scales comparable to individual microswimmers, the agreement is semi-quantative, but the overall decay is again close to the tumbling-dominated decay $e^{-\tau}$.
In Appendix \[appendix:dilt\], we assess the quality of our approximation, when used in Eq., which is its ultimate purpose.
Double inverse Laplace transform {#appendix:dilt}
================================
In Section \[subsection:approximateDILT\], we performed the double inverse Laplace transform in Eq. analytically by replacing $\psi(z)$ with $\psi_a(z)$, which led to Eq.. Here, we assess the quality of that approximation by performing the double inverse Laplace transform in Eq. numerically. The relevant part of Eq. reads $$\begin{aligned}
&{\mathcal L}^{-1}_{\tilde{s}_1,\tilde{t}} {\mathcal L}^{-1}_{\tilde{s}_2,\tilde{t}+\tau}
\frac{1}{1 + \tilde{s}_1}\frac{1}{1 + \tilde{s}_2}\frac{1}{\tilde{s}_1 + \tilde{s}_2} \frac{\tilde{z}_1 \psi(\tilde{z}_1) + \tilde{z}_2 \psi(\tilde{z}_2)}{\tilde{z}_1+\tilde{z}_2}
\nonumber \\
&\qquad \times\Bigg[ \frac{\tilde{z}_1 \psi(\tilde{z}_1)}{\tilde{\omega} - \tilde{z}_1 \psi(\tilde{z}_1)} + \frac{\tilde{z}_2 \psi(\tilde{z}_2)}{\tilde{\omega} - \tilde{z}_2 \psi(\tilde{z}_2)} \nonumber \\
&\qquad\qquad
+ \frac{\tilde{z}_1 \tilde{z}_2 \psi(\tilde{z}_1)\psi(\tilde{z}_2)}{\left(\tilde{\omega} - \tilde{z}_1 \psi(\tilde{z}_1)\right)\left(\tilde{\omega} - \tilde{z}_2 \psi(\tilde{z}_2)\right)} \Bigg],
\label{DILTeqnum}\end{aligned}$$ where, in anticipation of performing numerical calculations, we introduced the dimensionless times $\tau = \lambda T$ and ${\tilde t} = \lambda t$, Laplace frequencies ${\tilde s}_{1,2} = s_{1,2}/\lambda$, ${\tilde z}_{1,2} = \beta/(1+{\tilde s}_{1,2})$, and $\tilde{\omega} = \beta/\Delta$, where we absorbed $A(k\epsilon)$ into $\Delta$, as in Appendix \[appendix:psia\]. In what follows, we set $\tilde t=20$ to imitate the limit $\tilde{t}\rightarrow\infty$. The calculations are performed in Mathematics using the combined Fixed-Talbot and Gaver–Wynn–Rho algorithm described by Valko and Abate [@Valko2005]. A Mathematica notebook with the details of this calculation can be found here [@SI]. The results are compared to the relevant part of Eq., recast in the same dimensionless variables $$\begin{aligned}
&e^{-\tau} \Bigg[ -\cos{\left(\sqrt{\frac{3}{7}}\beta\tau\right)} + \frac{e^{\frac{1}{2} \Delta \tau}}{1-\Delta + \frac{3}{7}\beta^2} \Bigg\{ \nonumber \\
& \qquad\qquad \frac{2-\Delta + \frac{6}{7}\beta^2}{2-\Delta}\cosh{\left(\frac{1}{2}\Delta \tau \sqrt{1-\frac{12 \beta^2}{7 \Delta^2}}\right)} \nonumber \\
&\qquad\qquad\qquad +\frac{\sinh{\left(\frac{1}{2} \Delta \tau \sqrt{1-\frac{12 \beta^2}{7 \Delta^2}}\right)}}{\sqrt{1-\frac{12 \beta^2}{7 \Delta^2}}}
\Bigg\}
\Bigg].
\label{DILTeqanal}\end{aligned}$$
{width="1.0\linewidth"}
The results of the numerical double inverse Laplace transform and its analytical counterpart are shown in Fig.\[DILTcombined\]. As in Appendix \[appendix:psia\], we focus on high values of $\Delta$, which provide the most stringent test of our results. For $\beta\le1$, the analytic approximation agrees quite well with the numerical data, capturing not only the decay rate, but also the oscillatory behaviour, as can be seen from the $\beta=1$ case. These calculations required a very high number of terms, $O(100)$, in the combined Fixed-Talbot and Gaver–Wynn–Rho algorithm [@Valko2005]. For $\beta > 1$, we were unable to obtain converged results for the numerical Laplace transform for any viable number of terms in the numerical algorithm. Nevertheless, the results of Appendix \[appendix:psia\], and the degree of agreement exhibited in Fig.\[DILTcombined\] for the physically most relevant case of $\beta<1$ make us confident that Eq. faithfully reproduces the long-time behaviour of Eq..
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper studies the first moment of symmetric-square $L$-functions at the critical point in the weight aspect. Asymptotics with the best known error term $O(k^{-1/2})$ were obtained independently by Fomenko in 2005 and by Sun in 2013. We prove that there is an extra main term of size $k^{-1/2}$ in the asymptotic formula and show that the remainder term decays exponentially in $k$. The twisted first moment was evaluated asymptotically by Ng Ming Ho with the error bounded by $lk^{-1/2+\epsilon}$. We improve the error bound to $l^{5/6+\epsilon}k^{-1/2+\epsilon}$ unconditionally and to $l^{1/2+\epsilon}k^{-1/2}$ under the Lindelöf hypothesis for quadratic Dirichlet $L$-functions.'
address:
- 'University of Turku, Department of Mathematics and Statistics, Turku, 20014, Finland'
- 'Steklov Mathematical Institute of Russian Academy of Sciences, 8 Gubkina st., Moscow, 119991, Russia'
author:
- Olga Balkanova
- Dmitry Frolenkov
nocite: '\nocite{}'
title: 'On the mean value of symmetric square $L$-functions'
---
[^1]
Introduction
============
Asymptotic behavior of high moments of $L$-functions within different families can be predicted using random matrix theory [@CFKRS] or multiple Dirichlet series [@DGH]. However, obtaining asymptotic formulas with sharp error bounds is a hard problem even in case of small moments.
One of the most challenging families is symmetric square $L$-functions in weight aspect. Gelbart and Jacquet [@GJ] proved that these are $L$-function attached to $GL(3)$ cusp forms.
Despite numerous efforts, even an upper bound for the second moment of symmetric square $L$-functions remains an open problem. See Conjecture $1.2$ [@K2].
The first moment has been studied intensively during the last decades. See [@F; @K; @KS; @Lau; @NMH; @Sun]. Nethertheless, even the best known asymptotic error estimates do not appear to be sharp.
The present paper aims to optimize error bounds in existing asymptotic formulas. With this goal, we prove an exact formula for the twisted first moment of symmetric square $L$-functions, and apply the Liouville-Green method (also called WKB approximation) to estimate remainder terms. This technique, originating from the theory of approximation of second-order differential equations, is quite unusual for analytic number theory, yet very effective. See, for example, [@BF; @Zav].
Main results
============
Let $S_{2k}(1)$ denote the space of holomorphic cusp forms of weight $2k\geq 2$ with respect to the full modular group. Denote by $H_{2k}$ the normalized Hecke basis for $S_{2k}(1)$. Every $f \in H_{2k}$ has a Fourier expansion of the form $$f(z)=\sum_{n\geq 1}\lambda_f(n)n^{k-1/2}\exp(2\pi inz),$$ $$\lambda_f(1)=1.$$ For $\Re{s}>1$ the associated symmetric square $L$-function is given by $$L({\operatorname{sym}}^2f,s)=\zeta(2s)\sum_{n=1}^{\infty}\frac{\lambda_f(n^2)}{n^s}.$$ Let $\Gamma(s)$ be the Gamma function and $$L_{\infty}(s):=\pi^{-3s/2}\Gamma\left(\frac{s+1}{2}\right)\Gamma\left(\frac{s+k-1}{2}\right)\Gamma\left(\frac{s+k}{2}\right).$$ Shimura [@S] showed that the completed $L$-function $$\Lambda({\operatorname{sym}}^2f,s):=L_{\infty}(s)L({\operatorname{sym}}^2f,s)$$ is entire and satisfies the functional equation $$\Lambda({\operatorname{sym}}^2f,s)=\Lambda({\operatorname{sym}}^2f,1-s).$$
Consider $$M_1(l,s):=\sum_{f \in H_{2k}}^{h}\lambda_f(l^2)L({\operatorname{sym}}^2f,s).$$ The subscript $h$ in the formula above indicates that the expression in the sum is multiplied by the harmonic weight $\Gamma(2k-1)/((4\pi)^{2k-1} \langle f,f\rangle_1)$, where $\langle f,f\rangle_1$ is the Petersson inner product on the space of level $1$ holomorphic modular forms.
Denote by $\gamma$ the Euler constant and by $\psi(s)$ logarithmic derivative of the Gamma function. Let ${}_2F_{1}(a,b,c;x)$ be the Gauss hypergeometric function and $$\label{defphi}
\Phi_k(x):=\frac{\Gamma(k-1/4)\Gamma(3/4-k)}{\Gamma(1/2)}{}_2F_{1}\left(k-\frac{1}{4},\frac{3}{4}-k,1/2;x \right),$$ $$\Psi_k(x):=x^k\frac{\Gamma(k-1/4)\Gamma(k+1/4)}{\Gamma(2k)}{}_2F_{1}\left(k-\frac{1}{4},k+\frac{1}{4},2k;x \right).$$ We prove the following exact formula for the twisted first moment.
\[thm:explicitformula\] For any $l \geq 1$ one has $$\begin{gathered}
\label{mainformula}
M_1(l,1/2)=\frac{1}{2\sqrt{l}}\biggl(-2\log{l}-3\log{2\pi}+\frac{\pi}{2}+4\gamma+\psi(1)+\\ \psi(k-1/4)+ \psi(k+1/4) \biggr)+
\frac{\sqrt{2\pi}(-1)^k}{2\sqrt{l}}\frac{\Gamma(k-1/4)}{\Gamma(k+1/4)}\mathscr{L}_{-4l^2}(1/2)+\\
\frac{1}{\sqrt{l}} \sum_{1\leq n<2l}
\mathscr{L}_{n^2-4l^2}(1/2)\Phi_k\left(\frac{n^2}{4l^2}\right)+
\frac{1}{l\sqrt{2}} \sum_{n>2l}\mathscr{L}_{n^2-4l^2}(1/2)\sqrt{n}\Psi_k\left( \frac{4l^2}{n^2}\right),\end{gathered}$$ where $$\label{Lbyk}
\mathscr{L}_n(s)=\frac{\zeta(2s)}{\zeta(s)}\sum_{q=1}^{\infty}\frac{1}{q^s}\left( \sum_{1\leq t \leq 2q;t^2 \equiv n {\ (\textup{mod}\ 4q)}}1\right).$$
A similar formula, where the last two summands are expressed in terms of Legendre function of the first kind, was established by a different method by Zagier [@Z Theorem 1]. Zagier’s formula was applied by Kohnen and Sengupta [@KS] to prove an upper bound for $M_1(1,1/2)$, by Fomenko [@F] to obtain an asymptotic formula for $M_1(1,1/2)$, and by Luo [@L] to estimate the second moment of $L({\operatorname{sym}}^2f,1/2)$ over short intervals.
The proof of Theorem \[thm:explicitformula\] is quite simple and make use of Petersson’s trace formula and the functional equation for the Lerch zeta function.
When $l=1$, exact formula allows isolating the second main term of size $k^{-1/2}$ in the asymptotic formula so that the remainder term decays exponentially.
\[thm:asympformula2\] For some $c>0$ one has $$\begin{gathered}
\label{mainformula3}
M_1(1,1/2)=\frac{1}{2}\biggl(-3\log{2\pi}+\frac{\pi}{2}+4\gamma+\psi(1)+\\ \psi(k-1/4)+ \psi(k+1/4) \biggr)+
\frac{\sqrt{2\pi}(-1)^k}{2}\frac{\Gamma(k-1/4)}{\Gamma(k+1/4)}L(1/2,\chi_{-4})+\\ \Phi_k\left(\frac{1}{4}\right)L(1/2,\chi_{-3})+O\left(\frac{1}{\sqrt{k}}exp(-ck)\right),\end{gathered}$$ where $L(1/2,\chi_D)$ is a Dirichlet $L$-function for primitive quadratic character of conductor $D$.
After posting the first version of this paper to the arXiv, the authors have been informed by Shenhui Liu that he has independently obtained an asymptotic formula similar to by using an approximate functional equation. See [@Liu].
Corollary \[thm:asympformula2\] improves the series of previously known results with the following error bounds:
- $k^{-0.008}$ proved by Lau [@Lau] in 2002;
- $k^{-1/20}$ proved by Khan [@K] in 2007;
- $k^{-1/2}$ proved by Fomenko [@F] in 2005 and by Sun [@Sun] in 2013.
\[thm:asympformula\] For any $\epsilon>0$, $l >1$ one has $$\begin{gathered}
\label{mainformula2}
M_1(l,1/2)=\frac{1}{2\sqrt{l}}\biggl(-2\log{l}-3\log{2\pi}+\frac{\pi}{2}+4\gamma+\psi(1)+\\ \psi(k-1/4)+ \psi(k+1/4) \biggr)+O\left(\frac{l^{5/6+\epsilon}}{\sqrt{k}}\right).\end{gathered}$$ Assuming the Lindelöf hypothesis for quadratic Dirichlet $L$-functions, the error term above can be replaced by $O\left(l^{1/2+\epsilon}k^{-1/2}\right).$
This improves the error bound $lk^{-1/2+\epsilon}$ proved by Ng Ming Ho [@NMH Theorem 2.1.1].
Notations and tools
===================
Let $e(x)=exp(2\pi ix)$. For $ v \in {\mathbf{C}}$ let $$\tau_v(n)=\sum_{n_1n_2=n}\left( \frac{n_1}{n_2}\right)^v.$$ The classical Kloosterman sum is defined by $$S(n,m;c)=\sum_{\substack{a{\ (\textup{mod}\ c)}\\ (a,c)=1}}e\left( \frac{an+a^*m}{c}\right), \quad aa^*\equiv 1{\ (\textup{mod}\ c)}.$$
(Weil’s bound [@W]) One has $$|S(m,n;c)|\leq \tau_0(c)\sqrt{(m,n,c)}\sqrt{c}.$$
Let $J_{v}(x)$ be the Bessel function of the first kind.
(Petersson’s trace formula, [@P]) For $2k \geq 12$ and integral $l,n \geq 1$ one has $$\label{Pet}
\sum_{f \in H_{2k}}^{h}\lambda_f(l)\lambda_f(n)=\delta_{l,n}+2\pi i^{2k}\sum_{c =1}^{\infty}\frac{S(l,n;c)}{c}J_{2k-1}\left( \frac{4\pi \sqrt{ln}}{c}\right).$$
The Lerch zeta function $$\zeta(\alpha,\beta,s)=\sum_{n+\alpha>0}\frac{e(n\beta)}{(n+\alpha)^s}$$ was introduced by Lipschitz [@Lip] in $1857$ and was named after Lerch, who proved in $1887$ the following functional equation.
([@Ler]) One has $$\label{LerchFE}
\zeta(\alpha,0,s)=\frac{\Gamma(1-s)}{(2\pi)^{1-s}}\biggl(-ie\left(\frac{s}{4}\right)\zeta(0,\alpha,1-s)+
ie\left(-\frac{s}{4}\right)\zeta(0,-\alpha,1-s)\biggr).$$
Some properties of $\mathscr{L}_{n}(s)$
=======================================
The main references for this section are [@B; @SY; @Z]. Function can be written as follows $$\mathscr{L}_{n}(s)=\frac{\zeta(2s)}{\zeta(s)}\sum_{q=1}^{\infty}\frac{\rho_q(n)}{q^{s}}=\sum_{q=1}^{\infty}\frac{\lambda_q(n)}{q^{s}},$$ where $$\rho_q(n):=\#\{x{\ (\textup{mod}\ 2q)}:x^2\equiv n{\ (\textup{mod}\ 4q)}\},$$ $$\lambda_q(n):=\sum_{q_{1}^{2}q_2q_3=q}\mu(q_2)\rho_{q_3}(n).$$ For a fixed $n$ both $\rho_q(n)$ and $\lambda_q(n)$ are multiplicative functions of $q$. Furthermore, for $n \equiv 2,3 {\ (\textup{mod}\ 4)}$ the function $\rho_q(n)$ is identically zero. Therefore, $\mathscr{L}_n(s)$ does not vanish only for $n \equiv 0,1 {\ (\textup{mod}\ 4)}.$ If $n=0$ then $$\mathscr{L}_{n}(s)=\zeta(2s-1).$$ Otherwise, for $n=Dl^2$ with $D$ fundamental discriminant we have $$\label{ldecomp}
\mathscr{L}_{n}(s)=l^{1/2-s}T_{l}^{(D)}(s)L(s,\chi_D),$$ where $L(s,\chi_D)$ is a Dirichlet L-function for primitive quadratic character $\chi_D$ and $$\label{eq:td}
T_{l}^{(D)}(s)=\sum_{l_1l_2=l}\chi_D(l_1)\frac{\mu(l_1)}{\sqrt{l_1}}\tau_{s-1/2}(l_2).$$
The completed $L$-function $$\mathscr{L}_{n}^{*}(s)=(\pi/|n|)^{-s/2}\Gamma(s/2+1/4-{\operatorname{sgn}}{n}/4)\mathscr{L}_{n}(s)$$ satisfies the functional equation $$\label{functlstar}
\mathscr{L}_{n}^{*}(s)=\mathscr{L}_{n}^{*}(1-s).$$
One has $$\label{eq:sumofklsums}
\sum_{q=1}^{\infty}\frac{1}{q^{1+s}}\sum_{c{\ (\textup{mod}\ q)}}S(l^2,c^2;q)e\left(\frac{nc}{q}\right)=\frac{1}{\zeta(2s)}
\mathscr{L}_{n^2-4l^2}(s).$$
Consider $$\begin{gathered}
S:=\sum_{c{\ (\textup{mod}\ q)}}S(l^2,c^2;q)e\left(n\frac{c}{q}\right)=\\
\sum_{c{\ (\textup{mod}\ q)}}\sum_{\substack{a {\ (\textup{mod}\ q)}\\(a,q)=1}}e\left( \frac{ac^2+a^*l^2+nc}{q}\right),\end{gathered}$$ where $aa^*\equiv 1{\ (\textup{mod}\ q)}$. Making the change of variables $c=c_1a^*$, we have $$\begin{gathered}
S=
\sum_{\substack{a {\ (\textup{mod}\ q)}\\(a,q)=1}}\sum_{c_1 {\ (\textup{mod}\ q)}}e\left(\frac{c_{1}^{2}a^*+l^2a^*+nc_1a^*}{q}\right)=\\
\sum_{c_1 {\ (\textup{mod}\ q)}}S(0,c_{1}^{2}+l^2+nc_1;q)=\sum_{c{\ (\textup{mod}\ q)}}\sum_{\substack{bd=q\\b|c^2+l^2+nc}}\mu(d)b=\\
\sum_{bd=q}\mu(d)b\sum_{\substack{c {\ (\textup{mod}\ q)}\\ c^2+l^2+nc\equiv 0 {\ (\textup{mod}\ b)}}}1=\sum_{bd=q}\mu(d)b
\sum_{\substack{c{\ (\textup{mod}\ b)}\\ c^2+l^2+nc\equiv 0{\ (\textup{mod}\ b)}}}\frac{q}{b}.\end{gathered}$$ The condition $$c^2+l^2+nc\equiv 0 {\ (\textup{mod}\ b)}$$ is equivalent to $$(2c+n)^2+4l^2-n^2\equiv 0{\ (\textup{mod}\ 4b)}.$$ Hence $$S=q\sum_{bd=q}\mu(d)\sum_{\substack{c{\ (\textup{mod}\ 2b)}\\ c^2\equiv n^2-4l^2{\ (\textup{mod}\ 4b)}}}1=q\sum_{bd=q}\mu(d)\rho_b(n^2-4l^2).$$ Consequently, $$\begin{gathered}
\sum_{q=1}^{\infty}\frac{1}{q^{1+s}}\sum_{c{\ (\textup{mod}\ q)}}S(l^2,c^2;q)e\left(\frac{nc}{q}\right)=
\sum_{q=1}^{\infty}\frac{1}{q^s}\sum_{bd=q}\mu(d)\rho_b(n^2-4l^2)=\\
\sum_{b=1}^{\infty}\frac{\rho_b(n^2-4l^2)}{b^s}\sum_{q=1}^{\infty}\frac{\mu(q)}{q^s}=
\frac{1}{\zeta(s)}\sum_{q=1}^{\infty}\frac{\rho_q(n^2-4l^2)}{q^s}=\frac{\mathscr{L}_{n^2-4l^2}(s)}{\zeta(2s)}
.\end{gathered}$$
\[lem:subconvexity\] Assume that $d\neq 0$. For any $\epsilon>0$ one has $$\label{eq:subconvexity}
\mathscr{L}_d(1/2)\ll d^{1/6+\epsilon}.$$ If the Lindelöf hypothesis for Dirichlet $L$-functions is true, then $$\label{eq:lindelof}
\mathscr{L}_d(1/2)\ll d^{\epsilon}.$$
For $d=Dl^2$ with $D$ fundamental discriminant one has $$\mathscr{L}_{d}(1/2)=T_{l}^{(D)}(1/2)L(1/2,\chi_D)$$ by equation . It follows from equality that $$T_{l}^{(D)}(1/2)\ll \sum_{l_1l_2=l}\frac{\sigma(l_2)}{\sqrt{l_1}}\ll \sum_{l_1|l}\frac{(l/l_1)^{\epsilon}}{\sqrt{l_1}}\ll l^{\epsilon}.$$ By [@CI Corollary 1.5] for any $\epsilon>0$ one has $$L(1/2, \chi_D)\ll D^{1/6+\epsilon}.$$ This implies the required bounds for the function $\mathscr{L}_{d}(1/2)$.
Exact formula
=============
For $\Re{s}>3/2$ one has $$\begin{gathered}
\label{eq:M1ls}
M_1(l,s)=\frac{\zeta(2s)}{l^{s}}+
\frac{(2\pi)^{s}i^{2k}}{2l^{1-s}}\frac{\Gamma(k-s/2)}{\Gamma(k+s/2)}\mathscr{L}_{-4l^2}(s)+\\
(2\pi)^{s}i^{2k}\sum_{n=1}^{\infty}\frac{1}{n^{1-s}}\mathscr{L}_{n^2-4l^2}(s)I\left( \frac{n}{l}\right),\end{gathered}$$ where $$\label{eq:integralI}
I(x):=\frac{1}{2\pi i}\int_{(\Delta)}\frac{\Gamma(k-1/2+t/2)}{\Gamma(k+1/2-t/2)}\Gamma(1-s-t)\sin\left( \frac{s+t}{2}\right)x^tdt$$ with $1-2k<\Delta<1-\Re{s}.$
By the Petersson trace formula $$\begin{gathered}
M_1(l,s)=\zeta(2s)\sum_{n=1}^{\infty}\frac{1}{n^{s}}\sum_{f \in H_{2k}}^{h}\lambda_f(l^2)\lambda_f(n^2)=\\
\frac{\zeta(2s)}{l^{s}}+2\pi i^{2k}\zeta(2s)\sum_{q=1}^{\infty}\frac{1}{q}
\sum_{n=1}^{\infty}\frac{S(l^2,n^2;q)}{n^{s}}J_{2k-1}\left(4\pi \frac{ln}{q}\right).\end{gathered}$$
The change of order of summation above is justified by the absolute convergence for $\Re{s}>3/2$, which follows from the standard estimates $$S(l^2,n^2;q)\ll q^{1/2+\epsilon}(l^2,n^2,q)^{1/2} \text{ for any } \epsilon>0$$ and $$J_{2k-1}\left(4\pi \frac{ln}{q}\right)\ll \begin{cases}
(ln/q)^{2k-1} & q>ln\\
(ln/q)^{-1/2} & q<ln.
\end{cases}$$ Next, we use the Mellin-Barnes representation for the Bessel function $$\begin{gathered}
M_1(l,s)=\frac{\zeta(2s)}{l^{s}}+2\pi i^{2k}\zeta(2s)\sum_{q=1}^{\infty}\frac{1}{q}\times
\\ \frac{1}{4\pi i}
\int_{(\Delta)}\frac{\Gamma(k-1/2+t/2)}{\Gamma(k+1/2-t/2)}\sum_{n=1}^{\infty}\frac{S(l^2,n^2;q)}{n^{t+s}}
\left(\frac{q}{2\pi l} \right)^tdt,\end{gathered}$$ where $1-2k<\Delta<0$. To guarantee the absolute convergence of the integral over $t$ and the sums over $q,n$ we require that $$\max(1-2k,1-\Re{s})<\Delta<-1/2,$$ which is true for $\Re{s}>3/2$. Consider $$\begin{gathered}
\sum_{n=1}^{\infty}\frac{S(l^2,n^2;q)}{n^{t+s}}=\sum_{c{\ (\textup{mod}\ q)}}\sum_{n \equiv c{\ (\textup{mod}\ q)}}\frac{S(l^2,c^2;q)}{n^{t+s}}=\\ \sum_{c{\ (\textup{mod}\ q)}}S(l^2,c^2;q) \sum_{n=1}^{\infty}\frac{1}{(c+nq)^{t+s}}=
\sum_{c{\ (\textup{mod}\ q)}}\frac{S(l^2,c^2;q)}{q^{t+s}}\zeta\left(\frac{c}{q},0,t+s\right).\end{gathered}$$ Note that the Lerch zeta function has a pole at $t=1-s$ for all $c$. The next step is to apply functional equation for the Lerch zeta function which is only possible when $\Re(s+t)<0$. Accordingly, we move the $t$-contour to the left up to $\Delta_1:=-s-\epsilon$, crossing a simple pole at $t=1-s$. Therefore, $$\begin{gathered}
M_1(l,s)=\frac{\zeta(2s)}{l^s}+2\pi i^{2k}\zeta(2s)\frac{\Gamma(k-s/2)}{\Gamma(k+s/2)}\sum_{q=1}^{\infty}
\sum_{c{\ (\textup{mod}\ q)}}\frac{S(l^2,c^2;q)}{2q^2}\times \\ \left( \frac{q}{2\pi l}\right)^{1-s}+2\pi i^{2k}\zeta(2s)
\sum_{q=1}^{\infty}\frac{1}{q} \frac{1}{4\pi i}\times \\\int_{(\Delta_1)}\frac{\Gamma(k-1/2+t/2)}{\Gamma(k+1/2-t/2)}
\left(\frac{q}{2\pi l} \right)^t\sum_{c {\ (\textup{mod}\ q)}}\frac{S(l^2,c^2;q)}{q^{s+t}}\zeta(c/q,0;s+t)ds.\end{gathered}$$ Using functional equation , we obtain $$\begin{gathered}
\sum_{c {\ (\textup{mod}\ q)}}S(l^2,c^2;q)\zeta(c/q,0;s+t)=2(2\pi)^{s+t-1}\Gamma(1-s-t)\times \\\sin\left(\pi \frac{s+t}{2}\right)
\sum_{c{\ (\textup{mod}\ q)}}S(l^2,c^2;q)\zeta(0,c/q;1-s-t).\end{gathered}$$ Substituting this into $M_1(l,s)$ and opening the Lerch zeta function, one has $$\begin{gathered}
M_1(l,s)=\frac{\zeta(2s)}{l^s}+\frac{(2\pi)^s i^{2k}}{2l^{1-s}}\zeta(2s)\frac{\Gamma(k-s/2)}{\Gamma(k+s/2)}
\sum_{q=1}^{\infty}\sum_{c{\ (\textup{mod}\ q)}}\frac{S(l^2,c^2;q)}{q^{1+s}}+\\
(2\pi)^si^{2k}\zeta(2s)\sum_{n=1}^{\infty}\frac{1}{n^{1-s}}\sum_{q=1}^{\infty}\frac{1}{q^{1+s}}
\sum_{c{\ (\textup{mod}\ q)}}S(l^2,c^2;q)e\left(n\frac{c}{q}\right)I\left(\frac{n}{l} \right),\end{gathered}$$ where $I(x)$ is defined by equation . Finally, computing the sums over $c$ and $q$ using formula , we prove the Lemma.
If $x \geq 2$, then $$\begin{gathered}
\label{integralIgeq2}
I(x)=\frac{2^{2k}(-1)^k}{2^s\sqrt{\pi}}\cos\left( \frac{\pi s}{2}\right)x^{1-2k}\frac{\Gamma(k-s/2)\Gamma(k+1/2-s/2)}{\Gamma(2k)}\times \\ {}_2F_{1}\left(k-s/2,k+1/2-s/2,2k;\frac{4}{x^2} \right).\end{gathered}$$
Moving the contour of integration in to the left, we cross simple poles at $t=1-2k-2j$, $j=0,1,2,\ldots$ Therefore, $$I(x)=2(-1)^k\cos \left(\frac{\pi s}{2} \right)x^{1-2k}\sum_{j=0}^{\infty}\frac{1}{j!}\frac{\Gamma(2k-s+2j)}{\Gamma(2k+j)}x^{-2j}.$$ By [@HMF Eq. 5.5.5] we have $$\Gamma(2(k+j-s/2))=\frac{2^{2k-1-s+2j}}{\sqrt{\pi}}\Gamma(k+j-s/2)\Gamma(k+j+1/2-s/2).$$ This yields $$\begin{gathered}
I(x)=\frac{2(-1)^k}{\sqrt{\pi}}2^{2k-1-s}\cos \left(\frac{\pi s}{2} \right)x^{1-2k}\sum_{j=0}^{\infty}\frac{1}{j!}\frac{\Gamma(k-s/2+j)}{\Gamma(2k+j)}\times\\ \Gamma(k+1/2-s/2+j)\left(\frac{4}{x^2} \right)^j=
\frac{2^{2k}(-1)^k}{2^s\sqrt{\pi}}\cos\left( \frac{\pi s}{2}\right)x^{1-2k}\times \\ \frac{\Gamma(k-s/2)\Gamma(k+1/2-s/2)}{\Gamma(2k)}{}_2F_{1}\left(k-s/2,k+1/2-s/2,2k;\frac{4}{x^2} \right).\end{gathered}$$
One has $$\label{integralIeq2}
I(2)=\frac{2(-1)^k}{2^s\sqrt{\pi}}\cos{\left(\frac{\pi s}{2}\right)}\frac{\Gamma(k-s/2)\Gamma(k+1/2-s/2)}{\Gamma(k+s/2)\Gamma(k-1/2+s/2)}\Gamma(s-1/2).$$
Letting $x=2$ in and applying [@HMF Eq. 15.4.20], we find $${}_2F_{1}\left(k-s/2,k+1/2-s/2,2k;1 \right)=\frac{\Gamma(2k)\Gamma(s-1/2)}{\Gamma(k+s/2)\Gamma(k-1/2+s/2)}.$$ The assertion follows.
If $x<2$, then $$\begin{gathered}
\label{integralIl2}
I(x)=\frac{(-1)^k}{\sqrt{\pi}}\sin\left( \frac{\pi s}{2}\right)x^{1-s}\frac{\Gamma(k-s/2)\Gamma(1-k-s/2)}{\Gamma(1/2)}\times \\
{}_2F_{1}\left( k-\frac{s}{2},1-k-\frac{s}{2},1/2;\frac{x^2}{4}\right).\end{gathered}$$
Moving the contour of integration in to the right we cross simple poles at $t=1-s+j$, $j=0,1,2,\ldots$ Accordingly, $$I(x)=\sum_{j=0}^{\infty}\frac{(-1)^j}{j!}\frac{\Gamma(k-s/2+j/2)}{\Gamma(k+s/2-j/2)}\sin \left(\pi \frac{1+j}{2}\right)x^{1-s+j}.$$ Note that $$\sin \left(\pi \frac{1+j}{2}\right)=\cos \left( \frac{\pi j}{2}\right)=\begin{cases}
0 & j \text{ is odd,}\\
(-1)^m & j=2m.
\end{cases}$$ Thus $$I(x)=\sum_{m=0}^{\infty}\frac{1}{(2m)!}\frac{\Gamma(k-s/2+m)}{\Gamma(k+s/2-m)}(-1)^m x^{1-s+2m}.$$ In order to express $I(x)$ in terms of the Gauss hypergeometric function we apply [@HMF Eq. 5.5.5], obtaining $$(2m)!=\Gamma(2(m+1/2))=\frac{1}{\sqrt{\pi}}2^{2m}\Gamma(m+1/2)\Gamma(m+1).$$ Furthermore, by Euler’s reflection formula $$\Gamma(k+s/2-m)=\frac{\pi}{(-1)^{k-m}\sin{(\pi s/2)}\Gamma(1-k-s/2+m)}.$$ Finally, $$\begin{gathered}
I(x)=\frac{(-1)^k}{\sqrt{\pi}}\sin\left( \frac{\pi s}{2}\right)x^{1-s}
\sum_{m=0}^{\infty}\frac{1}{m!}\frac{\Gamma(k-s/2+m)}{\Gamma(m+1/2)}\times
\\ \Gamma(1-k-s/2+m)\left(\frac{x^2}{4}\right)^{m}=
\frac{(-1)^k}{\sqrt{\pi}}\sin\left( \frac{\pi s}{2}\right)x^{1-s}
\times \\ \frac{\Gamma(k-s/2)\Gamma(1-k-s/2)}{\Gamma(1/2)}
{}_2F_{1}\left( k-\frac{s}{2},1-k-\frac{s}{2},1/2;\frac{x^2}{4}\right).\end{gathered}$$
Next, we substitute equations , , into expression , proving the exact formula for the shifted first moment.
For any $l \geq 1$ and $2-2k<\Re{s}<2k-1$ one has $$\begin{gathered}
\label{eq:M1ls2}
M_1(l,s)=\frac{\zeta(2s)}{l^{s}}+
\frac{(2\pi)^{s}i^{2k}}{2l^{1-s}}\frac{\Gamma(k-s/2)}{\Gamma(k+s/2)}\mathscr{L}_{-4l^2}(s)+\\
\frac{(2\pi)^s}{\sqrt{\pi}}\frac{\zeta(2s-1)}{l^{1-s}}\cos\left(\frac{\pi s}{2}\right)\frac{\Gamma(k-s/2)\Gamma(k+1/2-s/2)}{\Gamma(k+s/2)\Gamma(k-1/2+s/2)}\Gamma(s-1/2)+\\
\frac{(2\pi)^{s}\sin(\pi s/2)}{\sqrt{\pi}l^{1-s}}\sum_{1\leq n<2l}\mathscr{L}_{n^2-4l^2}(s)
\frac{\Gamma(k-s/2)\Gamma(1-k-s/2)}{\Gamma(1/2)}\times\\
{}_2F_1\left( k-\frac{s}{2},1-k-\frac{s}{2},1/2;\left(\frac{n}{2l}\right)^2\right)+
\frac{2^{2k}\pi^{s}\cos(\pi s/2)}{\sqrt{\pi}}\times \\\sum_{n>2l}\mathscr{L}_{n^2-4l^2}(s)
\frac{\Gamma(k-s/2)\Gamma(k+1/2-s/2)}{\Gamma(2k)} \times \\
\frac{1}{n^{1-s}}\left(\frac{n}{l}\right)^{1-2k}
{}_2F_1\left( k-\frac{s}{2},k+1/2-\frac{s}{2},2k;\left( \frac{2l}{n}\right)^2\right).\end{gathered}$$
Note that in equation only the first and the third summands have poles at $s=1/2$. Computing the limit as $s \rightarrow 1/2$ we find that these poles cancel each other. This allows proving the exact formula for the first moment of symmetric square $L$-functions at the critical point given by Theorem \[thm:explicitformula\].
Liouville-Green approximation of hypergeometric functions
=========================================================
To estimate the functions $\Phi_k(x)$ and $\Psi_k(x)$, appearing in exact formula , we apply the Liouville-Green method. It turns out that these special functions have a similar behavior with the ones occurring in the exact formula for the second moment of cusp form $L$-functions. See [@BF Theorem 4.2].
Properties of $\Phi_k$
----------------------
Consider the function $\Phi_k(x)$ for $0<x<1$.
One has $$\begin{gathered}
\label{eq:decompphi}
\Phi_k(x)=-\pi\Biggl({}_2F_1\left( 2k-\frac{1}{2}, \frac{3}{2}-2k,1;\frac{1-\sqrt{x}}{2}\right)
+\\{}_2F_1\left( 2k-\frac{1}{2}, \frac{3}{2}-2k,1;\frac{1+\sqrt{x}}{2}\right)\Biggr).\end{gathered}$$
Applying the quadratic transformation given by [@HMF Eq. 15.8.27] we obtain $$\begin{gathered}
\Phi_k(x)=\Gamma(k-1/4)\Gamma(3/4-k)\Gamma(k+1/4)\Gamma(5/4-k)\times
\\ \frac{1}{2\Gamma^2(1/2)}\Biggl({}_2F_1\left( 2k-\frac{1}{2}, \frac{3}{2}-2k,1;\frac{1-\sqrt{x}}{2}\right)
+\\{}_2F_1\left( 2k-\frac{1}{2}, \frac{3}{2}-2k,1;\frac{1+\sqrt{x}}{2}\right)\Biggr).\end{gathered}$$ Note that $\Gamma(1/2)=\sqrt{\pi}$. Euler’s reflection formula yields $$\Gamma(k-1/4)\Gamma(k+1/4)\Gamma(3/4-k)\Gamma(5/4-k)=-2\pi^2.$$ The assertion follows.
Making the change of variables $$m:=2k-1/2,\quad k \in \mathbb{N},$$ $$y:=\frac{1-\sqrt{x}}{2}, \quad 0<y<1/2,$$ one has $$\label{eq:reprphi}
\Phi_k(y)=-\pi\left({}_2F_1\left( m, 1-m,1;y\right)
+{}_2F_1\left( m, 1-m,1;1-y\right)\right).$$
At the point $y=0$ the function ${}_2F_1\left( m, 1-m,1;y\right)$ is recessive and ${}_2F_1\left( m, 1-m,1;1-y\right)$ is dominant. Therefore, further transformations are required to apply the Liouville-Green method to the second function. In particular, we show that ${}_2F_1\left( m, 1-m,1;1-y\right)$ has a similar shape with $\phi_k(x)$ studied in [@BF]. Note that the parameter $m$ is now half-integral.
Let $m:=2k-1/2$, $k \in \mathbb{N}$. Then $$\begin{gathered}
{}_2F_1\left( m, 1-m,1;1-y\right)=(-\log{y}+2\psi(1)-2\psi(m))\times \\
{}_2F_1\left( m, 1-m,1;y\right)+
\frac{1}{\pi}\left(\frac{\partial}{\partial a}+\frac{\partial}{\partial b}+2\frac{\partial}{\partial c}\right){}_2F_1\left( a, b, c;y\right)\Bigg{|}_{\substack{a=m\\b=1-m\\c=1}}.\end{gathered}$$
By [@BE Eq. 33, p. 107] we have $$\begin{gathered}
{}_2F_1\left( m+u, 1-m+u,1;1-y\right)=\\
{}_2F_1\left( m+u, 1-m+u,1+2u;y\right)\frac{\Gamma(1)\Gamma(-2u)}{\Gamma(1-m-u)\Gamma(m-u)}+\\
{}_2F_1\left( m-u, 1-m-u,1-2u;y\right)\frac{\Gamma(1)\Gamma(2u)}{\Gamma(1-m+u)\Gamma(m+u)}y^{-2u}.\end{gathered}$$ Computing the limit as $u \rightarrow 0$, we prove the assertion.
\[lem:hyp\] For $m=2k-1/2$, $k \in \mathbb{N}$ one has $$\label{eq:hyp12}
{}_2F_1\left( m, 1-m,1;1/2\right)=-\frac{(-1)^k}{\sqrt{2\pi}}\frac{\Gamma(k-1/4)}{\Gamma(k+1/4)},$$ $$\label{eq:derhyp12}
\frac{d}{dx}\biggl({}_2F_1\left( m, 1-m,1;x\right)\biggr)\bigg|_{x=1/2}=\frac{4(-1)^k}{\sqrt{2\pi}}\frac{\Gamma(k+1/4)}{\Gamma(k-1/4)}.$$
On the one hand, by equation we have $${}_2F_1\left( m, 1-m,1;1/2\right)=-\frac{1}{2\pi}\Phi_k(0).$$ On the other hand, equation yields $$\Phi_k(0)=\frac{\Gamma(k-1/4)\Gamma(3/4-k)}{\Gamma(1/2)}=(-1)^k\sqrt{2\pi}\frac{\Gamma(k-1/4)}{\Gamma(k+1/4)}.$$ The last two equalities imply .
As a consequence of [@HMF Eq. 15.8.25] we obtain $$\begin{gathered}
{}_2F_1\left( m, 1-m,1;x\right)=\frac{\Gamma(1/2)}{\Gamma(m/2+1/2)\Gamma(1-m/2)} \times\\
{}_2F_1\left( m/2, 1/2-m/2,1/2;(1-2x)^2\right)+\\
(1-2x)\frac{\Gamma(-1/2)}{\Gamma(m/2)\Gamma(1/2-m/2)} {}_2F_1\left( m/2+1/2, 1-m/2,3/2;(1-2x)^2\right).\end{gathered}$$ Then equality follows by differentiating the last expression in $x$ and setting $x=1/2$.
\[lem:diffur\] The functions $${}_2F_1\left( m, 1-m,1;y\right)\text{ and }{}_2F_1\left( m, 1-m,1;1-y\right)$$ are solutions of differential equation $$\label{eq:diffur}
y(1-y)F''(y)+(1-2y)F'(y)+m(m-1)F(y)=0.$$
This follows from the differential equation for hypergeometric functions.
Approximation of $\Phi_k$
-------------------------
In order to find a Liouville-Green approximation for $\Phi_k(y)$ we use formula and study separately each of the hypergeometric functions $${}_2F_1\left( m, 1-m,1;y\right)\text{
and }{}_2F_1\left( m, 1-m,1;1-y\right).$$ As shown in Lemma \[lem:diffur\], these functions are solutions of differential equation that was already approximated in [@BF Section 5.2]. So our problem reduces to computation of the Liouville-Green constants $C_Y$ and $C_J$ in the approximation of ${}_2F_1\left( m, 1-m,1;1-y\right).$
For the reader’s convenience, we briefly recall the required results of [@BF Section 5.2]. It follows from Lemma \[lem:diffur\] that the functions $$G_1(y):={}_2F_1\left( m, 1-m,1;y\right)\sqrt{y(1-y)},$$ $$G_2(y):={}_2F_1\left( m, 1-m,1;1-y\right)\sqrt{y(1-y)}$$ are solutions of differential equation $$\label{diffurufg}
G''(y)=(u^2f(y)+g(y))G(y),$$ where $$u:=2k-1,\quad f(y):=-\frac{1}{y(1-y)},$$ $$g(y):=-\frac{1}{4y^2(1-y)^2}+\frac{1}{4y(1-y)}.$$ Making the change of variables $$\label{req2}
Z(y):=\frac{G(y)}{\alpha(y)}, \quad \alpha(y):=\frac{(y-y^2)^{1/4}}{2(\arcsin{\sqrt{y}})^{1/2}},$$ $$\xi:=4\arcsin^2{\sqrt{y}},$$ we transform equation into the following shape $$\label{diffurzxi}
\frac{d^2Z}{d \xi^2}+\left[\frac{u^2}{4\xi}+\frac{1}{4\xi^2}+\frac{\psi(\xi)}{\xi} \right]Z=0$$ with $$\psi(\xi):=\frac{1}{16 \sin^2{\sqrt{\xi}}}-\frac{1}{16\xi}.$$
Removing the summand with $\psi(\xi)/\xi$ in equation , we have $$\label{diffurjy}
\frac{d^2Z}{d \xi^2}+\left[\frac{u^2}{4\xi}+\frac{1}{4\xi^2} \right]Z=0.$$ The solutions of are defined by $$Z_C=\sqrt{\xi}C_0(u\sqrt{\xi}),$$ where $C_i$ is either $J$ or $Y$ Bessel function of index $i$.
Then according to [@O Chapter 12] solutions of original differential equation can be found in the form $$\label{solution}
Z_C(\xi)=\sqrt{\xi}C_0(u\sqrt{\xi})\sum_{n=0}^{\infty}\frac{A(n;\xi)}{u^{2n}}-\frac{\xi}{u}C_1(u\sqrt{\xi})\sum_{n=0}^{\infty}\frac{B(n;\xi)}{u^{2n}}.$$ In order to determine coefficients $A(n;\xi)$, $B(n;\xi)$ we use differential equations (see [@GR Eq. 8.491(3)]) for functions $$W(\xi):=\sqrt{\xi}C_0(u\sqrt{\xi}), \quad V(\xi):=\xi C_1(u\sqrt{\xi})$$ and substitute in equation . This yields $$W(\xi)\sum_{n=0}^{\infty}\frac{C_n(\xi)}{u^{2n}}-V(\xi)\sum_{n=0}^{\infty}\frac{D_n(\xi)}{u^{2n-1}}=0,$$ where $$C_n(\xi):=A''(n;\xi)+\frac{1}{\xi}A'(n;\xi)-\frac{\psi(\xi)}{\xi}A(n;\xi)-B'(n;\xi)-\frac{B(n;\xi)}{2\xi},$$ $$D_n(\xi):=B''(n-1;\xi)+\frac{1}{\xi}B'(n-1;\xi)-\frac{\psi(\xi)}{\xi}B(n-1;\xi)+\frac{1}{\xi}A'(n;\xi).$$
Letting $C_n(\xi)=D_n(\xi)=0$, we find the required recurrence relations $$\label{recurrence1}
A(n;\xi)=-\xi B'(n-1;\xi)+\int_{0}^{\xi}\psi(x)B(n-1;x)dx+\lambda_n,$$ $$\label{recurrence2}
\sqrt{\xi}B(n;\xi)=\int_{0}^{\xi}\frac{1}{\sqrt{x}}\left(xA''(n;x)+A'(n;x)-\psi(x)A(n;x) \right)dx$$ for some real constants of integration $\lambda_n$.
Assume that $A(0;\xi)=1$. Then $$\label{eq:bo}
B(0;\xi)=-\frac{1}{8\sqrt{\xi}}\left(\cot{\sqrt{\xi}}-\frac{1}{\sqrt{\xi}}\right),$$ $$\begin{gathered}
\label{eq:a1}
A(1;\xi)=\frac{1}{8}\left(\frac{1}{\xi}-\frac{\cot{\sqrt{\xi}}}{2\sqrt{\xi}}-\frac{1}{2\sin^2{\sqrt{\xi}}} \right)\\
-\frac{1}{128}\left(\cot{\sqrt{\xi}}-\frac{1}{\sqrt{\xi}} \right)^2+\lambda_1.\end{gathered}$$
Furthermore, solutions can be approximated by finite series using [@O Theorem 4.1, p. 444] or [@BD Theorem 1].
\[LGphi\] Let $\xi_2=\pi^2/4$. For each value of $u$ and each nonnegative integer $N$ equation has solutions $Z_Y(\xi)$, $Z_J(\xi)$ which are infinitely differentiable in $\xi$ on interval $(0, \xi_2)$, and are given by $$\begin{gathered}
\label{zyxi}
Z_Y(\xi)=\sqrt{\xi}Y_0(u\sqrt{\xi})\sum_{n=0}^{N}\frac{A_Y(n;\xi)}{u^{2n}}-\\
\frac{\xi}{u}Y_1(u\sqrt{\xi})\sum_{n=0}^{N-1}\frac{B_Y(n;\xi)}{u^{2n}}+\epsilon_{2N+1,1}(u,\xi),\end{gathered}$$ $$\begin{gathered}
\label{eq:zjapprox}
Z_J(\xi)=\sqrt{\xi}J_0(u\sqrt{\xi})\sum_{n=0}^{N}\frac{A_J(n;\xi)}{u^{2n}}-\\
\frac{\xi}{u}J_1(u\sqrt{\xi})\sum_{n=0}^{N-1}\frac{B_J(n;\xi)}{u^{2n}}+\epsilon_{2N+1,2}(u,\xi),\end{gathered}$$ where $$\label{zyxi2}
\epsilon_{2N+1,1}(u,\xi)\ll \frac{\sqrt{\xi}|Y_0(u\sqrt{\xi})|}{u^{2N+1}}\sqrt{\xi_2-\xi},$$ $$\epsilon_{2N+1,2}(u,\xi)\ll \frac{\sqrt{\xi}|J_0(u\sqrt{\xi})|}{u^{2N+1}}\min{(\sqrt{\xi},1)}$$ and coefficients $(A_Y(n;\xi),B_Y(n;\xi))$, $(A_J(n;\xi),B_J(n;\xi))$ are defined by -.
Functions $\xi^{1/4}(\sin{\sqrt{\xi}})^{1/2}G_1(\sin^2{\sqrt{\xi}/2})$ and $Z_J(\xi)$ are recessive solutions of equation as $\xi \rightarrow 0$. Therefore, there is $c_0$ such that $$\label{f1coeff}
\xi^{1/4}(\sin{\sqrt{\xi}})^{1/2}G_1(\sin^2{\sqrt{\xi}/2})=c_0Z_J(\xi).$$ The value of constant $c_0$ is determined by computing the limit of the left and right-hand sides of equation as $\xi \rightarrow 0$. On the one hand, $$\lim_{\xi \rightarrow 0}{}_2F_{1}(k,1-k,1;\sin^2{\sqrt{\xi}/2})=1.$$ On the other hand, $$Z_J(\xi)=\sqrt{\xi}\sum_{n=0}^{N}\frac{A_J(n;\xi)}{u^{2n}}+O(\xi) \text{ as } \xi \rightarrow 0.$$ Choosing $A_J(n;\xi)$ such that $A_J(0;0)=1$ and $A_J(n;0)=0$ for $n \geq 1$ we find that $c_0=1$.
To sum up, we proved the following Lemma.
\[lem:g1\] Let $\xi_2=\pi^2/4$. For $\xi \in (0,\xi_2)$ one has $$\xi^{1/4}(\sin\sqrt{\xi})^{1/2}G_1\left( \sin^2{\frac{\sqrt{\xi}}{2}}\right)=Z_J(\xi),$$ where $Z_J(\xi)$ is given by .
This concludes the summary of results of [@BF Section 5.2]. Our final goal is to compute $C_Y=C_Y(u)$ and $C_J=C_J(u)$ such that $$\label{eq:cycj2}
\xi^{1/4}(\sin{\sqrt{\xi}})^{1/2}G_2(\sin^2{\sqrt{\xi}/2})=C_YZ_Y(\xi)+C_JZ_J(\xi).$$
Note that there exist $c_1,c_2$ such that $$\begin{gathered}
\label{eq:zyxi}
Z_Y(\xi)=\left(c_1G_1\left( \sin^2{\frac{\sqrt{\xi}}{2}}\right)+c_2G_2\left( \sin^2{\frac{\sqrt{\xi}}{2}}\right) \right)\times \\ \xi^{1/4}(\sin\sqrt{\xi})^{1/2} =
c_1Z_J(\xi)+\xi^{1/4}(\sin\sqrt{\xi})^{1/2}c_2G_2\left( \sin^2{\frac{\sqrt{\xi}}{2}}\right).\end{gathered}$$ The last two equalities imply $$\label{eq:CYCJ}
C_Y=\frac{1}{c_2},\quad C_J=-\frac{c_1}{c_2}.$$
\[lem:c1c2\] One has $$\begin{gathered}
2c_1=(-1)^{k+1}\sqrt{2\pi}\frac{\Gamma(k+1/4)}{\Gamma(k-1/4)}\frac{Z_Y(\xi_2)}{\xi_{2}^{1/4}}+(-1)^k\sqrt{2\pi}\times \\\frac{\Gamma(k-1/4)}{\Gamma(k+1/4)} \xi_{2}^{1/4}\left(Z_Y'(\xi_2)-\frac{Z_Y(\xi_2)}{4\xi_2} \right),\end{gathered}$$ $$\begin{gathered}
2c_2=(-1)^{k+1}\sqrt{2\pi}\frac{\Gamma(k+1/4)}{\Gamma(k-1/4)}\frac{Z_Y(\xi_2)}{\xi_{2}^{1/4}}-(-1)^k\sqrt{2\pi}\times \\\frac{\Gamma(k-1/4)}{\Gamma(k+1/4)} \xi_{2}^{1/4}\left(Z_Y'(\xi_2)-\frac{Z_Y(\xi_2)}{4\xi_2} \right).\end{gathered}$$
To determine coefficients $c_1,c_2$ we consider the pair of equations $$Z_Y(\xi_2)=\xi_{2}^{1/4}\left( c_1G_1(1/2)+c_2G_2(1/2)\right),$$ $$Z_Y'(\xi_2)=\frac{Z_Y(\xi_2)}{4\xi_2}+\frac{1}{4\xi_{2}^{1/4}}\left( c_1G_1'(1/2)+c_2G_2'(1/2)\right).$$ Note that $$G_1(1/2)=G_2(1/2)\text{ and }G_1'(1/2)=-G_2'(1/2).$$ Therefore, $$(c_1+c_2)G_1(1/2)=\frac{Z_Y(\xi_2)}{\xi_{2}^{1/4}},$$ $$(c_1-c_2)G_1'(1/2)=4\xi_{2}^{1/4}\left(Z_Y'(\xi_2)-\frac{Z_Y(\xi_2)}{4\xi_{2}}\right).$$ The assertion follows by Lemma \[lem:hyp\].
\[lem:zyzyd\] For $\xi_2=\pi^2/4$ one has $$\label{eq:zyxi2}
Z_Y(\xi_2)=\frac{(-1)^{k+1}}{\sqrt{2u}}\left[1+\frac{1}{u^2}\left( \lambda_1-\frac{1}{16}\right)+O\left( \frac{1}{u^3}\right)\right],$$ $$\label{eq:zyxi2d}
Z_{Y}'(\xi_2)=\frac{(-1)^{k+1}}{\sqrt{2u}}\left[\frac{u}{\pi}+\frac{1}{\pi^2}+\frac{2\lambda_1+1/8}{2\pi u}
+O\left(\frac{1}{u^2} \right)\right].$$
Applying [@BD Theorem 1] with $N=1$ we obtain $$\epsilon_{3,1}(u;\xi_2)=0 \text{ and }\frac{\partial}{\partial \xi}\epsilon_{3,1}(u;\xi)\bigg|_{\xi=\xi_2}=0.$$ Therefore, $$Z_Y(\xi_2)=\sqrt{\xi_2}Y_0(u\sqrt{\xi_2})\left(1+\frac{A_Y(1,\xi_2)}{u^2}\right)-
\xi_2Y_1(u\sqrt{\xi_2})\frac{B_Y(0,\xi_2)}{u}$$ and $$\begin{gathered}
Z_Y'(\xi_2)=\sqrt{\xi_2}Y_0(u\sqrt{\xi_2})\biggl(\frac{1}{2\xi_2}\left[1+\frac{A_Y(1;\xi_2)}{u^{2}}
\right]+\frac{A_Y'(1;\xi_2)}{u^{2}}
-\\ \frac{1}{2}B_Y(0;\xi_2) \biggr)-
\xi_2Y_1(u\sqrt{\xi_2})\biggl( \frac{u}{2\xi_2}\left[1+\frac{A_Y(1;\xi_2)}{u^{2}}
\right]+\\\frac{1}{2\xi_2 u}B_Y(0;\xi_2)
+\frac{1}{u}B_Y'(0;\xi_2) \biggr).\end{gathered}$$
By means of the Hankel asymptotic expansion (see [@HMF Eq. 10.17.1, 10.17.4] and [@GR Eq. 8.451(1,7,8)]) we evaluate $$\begin{gathered}
\sqrt{\xi_2}Y_0(u\sqrt{\xi_2})=\frac{\pi}{2}Y_0\left((2k-1)\frac{\pi}{2}\right)=\\
\frac{(-1)^{k+1}}{\sqrt{2u}}\left[\sum_{j=0}^{\infty}(-1)^j\frac{a_{2j}(0)}{(\pi u/2)^{2j}}
+\sum_{j=0}^{\infty}(-1)^j\frac{a_{2j+1}(0)}{(\pi u/2)^{2j+1}}\right],\end{gathered}$$ $$\xi_2Y_1(u\sqrt{\xi_2})=\frac{\pi}{2}\frac{(-1)^{k}}{\sqrt{2u}}\left[\sum_{j=0}^{\infty}(-1)^j\frac{a_{2j}(1)}{(\pi u/2)^{2j}}
-\sum_{j=0}^{\infty}(-1)^j\frac{a_{2j+1}(1)}{(\pi u/2)^{2j+1}}\right],$$ where $$a_j(v)=\frac{\Gamma(v+j+1/2)}{2^j j! \Gamma(v-j+1/2)}.$$ This yields $$\begin{gathered}
Z_Y(\xi_2)=\frac{(-1)^{k+1}}{\sqrt{2u}}\left[1+\frac{a_1(0)}{\pi u/2}-\frac{a_2(0)}{(\pi u/2)^2}+O(u^{-3})\right]
\left(1+\frac{A_Y(1;\xi_2)}{u^2} \right)-\\
\frac{\pi}{2}\frac{(-1)^k}{\sqrt{2u}}\left[1-\frac{a_1(1)}{\pi u/2}-\frac{a_2(1)}{(\pi u/2)^2}+O(u^{-3})\right]\frac{B_Y(0,\xi_2)}{u}.\end{gathered}$$ Simplifying the expression above, one has $$\begin{gathered}
Z_Y(\xi_2)=\frac{(-1)^{k+1}}{\sqrt{2u}}\biggl[1+\frac{1}{u}\left(\frac{2}{\pi}a_1(0)+\frac{\pi}{2}B_Y(0,\xi_2) \right)+\\ \frac{1}{u^2}\left(A_Y(1,\xi_2)-\frac{4}{\pi^2}a_2(0)-a_1(1)B_Y(0,\xi_2) \right)\biggr]+O(u^{-7/2}).\end{gathered}$$ Using formulas and , we find $$a_1(0)=-\frac{1}{8},\quad \frac{2}{\pi}a_1(0)+\frac{\pi}{2}B_Y(0,\xi_2)=0,$$ $$A_Y(1,\xi_2)-\frac{4}{\pi^2}a_2(0)-a_1(1)B_Y(0,\xi_2)=\lambda_1-\frac{1}{16}.$$ This gives equation .
Similarly, we obtain $$\begin{gathered}
Z_Y'(\xi_2)=\sqrt{\xi_2}Y_0(u\sqrt{\xi_2})\left[\frac{7}{4\pi^2}+\frac{1}{u^2}\left( \frac{1}{\pi^2}\left(2\lambda_1-\frac{1}{32}\right)-\frac{15}{8\pi^4}\right)\right]-\\
\xi_2Y_1(u\sqrt{\xi_2})\left[u\frac{2}{\pi^2}+\frac{1}{u}\left(\frac{1}{\pi^2}\left( 2\lambda_1+\frac{1}{8}\right)-\frac{1}{16\pi^4} \right)\right].\end{gathered}$$ Hankel’s expansion for Bessel functions yields the following asymptotics $$\begin{gathered}
Z_Y'(\xi_2)=\frac{(-1)^{k+1}}{\sqrt{2u}}\biggl[u\frac{1}{\pi}+\left(\frac{7}{4\pi^2}-\frac{2}{\pi^2}a_1(1)\right)+\\ \frac{1}{u}\left( \frac{7}{2\pi^3}a_1(0)+\frac{\pi}{2}\left(\frac{2\lambda_1+1/8}{\pi^2}-\frac{1}{16\pi^4} \right)-\frac{4}{\pi^3}a_2(1)\right)+\\
\frac{1}{u^2}\left( -\frac{7}{\pi^4}a_2(0)-a_1(1)\left( \frac{2\lambda_1+1/8}{\pi^2}-\frac{1}{16\pi^4}\right)+\frac{8}{\pi^4}a_3(1)\right)+O(u^{-3}) \biggr].\end{gathered}$$ Finally, substituting $$a_1(0)=-\frac{1}{8}, \quad a_1(1)=\frac{3}{8}, \quad a_2(1)=-\frac{15}{128}$$ we prove equation .
\[corcjcy\] One has $$C_Y=1+O\left(\frac{1}{k}\right), \quad C_J=O\left(\frac{1}{k^2} \right).$$
By Lemma \[lem:zyzyd\] $$Z_Y'(\xi_2)-\frac{Z_Y(\xi_2)}{4\xi_2}=\frac{(-1)^{k+1}}{\sqrt{2u}}\left(\frac{u}{\pi}+\frac{2\lambda_1+1/8}{2\pi u}+O(u^{-2})\right).$$
It follows from [@HMF Eq. 5.11.13] that $$\frac{\Gamma(k+1/4)}{\Gamma(k-1/4)}=k^{1/2}-\frac{1}{4}k^{-1/2}+O(k^{-3/2})$$ and $$\frac{\Gamma(k-1/4)}{\Gamma(k+1/4)}=k^{-1/2}+O(k^{-3/2}).$$ Then Lemma \[lem:c1c2\] gives $$c_1=O(k^{-2}), \quad c_2=1+O(k^{-1}).$$ Equations yield the assertion.
As a consequence of equation , Lemma \[lem:g1\], Theorem \[LGphi\] and Corollary \[corcjcy\] we obtain the main result.
\[thm:apprphi\] Let $u=2k-1$, $\xi_2=\pi^2/4$. Then for $\xi \in (0,\xi_2)$ one has $$\Phi_k(\cos^2\sqrt{\xi})=\frac{-\pi}{\xi^{1/4}(\sin\sqrt{\xi})^{1/2}}\left[Z_J(\xi)+C_YZ_Y(\xi)+C_JZ_J(\xi)\right],$$ where $Z_Y$, $Z_J$ are given by , and $$C_Y=1+O\left(\frac{1}{k}\right), \quad C_J=O\left(\frac{1}{k^2} \right).$$
Approximation of $\Psi_k$
-------------------------
Next, we find a Liouville-Green approximation for the function $\Psi_k$. With this goal, we follow the arguments of [@BF Section 5.3] with minor changes. In particular, the differential equation for $\Psi_k(x)$ is slightly different, and, therefore, one requires to recompute various functions and constants appearing in the Liouville-Green approximation. We provide all details here to make the presentation self-contained.
Consider the function $$y(x):=\sqrt{1-x}\Psi_k(x).$$ Let $u:=k-1/2$ and $$f(x):=\frac{1}{x^2(1-x)}, \quad g(x):=-\frac{1}{4x^2(1-x)^2}+\frac{3}{16x(1-x)}.$$
The function $y=y(x)$ is a solution of equation $$\label{eq:diffurf21}
y''(x)-(u^2f(x)+g(x))y(x)=0.$$
Using the differential equation for the hypergeometric function, we find that $y=y(x)$ satisfies the following differential equation $$y''+\biggl( \frac{1-(2k-1)^2}{4x^2} +\frac{1}{4(1-x)^2}+\frac{5/4-(2k-1)^2}{4x(1-x)}\biggr)y=0.$$ The assertion follows.
Making the change $$Z(x):=\frac{y(x)}{\alpha(x)}, \quad \alpha(x):=\frac{(x^2-x^3)^{1/4}}{2(\operatorname{artanh}{\sqrt{1-x}})^{1/2}}$$ and the substitution $$\xi:=4 \operatorname{artanh}^2{\sqrt{1-x}},$$ we transform equation to the type $$\label{eq:diffurZpsi}
\frac{d^2Z}{d\xi^2}+\left[ -\frac{u^2}{4\xi}+\frac{1}{4\xi^2}-\frac{\psi(\xi)}{\xi}\right]Z=0,$$ where $$\psi(\xi)=\frac{1}{16}\left(\frac{1}{\xi}-\frac{1}{4\sinh^2{\sqrt{\xi}/2}} \right)$$ is an analytic function as $\xi \rightarrow 0$.
In order to find a Liouville-Green approximation to equation , we remove the term with $\psi(\xi)/\xi$ in . The resulting equation $$Z''+\left(-\frac{u^2}{4\xi}+\frac{1}{4\xi} \right)Z=0$$ has $I$ and $K$ Bessel functions as solutions (see [@HMF Eq. 10.13.2]), namely $$Z_L=\sqrt{\xi}L_0(u\sqrt{\xi}),$$ where $$L_v(x):=\begin{cases}
I_v(x)\\e^{\pi i v}K_v(x)
\end{cases}.$$ This suggests that solutions of the original differential equation can be written in the form (see [@O Eq. 2.09, Chapter 12]) $$\label{eq:solZ}
Z_L=\sqrt{\xi}L_0(u\sqrt{\xi})\sum_{n=0}^{\infty}\frac{A(n;\xi)}{u^{2n}}+\frac{\xi}{u}L_1(u\sqrt{\xi})
\sum_{n=0}^{\infty}\frac{B(n;\xi)}{u^{2n}}.$$
Coefficients $A(n;\xi)$ and $B(n;\xi)$ are given by $$\begin{gathered}
\label{rec:bnxi}
\sqrt{\xi}B(n;\xi)=-\sqrt{\xi}A'(n;\xi)+\\\int_{0}^{\xi}\left(\psi(x) A(n;x)-\frac{1}{2}A'(n;x)\right)\frac{dx}{\sqrt{x}},\end{gathered}$$ $$A(n;\xi)=-\xi B'(n-1;\xi)+\int_{0}^{\xi}\psi(x)B(n-1;x)dx+\lambda_n$$ for some real constants of integration $\lambda_n$.
By [@HMF Eq. 10.13.2, 10.13.5, 10.36, 10.29.2, 10.29.3] the functions $$W(\xi):=\sqrt{\xi}L_0(u\sqrt{\xi}),\quad V(\xi):=\xi L_1(u\sqrt{\xi})$$ satisfy the following relations $$W''+\left( -\frac{u^2}{4\xi}+\frac{1}{4\xi^2}\right)W=0,$$ $$V''-\frac{1}{\xi}V'+\left( -\frac{u^2}{4\xi}+\frac{3}{4\xi^2}\right)V=0,$$ $$V'=\frac{1}{2\xi}V+\frac{u}{2}W, \quad W'=\frac{1}{2\xi}W+\frac{u}{2\xi}V.$$
Using this and substituting solution into equation , we obtain that $$W(\xi)\sum_{n=0}^{\infty}\frac{C(n;\xi)}{u^{2n}}+V(\xi)\sum_{n=0}^{\infty}\frac{D(n;\xi)}{u^{2n+1}}=0,$$ where $$C(n;\xi)=A''(n;\xi)+\frac{A'(n;\xi)}{\xi}-\frac{\psi(\xi)}{\xi}A(n;\xi)+B'(n;\xi)+\frac{B(n;\xi)}{2\xi},$$ $$D(n;\xi)=B''(n-1;\xi)+\frac{B'(n-1;\xi)}{\xi}-\frac{\psi(\xi)}{\xi}B(n-1;\xi)+\frac{A'(n;\xi)}{\xi}.$$
Setting $C(n;\xi)=D(n;\xi)=0$ we find the required recurrence relations.
Let $A(0;\xi)=1$. Then $$B(0;\xi)=\frac{1}{16}\left( \frac{\coth{\sqrt{\xi}}}{\sqrt{\xi}}-\frac{2}{\xi}\right),$$
$$\begin{gathered}
A(1;\xi)=-\frac{1}{32}\left( \frac{4}{\xi}-\frac{\coth{\sqrt{\xi/4}}}{\sqrt{\xi}}-\frac{1}{2\sinh^2{\sqrt{\xi/4}}}\right)+\\
\frac{1}{512}\left( \coth{\sqrt{\xi/4}}-\frac{2}{\sqrt{\xi}}\right)^2+\lambda_1.\end{gathered}$$
Note that $$\label{eq:baxi0}
\lim_{\xi \rightarrow \infty}\sqrt{\xi}B(0;\xi)=\frac{1}{16}, \quad \lim_{\xi \rightarrow \infty}A(1;\xi)=\frac{1}{512}+\lambda_1.$$
The variation of the function is given by $$V_{a,b}(f(x)):=\int_{a}^{b}|(f(x))'|dx.$$
\[lem:variation\] The function $V_{\xi,\infty}(\sqrt{x}B(1;x))$ is bounded. For $n>1$ the function $V_{\xi,\infty}(\sqrt{x}B(n;x))$ converges.
As a consequence of recurrence relation , we find $$(\sqrt{x}B(1;x))'=O(x^{-1/2}) \text{ as } x \rightarrow 0$$ and $$(\sqrt{x}B(1;x))'=O(x^{-2}) \text{ as } x \rightarrow \infty.$$ Thus $V_{\xi,\infty}(\sqrt{x}B(1;x))$ is bounded. Note that $$\psi^{(s)}(\xi)=O\left(\frac{1}{|\xi|^{s+1}} \right).$$ The convergence of variation for $n>1$ then follows by [@O Exercise 4.2, p. 445].
Using Lemma \[lem:variation\], we can truncate the infinite summation in up to $N$ summands with a negligible error term. The value of $N$ determines the quality of approximation: the error is smaller for larger $N$.
\[thm:zk\] For each value of $u$ and each nonnegative integer $N$ equation has solution $Z_K(\xi)$ which is infinitely differentiable in $\xi$ on interval $(0, \infty)$ and is given by $$\begin{gathered}
\label{eq:zkxi}
Z_K(\xi)=\sqrt{\xi}K_0(u\sqrt{\xi})\sum_{n=0}^{N}\frac{A_K(n;\xi)}{u^{2n}}-\\\frac{\xi}{u}K_1(u\sqrt{\xi})
\sum_{n=0}^{N-1}\frac{B_K(n;\xi)}{u^{2n}}+\epsilon_{2N+1,3}(u,\xi),\end{gathered}$$ where $$\begin{gathered}
|\epsilon_{2N+1,3}(u,\xi)|\leq \frac{\sqrt{\xi}K_0(u\sqrt{\xi})}{u^{2N+1}}\times \\ V_{\xi,\infty}(\sqrt{\xi}B_K(N;\xi))exp\left( \frac{1}{u}V_{\xi, \infty}(\sqrt{\xi}B_K(0;\xi))\right).\end{gathered}$$ In particular, for $N=1$ $$\epsilon_{3,3}(u,\xi)\ll \frac{\sqrt{\xi}K_0(u\sqrt{\xi})}{u^{3}}\min\left(\sqrt{\xi}, \frac{1}{\xi}\right).$$
As $\xi \rightarrow \infty$, differential equation has two recessive solutions, namely $$Z(\xi)=
\Psi_k\left(\frac{1}{\cosh^2{\sqrt{\xi}/2}} \right)\left( \xi\sinh^2{\sqrt{\xi}}\right)^{1/4}$$ and $Z_K(\xi)$ given by . Thus there is $C_K=C_K(u)$ such that $$\label{eq:phiklim}
\Psi_k\left(\frac{1}{\cosh^2{\sqrt{\xi}/2}} \right)\left( \xi\sinh^2{\sqrt{\xi}}\right)^{1/4}=
C_KZ_K(\xi).$$
The last step is to compute $C_K=C_K(u)$.
One has $$\label{asymp:ck}
C_K=2+O(k^{-1}).$$
To determine $C_K$, we compute the limit of the left and right- hand sides of equation as $\xi \rightarrow \infty$. This implies $$\label{eq:ckexpl}
C_K=\frac{\Gamma(k-1/4)\Gamma(k+1/4)}{\Gamma(2k)}\frac{2^{2k}\sqrt{u}}{\sqrt{\pi}}\left[\sum_{n=0}^{N}\frac{a_n}{u^{2n}}-\sum_{n=0}^{N-1}\frac{b_n}{u^{2n+1}}\right]^{-1},$$ where $$a_n=\lim_{\xi \rightarrow \infty}A(n;\xi), \quad b_n=\lim_{\xi \rightarrow \infty}B(n;\xi)\sqrt{\xi}.$$ According to we know that $$a_0=1, \quad a_1=\frac{1}{512}+\lambda_1, \quad b_0=\frac{1}{16}.$$ Furthermore, $$\frac{\Gamma(k-1/4)\Gamma(k+1/4)}{ \Gamma(2k)}= \frac{\Gamma^2(k)}{\Gamma(2k)} (1+O(1/k))=\frac{2\sqrt{\pi}}{\sqrt{k}2^{2k}}(1+O(1/k)).$$ The assertion follows.
Finally, we obtain the main Theorem.
\[thm:approxPsi\] For $\xi \in (0, \infty)$ the following equality holds $$\label{eq:phiklim1}
\Psi_k\left(\frac{1}{\cosh^2{\sqrt{\xi}/2}} \right)\left( \xi\sinh^2{\sqrt{\xi}}\right)^{1/4}=
C_KZ_K(\xi),$$ where $Z_K(\xi)$ is defined by and $C_K=2+O(k^{-1})$.
Asymptotic formula
==================
Corollaries \[thm:asympformula2\] and \[thm:asympformula\] are derived from Theorem \[thm:explicitformula\] by estimating the last two summands in exact formula , as we now show.
For any $\epsilon>0$ one has $$E_1(k,l):=\frac{1}{\sqrt{l}}\sum_{1\leq n<2l}\mathscr{L}_{n^2-4l^2}(1/2)\Phi_k\left( \frac{n^2}{4l^2}\right)\ll
\frac{l^{5/6+\epsilon}}{\sqrt{k}}.$$
Using subconvexity bound we obtain $$\begin{gathered}
E_1(k,l)\ll \frac{l^{\epsilon}}{\sqrt{l}}\sum_{1\leq n<2l}(2l-n)^{1/6}(2l+n)^{1/6}
\left|\Phi_k\left( \frac{n^2}{4l^2}\right)\right|\ll\\
\frac{l^{1/6+\epsilon}}{\sqrt{l}}\sum_{1\leq n<2l}n^{1/6}
\left|\Phi_k\left(\left(1- \frac{n}{2l}\right)^2\right)\right|\end{gathered}$$
Let $$\xi=4\left(\arcsin{\sqrt{\frac{n}{4l}}} \right)^2, \quad u=2k-1.$$ Then by Theorem \[thm:apprphi\] one has $$\begin{gathered}
\Phi_k\left(\left(1- \frac{n}{2l}\right)^2\right)\ll \frac{\sqrt{\xi}Y_0(u\sqrt{\xi})}{(\arcsin(\sqrt{n/4l}))^{1/2}(n/l)^{1/4}}\ll\\
\frac{(\arcsin(\sqrt{n/4l}))^{1/2}}{(n/l)^{1/4}}Y_0(u\sqrt{\xi}).\end{gathered}$$ If $l\ll k^2$, one has $u\sqrt{\xi}\gg 1$. Then the estimate for the Bessel function $$Y_0(u\sqrt{\xi})\ll \frac{1}{u^{1/2}\xi^{1/4}}$$ yields $$\Phi_k\left(\left(1- \frac{n}{2l}\right)^2\right)\ll\frac{1}{k^{1/2}(n/l)^{1/4}}.$$ Consequently, $$E_1(k,l)\ll \frac{l^{1/6+\epsilon}}{\sqrt{l}}\sum_{n<2l}\frac{n^{1/6}l^{1/4}}{k^{1/2}n^{1/4}}\ll
\frac{l^{5/6+\epsilon}}{\sqrt{k}}.$$
If the Lindelöf hypothesis for Dirichlet L-functions is true, then for any $\epsilon>0$ one has $$E_1(k,l)\ll \frac{l^{1/2+\epsilon}}{\sqrt{k}}.$$
For some $c>0$ one has $$\begin{gathered}
E_2(k,l):=\frac{1}{l\sqrt{2}}\sum_{n>2l}\mathscr{L}_{n^2-4l^2}(1/2)\sqrt{n}\Psi_k\left( \frac{4l^2}{n^2}\right)\ll\\
\frac{l^{-1/12}}{\sqrt{k}}\exp\left(-\frac{ck}{\sqrt{l}} \right).\end{gathered}$$
It follows from subconvexity bound that $$\begin{gathered}
E_2(k,l)\ll \frac{1}{l}\sum_{n>2l}(n^2-4l^2)^{1/6}\sqrt{n}\left|\Psi_k\left( \frac{4l^2}{n^2}\right)\right|\ll\\
\frac{1}{l}\int_{2l+1}^{\infty}x^{1/2}(x^2-4l^2)^{1/6}\left|\Psi_k\left( \frac{4l^2}{x^2}\right)\right|dx+
l^{-1/3}\left| \Psi_k\left( \frac{4l^2}{(2l+1)^2}\right)\right|.\end{gathered}$$ Next, we make the change of variables $$x=2l\cosh{\frac{\sqrt{\xi}}{2}}$$ and estimate $E_2(k,l)$ using Theorem \[thm:approxPsi\] with $N=0$. Consider the first summand $$\begin{gathered}
E_{2,1}(k,l):=\frac{1}{l}\int_{2l+1}^{\infty}x^{1/2}(x^2-4l^2)^{1/6}\left|\Psi_k\left( \frac{4l^2}{x^2}\right)\right|dx\ll
l^{5/6}\times \\\int_{\xi_0}^{\infty}\left(\sinh{\frac{\sqrt{\xi}}{2}}\right)^{5/6}\frac{|Z_k(\xi)|}{\xi^{3/4}}d\xi\ll
l^{5/6}\int_{\xi_0}^{\infty}\left(\sinh{\frac{\sqrt{\xi}}{2}}\right)^{5/6}\frac{|K_0(u\sqrt{\xi})|}{\xi^{1/4}}d\xi,\end{gathered}$$ where $u=k-1/2$ and the limit of integration $\xi_0$ is defined by $$\cosh{\frac{\sqrt{\xi_0}}{2}}=1+\frac{1}{2l}.$$
Making the change of variables $\sqrt{\xi}=t$, one has $$t_0=4\operatorname{arcsinh}{\frac{1}{\sqrt{4l}}}.$$ Since $t_0 \gg 1/\sqrt{l}$ and $ut\geq ut_0\gg k/\sqrt{l}\gg 1$, we estimate the Bessel function as follows $$K_0(ut)\ll \frac{\exp(-ut)}{\sqrt{ut}}.$$ Finally, $$\begin{gathered}
E_{2,1}(k,l)\ll
l^{5/6}\int_{t_0}^{\infty}\left(\sinh{\frac{t}{2}}\right)^{5/6}\frac{|K_0(ut)|}{\sqrt{t}}tdt\ll\\
\frac{l^{5/6}}{\sqrt{k}}\int_{t_0}^{\infty}\left(\sinh{\frac{t}{2}}\right)^{5/6}e^{-ut}dt\ll
\frac{l^{5/6}}{\sqrt{k}}\times \\\left( \int_{t_0}^{1}t^{5/6}\exp{(-ut)}dt+\int_{1}^{\infty}\exp{(-ut+5t/12)}dt\right)
\ll \frac{l^{5/12}}{u\sqrt{k}}\exp(-ut_0).\end{gathered}$$ The second summand can be estimated similarly: $$\begin{gathered}
E_{2,2}(k,l):=l^{-1/3}\left| \Psi_k\left( \frac{4l^2}{(2l+1)^2}\right)\right|\ll
l^{-1/3}\frac{C_K|Z_k(\xi)|}{\xi^{1/4}(\sinh{\sqrt{\xi}})^{1/2}}\ll\\
l^{-1/3}l^{1/2}|Z_K(\xi)|\ll l^{1/6}\sqrt{\xi}|K_0(u\sqrt{\xi})|\ll \frac{\exp{(-ut_0)}}{k^{1/2}l^{1/12}}.\end{gathered}$$ To sum up, $$E_{2}(k,l)\ll E_{2,1}(k,l)+E_{2,2}(k,l)\ll \frac{1}{l^{1/12}\sqrt{k}}\exp\left(\frac{-ck}{\sqrt{l}}\right)$$ for some $c>0$.
O. Balkanova, D. Frolenkov, *Moments of $L$-functions and the Liouville-Green method*, arXiv:1610.03465 \[math.NT\] .
H. Beitman and A. Erdelyi, *Higher transcendental functions*, Vol. 1, McGraw-Hill, New York, 1953.
W. G .C. Boyd and T. M. Dunster, *Uniform asymptotic solutions of a class of second-order linear differential equations having a turning point and a regular singularity, with an application to Legendre functions*, SIAM J. Math. Anal. 17-2 (1986), 422–450.
V.A. Bykovskii, *Density theorems and the mean value of arithmetic functions on short intervals*. (Russian) Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI) 212 (1994), Anal. Teor. Chisel i Teor. Funktsii. 12, 56–70, 196; translation in J. Math. Sci. (New York) 83 (1997), no. 6, 720–730.
B. Conrey, D. W. Farmer, J. P. Keating, M. O. Rubinstein, and N. C. Snaith, *Integral moments of $L$-functions*, Proc. London Math. Soc. 91 (2005), 33–104.
J. B. Conrey and H. Iwaniec, *The cubic moment of central values of automorphic $L$-functions*, Ann. of Math. (2) 151 (2000), 1175–1216.
A. Diaconu, D. Goldfeld, and J. Hoffstein, *Multiple Dirichlet series and moments of zeta and L-functions*, Compositio Math. 139 (2003), 297–360.
O.M. Fomenko, *Behavior of Automorphic L-Functions at the Points s=1 and s=1/2*, J Math Sci (2005) 129: 3898–3909.
S. Gelbart and H. Jacquet, *A relation between automorphic representations of $GL(2)$ and $GL(3)$*, Ann. Sci. École Norm. Sup. (4) 11 (1978), no. 4, 471–542.
I. S. Gradshteyn and I. M. Ryzhik, *Table of Integrals, Series, and Products*. Edited by A. Jeffrey and D. Zwillinger. Academic Press, New York, 7th edition, 2007.
R. Khan, *The first moment of the symmetric-square L-function*, J. Number Theory 124 (2007), 259–266.
R. Khan, *Non-vanishing of the symmetric square $L$-function at the central point*, Proc. London Math. Soc. (2010) 100 (3): 736–762.
W. Kohnen, J. Sengupta, *On the average of central values of symmetric square L-functions in weight aspect*, Nagoya Math. J. 167, 2002, 95–100.
Y.-K. Lau, *Non-vanishing of symmetric square L-functions*, Proc. Amer. Math. Soc. 130 (11) (2002) 3133–3139.
M. Lerch, *Note sur la fonction $R(w,x,s)=\sum_{k=0}^{\infty}\frac{e^{2\pi ina}}{(n+c)^s}$*, Acta Math. 11 (1887), 19–24.
M. Lipschitz, *Untersuchung einer aus vier Elementen gebildeten Reihe*, J. Reine Angew. 54 (1857), 313–328.
S. Liu, *The first moment of central values of symmetric-square L-functions in the weight aspect*, arXiv:1610.07652 \[math.NT\].
W. Luo, *Central values of the symmetric square L-functions*, Proc. Amer. Math. Soc., 140, (2012), no.10, 3313–3322.
Ng Ming Ho, *Moments of automorphic $L$-functions*, PhD thesis, University of Hong Kong, 2016.
F. W. J. Olver, *Asymptotics and Special Functions*, Academic Press, New York, 1974.
F.W.J. Olver , D.W. Lozier, R.F. Boisvert and C.W. Clarke, *[NIST]{} [H]{}andbook of [M]{}athematical [F]{}unctions*, Cambridge University Press, Cambridge $(2010)$.
H. Petersson, *Über die Entwicklungskoeffizienten der automorphen Formen*, Acta Math. 58 (1932), no. 1, 169–215.
G. Shimura, *On the holomorphy of certain Dirichlet series*, Proc. London Math. Soc. (3) 31 (1975) 79–98.
K. Soundararajan, M. P. Young, *The prime geodesic theorem*, J. Reine Angew. Math. 676 (2013), 105–120.
Q. Sun, *On the first moment of symmetric-square L-functions*, Proc. Amer. Math. Soc., 141, (2013), no.2, 369–375.
A. Weil, *On some exponential sums*, Proc. Natl. Acad. Sci. USA 34, 1948, 204–207.
D. Zagier, *Modular forms whose Fourier coefficients involve zeta-functions of quadratic fields*. Modular functions of one variable, VI (Proc. Second Internat. Conf., Univ. Bonn, Bonn, 1976), pp. 105–169. Lecture Notes in Math., Vol. 627, Springer, Berlin, 1977.
N. I. Zavorotny, *Automorphic functions and number theory*, part I, II (Russian), Akad. Nauk SSSR, Dal’nevostochn. Otdel., Vladivostok 254 (1989), p. 69–124.
[^1]: Research of the first author is supported by Academy of Finland project no. $293876$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Encoder-decoder models typically only employ words that are frequently used in the training corpus to reduce the computational costs and exclude noise. However, this vocabulary set may still include words that interfere with learning in encoder-decoder models. This paper proposes a method for selecting more suitable words for learning encoders by utilizing not only frequency but also co-occurrence information, which we capture using the HITS algorithm.[^1] We apply our proposed method to two tasks: machine translation and grammatical error correction. For Japanese-to-English translation, this method achieves a BLEU score that is 0.56 points more than that of a baseline. Furthermore, it outperforms the baseline method for English grammatical error correction, with an $\mathrm{F_{0.5}}$-measure that is 1.48 points higher.'
author:
- 'Satoru Katsumata, Yukio Matsumura[^2] , Hayahide Yamagishi'
- |
Mamoru Komachi\
Tokyo Metropolitan University\
[{katsumata-satoru, matsumura-yukio, yamagishi-hayahide}@ed.tmu.ac.jp,]{}\
[[email protected]]{}
bibliography:
- 'b4\_ref.bib'
title: |
Graph-based Filtering of Out-of-Vocabulary Words\
for Encoder-Decoder Models
---
Introduction
============
Encoder-decoder models [@sutskever2014] are effective in tasks such as machine translation (@cho2014, [-@cho2014]; @bahdanau2014, [-@bahdanau2014]) and grammatical error correction [@yuan-briscoe2016]. Vocabulary in encoder-decoder models is generally selected from the training corpus in descending order of frequency, and low-frequency words are replaced with an unknown word token `<unk>`. The so-called out-of-vocabulary (OOV) words are replaced with `<unk>` to not increase the decoder’s complexity and to reduce noise. However, naive frequency-based OOV replacement may lead to loss of information that is necessary for modeling context in the encoder.
This study hypothesizes that vocabulary constructed using unigram frequency includes words that interfere with learning in encoder-decoder models. That is, we presume that vocabulary selection that considers co-occurrence information selects fewer noisy words for learning robust encoders in encoder-decoder models. We apply the hyperlink-induced topic search (HITS) algorithm to extract the co-occurrence relations between words. Intuitively, the removal of words that rarely co-occur with others yields better encoder models than ones that include noisy low-frequency words.
This study examines two tasks, machine translation (MT) and grammatical error correction (GEC) to confirm the effect of decreasing noisy words, with a focus on the vocabulary of the encoder side, because the vocabulary on the decoder side is relatively limited. In a Japanese-to-English MT experiment, our method achieves a BLEU score that is 0.56 points more than that of the frequency-based method. Further, it outperforms the frequency-based method for English GEC, with an $\mathrm{F_{0.5}}$-measure that is 1.48 points higher.
The main contributions of this study are as follows:
1. The simple but effective preprocessing method we propose for vocabulary selection improves encoder-decoder model performance.
2. This study is the first to address noise reduction in the source text of encoder-decoder models.
Related Work
============
There is currently a growing interest in applying neural models to MT (@sutskever2014, [-@sutskever2014]; @cho2014, [-@cho2014]; @bahdanau2014, [-@bahdanau2014]; @wu2016, [-@wu2016]) and GEC (@yuan-briscoe2016, [-@yuan-briscoe2016]; @xie2016, [-@xie2016]; @ji2017, [-@ji2017]); hence, this study focuses on improving the simple attentional encoder-decoder models that are applied to these tasks.
In the investigation of vocabulary restriction in neural models, @sennrich2016 applied byte pair encoding to words and created a partial character string set that could express all the words in the training data. They increased the number of words included in the vocabulary to enable the encoder-decoder model to robustly learn contextual information. In contrast, we aim to improve neural models by using vocabulary that is appropriate for a training corpus—not to improve neural models by increasing their vocabulary.
@jean2015 proposed a method of replacing and copying an unknown word token with a bilingual dictionary in neural MT. They automatically constructed a translation dictionary from a training corpus using a word-alignment model (GIZA++), which finds a corresponding source word for each unknown target word token. They replaced the unknown word token with the corresponding word into which the source word was translated by the bilingual dictionary. @yuan-briscoe2016 used a similar method for neural GEC. Because our proposed method is performed as preprocessing, it can be used simultaneously with this replace-and-copy method.
Algorithms that rank words using co-occurrence are employed in many natural language processing tasks. For example, [@mihalcea2004] uses PageRank [@brin_and_page1998] for keyword extraction. TextRank constructs a word graph in which nodes represent words, and edges represent co-occurrences between words within a fixed window; TextRank then executes the PageRank algorithm to extract keywords. Although this is an unsupervised method, it achieves nearly the same precision as one state-of-the-art supervised method [@hulth2003]. @kiso2011 used HITS [@kleinberg1999] to select seeds and create a stop list for bootstrapping in natural language processing. They reported significant improvements over a baseline method using unigram frequency. Their graph-based algorithm was effective at extracting the relevance between words, which cannot be grasped with a simple unigram frequency. In this study, we use HITS to retrieve co-occurring words from a training corpus to reduce noise in the source text.
hubness vector adjacency matrix iteration number $\tau$ hubness vector authority vector $\bm{i} \gets \bm{i_0}$ $\bm{p} \gets \bm{A^{\mathrm{T}}} \bm{i}$ $\bm{i} \gets \bm{A} \bm{p}$ normalize and and
\[tab:algohits\]
Graph-based Filtering of OOV Words
==================================
Hubness and authority scores from HITS
--------------------------------------
HITS, which is a web page ranking algorithm proposed by @kleinberg1999, computes hubness and authority scores for a web page (node) using the adjacency matrix that represents the web page’s link (edge) transitions. A web page with high authority is linked from a page with high hubness scores, and a web page with a high hubness score links to a page with a high authority score. Algorithm \[tab:algohits\] shows pseudocode for the HITS algorithm. Hubness and authority scores converge by setting the iteration number $\tau$ to a sufficiently large value.
Vocabulary selection using HITS
-------------------------------
In this study, we create an adjacency matrix from a training corpus by considering a word as a node and the co-occurrence between words as an edge. Unlike in web pages, co-occurrence between words is nonbinary; therefore, several co-occurrence measures can be used as edge weights. Section 3.3 describes the co-occurrence measures and the context in which co-occurrence is defined.
The HITS algorithm is executed using the adjacency matrix created in the way described above. As a result, it is possible to obtain a score indicating importance of each word while considering contextual information in the training corpus.
Figure \[tab:graphwords\] shows a word graph example. A word that obtains a high score in the HITS algorithm is considered to co-occur with a variety of words. Figure \[tab:graphwords\] demonstrates that second order co-occurrence scores (the scores of words co-occurring with words that co-occur with various words [@schutze98]) are also high.
In this study, words with high hubness scores are considered to co-occur with an important word, and low-scoring words are excluded from the vocabulary. Using this method appears to generate a vocabulary that includes words that are more suitable for representing a context vector for encoder models.
Word graph construction
-----------------------
To acquire co-occurrence relations, we use a combination of each word and its peripheral words. Specifically, we combine the target word with surrounding words within window width $N$ and count the occurrences. When defining the context in this way, because the adjacency matrix becomes symmetric, the same hubness and authority scores can be obtained. Figure \[tab:explain\] shows an example of co-occurrence in which $N$ is set to two.
We use raw co-occurrence frequency (Freq) and positive pointwise mutual information (PPMI) between words as the ($x, y$) element $A_{xy}$ of the adjacency matrix. However, naive PPMI reacts sensitively to low-frequency words in a training corpus. To account for high-frequency, we weight the PMI by the logarithm of the number of co-occurrences and use PPMI based on this weighted PMI (Equation \[ppmi\_equation\]). $$\begin{aligned}
\label{freq_equation} \lefteqn{\hspace{-64mm}A_{xy}^{freq} = |x, y| }\\
\label{ppmi_equation} A_{xy}^{ppmi} = \mathrm{max}(0, \mathrm{pmi}(x, y) + \log_2|x, y|) \end{aligned}$$ Equation \[pmi\_equation\] is the PMI of target word $x$ and co-occurrence word $y$. $M$ is the number of tokens of the combination, $|x, *|$ and $|*, y| $ are the number of token combinations when fixing target word $x$ and co-occurrence word $y$, respectively. $$\begin{aligned}
\label{pmi_equation} \mathrm{pmi}(x, y) = \log_2 \frac{M \cdot |x, y|}{|x, *||*, y|} \end{aligned}$$
Machine Translation
===================
Experimental setting
--------------------
In the first experiment, we conduct a Japanese-to-English translation using the Asian Scientific Paper Excerpt Corpus (ASPEC; @nakazawa, [-@nakazawa]). We follow the official split of the train, development, and test sets. As training data, we use only the first 1.5 million sentences sorted by sentence alignment confidence to obtain a Japanese–English parallel corpus (sentences of more than 60 words are excluded). Our training set consists of 1,456,278 sentences, development set consists of 1,790 sentences, and test set consists of 1,812 sentences. The training set has 247,281 Japanese word types and 476,608 English word types.
The co-occurrence window width $N$ is set to two. For combinations that co-occurred only once within the training corpus, we set the value of element $A_{xy}$ of the adjacency matrix to zero. The iteration number $\tau$ of the HITS algorithm is set to 300. As mentioned in Section 1, we only use the proposed method on the encoder side.
For this study’s neural MT model, we implement global dot attention [@luong2015b]. We train a baseline model that uses vocabulary that is determined by its frequency in the training corpus. Vocabulary size is set to 100K on the encoder side and 50K on the decoder side. Additionally, we conduct an experiment of varying vocabulary size of the encoder to 50K in the baseline and PPMI to investigate the effect of vocabulary size. Unless otherwise noted, we conduct an analysis of the model using the vocabulary size of 100K. The number of dimensions for each of the hidden and embedding layers is 512. The mini-batch size is 150. AdaGrad is used as an optimization method with an initial learning rate of 0.01. Dropout is applied with a probability of 0.2.
For this experiment, a bilingual dictionary is prepared for postprocessing unknown words [@jean2015]. When the model outputs an unknown word token, the word with the highest attention score is used as a query to replace the unknown token with the corresponding word from the dictionary. If not in the dictionary, we replace the unknown word token with the source word (unk\_rep). This dictionary is created based on word alignment obtained using fast\_align [@dyer2013] on the training corpus.
We evaluate translation results using BLEU scores [@papineni2002].
baseline HITS (Freq) HITS (PPMI)
------------- ---------- ------------- -------------
BLEU (50K) 22.24 - 22.40
BLEU (100K) 22.21 22.25 **22.77**
$p$-value - 0.35 0.01
: BLEU scores for Japanese-to-English translation. The parentheses indicate vocabulary size of the encoder.[]{data-label="tab:result1"}
------ ---------- ----------- ---------- -----------
baseline PPMI baseline PPMI
BLEU 22.33 **22.98** 21.44 **21.98**
------ ---------- ----------- ---------- -----------
: BLEU scores of the COMMON and DIFF outputs.[]{data-label="tab:incorexc"}
Results
-------
Table \[tab:result1\] shows the translation accuracy (BLEU scores) and $p$-value of a significance test ($p < 0.05$) by bootstrap resampling [@koehn2004]. The PPMI model improves translation accuracy by 0.56 points in Japanese-to-English translation, which is a significant improvement.
Next, we examine differences in vocabulary by comparing each model with the baseline. Compared to the vocabulary of the baseline in 100K setting, Freq and PPMI replace 16,107 and 17,166 types, respectively; compared to the vocabulary of the baseline in 50K setting, PPMI replaces 4,791 types.
src 有用 物質 の 分離 ・ 抽出 , 反応 性 向上 , 新 材料 創製 , 廃棄 物 処理 , 分析 等 の 分野 が ある 。
---------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------
baseline there are fields such as separation , extraction , extraction , improvement of new material creation , waste treatment , analysis , etc .
Freq there are separation and extraction of useful substances , the improvement of reactivity , new material creation , waste treatment and analysis .
PPMI there are the fields such as separation and extraction of useful materials , the reaction improvement , new material creation , waste treatment , analysis , etc ...
ref the application fields are separation and extraction of useful substances , reactivity improvement , creation of new products , waste treatment , and chemical analysis .
Analysis
--------
According to Table \[tab:result1\], the performance of Freq is almost the same as that of the baseline. When examining the differences in selected words in vocabulary between PPMI and Freq, we find that PPMI selects more low-frequency words in the training corpus compared to Freq, because PPMI deals with not only frequency but also co-occurrence.
The effect of unk\_rep is almost the same in the baseline as in the proposed method, which indicates that the proposed method can be combined with other schemes as a preprocessing step.
As a comparison of the vocabulary size 50K and 100K, the BLEU score of 100K is higher than that of 50K in PPMI. Moreover, the BLEU scores are almost the same in the baseline. We suppose that the larger the vocabulary size of encoder, the more noisy words the baseline includes, while the PPMI filters these words. That is why the proposed method works well in the case where the vocabulary size is large.
To examine the effect of changing the vocabulary on the source side, the test set is divided into two subsets: COMMON and DIFF. The former (1,484 sentences) consists of only the common vocabulary between the baseline and PPMI, whereas the latter (328 sentences) includes at least one word excluded from the common vocabulary.
Table \[tab:incorexc\] shows the translation accuracy of the COMMON and DIFF outputs. Translation performance of both corpora is improved.
In order to observe how PPMI improves COMMON outputs, we measure the similarity of the baseline and PPMI output sentences by counting the exact same sentences. In the COMMON outputs, 72 sentence pairs (4.85%) are the same, whereas 9 sentence pairs are the same in the DIFF outputs (2.74%). Surprisingly, even though it uses the same vocabulary, PPMI often outputs different but fluent sentences.
Table \[tab:resultexam\] shows an example of Japanese-to-English translation. The outputs of the proposed method (especially PPMI) are improved, despite the source sentence being expressed with common vocabulary; this is because the proposed method yielded a better encoder model than the baseline.
Grammatical Error Correction
============================
Experimental setting
--------------------
The second experiment addresses GEC. We combine the FCE public dataset [@yann2011], NUCLE corpus [@dahlmeier2013], and English learner corpus from the Lang-8 learner corpus [@mizumoto2011] and remove sentences longer than 100 words to create a training corpus. From the Lang-8 learner corpus, we use only the pairs of erroneous and corrected sentences. We use 1,452,584 sentences as a training set (502,908 types on the encoder side and 639,574 types on the decoder side). We evaluate the models’ performances on the standard sets from the CoNLL-14 shared task [@ng2014] using CoNLL-13 data as a development set (1,381 sentences) and CoNLL-14 data as a test set (1,312 sentences). We employ $\mathrm{F_{0.5}}$ as an evaluation measure for the CoNLL-14 shared task.
-------------------- ------- ------- ----------- -----------
50K 150K 50K 150K
Precision 48.09 46.53 **49.45** 49.23
Recall 8.30 8.50 8.61 **9.02**
$\mathrm{F_{0.5}}$ 24.55 24.55 25.37 **26.03**
-------------------- ------- ------- ----------- -----------
: $\mathrm{F_{0.5}}$ results on the CoNLL-14 test set.[]{data-label="tab:gec_result"}
-------------------- ---------- ------- ---------- -------
baseline PPMI baseline PPMI
P 48.26 60.07 9.40 17.32
R 0.01 0.01 0.01 0.02
$\mathrm{F_{0.5}}$ 0.04 0.04 0.04 0.08
-------------------- ---------- ------- ---------- -------
: $\mathrm{F_{0.5}}$ of COMMON and DIFF outputs.[]{data-label="tab:incorexc4gec"}
We use the same model as in Section 4.1 as a neural model for GEC. The models’ parameter settings are similar to the MT experiment, except for the vocabulary and batch sizes. In this experiment, we set the vocabulary size on the encoder and decoder sides to 150K and 50K, respectively. Additionally, we conduct the experiment of changing vocabulary size of the encoder to 50K to investigate the effect of the vocabulary size. Unless otherwise noted, we conduct an analysis of the model using the vocabulary size of 150K. The mini-batch size is 100.
Result
------
Table \[tab:gec\_result\] shows the performance of the baseline and proposed method. The PPMI model improves precision and recall; it achieves a $\mathrm{F_{0.5}}$-measure 1.48 points higher than the baseline method.
In setting the vocabulary size of encoder to 150K, PPMI replaces 37,185 types from the baseline; in the 50K setting, PPMI replaces 10,203 types.
src Genetic refers the chance of inheriting a disorder or disease .
---------- ---------------------------------------------------------------------------------
baseline Genetic refers the chance of inheriting a disorder or disease .
PPMI Genetic refers **to** the chance of inheriting a disorder or disease .
gold Genetic **risk** refers **to** the chance of inheriting a disorder or disease .
: An example of GEC using a source sentence from COMMON.[]{data-label="tab:resultexam4gec"}
Analysis
--------
The $\mathrm{F_{0.5}}$ of the baseline is almost the same while the PPMI model improves the score in the case where the vocabulary size increases. Similar to MT, we suppose that the PPMI filters noisy words.
As in Section 4.3, we perform a follow-up experiment using two data subsets: COMMON and DIFF, which contain 1,072 and 240 sentences, respectively.
Table \[tab:incorexc4gec\] shows the accuracy of the error correction of the COMMON and DIFF outputs. Precision increases by 11.81 points, whereas recall remains the same for the COMMON outputs.
In GEC, approximately 20% of COMMON’s output pairs differ, which is caused by the differences in the training environment. Unlike MT, we can copy OOV in the target sentence from the source sentence without loss of fluency; therefore, our model has little effect on recall, whereas its precision improves because of noise reduction.
Table \[tab:resultexam4gec\] shows an example of GEC. The proposed method’s output improves when the source sentence is expressed using common vocabulary.
Discussion
==========
We described that the proposed method has a positive effect on learning the encoder. However, we have a question; what affects the performance? We conduct an analysis of this question in this section.
First, we count the occurrence of the words included only in the baseline or PPMI in the training corpus. We also show the number of the tokens per types (“Ave. tokens”) included only in either the baseline or PPMI vocabulary.
------------- -------- -------- --------- --------
tokens 52,700 27,364 126,884 70,003
Ave. tokens 3.07 1.59 3.36 1.85
------------- -------- -------- --------- --------
: Number of words included only in either the baseline or PPMI vocabulary. []{data-label="tab:wordCount"}
The result is shown in Table \[tab:wordCount\]. We find that the proposed method uses low-frequency words instead of high-frequency words in the training corpus. This result suggests that the proposed method works well despite the fact that the encoder of the proposed method encounters more `<unk>` than the baseline. This is because the proposed method excludes words that may interfere with the learning of encoder-decoder models.
Second, we conduct an analysis of the POS of the words in GEC to find why increasing OOV improves the learning of encoder-decoder models. Specifically, we apply POS tagging to the training corpus and calculate the occurrence of the POS of the words only included in the baseline or PPMI. We use NLTK as a POS tagger.
Table \[tab:posCount\] shows the result. It is observed that NOUN is the most affected POS by the proposed method and becomes often represented by `<unk>`. NOUN words in the vocabulary of the baseline contain some non-English words, such as Japanese or Korean. These words should be treated as OOV but the baseline fails to exclude them using only the frequency. According to Table \[tab:posCount\], NUM is also affected by the proposed method. NUM words of the baseline include a simple numeral such as “`119`”, in addition to incorrectly segmented numerals such as “`514&objID`”. This word appears 25 times in the training corpus owing to the noisy nature of Lang-8. We suppose that the proposed method excludes these noisy words and has a positive effect on training.
POS
------- --------- -------- ------------
NOUN 92,693 44,472 4,644,478
VERB 11,066 10,099 3,597,895
PRON 127 107 1,869,422
ADP 626 685 1,836,193
DET 128 202 1,473,391
ADJ 13,855 12,270 1,429,056
ADV 2,032 1,688 931,763
PRT 319 75 615,817
CONJ 62 28 537,346
PUNCT 110 11 223,573
NUM 5,585 299 207,487
OTHER 281 67 5,209
Total 126,884 70,003 17,371,630
: Number of the POS of words only included in the baseline or PPMI.[]{data-label="tab:posCount"}
Conclusion
==========
In this paper, we proposed an OOV filtering method, which considers word co-occurrence information for encoder-decoder models. Unlike conventional OOV handling, this graph-based method selects the words that are more suitable for learning encoder models by considering contextual information. This method is effective for not only machine translation but also grammatical error correction.
This study employed a symmetric matrix (similar to skip-gram with negative sampling) to express relationships between words. In future research, we will develop this method by using vocabulary obtained by designing an asymmetric matrix to incorporate syntactic relations.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Yangyang Xi of Lang-8, Inc. for allowing us to use the Lang-8 learner corpus. We also thank Masahiro Kaneko and anonymous reviewers for their insightful comments.
[^1]: <https://github.com/Katsumata420/HITS_Ranking>
[^2]: Both authors equally contributed to the paper.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For a given curve $X$ and divisor class $C$, we give lower bounds on the degree of a divisor $A$ such that $A$ and $A-C$ belong to specified semigroups of divisors. For suitable choices of the semigroups we obtain (1) lower bounds for the size of a party $A$ that can recover the secret in an algebraic geometric linear secret sharing scheme with adversary threshold $C$, and (2) lower bounds for the support $A$ of a codeword in a geometric Goppa code with designed minimum support $C$. Our bounds include and improve both the order bound and the floor bound. The bounds are illustrated for two-point codes on general Hermitian and Suzuki curves.'
author:
- 'Iwan M. Duursma and Seungkook Park'
date: 'October 8, 2008'
nocite:
- '[@Bra04]'
- '[@BraSul07]'
- '[@SKP07]'
- '[@Sul01]'
- '[@Pre98]'
- '[@HufPle03]'
- '[@Lin99]'
- '[@Pre98]'
- '[@Ste99]'
- '[@Sti93]'
- '[@TsfVla07]'
- '[@CraetSix05]'
- '[@HanSti90]'
- '[@Mat04]'
- '[@CarTor05]'
- '[@MahMat06]'
- '[@LunMcc06]'
- '[@Gar93]'
- '[@KirPel95]'
- '[@BeeNes06JPA]'
- '[@Bee07FF]'
- '[@HomKim06]'
- '[@Kim94]'
- '[@Mat01]'
- '[@CheCra06]'
- '[@Duu93]'
- '[@Duu08]'
title: Coset bounds for algebraic geometric codes
---
Introduction {#S:Introduction .unnumbered}
============
Two recent results motivated this paper. The first is the complete description of the minimum distance of Hermitian two-point codes by Homma and Kim [@HomKim06]. The second is the introduction of algebraic geometric linear secret sharing schemes by Chen and Cramer [@CheCra06].\
For algebraic geometric codes, the actual value of the minimum distance is not a priori known and needs to be determined or estimated from the data used in the construction. The best known lower bounds for the minimum distance of an algebraic geometric code are the order bound and the floor bound. Beelen [@Bee07FF], and independently the second author [@SKP07], have shown that the order bound agrees, for Hermitian two-point codes, with the actual minimum distances found by Homma and Kim. In this paper we improve both the order bound and the floor bound. We illustrate our results and the obtained improvements for two-point codes from the Suzuki curves.\
An important application of secret sharing schemes is secure multi-party computation, which requires linear secret sharing schemes with a multiplicative property [@CraDamMau00], [@CraetSix05]. Chen and Cramer proposed to use one-point algebraic geometric codes for secret sharing and they have shown that the obtained algebraic geometric linear secret sharing schemes can be used for efficient secure computation over small fields [@CheCra06]. Parties can reconstruct a secret uniquely from their shares only if the total number of shares exceeds the adversary threshold of the secret sharing scheme. The algebraic geometric construction of a linear secret sharing scheme guarantees a lower bound for the adversary threshold. The precise value of the threshold is in general not known. We show that the adversary threshold corresponds to the minimum distance between cosets of a code. Our results give improved lower bounds for distances between cosets of an algebraic geometric code, and therefore improved lower bounds for adversary thresholds of algebraic geometric linear secret sharing schemes.\
As our main results, we formulate an *ABZ bound for codes* and an *ABZ bound for cosets*. The bounds improve and generalize the floor bound and the order bound, respectively. For each of the bounds, we illustrate the improvements with examples from the Suzuki curves. The bounds can be used as tools for constructing improved codes as well as improved secret sharing schemes. Our *Main theorem* is an even more general bound. Its main advantage is that it has a short proof and that all other bounds can be obtained as special cases.\
The floor bound is independent of the order bound. Algorithms are available for decoding up to half the order bound but not for decoding up to half the floor bound. Beelen [@Bee07FF] gives an example where the floor bound exceeds the order bound. For our generalizations there is a strict hierarchy. The improved order bound, obtained with the ABZ bound for cosets, is at least the ABZ bound for codes, which improves the floor bound. We show that decoding is possible up to half the bound in our main theorem, and therefore up to half of all our bounds. In particular, we obtain for the first time an approach to decode up to half the floor bound.\
In Section \[S:CosetsLinearCodes\], we describe the use of linear codes for secret sharing and the relation between coset distances and adversary thresholds. Theorem \[T:cosetbound\] gives a general coset bound for linear codes. Appendix \[S:CosetDecoding\] gives a coset decoding procedure that decodes up to half the bound. Algebraic geometric codes are defined in Section \[S:AGCodes\]. Theorem \[T:ABZcodes\] gives the ABZ bound for algebraic geometric codes with a first proof based on the AB bound for linear codes. Section \[S:CosetsAGCodes\] gives a geometric characterization of coset distances for algebraic geometric codes. In Section \[S:SemigroupIdeals\] we define, for a divisor $C$ and for a point $P$, a semigroup ideal $$\Gamma_P(C) = \{ A : L(A) \neq L(A-P) \wedge L(A-C) \neq L(A-C-P) \}$$ such that the minimal degree for a divisor $A$ in $\Gamma_P(C)$ is a lower bound for the coset distance of an algebraic geometric code. In Section \[S:MainTheorem\], the main theorem gives a lower bound for the degree of a divisor in the semigroup ideal (Theorem \[T:cbdiv\]). In Section \[S:OrderFloor\], we formulate the ABZ bound for cosets (Theorem \[T:cbabz\]) and we describe its relation to both the order bound (Theorem \[T:order\]) and the floor bound (Theorem \[T:floor\]). The successful application of our bounds depends on the possibility to analyse the complement $$\Delta_P(C) = \{ A : L(A) \neq L(A-P) \wedge L(A-C) = L(A-C-P) \}$$ and to compare naturally defined subsets of $\Delta_P(C)$. Section \[S:DeltaSets\] gives important basic relations among delta sets. In Section \[S:Discrepancies\], we define a discrepancy, for given points $P$ and $Q$, as a divisor $A \in \Delta_P(Q) = \Delta_Q(P).$ Discrepancies are our main tool for analyzing and improving lower bounds for coset distances in large families of codes. In Section \[S:Hermitian\], we give two proofs, one due to [@Bee07FF], [@SKP07], and one new, for lower bounds for the minimum distance of Hermitian two-point codes. In Section \[S:Suzuki\], we determine discrepancies for Suzuki curves, and we give examples of the ABZ bound for codes, the ABZ bound for cosets, and the main theorem, that improve previously known bounds.
Cosets of linear codes {#S:CosetsLinearCodes}
======================
Let ${{\mathbb F}}$ be a finite field. A [${{\mathbb F}}$-linear code]{} ${{\mathcal C}}$ of [length]{} $n$ is a linear subspace of ${{\mathbb F}}^n$. The [Hamming distance]{} between two vectors $x, y \in {{\mathbb F}}^n$ is $d(x,y) = | \{ i : x_i \neq y_i \} |.$ The [minimum distance]{} of a nontrivial linear code ${{\mathcal C}}$ is $$\begin{aligned}
d({{\mathcal C}}) &= \min\, \{ d(x,y) : x,y \in {{\mathcal C}}, x \neq y \} \\
&= \min \, \{ d(x,0) : x \in {{\mathcal C}}, x \neq 0 \}.\end{aligned}$$ If $d({{\mathcal C}}) \geq 2t+1$ and if $y \in {{\mathbb F}}^n$ is at distance at most $t$ from ${{\mathcal C}}$ then there exists a unique word $c \in {{\mathcal C}}$ with $d(c,y) \leq t.$\
The Hamming distance between two nonempty subsets $X, Y\subset {{\mathbb F}}^n$ is the minimum of $\{ d(x,y) : x \in X, y \in Y \}$. For a proper subcode ${{\mathcal C}}' \subset {{\mathcal C}}$, the minimum distance of the collection of cosets ${{\mathcal C}}/{{\mathcal C}}'$ is $$\begin{aligned}
d({{\mathcal C}}/{{\mathcal C}}') &= \min\, \{ d(x+{{\mathcal C}}',y+{{\mathcal C}}') : x,y \in {{\mathcal C}}, x-y \not \in {{\mathcal C}}' \} \\
&= \min\, \{ d(x,0) : x \in {{\mathcal C}}, x \not \in {{\mathcal C}}' \}.\end{aligned}$$
If $d({{\mathcal C}}/{{\mathcal C}}') \geq 2t+1$ and if $y \in {{\mathbb F}}^n$ is at distance at most $t$ from ${{\mathcal C}}$ then there exists a unique coset $c + {{\mathcal C}}' \in {{\mathcal C}}/ {{\mathcal C}}'$ with $d(c+{{\mathcal C}}',y+{{\mathcal C}}') \leq t.$
The [dual code]{} ${{\mathcal D}}$ of ${{\mathcal C}}$ is the maximal subspace of ${{\mathbb F}}^n$ that is orthogonal to ${{\mathcal C}}$ with respect to the standard inner product. To the extension of codes ${{\mathcal C}}/{{\mathcal C}}'$ corresponds an extension of dual codes ${{\mathcal D}}'/{{\mathcal D}}$ with distance parameter $d({{\mathcal D}}'/{{\mathcal D}}).$ For two vectors $x, y \in {{\mathbb F}}^n$, let $x \ast y \in {{\mathbb F}}^n$ denote the Hadamard or coordinate-wise product of the two vectors.
(Shift bound or Coset bound) \[T:cosetbound\] Let ${{\mathcal C}}/{{\mathcal C}}_1$ be an extension of ${{\mathbb F}}$-linear codes with corresponding extension of dual codes ${{\mathcal D}}_1/{{\mathcal D}}$ such that $\dim {{\mathcal C}}/{{\mathcal C}}_1 = \dim {{\mathcal D}}_1/{{\mathcal D}}= 1$. If there exist vectors $a_1, \ldots, a_w$ and $b_1, \ldots, b_w$ such that $$\begin{cases}
a_i \ast b_j \in {{\mathcal D}}&\text{for $i+j \leq w$}, \\
a_i \ast b_j \in {{\mathcal D}}_1 \backslash {{\mathcal D}}&\text{for $i+j = w+1$},
\end{cases}$$ then $d({{\mathcal C}}/{{\mathcal C}}_1) \geq w.$
For all $c \in {{\mathcal C}}\backslash {{\mathcal C}}_1$ and $a \ast b \in {{\mathcal D}}_1 \backslash {{\mathcal D}}$, $\sum_i a_i b_i c_i \neq 0$. To show the nonexistence of a vector $c \in {{\mathcal C}}\backslash {{\mathcal C}}_1$ with $d(c,0)<w$, it suffices to show, for any choice of $w-1$ coordinates, the existence of a vector $a \ast b \in {{\mathcal D}}_1 \backslash {{\mathcal D}}$ that is zero in those coordinates. The conditions show that the vectors $a_1, \ldots, a_w$ are linearly independent, and there exists a nonzero linear combination $a$ of the vectors $a_1, \ldots, a_w$ vanishing at $w-1$ given coordinates. If $i$ is maximal such that $a_i$ has a nonzero coefficient in the linear combination $a$ then $a \ast b_{w+1-i} \in {{\mathcal D}}_1 \backslash {{\mathcal D}}$ is zero in the $w-1$ coordinates.
Let $y \in {{\mathbb F}}^n$ be a word at distance at most $t$ from ${{\mathcal C}}$. For given vectors $a_1, \ldots, a_w$ and $b_1, \ldots, b_w$ such that $w > 2t$, the unique coset $c + {{\mathcal C}}_1 \in {{\mathcal C}}/ {{\mathcal C}}_1$ with $d(c+{{\mathcal C}}_1,y+{{\mathcal C}}_1) \leq t$ can be computed efficiently with the coset decoding procedure in Appendix \[S:CosetDecoding\]. Theorem \[T:cosetbound\] can be used to estimate the minimum distance $d({{\mathcal C}}/{{\mathcal C}}')$ of an extension ${{\mathcal C}}/{{\mathcal C}}'$ with $\dim {{\mathcal C}}/{{\mathcal C}}' > 1,$ after dividing ${{\mathcal C}}/{{\mathcal C}}'$ into subextensions.
\[T:itercosetbound\] Let ${{\mathcal C}}/{{\mathcal C}}'$ be an extension of ${{\mathbb F}}$-linear codes of length $n$. For ${{\mathcal C}}\supset {{\mathcal C}}'' \subset {{\mathcal C}}'$, $$d({{\mathcal C}}/{{\mathcal C}}') = \min \{ d({{\mathcal C}}/{{\mathcal C}}''), d({{\mathcal C}}''/{{\mathcal C}}') \}.$$
We will now describe the use of code extensions for secret sharing. Our description focuses on the connection between secret sharing thresholds and coset distances that will be established in Corollary \[C:ssbounds\]. The main properties that we need are described in the following two lemmas.
\[L:dual\] Let $\{ 1, 2, \ldots, n \} = I \cup J$ be a partition of the coordinates. There exists a word $r \in {{\mathcal C}}\backslash {{\mathcal C}}_1$ with support in $I$ if and only if there exists no word $s \in {{\mathcal D}}_1 \backslash {{\mathcal D}}$ with support in $J$.
Let $E_I$ (resp. $E_J$) be the subspace of ${{\mathbb F}}^n$ of all vectors with support in $I$ (resp. $J$). The exact sequences $$\begin{aligned}
0 \longrightarrow {{\mathcal C}}\cap E_I / {{\mathcal C}}_1 \cap E_I \longrightarrow &{{\mathcal C}}/ {{\mathcal C}}_1
\longrightarrow {{\mathcal C}}+ E_I / {{\mathcal C}}_1 + E_I \longrightarrow 0, \\
0 \longrightarrow {{\mathcal D}}_1 \cap E_J / {{\mathcal D}}\cap E_J \longrightarrow &{{\mathcal D}}_1 / {{\mathcal D}}\longrightarrow {{\mathcal D}}_1 + E_J / {{\mathcal D}}+ E_J \longrightarrow 0, \end{aligned}$$ are in duality via $V \mapsto V^\ast = \text{\rm{Hom}}(V,{{\mathbb F}})$. And $$(\dim {{\mathcal C}}\cap E_I / {{\mathcal C}}_1 \cap E_I) + (\dim {{\mathcal D}}_1 \cap E_J / {{\mathcal D}}\cap E_J) = \dim {{\mathcal C}}/ {{\mathcal C}}_1 = 1.$$
\[L:DD1\] Let $y_1 \in {{\mathcal D}}_1 \backslash {{\mathcal D}}$. For a given vector $s \in {{\mathcal D}}_1 = {{\mathcal D}}\oplus \langle y_1 \rangle$, the projection of $s$ on $\langle y_1 \rangle$ is uniquely determined by the subset of coordinates $\{ s_i : i \in A \}$ if and only if ${{\mathcal D}}_1 \backslash {{\mathcal D}}$ contains no word that is zero in the positions $A$.
The only if part is clear. For the if part we may assume with the previous lemma that there exists $r \in {{\mathcal C}}\backslash {{\mathcal C}}_1$ with support in $A$. For any such $r$, and for $s = y + \lambda y_1$, $y \in {{\mathcal D}}$, $$r \cdot s = r \cdot (y + \lambda y_1) = \lambda (r \cdot y_1).$$ Since $(r \cdot y_1) \neq 0$, we obtain $\lambda = (r \cdot s) / (r \cdot y_1).$
Let $y_1 \in {{\mathcal D}}_1 \backslash {{\mathcal D}}$. For a secret $\lambda \in {{\mathbb F}}$, and for a random vector $y \in {{\mathcal D}}$, the vector $s = y + \lambda y_1$ is called a vector of shares for $\lambda$. A subset $A \subset \{1,2,\ldots,n\}$ is called qualified if the shares $\{ s_i : i \in A \}$ determine $\lambda$ uniquely. Whether $A$ is qualified depends on ${{\mathcal D}}_1 / {{\mathcal D}}$ but not on the vector of shares $s \in {{\mathcal D}}_1$. Let $\Gamma({{\mathcal D}}_1 / {{\mathcal D}})$ denote the collection of all subsets $A \subset \{1,2,\ldots,n\}$ that are qualified for ${{\mathcal D}}_1 / {{\mathcal D}}$ and let $\Delta({{\mathcal D}}_1 / {{\mathcal D}})$ denote the collection of all subsets $A \subset \{1,2,\ldots,n\}$ that are not qualified for ${{\mathcal D}}_1 / {{\mathcal D}}$. For the definition and main properties of a general linear secret sharing scheme we refer to [@CraetSix05].
\[T:ssbounds\] Let ${{\mathcal C}}/{{\mathcal C}}_1$ and ${{\mathcal D}}_1/{{\mathcal D}}$ be dual extensions of ${{\mathbb F}}$-linear codes of length $n$. Let $E_A$ be the subset of ${{\mathbb F}}^n$ of all vectors with support in $A$. $$\begin{aligned}
\Gamma({{\mathcal D}}_1 / {{\mathcal D}}) &= \{ A : {{\mathcal C}}\cap E_A \neq {{\mathcal C}}_1 \cap E_A \}, \\
\Delta({{\mathcal D}}_1 / {{\mathcal D}}) &= \{ A : {{\mathcal C}}\cap E_A = {{\mathcal C}}_1 \cap E_A \}.\end{aligned}$$ Moreover, for ${\bar A} = \{1,2,\ldots,n\} \backslash A,$ $$\begin{aligned}
\Gamma({{\mathcal C}}/ {{\mathcal C}}_1) &= \{ A : {\bar A} \in \Delta({{\mathcal D}}_1 / {{\mathcal D}}) \}, \\
\Delta({{\mathcal C}}/ {{\mathcal C}}_1) &= \{ A : {\bar A} \in \Gamma({{\mathcal D}}_1 / {{\mathcal D}}) \}.\end{aligned}$$
Lemma \[L:DD1\] and Lemma \[L:dual\].
\[C:ssbounds\] The smallest qualified subset for ${{\mathcal D}}_1/{{\mathcal D}}$ is of size $$\min \{ |A| : A \in \Gamma({{\mathcal D}}_1 / {{\mathcal D}}) \} = d({{\mathcal C}}/{{\mathcal C}}_1).$$ The largest unqualified subset for ${{\mathcal D}}_1/{{\mathcal D}}$ is of size $$\max \{ |A| : A \in \Delta({{\mathcal D}}_1 / {{\mathcal D}}) \} = n-d({{\mathcal D}}_1/{{\mathcal D}}),$$
Algebraic geometric codes {#S:AGCodes}
=========================
Let $X/{{\mathbb F}}$ be an algebraic curve (absolutely irreducible, smooth, projective) of genus $g$ over a finite field ${{\mathbb F}}$. Let ${{\mathbb F}}(X)$ be the function field of $X/{{\mathbb F}}$ and let $\Omega(X)$ be the module of rational differentials of $X/{{\mathbb F}}$. Given a divisor $E$ on $X$ defined over ${{\mathbb F}}$, let $L(E)$ denote the vector space over ${{\mathbb F}}$ of functions $f \in {{\mathbb F}}(X) \backslash\{0\}$ with $(f)+E\geq 0$ together with the zero function. Let $\Omega(E)$ denote the vector space over ${{\mathbb F}}$ of differentials $\omega \in \Omega(X) \backslash\{0\}$ with $(\omega) \geq E$ together with the zero differential. Let $K$ represent the canonical divisor class.\
For $n$ distinct rational points $P_1, \ldots, P_n$ on $X$ and for disjoint divisors $D=P_1+\cdots+P_n$ and $G$, the geometric Goppa codes $C_L(D,G)$ and $C_\Omega(D,G)$ are defined as the images of the maps $$\begin{aligned}
\alpha _L ~:~ &L(G)~~\longrightarrow~~{{\mathbb F}}^{\,n}, ~~f \mapsto (\,f(P_1), \ldots, f(P_n) \,), \\
\alpha _\Omega ~:~ &\Omega(G-D)~~\longrightarrow~~{{\mathbb F}}^{\,n},
~~\omega \mapsto (\,\Res_{P_1}(\omega), \ldots, \Res_{P_n}(\omega) \,).\end{aligned}$$ The maps establish isomorphisms $L(G)/L(G-D) \simeq C_L(D,G)$ and $\Omega(G-D)/\Omega(G) \simeq C_\Omega(D,G).$ With the Residue theorem, the images are orthogonal subspaces of ${{\mathbb F}}^n$. With the Riemann-Roch theorem they are maximal orthogonal subspaces.\
There exists a nonzero word in $C_L(D,G)$ with support in $A$, for $0 \leq A \leq D$, if and only if $L(G-D+A)/L(G-D) \neq 0.$ There exists a nonzero word in $C_\Omega(D,G)$ with support in $A$, for $0 \leq A \leq D$, if and only if $\Omega(G-A)/\Omega(G) \neq 0$ if and only if $L(K-G+A) / L(K-G) \neq 0.$
\[P:agd\] $$\begin{aligned}
&d(C_L(D,G)) = \min \{ \deg A : 0 \leq A \leq D \;|\; L(A-C) \neq L(-C) \}, \quad \text{for $C = D-G.$} \\
&d(C_\Omega(D,G)) = \min \{ \deg A : 0 \leq A \leq D \;|\; L(A-C) \neq L(-C) \}, \quad \text{for $C = G-K.$} \end{aligned}$$
(Goppa bound) \[T:Goppa\] A nonzero word in $C_L(D,G)$ has weight $w \geq \deg\,(D-G)$. A nonzero word in $C_\Omega(D,G)$ has weight $w \geq \deg\,(G-K)$.
The following bound improves on the Goppa bound in special cases ([@CarTor05], [@MahMat06], [@LunMcc06]).
\[T:floor\] (Floor bound) Let $G = K+C = A+B+Z,$ for $Z \geq 0$ such that $L(A+Z) = L(A)$ and $L(B+Z) = L(B)$. For $D$ with $D \cap Z = \emptyset$, a nonzero word in $C_\Omega(D,G)$ has weight at least $\deg\, C + \deg\, Z$.
Most algebraic bounds for the minimum distance of a linear code rely on one of two basic arguments. In the paper [@LinWil86] on cyclic codes they were named the AB bound and the Shift bound. We obtain the following bound, which includes the floor bound, using the AB bound argument in combination with the Goppa bound.
\[T:ABZcodes\] (ABZ bound for codes) Let $G = K+C = A+B+Z,$ for $Z \geq 0$. For $D$ with $D \cap Z = \emptyset$, a nonzero word in $C_\Omega(D,G)$ has weight $w \geq l(A)-l(A-C)+l(B)-l(B-C).$
We may assume that $A$ and $B$ are disjoint from $D$. Since $Z \geq 0$ and $D \cap Z = \emptyset$, the code $C_L(D,G)$ contains the code $C_L(D,A+B)$ as a subcode. Thus, for a word $c \in C_\Omega(D,G)$, and for words $a \in C_L(D,A)$ and $b \in C_L(D,B)$, if $c$ has support $D'$ then $\sum_{P \in D'} a_P b_P c_P = 0.$ The last orthogonality holds for all $a \in C_L(D',A)$ and $b \in C_L(D',B)$, so that $\dim C_L(D',A) + \dim C_L(D',B) \leq \deg D',$ and $\deg D' \geq l(A)-l(A-D') + l(B)-l(B-D').$ Together with $L(D'-C) \neq 0$, $\deg D' \geq l(A)-l(A-C) + l(B) - l(B-C).$
Example \[E:abz\] gives a code for which the ABZ bound improves both the floor bound and the order bound. It is easy to see, using the Riemann-Roch theorem, that the choice $Z=0$ returns the Goppa bound. Improvements of the Goppa bound are obtained only if the divisors $A, B,$ and $Z,$ are carefully chosen. For the special case $L(A+Z)=L(A)$ and $L(B+Z)=L(B)$, we recover the floor bound. In that case, for $K+C=A+B+Z$, $$\begin{aligned}
&l(A)-l(A-C)+l(B)-l(B-C) \\
=~ &l(A+Z)-l(K-B-Z)+l(B+Z)-l(K-A-Z) \\
=~ &\deg (A+Z) + \deg (B+Z) + 2 - 2g = \deg C + \deg Z.\end{aligned}$$
Cosets of algebraic geometric codes {#S:CosetsAGCodes}
===================================
Let ${{\mathcal D}}= C_\Omega(D,G)$ and ${{\mathcal C}}= C_L(D,G)$ be dual algebraic geometric codes. For a point $P$ disjoint from $D$, let $$\begin{aligned}
{{\mathcal D}}_1 / {{\mathcal D}}= C_\Omega(D,G-P) / C_\Omega(D,G), \\
{{\mathcal C}}/ {{\mathcal C}}_1 = C_L(D,G) / C_L(D,G-P), \end{aligned}$$ be dual extensions of codes. When $\dim {{\mathcal C}}/ {{\mathcal C}}_1 = \dim {{\mathcal D}}_1 / {{\mathcal D}}= 1$, the extensions can be used for secret sharing as described in Section \[S:CosetsLinearCodes\]. Theorem \[T:ssbounds\] describes the parties that can recover the secret for the extension ${{\mathcal D}}_1 / {{\mathcal D}}$ as the subsets $0 \leq A \leq D$ that support a word in ${{\mathcal C}}/ {{\mathcal C}}_1.$ The formulation in terms of divisors is given in Proposition \[P:O\], with a similar result for the extension ${{\mathcal C}}/ {{\mathcal C}}_1$ in Proposition \[P:L\]. As additional motivation, we give a natural choice for the secret for each of the extensions ${{\mathcal D}}_1 / {{\mathcal D}}$ and ${{\mathcal C}}/ {{\mathcal C}}_1$, and we describe directly the qualified parties that can determine the secret, in Lemma \[L:O\] and Lemma \[L:L\], respectively. The propositions can then also be obtained from the lemmas.\
Let $P$ have multiplicity $e$ in $G$, and let $t$ be a fixed local parameter for $P$. For $\dim {{\mathcal D}}_1 / {{\mathcal D}}= 1$, there exists a natural isomorphism $\Omega(G-D-P)/\Omega(G-D) \simeq {{\mathcal D}}_1 / {{\mathcal D}}\simeq {{\mathbb F}}$ that maps $\omega \in \Omega(G-D-P)/\Omega(G-D)$ to $\Res_P(t^{-e} \omega).$ For $\dim {{\mathcal C}}/ {{\mathcal C}}_1 = 1,$ there exists a natural isomorphism $L(G)/L(G-P) \simeq {{\mathcal C}}/ {{\mathcal C}}_1 \simeq {{\mathbb F}}$, that maps $f \in L(G)/L(G-P)$ to $(f t^e)(P).$
\[L:O\] For $\omega \in \Omega(G-D-P)$, the residue $\Res_P(t^{-e}\omega)(P)$ is uniquely determined by the values $\{ f(P) : P \in A \}$, for $0 \leq A \leq D,$ if and only if $\Omega(G-D+A-P) = \Omega(G-D+A)$.
\[P:O\] For the extension of codes ${{\mathcal D}}_1 / {{\mathcal D}}= C_\Omega(D,G-P) / C_\Omega(D,G)$, where $D=P_1+\cdots+P_n$ is a sum of $n$ distinct points, $G$ is a divisor disjoint from $D$, and $P$ is a point disjoint from $D$, $$\begin{aligned}
\Gamma({{\mathcal D}}_1 / {{\mathcal D}}) &= \{ 0 \leq A \leq D : {{\mathcal C}}\cap E_A \neq {{\mathcal C}}_1 \cap E_A \}, \\
&= \{ 0 \leq A \leq D : L(G-D+A) \neq L(G-D+A-P) \}. \\
\Delta({{\mathcal D}}_1 / {{\mathcal D}}) &= \{ 0 \leq A \leq D : {{\mathcal C}}\cap E_A = {{\mathcal C}}_1 \cap E_A \}, \\
&= \{ 0 \leq A \leq D : L(G-D+A) = L(G-D+A-P) \}. \end{aligned}$$
In each case, the two descriptions are clearly equivalent. The first description of $\Gamma({{\mathcal D}}_1 / {{\mathcal D}})$ uses Theorem \[T:ssbounds\]. The second description uses Lemma \[L:O\].
\[L:L\] For $f \in L(G)$, the value $(t^e f)(P)$ is uniquely determined by the values $\{ f(P) : P \in A \}$, for $0 \leq A \leq D,$ if and only if $L(G-A) = L(G-A-P)$.
\[P:L\] For the extension of codes ${{\mathcal C}}/ {{\mathcal C}}_1 = C_L(D,G) / C_L(D,G-P)$, where $D=P_1+\cdots+P_n$ is a sum of $n$ distinct points, $G$ is a divisor disjoint from $D$, and $P$ is a point disjoint from $D$, $$\begin{aligned}
\Gamma({{\mathcal C}}/ {{\mathcal C}}_1) &= \{ 0 \leq A \leq D : {{\mathcal D}}_1 \cap E_A \neq {{\mathcal D}}\cap E_A \}, \\
&= \{ 0 \leq A \leq D : \Omega(G-A-P) \neq \Omega(G-A) \}. \\
\Delta({{\mathcal C}}/ {{\mathcal C}}_1) &= \{ 0 \leq A \leq D : {{\mathcal D}}_1 \cap E_A = {{\mathcal D}}\cap E_A \}, \\
&= \{ 0 \leq A \leq D : \Omega(G-A-P) = \Omega(G-A) \}. \end{aligned}$$
As in Proposition \[P:O\] but use Lemma \[L:L\].
The propositions are related via the dualities $A \in \Gamma({{\mathcal D}}_1 / {{\mathcal D}})$ if and only if $D-A \in \Delta({{\mathcal C}}/ {{\mathcal C}}_1)$ and $A \in \Gamma({{\mathcal C}}/ {{\mathcal C}}_1)$ if and only if $D-A \in \Delta({{\mathcal D}}_1 / {{\mathcal D}})$ (as Theorem \[T:ssbounds\]). The minimal degree of a divisor $A \in \Gamma({{\mathcal D}}_1/{{\mathcal D}})$ or $A \in \Gamma({{\mathcal C}}/{{\mathcal C}}_1)$ is given by the coset distance $d({{\mathcal C}}/ {{\mathcal C}}_1)$ or $d({{\mathcal D}}_1/{{\mathcal D}})$, respectively (as in Corollary \[C:ssbounds\]).
\[P:cb\] $$\begin{aligned}
d({{\mathcal C}}/ {{\mathcal C}}_1) &= \min \{ \deg A : 0 \leq A \leq D \;|\; L(A-C) \neq L(A-C-P) \}, \quad \text{for $C = D-G,$} \\
d({{\mathcal D}}_1 / {{\mathcal D}}) &= \min \{ \deg A : 0 \leq A \leq D \;|\; L(A-C) \neq L(A-C-P) \}, \quad \text{for $C = G-K-P.$} \end{aligned}$$
For a given divisor $C$ and a point $P$, let $$\Gamma_{P}(C) = \{ A : L(A) \neq L(A-P) \wedge L(A-C) \neq L(A-C-P) \},$$ and let $\gamma_P(C)$ be the minimal degree for a divisor $A \in \Gamma_P(C).$ So that $\gamma_P(C) \geq \max \{ 0, \deg C \}.$
\[T:gammacoset\] For the extensions of codes ${{\mathcal D}}_1 / {{\mathcal D}}= C_\Omega(D,G-P) / C_\Omega(D,G)$ and ${{\mathcal C}}/ {{\mathcal C}}_1 = C_L(D,G) / C_L(D,G-P)$, $$\begin{aligned}
A \in \Gamma({{\mathcal D}}_1 / {{\mathcal D}}) &~\Rightarrow~ \deg A \,\geq\, \gamma_P(D-G) \,\geq\, n - \deg G. \\
A \in \Delta({{\mathcal D}}_1 / {{\mathcal D}}) &~\Rightarrow~ \deg A \,\leq\, n - \gamma_P(G-K-P) \,\leq\, n - \deg G + 2g - 1. \\[1ex]
A \in \Gamma({{\mathcal C}}/ {{\mathcal C}}_1) &~\Rightarrow~ \deg A \,\geq\, \gamma_P(G-K-P) \,\geq\, \deg G - 2g +1. \\
A \in \Delta({{\mathcal C}}/ {{\mathcal C}}_1) &~\Rightarrow~ \deg A \,\leq\, n - \gamma_P(D-G) \,\leq\, \deg G. \end{aligned}$$
The lower bounds for $\deg A$ that are obtained with $\gamma_P(D-G)$ and $\gamma_P(G-K-P)$ use the assumption $L(A) \neq L(A-P)$ instead of the stronger assumption $0 \leq A \leq D.$ Thus, when the bound for $\deg A$ is not attained by divisors $A$ of the form $0 \leq A \leq D$, the bounds will not be optimal. Essentially, we separate the problem of finding a small $A \in \Gamma({{\mathcal D}}_1/{{\mathcal D}})$ into two parts: a geometric part that considers all effective divisors $A$ not containing $P$, and an arithmetic part that verifies if $A$ can be represented by a divisor with $0 \leq A \leq D$. Only the first part is considered in this paper. In other words, the bounds that we obtain apply to a different and more general problem, that of recovering local data at a point $P$ from given local data at a divisor $A$, for any divisor $A$ with no base point at $P$. We briefly outline this setting.
Let $X/{{\mathbb F}}$ be a curve, and let $C$ be a divisor on $X$. For a given point $P$ on $X$ define the collection $\Sigma_P(C) = \{ \pi_A : A \geq 0 \}$ of surjective maps $$\pi_A : \Omega(-C-P) \longrightarrow \Omega(-C-P)/\Omega(A-C-P), \quad \text{$A \geq 0.$}$$
The map $\pi_A$ assigns to a differential $\omega \in \Omega(-C-P)$ the local information $\omega$ modulo $\Omega(A-C-P)$, in short the local information of $\omega$ at $A$. Given that $\omega \in \Omega(-C-P)$, any sufficiently large amount of local information determines $\omega$ uniquely. Indeed, for any divisor $A$ of sufficiently large degree, $\Omega(A-C-P) = 0$ and $\pi_A$ is a bijection. For a divisor $A$ with the weaker property $\Omega(A-C-P) = \Omega(A-C),$ the maps $\pi_A = \pi_{A+P}$ agree. In that case, the local information of $\omega$ at $A$ determines uniquely the local information of $\omega$ at $A+P$. If $P$ occurs in the support of $A$ then this means that the local information can be determined with increased precision. For secret sharing we assume that the secret corresponds to a fixed map $\pi_P$. Then the parties that do not know $\pi_P$ a priori are those with $L(A) \neq L(A-P).$ Among those, the parties that can determine $\pi_P$ from $\pi_A$ are those that satisfy $\Omega(A-C-P) = \Omega(A-C)$, or, equivalently, $L(A-C) \neq L(A-C-P).$ Together the conditions define the set $\Gamma_P(C).$ In this setting, the access structure $\Gamma_P(C)$ can be analysed without further assumptions on the representation of the maps $\pi_A$. The image under $\pi_A$ of a differential $\omega \in \Omega(-C-P)$ might be written out explicitly in terms of local parameters and residues, much like an algebraic geometric code, or it might simply be given as a differential $\omega+\eta$ for $\eta \in \Omega(A-C-P).$
Semigroup ideals {#S:SemigroupIdeals}
================
Let $X/{{\mathbb F}}$ be a curve over a field ${{\mathbb F}}$ and let $\Pic(X)$ be the group of divisor classes. Let $\Gamma = \{ A : L(A) \neq 0 \}$ be the semigroup of effective divisor classes. For a given point $P \in X$, let $\Gamma_P = \{ A : L(A) \neq L(A-P) \}$ be the semigroup of effective divisor classes with no base point at $P$. Call $A \in \Gamma_P$ a $P$-denominator for the divisor class $C \in \Pic(X)$ if $A-C \in \Gamma_P.$ So that $A-(A-C)$ expresses $C$ as the difference of two effective divisor classes without base point at $P$. The $P$-denominators for $C$ form the $\Gamma_P$-ideal $$\Gamma_P(C) = \{ A \in \Gamma_P : A-C \in \Gamma_P \}.$$ The ideal structure of the semigroup $\Gamma_P(C)$ amounts to the property $A+E \in \Gamma_P(C)$ whenever $A \in \Gamma_P(C)$ and $E \in \Gamma_P.$ The $\Gamma_P$-ideal of $P$-numerators for $C$ is the ideal $$\Gamma_P(-C) = \{ A \in \Gamma_P : A+C \in \Gamma_P \}.$$ Clearly, $A$ is a $P$-denominator for $C$ if and only if $A-C$ is a $P-$numerator for $C$, that is $$A \in \Gamma_P(C) ~\Leftrightarrow~ A-C \in \Gamma_P(-C),$$ The minimal degree $\gamma_P(C)$ of a $P$-denominator for $C$ is defined as $$\gamma_P(C) = \min \{ \deg A : A \in \Gamma_P(C) \}.$$ The minimal degrees satisfy $$\gamma_P(C) - \gamma_P(-C) = \deg C.$$ The denominator and numerator terminology is borrowed from the ideal interpretation of divisors. Let $O$ be the ring of rational functions in ${{\mathbb F}}(X)$ that are regular outside $P$. For effective divisors $A$ and $B$ disjoint from $P$, the fractional $O$-ideal $\cup_{i \geq 0} L(i P-(B-A)) = J I^{-1}$ is the quotient of the integral $O$-ideals $J = \cup_{i \geq 0} L(i P - B)$ and $I = \cup_{i \geq 0} L(i P-A).$ To a denominator $A$ of smallest degree corresponds an ideal $I$ of smallest norm.\
If either $C \in \Gamma_P$ or $-C \in \Gamma_P$ then the conditions $A \in \Gamma_P$ and $A-C \in \Gamma_P$ are dependent.
\[P:gammap\] For a divisor $C$ on a curve $X$ of genus $g$, $\gamma_P(C) \geq \max \{ 0, \deg C\}.$ Moreover, $$\begin{aligned}
\gamma_P(C) = 0 ~\Leftrightarrow~ -C \in \Gamma_P ~&\Leftrightarrow~ \Gamma_P(C) = \Gamma_P. \\
\gamma_P(C) = \deg C ~\Leftrightarrow~ C \in \Gamma_P ~&\Leftrightarrow~ \Gamma_P(-C) = \Gamma_P.\end{aligned}$$ The inequality is strict if and only if $C, -C \not \in \Gamma_P$ only if $|\deg C| < 2g.$
For suitable choices of the divisor $C$, the parameter $\gamma_P(C)$ gives a lower bound for the coset distance of an algebraic geometric code (Proposition \[P:cb\]) and therefore bounds for the access structure of an algebraic geometric linear secret sharing scheme (Theorem \[T:gammacoset\]). Proposition \[P:gammap\] shows that we can expect improvements over the trivial lower bound $\gamma_P(C) \geq \deg C$ that is used for Theorem \[T:gammacoset\] only if $P$ is a base point for the divisor $C$.\
Let $S$ be a finite set of rational points that includes $P$. For $\Gamma_S = \cap_{P \in S} \Gamma_P,$ let $\Gamma_P(C;S) = \Gamma_P(C) \cap \Gamma_S = \{ A \in \Gamma_S : A-C \in \Gamma_P \}$, and let $\gamma_P(C;S)$ be the minimal degree for a divisor $A \in \Gamma_P(C;S).$
\[L:cosetgammas\] For a given set of rational points $S$ that includes $P$, and for extensions of algebraic geometric codes $C_\Omega(D,G-P) / C_\Omega(D,G)$ and $C_L(D,G) / C_L(D,G-P)$ defined with a divisor $D=P_1+\cdots+P_n$ disjoint from $S$, $$\begin{aligned}
d(C_L(D,G) / C_L(D,G-P)) &\geq \gamma_P(C;S), \quad \text{for $C = D-G,$} \\
d(C_\Omega(D,G-P) / C_\Omega(D,G)) &\geq \gamma_P(C;S), \quad \text{for $C = G-K-P.$} \end{aligned}$$
Proposition \[P:cb\].
To obtain similar estimates for the minimum distance of an algebraic geometric code, we use Proposition \[P:agd\]. Define the $\Gamma_S$-ideals $\Gamma^\ast(C;S) \subseteq \Gamma(C;S)$, $$\begin{aligned}
\Gamma^\ast(C;S) &= \{ A \in \Gamma_S : L(A-C) \neq L(-C) \}, \\
\Gamma(C;S) &= \{ A \in \Gamma_S : L(A-C) \neq 0 \}.\end{aligned}$$ Let $\gamma^\ast(C;S)$ (resp. $\gamma(C;S)$) denote the minimal degree for a divisor $A \in \Gamma^\ast(C;S)$ (resp. $A \in \Gamma(C,S)$).
\[L:codegammas\] For a given set of rational points $S$, and for algebraic geometric codes $C_L(D,G)$ and $C_\Omega(D,G)$ defined with a divisor $D=P_1+\cdots+P_n$ disjoint from $S$, $$\begin{aligned}
&d(C_L(D,G)) \geq \gamma^\ast(C;S) \geq \gamma(C;S), \quad \text{for $C=D-G$}, \\
&d(C_\Omega(D,G)) \geq \gamma^\ast(C;S) \geq \gamma(C;S), \quad \text{for $C=G-K$}. \end{aligned}$$ For $L(-C) = 0$, $\gamma^\ast(C;S) = \gamma(C;S)$.
Proposition \[P:agd\].
The condition $L(-C) = 0$ holds in all cases where the Goppa lower bound $d \geq \deg C$ (Theorem \[T:Goppa\]) is positive. We give lower bounds for $\gamma(C;S)$ using lower bounds for $\gamma_P(C;S)$. With a minor modification, we obtain lower bounds for $\gamma^\ast(C;S).$
\[L:gammas\] Let $S$ be a finite set of rational points. For a divisor $C$, and for a point $P \in S$, $$\begin{aligned}
&\Gamma(C;S) = \Gamma_P(C;S) \cup \Gamma(C+P;S). \\
&\Gamma^\ast(C;S) \subseteq \Gamma_P(C;S) \cup \Gamma^\ast(C+P;S). \end{aligned}$$ Moreover, for $-C \in \Gamma_P$, $$\Gamma^\ast(C;S) \subseteq \Gamma^\ast(C+P;S).$$
For the equality, $L(A-C) \neq 0$ if and only if $L(A-C) \neq L(A-C-P)$ or $L(A-C-P) \neq 0.$ For the inclusion, $L(A-C) \neq L(-C)$ only if $L(A-C) \neq L(A-C-P)$ or $L(A-C-P) \neq L(-C-P).$ Finally, for $A \in \Gamma^\ast(C;S)$ such that $-C \in \Gamma_P$, we have $\dim L(A-C) / L(-C-P) > 1$, and thus $L(A-C-P) \neq L(-C-P).$ So that $A \in \Gamma^\ast(C+P;S).$
\[P:gammas\] $$\begin{aligned}
\gamma(C;S) &\geq \min \{ \gamma_P(C;S), \gamma(C+P;S) \}. \\
\gamma^\ast(C;S) &\geq \min \{ \gamma_P(C;S), \gamma^\ast(C+P;S) \} \backslash \{ 0 \}. \end{aligned}$$
In general $\gamma^\ast(C;S) > 0.$ And $\gamma_P(C;S) = 0$ only if $\gamma_P(C) = 0$ if and only if $-C \in \Gamma_P$, in which case we can omit $\gamma_P(C;S)$ before taking the minimum.
Main theorem {#S:MainTheorem}
============
For a given curve $X / {{\mathbb F}}$, let $C \in \Pic(X)$ be a divisor class and let $P$ be a point on $X$. For the semigroup $\Gamma_P = \{ A : L(A) \neq L(A-P) \}$ and the $\Gamma_P$-ideal $$\Gamma_P(C) = \{ A \in \Gamma_P : A-C \in \Gamma_P \},$$ define the complement $$\Delta_P(C) = \{ A \in \Gamma_P : A-C \not \in \Gamma_P \}.$$
\[L:Dempty\] $$\Delta_P(C) = \emptyset ~\Leftrightarrow~ \Gamma_P(C) = \Gamma_P ~\Leftrightarrow~ -C \in \Gamma_P. \\$$
Let $X$ be of genus $g$ and let $K$ represent the canonical divisor class.
\[L:Ddegree\] In general, $$A \in \Delta_P(C) ~\Leftrightarrow~ K+C+P-A \in \Delta_P(C).$$ For $A \in \Delta_P(C)$, $$\min \{ 0, \deg C \} \;\leq\; \deg A \;\leq\; \max \{ 2g-1, \deg C + 2g-1 \}.$$
This follows from the definition together with the Riemann-Roch theorem.
The following is the analogue of Theorem \[T:cosetbound\] in the language of divisors.
(Coset bound for divisors) \[T:cbdiv\] Let $\{ A_1 \leq A_2 \leq \cdots \leq A_w \} \subset \Delta_P(C)$ be a sequence of divisors with $A_{i+1} \geq A_i + P$, for $i=1,\ldots,w-1.$ Then $\deg A \geq w$, for every divisor $A \in \Gamma_P(C)$ with support disjoint from $A_w-A_1,$ that is $$\gamma_P(C;A_w-A_1) \geq w.$$
After replacing the sequence with an equivalent sequence if necessary, we may assume that $A_1, A_2, \ldots, A_w$ are disjoint from $A$. We obtain two sequences of subspaces. $$\begin{gathered}
L(A_w) \supsetneq L(A_w - P) \supseteq L(A_{w-1}) \supsetneq L(A_{w-1}-P) \supseteq \cdots \\
\cdots \supseteq L(A_2) \supsetneq L(A_2-P) \supseteq L(A_1) \supsetneq L(A_1-P).\end{gathered}$$ $$\begin{gathered}
\Omega(A_w-C) \subsetneq \Omega(A_w-C-P) \subseteq \Omega(A_{w-1}-C) \subsetneq \Omega(A_{w-1}-C-P) \subseteq \cdots \\
\cdots \subset \Omega(A_2-C) \subsetneq \Omega(A_2-C-P) \subseteq \Omega(A_1-C) \subsetneq \Omega(A_1-C-P).\end{gathered}$$ For $i=1,2,\ldots,w,$ choose $$f_i \in L(A_i) \backslash L(A_i-P) ~~\text{and}~~ \eta_i \in \Omega(A_i-C-P) \backslash \Omega(A_i-C).$$ Let $A \in \Gamma_P$ be of degree $\deg A < w.$ Then there exists a linear combination $f$ of $f_1, f_2, \ldots, f_w$ that vanishes on $A$. If $f_i$ is the leading function in the linear combination then $f \in L(A_i - A) \backslash L(A_i-A-P)$ and $f \eta_i \in \Omega(-C-P+A) \backslash \Omega(-C+A)$. Thus $A-C \not \in \Gamma_P$ and $A \not \in \Gamma_P(C).$
For a divisor $B$, let $$\begin{aligned}
\Delta_P(B,C) &= \{ B+iP : i \in {{\mathbb Z}}\} \cap \Delta_P(C), \\
&= \{ B+iP \in \Gamma_P, B-C+iP \not \in \Gamma_P \}.\end{aligned}$$
\[L:DminD\] To the set $\Delta_P(B,C)$ corresponds a dual set $$\begin{aligned}
\Delta_P(B-C,-C) &= \{ B-C+iP : i \in {{\mathbb Z}}\} \cap \Delta_P(-C), \\
&= \{ B-C+iP \in \Gamma_P, B+iP \not \in \Gamma_P \},\end{aligned}$$ such that $\# \Delta_P(B,C) - \# \Delta_P(B-C,-C) = \deg C.$ Furthermore, $$\# \Delta_P(B,C) = \begin{cases} \deg C, &\quad \text{if $C \in \Gamma_P.$} \\
0, &\quad \text{if $-C \in \Gamma_P.$}
\end{cases}$$ In particular, $$\# \Delta_P(B,C) = \begin{cases} \deg C, &\quad \text{if $\deg C \geq 2g.$} \\
0, &\quad \text{if $\deg C \leq -2g.$}
\end{cases}$$
For $i_0$ large enough, $$\begin{aligned}
&\# \Delta_P(B,C) - \# \Delta_P(B-C,-C) \\
=~&\# \{ i \leq i_0 : B+iP \in \Gamma_P, B-C+iP \not \in \Gamma_P \} \\
&\quad - \# \{ i \leq i_0 : B+iP \not \in \Gamma_P, B-C+iP \in \Gamma_P \} \\
=~&\sum_{i \leq i_0} (l(B+iP)-l(B+iP-P))-(l(B-C+iP)-l(B-C+iP-P)) \\
=~&\dim L(B+i_0 P) - \dim L(B-C+i_0 P) = \deg C.\end{aligned}$$ For the remainder use Lemma \[L:Dempty\].
\[C:cbfrac\] For any choice of divisor $B$, there is a pair of equivalent bounds $$\begin{aligned}
&\gamma_P(C) \geq \# \Delta_P(B,C).
&\gamma_P(-C) \geq \# \Delta_P(B-C,-C).\end{aligned}$$
For the first inequality, the elements $A_1, A_2, \ldots, A_w \in \Delta_P(B,C)$, ordered from lowest to highest degree, meet the conditions of the theorem. Similar for the second inequality. Equivalence follows from $\gamma_P(C) - \gamma_P(-C) = \deg C$ and the previous lemma.
If $A \in \Gamma_P(E)$ and $E \in \Gamma_P(C)$ then $A \in \Gamma_P(C).$ For $E \in \Gamma_P(C)$, $$\Delta_P(C) \subset \Delta_P(E).$$
The first claim is immediate from the definitions, in particular $A-E \in \Gamma_P$ and $E-C \in \Gamma_P$ implies $A-C \in \Gamma_P.$ For $E \in \Gamma_P(C)$, the first claim shows that $A \not \in \Gamma_P(E)$ whenever $A \not \in \Gamma_P(C).$
Order bound and floor bound {#S:OrderFloor}
===========================
We unify and improve two known lower bounds for the minimum distance of an algebraic geometric code. Let $S$ be a given set of rational points, and let $C_L(D,G)$ and $C_\Omega(D,G)$ be algebraic geometric codes defined with a divisor $D=P_1+\cdots+P_n$ disjoint from $S$. With Lemma \[L:codegammas\], $$\begin{aligned}
&d(C_L(D,G)) \geq \gamma^\ast(C;S), \quad \text{for $C=D-G$}, \\
&d(C_\Omega(D,G)) \geq \gamma^\ast(C;S), \quad \text{for $C=G-K$}. \end{aligned}$$
\[P:order\] For points $Q_0, \ldots, Q_{r-1} \in S$, define divisors $C_0 \leq C_1 \leq \cdots \leq C_r$ such that $C_0=C$ and $C_{i+1}=C_i+Q_i$, for $i=0,\ldots,r-2.$ Then $$\gamma^\ast(C;S) \geq\; \min \{ \gamma_{Q_0}(C_0;S), \gamma_{Q_1}(C_1;S), \ldots, \gamma_{Q_{r-1}}(C_{r-1};S), \gamma^\ast(C_r;S) \} \backslash \{0\}, \\$$ In general, $\gamma^\ast(C_r;S) \geq \deg C + r.$
Proposition \[P:gammas\] gives $\gamma^\ast(C_i;S) \geq \min \{ \gamma_{Q_i}(C_i;S), \gamma^\ast(C_{i+1};S) \} \backslash \{0\}.$
We give a formulation of the order bound for an algebraic geometric code $C_\Omega(D,G).$ For the case that $G$ is supported in two points, a similar result formulated in terms of near order functions can be found in [@Caretal07 Theorem 1].
(Order bound [@Bee07FF Theorem 7]) Let ${\cal C}$ be an algebraic curve and $G$ a rational divisor. Let ${\cal P}$ be a set of rational points not occuring in the support of the divisor $G$. Then we have $$d ({C}_{\cal P}(G) ) \geq d_{\cal P}(G) \geq d(G).$$
Using [@Bee07FF Remark 5, Definition 6], we expand the theorem in the notation of the current paper. In comparison with the original theorem, we have removed the condition that the divisors $B_0, \ldots, B_r$ are disjoint from $D$.
(Order bound [@Bee07FF]) \[T:order\] Let $C_\Omega(D,G)$ be an algebraic geometric code, and let $G=K+C.$ For a sequence of points $Q_0, \ldots, Q_{r-1}$ disjoint from $D$, let $C_0=C$ and $C_{i+1}=C_i+Q_i$, for $i=0,\ldots,r-2.$ $${{\mathcal C}}_0 = C_\Omega(D,K+C) \supseteq {{\mathcal C}}_1 = C_\Omega(D,K+C_1) \supseteq \cdots \supseteq {{\mathcal C}}_r = C_\Omega(D,K+C_r).$$ If ${{\mathcal C}}_i \neq {{\mathcal C}}_{i+1}$ then a word in ${{\mathcal C}}_i \backslash {{\mathcal C}}_{i+1}$ has weight $w \geq \# \Delta_{Q_i}(0,C_i)$. For $r$ large enough, $$d(C_\Omega(D,G)) \geq \min \{ \# \Delta_{Q_i}(0,C_i) : {{\mathcal C}}_i \neq {{\mathcal C}}_{i+1} \}.$$ Moreover, for a sequence of divisors $B_0, \ldots, B_{r-1}$, $$d(C_\Omega(D,G)) \geq \min \{ \# \Delta_{Q_i}(B_i,C_i) : {{\mathcal C}}_i \neq {{\mathcal C}}_{i+1} \}.$$
The order bound for the minimum distance combines Proposition \[P:order\] with the estimates $\gamma_{Q_i}(C_i;S) \geq \gamma_{Q_i}(C_i) \geq \Delta_{Q_i}(B_i,C_i)$ in Corollary \[C:cbfrac\].
We analyse the choice of the points $Q_0, Q_1, \ldots, Q_r$. In [@Bee07FF], the choice of the points is unrestricted, and an example is given where the optimal lower bound is obtained with a choice of $Q_i$ outside $G$. On the other hand, Proposition \[P:gammap\] shows that $\gamma_{Q_i}(C_i) \geq \deg C_i$. Thus, we may assume that the minimum $\min \{ \gamma_{Q_i}(C_i) \} \backslash \{0\}$ is taken over an interval $i=0,1,\ldots,r$ such that, for all $i$ in the interval, either $\gamma_{Q_i}(C_i)=0$ or $\gamma_{Q_i}(C_i) > \deg C_i$. With Proposition \[P:gammap\] this implies that either $-C_i \in \Gamma_{Q_i}$ or $C_i \not \in \Gamma_{Q_i}.$ In both cases, we can conclude, for $C_i \neq 0$, that $C_i \not \in \Gamma_{Q_i}$, i.e. that $L(C_i) = L(C_i-Q_i).$ The same conclusion can be reached with Lemma \[L:DminD\] if the argument is repeated for $\Delta_{Q_i}(B_i,C_i)$ instead of $\gamma_{Q_i}(C_i).$ The following stronger result holds.
The maximum in the order bound is attained for a choice of points $Q_0, Q_1, \ldots, Q_{r-1}$ such that, for $i=0,1,\ldots,r-1$, either $C_i=0$, or $Q_i, \ldots, Q_{r-1}$ are base points of the divisor $C_i$. In particular, if $C_i$ is a nonzero effective divisor, we may restrict the choice for $Q_i, \ldots, Q_{r-1}$ to points in the support of $C_i$.
For $Q \in \{ Q_i, \ldots, Q_{r-1} \},$ let $j$ be minimal in $\{ i, \ldots, r-1 \}$ such that $Q_j=Q$. If $C_j \neq 0$, we may assume as explained above, that $C_j \not \in \Gamma_Q$. With $E = Q_i + \cdots + Q_{j-1} \in \Gamma_Q$ and $C_j=C_i+E$ it follows that $C_i \not \in \Gamma_Q$. If $C_j = 0$ then either $i=j$, in which case $C_i=0$, or $i<j$, in which case $\deg C_i < 0$ and $C_i \not \in \Gamma_Q.$
In [@Bee07FF Example 8], the minimum distance lower bound for a code $C_\Omega(D,5P)$ on the Klein curve is improved with a choice $Q_0=P, Q_1=Q \neq P.$ For the example, $5P=K+2P-Q$ and $6P=K+Q+R$, so that $C_0=2P-Q$ and $C_1=Q+R.$ Indeed, with the proposition, we can expect improvements only with $Q_1 = Q$ or with $Q_1 = R$.\
To improve the order bound we apply the main theorem with a different format for the divisors $A_1, \ldots, A_w.$ Let $$\begin{aligned}
\Delta_P(\leq B,C) &= \{ B+iP \in \Gamma_p : B-C-iP \not \in \Gamma_P \;\wedge\; i \leq 0\}, \\
\Delta_P(\geq B+P,C) &= \{ B+iP \in \Gamma_p : B-C-iP \not \in \Gamma_P \;\wedge\; i \geq 1 \},\end{aligned}$$ be a partition of the set $\Delta_P(B,C)$ into divisors of small and large degree.
\[L:cnt\] $$\# \Delta_P(\leq B,C) = \dim L(B) - \dim L(B-C) + \# \Delta_P(\leq B-C,-C).$$
Similar to the proof of Lemma \[L:DminD\], but use $i_0=0.$
\[T:cbabz\] (ABZ bound for cosets) Let $C$ be a divisor and let $P$ be a point. For $G=K+C=A+B+Z,$ $Z \geq 0$, $$\begin{aligned}
\gamma_P(C;Z \cup P) &\geq \# \Delta_P(\leq A,C) + \# \Delta_P(\leq B,C). \\\end{aligned}$$
With Lemma \[L:Ddegree\], a divisor $A' \in \Delta_P(C)$ if and only if $K+C+P-A' \in \Delta_P(C)$. And $A' \leq A$ if and only if $K+C+P-A' \geq K+C+P-A = B+P+Z.$ The elements $A_1, A_2, \ldots, A_w \in \Delta_P(\leq B,C) \cup \Delta_P(\geq B+P+Z,C)$, ordered from lowest to highest degree, meet the conditions of Theorem \[T:cbdiv\], with $w = \# \Delta_P(\leq A,C) + \# \Delta_P(\leq B,C).$
The lower bound $\# \Delta_P(B,C)$ that is used for the order bound takes into account only the number of divisors in a delta set $\Delta_P(B,C)$. The improved bounds in Theorem \[T:cbabz\] are possible by considering also the degree distribution of divisors in the delta set. For $Z=0$, the bounds in the theorem include those used in the order bound (Theorem \[T:order\]). The floor bound (Theorem \[T:floor\]) sometimes exceeds the order bound. The ABZ bound for codes (Theorem \[T:ABZcodes\]) gives an improvement and generalization of the floor bound. We show that the bounds in the theorem not only include those obtained with the order bound but also those obtained with the ABZ bound for codes. In each case, the coset decoding procedure in the appendix decodes efficiently up to half the bound.
\[T:Floor2\] (ABZ bound for codes) Let $G = K+C = A+B+Z,$ for $Z \geq 0$. For $D$ with $D \cap Z = \emptyset$, a nonzero word in $C_\Omega(D,G)$ has weight $w \geq l(A)-l(A-C)+l(B)-l(B-C).$
Let $P$ be a point on the curve not in the support of $D$, if necessary it can be chosen over an extension field. We use Proposition \[P:order\] with $S=Z \cup P$ and $Q_0=Q_1=\ldots=Q_{r-1}=P$. $$\gamma^\ast(C;S) \geq \min \{ \gamma_{P}(C;S), \gamma_{P}(C+P;S), \ldots, \gamma_{P}(C+(r-1)P;S), \gamma^\ast(C+rP;S) \} \backslash \{0\}.$$ Now use Theorem \[T:cbabz\] with $K+C+iP = A+B+(Z+iP)$, $$\gamma_P(C+iP;S) \geq \# \Delta_P(\leq A,C+iP) + \# \Delta_P(\leq B,C+iP).$$ With Lemma \[L:cnt\], $$\begin{aligned}
\gamma_P(C+iP;S) &\geq l(A)-l(A-C-iP)+l(B)-l(B-C-iP) \\
&\geq l(A)-l(A-C)+l(B)-l(B-C).\end{aligned}$$ Hence, by taking $r$ large enough, $\gamma^\ast(C;S) \geq l(A)-l(A-C)+l(B)-l(B-C).$
Neither the $ABZ$ bound for codes, nor the $ABZ$ bound for cosets gives an improvement in general. For $Z=0$, both bounds return previously known bounds, namely the Goppa bound and the order bound, respectively. For carefully chosen nontrivial $Z$, there are possible improvements. If we apply Lemma \[L:cnt\] with both $A$ and $B$, $$\begin{aligned}
\# \Delta_P(\leq A,C) = \dim L(A) - \dim L(A-C) + \# \Delta_P(\leq A-C,-C), \\
\# \Delta_P(\leq B,C) = \dim L(B) - \dim L(B-C) + \# \Delta_P(\leq B-C,-C),\end{aligned}$$ and add the two equations, then we see that the improvement of the ABZ coset bound applied to $G=K+C=A+B+Z$ over the floor bound applied to $G=K+C=A+B+Z$ is given by the ABZ coset bound applied to the dual decomposition $G'=K-C=(A-C)+(B-C)+Z.$ For $Z=0$, we recover that the improvement of the order bound applied to $G=K+C$ over the Goppa bound $\deg C$ is given by the order bound applied to $G'=K-C$ (Lemma \[L:DminD\] and Corollary \[C:cbfrac\]).\
We consider the special case of the order bound with $B_0 = \cdots = B_{r-1} = 0$ and $Q_0 = \cdots = Q_{r-1} = P.$ For codes of the form $C_L(D,\rho P)^\perp = C_\Omega(D,\rho P)$ or of the form $C_L(D,K+P+\rho P)^\perp = C_\Omega(D,K+P+\rho P)$ the resulting bound can be formulated entirely in terms of the numerical semigroup $S$ of Weierstrass $P$-nongaps. For the first code use $C = \rho P - K,$ and for the second $C = \rho P+P.$ For the delta sets we obtain $$\begin{aligned}
p P \in \Delta_P(\rho P -K) &~\Leftrightarrow~ p P \in \Gamma_P \wedge K + p P - \rho P \not \in \Gamma_P. \\
&~\Leftrightarrow~ p \in S \wedge \rho - p + 1 \in S, \\
p P \in \Delta_P(\rho P ) &~\Leftrightarrow~ p P \in \Gamma_P \wedge p P - \rho P-P \not \in \Gamma_P. \\
&~\Leftrightarrow~ p \in S \wedge p-\rho-1 \not \in S. \end{aligned}$$ The first of the two bounds in the following theorem is the Feng-Rao bound [@FenRao93], [@CamFarMun00]. The second bound is different when the canonical divisor $K \not \sim (2g-2)P.$
(Feng-Rao bound) Let $S$ be the semigroup of Weierstrass $P$-nongaps. $$d(C_L(D,\rho P)^\perp) \geq \min \{ \#A[\rho'] : \rho' > \rho \} \backslash \{ 0 \},$$ where $A[\rho] = \{ p \in S | \rho - p \in S \}$. $$d(C_L(D,K + \rho P)^\perp) \geq \min \{ \#B[\rho'] : \rho' > \rho \} \backslash \{ 0 \},$$ where $B[\rho] = \{ p \in S | p - \rho \not \in S \}$.
Apply Proposition \[P:order\] with the given delta sets.
Delta sets {#S:DeltaSets}
==========
For a divisor $C$ and a point $P$, we defined the $\Gamma_P$-ideal $\Gamma_P(C) = \{ A \in \Gamma_P : A-C \in \Gamma_P \}.$ Theorem \[T:cbdiv\] gives a lower bound for $\deg A$, for $A \in \Gamma_P(C)$, in terms of the complement $\Delta_P(C) = \{ A \in \Gamma_P : A-C \not \in \Gamma_P \}$. Theorem \[T:order\] and Theorem \[T:cbabz\] are formulated in terms of the subsets $\Delta_P(B,C)$ and $\Delta_P(\leq B, C)$, respectively, for a suitable choice of divisor $B$. The computation of optimal lower bounds requires either a complete description of the delta set (for the main theorem) or at least a description from which the size of the sets $\Delta_P(B,C)$ or $\Delta_P(\leq B,C)$ can be computed (for the other two theorems). We collect some straightforward relations that can be used to construct delta sets, to compare delta sets, or to compare sizes of delta sets. Most relations come in pairs such that $A \in \Delta_P(C)$ (i.e., $A \in \Gamma_P, A-C \not \in \Gamma_P$) corresponds to $A-C \in \Delta_P(-C)$ (i.e., $A \not \in \Gamma_P, A-C \in \Gamma_P$). The proofs in this section are entirely straightforward, in most cases applying the definition of $\Delta_P(C)$ is enough, and no proofs are included. In general, for $E \in \Gamma_P$, $\Delta_P(C) \subset \Delta_P(C+E).$ Lemma \[L:dset\] gives a precise version and its dual.
\[L:dset\] Let $C$ be a divisor and $P$ a point. For $E \in \Gamma_P$, $$\begin{aligned}
&A \in \Delta_P(C) \\
&\qquad ~\Leftrightarrow~ A \in \Delta_P(C+E) \;\wedge\; A+E \in \Delta_P(C+E). \\[1ex]
&A-C-E \in \Delta_P(-C-E) \\
&\qquad ~\Leftrightarrow~ A-C-E \in \Delta_P(-C) \;\wedge\; A-C \in \Delta_P(-C).\end{aligned}$$
For the four relations on the right we describe when the reverse implication fails.
\[L:diffA\] $$\begin{aligned}
&A \in \Delta_P(C+E) \;\wedge\; A \not \in \Delta_P(C), \\
&\qquad ~\Leftrightarrow~ A\in \Delta_P(C+E) \;\wedge\; A+E \not \in \Delta_P(C+E), \\
&\qquad ~\Leftrightarrow~ A\in \Delta_P(C+E) \;\wedge\; A-C \in \Gamma_P, \\
&\qquad ~\Leftrightarrow~ A-C \in \Delta_P(E) \;\wedge\; A \in \Gamma_P. \\[1ex]
&A-C \in \Delta_P(-C) \;\wedge\; A-C-E \not \in \Delta_P(-C-E), \\
&\qquad ~\Leftrightarrow~ A-C \in \Delta_P(-C) \;\wedge\; A-C-E \not \in \Delta_P(-C), \\
&\qquad ~\Leftrightarrow~ A-C \in \Delta_P(-C) \;\wedge\; A-C-E \not \in \Gamma, \\
&\qquad ~\Leftrightarrow~ A-C \in \Delta_P(E) \;\wedge\; A \not \in \Gamma_P.\end{aligned}$$
The second group follows with a substitution $A \mapsto A-C-E, C \mapsto -C-E$.
\[L:diffB\] $$\begin{aligned}
&A-C-E \in \Delta_P(-C) \;\wedge\; A-C-E \not \in \Delta_P(-C-E), \\
&\qquad ~\Leftrightarrow~ A-C-E\in \Delta_P(-C) \;\wedge\; A-C \not \in \Delta_P(-C), \\
&\qquad ~\Leftrightarrow~ A-C-E\in \Delta_P(-C) \;\wedge\; A \in \Gamma_P, \\
&\qquad ~\Leftrightarrow~ A \in \Delta_P(E) \;\wedge\; A-C-E \in \Gamma_P. \\[1ex]
&A \in \Delta_P(C+E) \;\wedge\; A-E \not \in \Delta_P(C), \\
&\qquad ~\Leftrightarrow~ A \in \Delta_P(C+E) \;\wedge\; A-E \not \in \Delta_P(C+E), \\
&\qquad ~\Leftrightarrow~ A \in \Delta_P(C+E) \;\wedge\; A-E \not \in \Gamma, \\
&\qquad ~\Leftrightarrow~ A \in \Delta_P(E) \;\wedge\; A-C-E \not \in \Gamma_P.\end{aligned}$$
As in Section \[S:MainTheorem\], let $$\begin{aligned}
\Delta_P(B,C) &= \{ B+iP : i \in {{\mathbb Z}}\} \cap \Delta_P(C), \\
&= \{ B+iP \in \Gamma_P, B-C+iP \not \in \Gamma_P \}.\end{aligned}$$ Furthermore, let $$\begin{aligned}
I_P(B,C) &= \{ i \in {{\mathbb Z}}: B+iP \in \Delta_P(C) \}, \\
&= \{ i \in {{\mathbb Z}}: B+iP \in \Gamma_P, B-C+iP \not \in \Gamma_P \}. \\
I^\ast_P(B,C) &= \{ i \in {{\mathbb Z}}: B-C+iP \in \Delta_P(-C) \}, \\
&= \{ i \in {{\mathbb Z}}: B-C+iP \in \Gamma_P, B+iP \not \in \Gamma_P \}.\end{aligned}$$ So that $I^\ast_P(B,C) = I_P(B-C,-C),$ and $\# I_P(B,C) - \# I^\ast_P(B,C) = \deg C$ (Lemma \[L:DminD\]). We rephrase some of the previous relations.
$$\begin{aligned}
&I_P(B,C) = I_P(B,C+E) \cap I_P(B+E,C+E). \\
&I^\ast_P(B,C+E) = I^\ast_P(B-E,C) \cap I^\ast_P(B,C). \end{aligned}$$
$$\begin{aligned}
&i \in I_P(B,C+E) \backslash I_P(B+E,C+E) ~\Leftrightarrow~ i \in I_P(B-C,E) \;\wedge\; B+iP \in \Gamma_P, \\
&i \in I_P(B,C+E) \backslash I_P(B-E,C+E) ~\Leftrightarrow~ i \in I_P(B,E) \;\wedge\; B-C+E+iP \not \in \Gamma_P.\end{aligned}$$
$$\begin{aligned}
&I_P(B,C+E) \backslash I_P(B,C) \cup I^\ast(B,C) \backslash I^\ast(B,C+E) = I_P(B-C,E). \\
&I_P(B,C+E) \backslash I_P(B-E,C) \cup I^\ast(B-E,C) \backslash I^\ast_P(B,C+E) = I_P(B,E).\end{aligned}$$
We describe the first partition for the following choice of divisors. For divisors $B_0$ and $C_0$ of degree zero, and for a point $Q$, let $B=B_0, C=C_0-2gQ, E=4gQ.$ Then $$I_P(B_0,C_0+2gQ) \cup I_P(B_0-C_0+2gQ,-C_0+2gQ) = I_P(B_0-C_0+2gQ,4gQ).$$ In general, $$\{ 0, \ldots, 2g-1 \} \subset I_P(B_0-C_0+2gQ,4gQ) \subset \{ -2g, \ldots, 4g-1 \}.$$ The first inclusion follows with the definition of $I_P(B,C)$. For the second inclusion, Lemma \[L:Ddegree\] gives $0 \leq i+2g \leq 6g-1.$
\[P:6g\] Let $B_0$ and $C_0$ be divisor classes of degree zero. Define partitions $\{ -2g, \ldots, -1 \} = N_1 \cup G_1,$ $\{ 0, \ldots,2g-1 \} = N_2 \cup G_2,$ and $\{ 2g, \ldots, 4g-1 \} = N_3 \cup G_3,$ such that $$\begin{aligned}
&k \in N_1 ~\Leftrightarrow~ B_0-C_0+2gQ+kP \in \Gamma_P, \\
&k \in N_2 ~\Leftrightarrow~ B_0+kP \in \Gamma_P, \\
&k \in N_3 ~\Leftrightarrow~ B_0-C_0-2gQ+kP \in \Gamma_P.\end{aligned}$$ Then $\# N_i = \# G_i = g$, for $i=1,2,3.$ Moreover $$I_P(B_0,C_0+2gQ) = N_2 \cup G_3 ~~\text{and}~~ I_P(B_0-C_0+2gQ,-C_0+2gQ) = N_1 \cup G_2.$$
$$\begin{array}{cccc}
&\{-2g, \ldots, -1\} &\{0, \ldots, 2g-1\} &\{2g, \ldots, 4g-1 \} \\[1ex]
\begin{array}{rl} \{ k :&B_0 + kP \in \Gamma_P \;\wedge \\
&B_0-C_0-2gQ+kP \in \Gamma_P \} \end{array} &- &- &N_3 \\[1.5ex]
\begin{array}{rl} \{ k :&B_0 + kP \in \Gamma_P \;\wedge \\
&B_0-C_0-2gQ+kP \not \in \Gamma_P \} \end{array} &- &N_2 &G_3 \\[1.5ex]
\begin{array}{rl} \{ k :&B_0 + kP \not \in \Gamma_P \;\wedge \\
&B_0-C_0+2gQ+kP \in \Gamma_P \} \end{array} &N_1 &G_2 &- \\[1.5ex]
\begin{array}{rl} \{ k :&B_0 + kP \not \in \Gamma_P \wedge \\
&B_0-C_0+2gQ+kP \not \in \Gamma_P \} \end{array} &G_1 &- &-
\end{array}$$
Discrepancies {#S:Discrepancies}
=============
We continue the description of a delta set $\Delta_P(C)$ in terms of other known delta sets. The results in the previous section show that differences between similar delta sets, such as $\Delta_P(C+E)$ and $\Delta_P(C)$, for $E \in \Gamma_P$, can be described in terms of the delta set $\Delta_P(E).$ In this section, we refine the results for the special case that $E = Q$ is a point different from $P$.
For distinct points $P$ and $Q$, $\Delta_P(Q)=\Delta_Q(P)$.
$$\begin{gathered}
A \in \Delta_P(Q) ~\Leftrightarrow~ L(A) \neq L(A-P) \;\wedge\; L(A-Q) = L(A-Q-P) \\
~\Leftrightarrow~ L(A) \neq L(A-Q) \;\wedge\; L(A-P) = L(A-P-Q) ~\Leftrightarrow~ A \in \Delta_Q(P).\end{gathered}$$
Let $D(P,Q) = \Delta_P(Q) = \Delta_Q(P).$ We call a divisor $A \in D(P,Q)$ a discrepancy for the points $P$ and $Q$.
\[L:discdeg\] A divisor $A \in D(P,Q)$ is of degree $0 \leq \deg A \leq 2g$. The cases $\deg A = 0$ and $\deg A = 2g$ correspond to unique divisor classes $A=0$ and $A=K+P+Q$, respectively. Furthermore, $$A \in D(P,Q) ~\Leftrightarrow~ K+P+Q-A \in D(P,Q).$$
Use Lemma \[L:Ddegree\].
The set $\Delta_P(B,Q)$ is defined as $\{ B+kP : k \in {{\mathbb Z}}\} \cap \Delta_P(Q)$. It follows from Lemma \[L:DminD\] that $\Delta_P(B,Q)$ is a singleton set. For a divisor $B$, and for a given choice of distinct points $P$ and $Q$, define $$B_Q = \Delta_P(B,Q), \qquad B_P = \Delta_Q(B,P).$$
\[L:AisAp\] $$A \in D(P,Q) ~\Leftrightarrow~ A = A_Q ~\Leftrightarrow~ A = A_P.$$
Clear after writing $A_Q = \Delta_P(A,Q)$ and $A_P = \Delta_Q(A,P).$
\[L:singleton\] For distinct points $P$ and $Q$, and for a divisor $B$, $B_Q = \Delta_P(B,Q) = B+kP,$ for $k$ minimal such that $L(B+kP) \neq L(B+kP-Q)$. For a general $k$, $B+kP \in \Gamma_Q$ if and only if $B+kP \geq B_Q.$
$$\begin{aligned}
B+kP \in \Delta_P(Q) &~\Leftrightarrow~ B+kP \in \Delta_Q(P) \\
&~\Leftrightarrow~ B+kP \in \Gamma_Q \;\wedge\; B+(k-1)P \not \in \Gamma_Q.\end{aligned}$$
For distinct points $P$ and $Q$, and for a divisor $B$, let $$D_B(P,Q) = \{ B+iP+jQ : i, j \in {{\mathbb Z}}\} \cap D(P,Q),$$ and define functions $\sigma = \sigma_B, \tau = \tau_B : {{\mathbb Z}}\longrightarrow {{\mathbb Z}}$ such that $$\begin{aligned}
&B+iP+jQ \in \Gamma_P ~\Leftrightarrow~ j \geq \sigma(i), \\
&B+iP+jQ \in \Gamma_Q ~\Leftrightarrow~ i \geq \tau(j).\end{aligned}$$ With the lemma, $$\begin{aligned}
&(B+iP)_P = \Delta_Q(B+iP,P) = B+iP+\sigma(i)Q. \\
&(B+jQ)_Q = \Delta_P(B+jQ,Q) = B+\tau(j)P+jQ.\end{aligned}$$
For a divisor $B$, $$D_B(P,Q) = \{ B+iP+\sigma(i)Q : i \in {{\mathbb Z}}\} = \{ B+\tau(j)P+jQ : j \in {{\mathbb Z}}\}.$$ In particular, the functions $\sigma = \sigma_B$ and $\tau = \tau_B$ are mutual inverses and describe permutations of the integers. For a divisor $B$ of degree zero, and for $i \in {{\mathbb Z}}$, $-i \leq \sigma(i), \tau(i) \leq 2g-i.$ For $m$ such that $mP \sim mQ$, the functions $i+\sigma(i), j+\tau(j)$ only depend on $i,j$ modulo $m$. The functions $\sigma, \tau$ are determined by their images on a full set of representatives for ${{\mathbb Z}}/m {{\mathbb Z}}.$
For the second claim use Lemma \[L:discdeg\]. Finally, $B+iP+jQ \in D(P,Q)$ only depends on the divisor class of $B+iP+jQ$ and therefore $$B+iP+jQ \in D(P,Q) ~\Leftrightarrow~ B+(i+m)P+(j-m)Q \in D(P,Q),$$ so that $\sigma(i+m) = \sigma(i)-m.$
The discrepancies $D_B(P,Q)$ serve as an index set for a common basis of the vector spaces $L(B+aP+bQ)$, for $a, b \in {{\mathbb Z}}.$
\[T:basis\] $$\dim L(B+aP+bQ) = \# \{ B+iP+jQ \in D(P,Q) : i \leq a \wedge j \leq b \}.$$
$\dim L(B+aP+bQ) \neq \dim L(B+aP+bQ-P)$ if and only if $B+aP+bQ \in \Gamma_P$ if and only if $B+aP+bQ \geq (B+aP)_P \in D_B(P,Q)$ if and only if there exists $B+iP+jP \in D_B(P,Q)$ with $i=a, j \leq b.$ Use induction on $a$ to complete the proof.
\[T:ApA\] For given distinct points $P$ and $Q$, and for divisors $A$ and $C$, $$A \in \Delta_P(C+Q) ~\Leftrightarrow~ A_P \leq A \leq (A-C)_P + C.$$ Moreover, $$\begin{aligned}
A_P = A \leq (A-C)_P + C &~\Leftrightarrow~ A \in \Delta_P(C+Q) \;\wedge\; A-Q \not \in \Delta_P(C+Q). \\
A_P \leq A = (A-C)_P + C &~\Leftrightarrow~ A \in \Delta_P(C+Q) \;\wedge\; A+Q \not \in \Delta_P(C+Q), \\
&~\Leftrightarrow~ A \in \Delta_P(C+Q) \;\wedge\; A \not \in \Delta_P(C). \end{aligned}$$
With Lemma \[L:singleton\], $A \in \Gamma_P$ if and only if $A \geq A_P,$ and $A-C-Q \not \in \Gamma_P$ if and only if $A-C \leq (A-C)_P.$ The last claims use Lemma \[L:diffB\] (part two) and Lemma \[L:diffA\] (part one), respectively.
Note that the divisors $A_P,$ $A$, and $(A-C)_P+C$, have the same multiplicities at any point other than $Q$. Let $B_0$ and $C_0$ be divisors of degree zero, and let $\sigma = \sigma_{B_0}$ and $\sigma' = \sigma_{B_0-C_0}$.
$$\begin{aligned}
&B_0+kP+\ell Q \in \Delta_P(C_0+iP+jQ+Q). \\
&\quad ~\Leftrightarrow ~k+\sigma(k) \;\leq\; k+\ell \;\leq\; (k-i)+\sigma'(k-i) + i+j. \\
&\quad ~\Leftrightarrow ~\sigma(k) \;\leq\; \ell \;\leq\; \sigma'(k-i)+j. \\\end{aligned}$$
For $A = B_0+kP+\ell Q,$ $A_P = B_0+kP+\sigma(k) Q$. For $C=C_0+iP+jQ$, $(A-C)_P = B_0-C_0+(k-i)P+\sigma'(k-i)Q.$ Now use the theorem, and compare either the degrees of the divisors $A_P$, $A$, and $(A-C)_P+C$, or their multiplicities at $Q$.
Let $\tau$ and $\tau'$ be the inverse functions for $\sigma=\sigma_{B_0}$ and $\sigma'=\sigma_{B_0-C_0}$, respectively. Let $$\begin{aligned}
&d_P(k) = \deg (B_0+kP)_P = k+\sigma(k), \\
&d_Q(\ell) = \deg (B_0+\ell Q)_Q = \tau(\ell)+\ell. \\[1ex]
&d'_P(k-i) = \deg (B_0-C_0+(k-i)P)_P = k-i+\sigma'(k-i), \\
&d'_Q(\ell-j) = \deg (B_0-C_0+(\ell-j)Q)_Q = \tau'(\ell-j)+\ell-j. \end{aligned}$$ For $mP \sim mQ,$ the functions $d_P, d_Q$ and $d'_P, d'_Q$ are defined modulo $m$.
\[P:tables\] Let $A=B_0+kP+\ell Q$, and let $C=C_0+iP+jQ+Q$. $$\begin{aligned}
&A_P = A \leq (A-C)_P + C \\
&\quad ~\Leftrightarrow~ k = k^+ = \tau(\ell) \;\wedge\; \ell-j \leq \sigma'(k^+-i). \\
&\quad ~\Leftrightarrow~ k = k^+ = d_Q(\ell) - \ell \;\wedge\; d_Q(\ell) \leq d'_P(k^+-i) + i+j. \\[1ex]
&A_P \leq A = (A-C)_P + C \\
&\quad ~\Leftrightarrow~ k = k^- = \tau'(\ell-j)+i \;\wedge\; \sigma(k^-) \leq \ell. \\
&\quad ~\Leftrightarrow~ k = k^- = d'_Q(\ell-j)-\ell+i+j \;\wedge\; d_P(k^-) \leq d'_Q(\ell-j)+i+j. \end{aligned}$$
We use the proposition to create tables for each of the three equivalences in Theorem \[T:ApA\]. The tables $N$ and $K$ are used to compute the size of a delta set or to construct a delta set, respectively (Example \[E:herm\]). The tables $N^+$ and $N^-$ are used in the optimization of the order bound (Example \[E:suz1\]). The tables $K^+$ and $K^-$ provide more information that can be used for further improvements with the ABZ bound (Example \[E:suz2\]). In all cases, let $A=B_0+kP+\ell Q$, and let $C=C_0+iP+jQ+Q$. $$A_P \leq A = (A-C)_P + C ~\Leftrightarrow~ A \in \Delta_P(C+Q) \;\wedge\; A \not \in \Delta_P(C).$$ $$K_\ell(i,j) = \tau'(\ell-j)+i , \qquad N_\ell(i,j) = \begin{cases} 1 &~~\text{if $\sigma(k) \leq \ell$} \\ 0 &~~\text{if $\sigma(k) > \ell$} \end{cases}$$ The table $N_\ell(i,j)$ indicates whether there exists $A \in \Delta_P(B_0+\ell Q,C_0+iP+jQ+Q) \backslash \Delta_P(B_0+\ell Q,C_0+iP+jQ)$. In the affirmative case (N=1), the table $K_\ell(i,j)$ gives the unique value of $k$ such that $A=B_0+kP+\ell Q.$ The table $N$ is sufficient for computing the size of a delta set, the table $K$ moreover provides the elements of a delta set. $$A_P = A \leq (A-C)_P + C ~\Leftrightarrow~ A \in \Delta_P(C+Q) \;\wedge\; A-Q \not \in \Delta_P(C+Q).$$ $$K^+_i(j,\ell) = \tau(\ell) , \qquad N^+_i(j,\ell) = \begin{cases} 1 &~~\text{if $\sigma'(k-i) \geq \ell-j$} \\ 0 &~~\text{if $\sigma'(k-i) < \ell-j$} \end{cases}$$ The table $N^+_i(j,\ell)$ indicates whether there exists $A \in \Delta_P(B_0+\ell Q,C_0+iP+jQ+Q)$ with $A-Q \not \in \Delta_P(B_0+\ell Q-Q,C_0+iP+jQ)$. In the affirmative case (N=1), the table $K^+_i(j,\ell)$ gives the unique value of $k$ such that $A=B_0+kP+\ell Q.$ The table $N^+$ is sufficient for comparing the sizes of two delta sets, the table $K^+$ moreover provides the elements for the difference in one direction. $$A_P \leq A = (A-C)_P + C ~\Leftrightarrow~ A \in \Delta_P(C+Q) \;\wedge\; A+Q \not \in \Delta_P(C+Q).$$ $$K^-_i(j,\ell) = \tau'(\ell-j)+i , \qquad N^-_i(j,\ell) = \begin{cases} 1 &~~\text{if $\sigma(k) \leq \ell$} \\ 0 &~~\text{if $\sigma(k) > \ell$} \end{cases}$$ The table $N^-_i(j,\ell)$ indicates whether there exists $A \in \Delta_P(B_0+\ell Q,C_0+iP+jQ+Q)$ with $A+Q \neq \Delta_P(B_0+\ell Q+Q,C_0+iP+jQ+Q)$. In the affirmative case (N=1), the table $K^-_i(j,\ell)$ gives the unique value of $k$ such that $A=B_0+kP+\ell Q.$ The table $N$ is sufficient for comparing the sizes of two delta sets, the table $K$ moreover provides the elements in the difference.
Hermitian curves {#S:Hermitian}
================
Let $X$ be the Hermitian curve over $\mathbb{F}_{q^2}$ defined by the equation $y^q+y=x^{q+1}$. The curve has $q^3+1$ rational points and genus $g = q(q-1)/2$. Let $P$ and $Q$ be two distinct rational points. We will give a description of the set $$D_0(P,Q) = D(P,Q) \cap \{ iP+jQ : i,j \in {{\mathbb Z}}\}.$$ We use this description to determine lower bounds for $\gamma_P(C),$ for $C \in \{ iP+jQ : i,j \in {{\mathbb Z}}\}.$ The only property of the two rational points that we use is that lines intersect the pair $(P,Q)$ with one of the multiplicities $$\begin{array}{rrr} (0,0) &(0,1) &(0,q+1) \\ (1,0) &\underline{(1,1)} \\ (q+1,0) \end{array}$$ The curve is a smooth plane curve and if $H$ is the intersection divisor of a line then $K=(q-2)H$ represents the canonical class. We have $H \sim (q+1)P \sim (q+1)Q$ and $m(P-Q)$ is principal for $m=q+1.$
\[P:discherm\] The $m$ inequivalent divisor classes in $D_0(P,Q)$ are represented by the divisors $$dH - d P - d Q, \quad \text{for $d=0,1,\ldots,q.$}$$
Since $m=q+1$ is minimal such that $mP \sim mQ$, the divisors are inequivalent. As multiples of $H-P-Q \in \Gamma_P$, each of the divisors $dH - d P - d Q \in \Gamma_P,$ for $d=0,1,\ldots,q.$ A divisor $A \in \Gamma_P$ is a discrepancy if and only if $K+P+Q-A \in \Gamma_P.$ Now use $K+P+Q=(q-2)H+P+Q=q(H-P-Q).$
The function $y$ has divisor $y=(q+1)(P_0-P_\infty)$, where $P_0 = (0,0)$ and $P_\infty=(0:1:0)$. Moreover $L(H-P_0-P_\infty) = \langle x \rangle.$
The ring $O$ of functions that are regular outside $P_0$ and $P_\infty$ has a basis $\langle\; x^i y^j \,|\, 0 \leq i \leq q, j \in {{\mathbb Z}}\;\rangle$ as vector space over $\mathbb{F}_{q^2}$.
Theorem \[T:basis\].
\[L:hermP\] Let $C = dH-aP-bQ,$ for $0 \leq a \leq q, 0 \leq b \leq q+1.$ $$C \in \Gamma_P ~\Leftrightarrow~ d > a ~\text{or}~ d = a \geq b.$$
We use Lemma \[L:singleton\] together with Proposition \[P:discherm\]. We may assume $H=(q+1)Q.$ Then, $dH-aP-bQ \in \Gamma_P$ if and only if $dH-aP-bQ \geq aH-aP-aQ$ if and only if $d>a$ or $d = a \geq b$.
\[P:short\] Let $C = dH-aP-bQ,$ for $0 \leq a, b \leq q.$ The set $\Delta_P(-C) = \{ A \in \Gamma_P : A+C \not \in \Gamma_P \}$ contains the following elements $$\begin{aligned}
&(q-1-d-r)H-(q-a)P, &\text{for $d \leq d+r < a.$} \\
&sH, &\text{for $d \leq d+s < a.$} \\
&sH-(d+s-a)P &\text{for $d \leq a \leq d+s < b.$} \end{aligned}$$
With the lemma, $A=(q-1-d-r)H-(q-a)P \in \Gamma_P$ for $q-a \leq q-1-d-r,$ and $A+C = (q-1-r)H-qP-bQ \not \in \Gamma_P$ for $r \geq 0.$ Clearly, $A=sH \in \Gamma_P$ for $s \geq 0$, and $A+C = (d+s)H-aP-bQ \not \in \Gamma_P$ for $a > d+s.$ Finally, $A=sH-(d+s-a)P \in \Gamma_P$ for $0 \leq d+s-a \leq s$, and $A+C = (d+s)H-(d+s)P-bQ \not \in \Gamma_P$ for $b > d+s.$
Let $0 \leq a, b \leq q.$ There exists a form of degree $d$ that intersects the curve in $(P,Q)$ with precise multiplicities $(a,b)$ if and only if $0 \leq a, b \leq d.$
Such a curve exists if and only if $dH-aP-bQ \in \Gamma_P \cap \Gamma_Q.$ With Lemma \[L:hermP\], the latter holds if and only $0 \leq a, b, \leq d.$
\[L:case1\] For $d \geq 0$, let $C$ be a divisor with $dH-dP-dQ \leq C \leq dH$. Then $C$ has no base points and $\gamma_P(C) = \deg C.$
Since $C$ is equivalent to an effective divisor with support in $P$ and $Q$, those two points are the only candidates for the base points. With Lemma \[L:hermP\], $C \in \Gamma_P \cap \Gamma_Q$, and therefore neither $P$ nor $Q$ is a base point. The last claim uses Proposition \[P:gammap\].
Let $$\begin{aligned}
&C = dH -aP -bQ, \qquad \text{for $d \in {{\mathbb Z}}, 0 \leq a,b \leq q,$} \\
&A = jH+i(H-P), \qquad \text{for $j \in {{\mathbb Z}}, 0 \leq i \leq q.$}\end{aligned}$$ Then $A \in \Delta_P(C)$ if and only if $$\begin{cases}
0 \leq j \leq (d-a+q-1), &\text{if $0 \leq i < a-b$} \\
0 \leq j \leq (d-a+q-2), &\text{if $a-b \leq i < a$} \\
0 \leq j \leq (d-a-1), &\text{if $a \leq i < a-b+q+1$} \\
0 \leq j \leq (d-a-2), &\text{if $a-b+q+1 \leq i \leq q$}
\end{cases}$$
For $A-C$ we write $$\begin{cases} &(j+i-d+1)H-(i-a)P-(q+1-b)Q, ~~\text{if $i-a \geq 0.$} \\
&(j+i-d+2)H-(i-a+q+1)P-(q+1-b)Q, ~~\text{if $i-a < 0.$}
\end{cases}$$ With Lemma \[L:hermP\], $A-C \not \in \Gamma_P$ if and only if $$\begin{cases}
&j < d-a-1, ~~\text{or}~~ j=d-a-1, i-a < q+1-b, ~~\text{if $i-a \geq 0.$} \\
&j < d-a-1+q ~~\text{or}~~ j=d-a-1+q, i-a < -b, ~~\text{if $i-a < 0.$}
\end{cases}$$ In combination with $A \in \Gamma_P$ if and only if $j \geq 0$, this proves the claim.
\[C:delt\] Let $C=dH-aP-bQ,$ for $0 \leq a, b \leq q.$\
For $a-d < 0$, $$\Delta_P(-C) = \emptyset, \qquad \# \Delta_P(0,C) = \deg C.$$ For $0 \leq a-d \leq q-1$, $$\begin{aligned}
&\# \Delta_P(0,C) = a(q-1-a+d)+\max\{0,a-b\}. \\
&\# \Delta_P(0,-C) = (q+1-a)(a-d)+\max\{0,b-a\}. \end{aligned}$$ For $a-d > q-1$, $$\Delta_P(C) = \emptyset, \qquad \# \Delta_P(0,-C) = - \deg C.$$
\[mult2\] Let $C = dH -aP -bQ,$ for $d \in {{\mathbb Z}},$ and for $0 \leq a,b \leq q.$ Then $$\begin{array}{lcl}
(\text{\rm{Case 1 : }} a, b \leq d) & &\gamma_{P}(C) \;=\; \gamma_Q(C) \;=\; \deg C. \\
(\text{\rm{Case 2a : }} b \leq d \leq a) & &\gamma_{P}(C) \;\geq\; \deg C + a-d. \\
(\text{\rm{Case 2b : }} a \leq d \leq b) & &\gamma_{Q}(C) \;\geq\; \deg C + b-d. \\
(\text{\rm{Case 3a : }} d \leq a \leq b, a<q) & &\gamma_{P}(C) \;\geq\; \deg C + a-d+b-d. \\
(\text{\rm{Case 3b : }} d \leq b \leq a, b<q) & &\gamma_{Q}(C) \;\geq\; \deg C + a-d+b-d. \\
(\text{\rm{Case 4 : }} d \leq a=b=q) & &\gamma_{P}(C) \;=\; \gamma_Q(C) \;\geq\; \deg C + q - d.
\end{array}$$
(Case 1) uses Lemma \[L:case1\]. The lower bounds follow from Proposition \[P:short\] by using $\gamma_{P}(C) = \deg C + \gamma_{P}(-C)$. Or we can obtain the lower bounds from Corollary \[C:delt\] in combination with $\gamma_P(C) \geq \# \Delta_P(0,C)$ (Corollary \[C:cbfrac\]). For $0 \leq a-d \leq q-1$, $$\# \Delta_{P_\infty}(0,C) = a(q-1-a+d)+\max\{0,a-b\}.$$ (Case 2a: $b \leq d \leq a$) $~a(q-1-a+d)+a-b - d(q-1)+b-d = (a-d)(q-a) \geq 0.$\
(Case 3a : $d \leq a \leq b$) $~a(q-1-a+d)+0 - d(q-1) = (a-d)(q-1-a) \geq 0.$\
(Case 4 : $d \leq a=b=q$) $~q(d-1) = d(q+1)-q-q+(q-d).$
\[thm:dist\] For $G=K+C$, and for $D \cap S=\emptyset$, the algebraic geometric code $C_\Omega(D,G)$ has minimum distance $d \geq \gamma(C;S)$. Let $C = dH -aP -bQ,$ for $d \in {{\mathbb Z}},$ and for $0 \leq a,b \leq q.$ Then $$\begin{array}{lcl}
(\text{\rm{Case 1 : }} a, b \leq d) & &\gamma(C) \;\geq\; \deg C. \\
(\text{\rm{Case 2a : }} b \leq d \leq a) & &\gamma(C;P) \;\geq\; \deg C + a-d. \\
(\text{\rm{Case 2b : }} a \leq d \leq b) & &\gamma(C;Q) \;\geq\; \deg C + b-d. \\
(\text{\rm{Case 3a : }} d \leq a \leq b, a<q) & &\gamma(C;P,Q) \;\geq\; \deg C + a-d+b-d. \\
(\text{\rm{Case 3b : }} d \leq b \leq a, b<q) & &\gamma(C;P,Q) \;\geq\; \deg C + a-d+b-d. \\
(\text{\rm{Case 4 : }} d \leq a=b=q) & &\gamma(C;P) = \gamma(C;Q) \;\geq\; \deg C + q - d.
\end{array}$$
Use the order bound with\
(Case 2a: $b \leq d \leq a$) $Q_0 = \ldots = Q_{a-d-1} = P.$\
(Case 3a : $d \leq a \leq b$) $Q_0 = \ldots = Q_{a-d-1} = P,$ $Q_{a-d} = \ldots = Q_{a-d+b-d-1} = Q.$\
(Case 4 : $d \leq a=b=q$) $Q_0 = \ldots = Q_{q-d-1} = P.$
The following example illustrates the use of the tables $K_\ell(i,j)$ and $N_\ell(i,j)$ for constructing a delta set $\Delta_P(\ell Q,iP+jQ+Q)$ (Proposition \[P:tables\]). In this case, the functions $d_P=d_Q=d'_P=d'_Q$ all agree and we can omit the index.
\[E:herm\] For the Hermitian curve of degree four, the genus $g=3.$ The discrepancies $D_0(P,Q)$ are represented by the divisors $0, H-P-Q, 2H-2P-2Q, 3H-3P-3Q.$ In particular, $d(k) = 0, 2, 4, 6,$ for $k = 0, -1, -2, -3$ modulo $4$, respectively. $$\begin{array}{lrrrrrrrrrrrrrr}
(\ell=0,i=1) &j &= -7 &-6 &-5 &-4 &-3 &-2 &-1 &0 &1 &2 &3 &4 \\[1ex] \hline
\multicolumn{2}{r}{d(\ell-j)} &2 &4 &6 &0 &2 &4 &6 &0 &2 &4 &6 &0 \\
(K) &k &(-4) &(-1) &(2) &(-3) &0 &3 &6 &(1) &4 &7 &10 &(5) \\[1ex]
&d(k) &0 &2 &4 &6 &0 &2 &4 &6 &0 &2 &4 &6 \\
\multicolumn{2}{r}{d(k)-d(\ell-j)} &-2 &-2 &-2 &6 &-2 &-2 &-2 &6 &-2 &-2 &-2 &6 \\
(N) &\leq i+j &0 &0 &0 &0 &1 &1 &1 &0 &1 &1 &1 &0 \\[1ex] \hline
\end{array} \\[1ex]$$ The value for $k = d(\ell-j)-\ell+i+j.$ Row (K) gives the difference $\Delta_P(\ell Q,iP+jQ+Q) \backslash \Delta_P(\ell Q, iP+jQ) = \{ kP+\ell Q \} \cap \Gamma_P$, with empty intersection if and only if $k$ appears in parentheses. Row (N) has the decision whether the difference is empty (N=0) or nonempty (N=1). As a special case, we see that $\Delta_P(0,P+2Q) = \{ 0, 3P, 6P, 4P \}$. The numbers in parentheses illustrate the duality in Proposition \[P:6g\]. $$\begin{aligned}
&\Delta_P(0,P+5Q) = 0 + \{ 0, 3P, 6P, 4P, 7P, 10P \}, \\
&\Delta_P(-P+7Q,-P+7Q) = (-P+7Q) + \{ -4P, -P, 2P, -3P, P, 5P \}. \end{aligned}$$ With $-P+7Q \sim 7P-Q,$ $$\Delta_P(-Q,-P+7Q) = \{ 3P-Q, 6P-Q, 9P-Q, 4P-Q, 8P-Q, 9P-Q \}.$$ The partition of the interval $\{ -2g, \ldots, 4g-1 \}$ in Proposition \[P:6g\] is given by $$\begin{array}{lll}
& &N_3 = \{ 8, 9, 11 \} \\
&N_2 = \{ 0, 3, 4 \} &G_3 = \{ 6, 7, 10 \} \\
N_1 = \{ -4, -3, -1 \} &G_2 = \{ 1, 2, 5 \} & \\
G_1 = \{ -6, -5, -2 \} & & \\
\end{array}$$
Suzuki curves {#S:Suzuki}
=============
The Suzuki curve over the field of $q = 2q_0^2$ elements is defined by the equation $y^q+y=x^q_0(x^q+x).$ The curve has $q^2+1$ rational points and genus $g = q_0(q-1)$. The semigroup of Weierstrass nongaps at a rational point is generated by $\{q, q+q_0, q+2q_0, q+2q_0+1\}$. For any two rational points $P$ and $Q$ there exists a function with divisor $(q+2q_0+1)(P-Q).$ Let $m = q+2q_0+1=(q_0+1)^2+{q_0}^2$, and let $H$ be the divisor class containing $mP \sim mQ$. The divisor $H$ is very ample and gives an embedding of the Suzuki curve in ${\mathbb P}^4$ as a smooth curve of degree $m$. The canonical divisor $K \sim 2(q_0-1)H.$ A hyperplane $H$ intersects $(P,Q)$ with one of the following multiplicities. $$\begin{array}{rrrrr}
(0,0) &(0,1) &(0,q_0+1) &(0,2q_0+1) &(0,q+2q_0+1) \\
(1,0) &(1,1) &(1,q_0+1) &\underline{(1,2q_0+1)} \\
(q_0+1,0) &(q_0+1,1) &\underline{(q_0+1,q_0+1)} \\
(2q_0+1,0) &\underline{(2q_0+1,1)} \\
(q+2q_0+1,0)
\end{array}$$ Let $$D_0 = H-(2q_0+1)P-Q, ~
D_1 = H-(q_0+1)(P+Q), ~
D_2 = H-P-(2q_0+1)Q.$$ Then $L(D_i) \neq 0,$ and $L(D_i-P) = \dim L(D_i-Q) = 0,$ for $i=0,1,2.$ And $D_i \in D(P,Q)$, for $i=0,1,2.$
\[L:q0\] For any nonnegative integer $q_0$, $$\{ -q_0(q_0+1), \ldots, +q_0(q_0+1) \} = \{ a(q_0+1)+bq_0 : |a| + |b| \leq q_0 \}.$$
\[T:descsuz\] The $m$ inequivalent divisor classes in $D_0(P,Q)$ are represented by $$\begin{aligned}
&i D_0 +j D_2, \quad \text{for $0 \leq i, j \leq q_0,$ and} \\
&D_1 + i' D_0 +j' D_2, \quad \text{for $0 \leq i', j' \leq q_0-1.$}\end{aligned}$$ The given representatives correspond one-to-one to the $m$ divisors $$D(a,b) = (a+q_0)H\,-\,((a+q_0)(q_0+1)+bq_0)P\,-\,((a+q_0)(q_0+1)-bq_0)Q,$$ for $|a|+|b| \leq q_0.$
$$\begin{aligned}
&i D_0 + j D_2 \\
=~&i(H-(2q_0+1)P-Q)\,+\,j(H-P-(2q_0+1)Q), \\
=~&(i+j)H-(i+j)(q_0+1)(P+Q)-(i-j)q_0(P-Q). \end{aligned}$$
Moreover, $0 \leq i,j \leq q_0$ if and only if $|i+j-q_0|+|i-j| \leq q_0.$ Thus $$\{ i D_0 + j D_2 : 0 \leq i, j \leq q_0 \} = \{ D(a,b) : |a|+|b| \leq q_0,~ a - b \equiv 0 \pmod{2} \}$$ $$\begin{aligned}
&H-(q_0+1)(P+Q) + i' D_0 + j' D_2 \\
=~&(i'+j'+1)H-(i'+j'+1)(q_0+1)(P+Q)-(i'-j')q_0(P-Q). \end{aligned}$$ Similarly, $0 \leq i',j' \leq q_0-1$ if and only if $|i'+j'+1-q_0|+|i'-j'| \leq q_0-1.$ And $$\begin{aligned}
&\{ H-(q_0+1)(P+Q) + i' D_0 + j' D_2 : 0 \leq i', j' \leq q_0-1 \} \\
=~&\{ D(a,b) : |a|+|b| \leq q_0,~ a - b \equiv 1 \pmod{2} \}\end{aligned}$$ We have constructed $m$ inequivalent divisors in $\Gamma_P$. A divisor $A \in \Gamma_P$ is a discrepancy if and only if $K+P+Q-A \in \Gamma_P.$ With $K=2(q_0-1)H$, we see that $$\begin{aligned}
D(a,b)+D(-a,-b) &= (2q_0)H-(2q_0(q_0+1)P -2q_0(q_0+1)Q \\
&= (2q_0-2)H+P+Q = K+P+Q.\end{aligned}$$
As an illustration, we give the discrepancies for the Suzuki curve $y^8+y=x^{10}+x^3$ over the field of eight elements ($q_0= 2, q=8, g=14, N=65, m=13=3^2+2^2$). $$\begin {array}{ccccc}
0&\cdot&H-5P-Q&\cdot&2H-10P-2Q\\
\cdot&H-3P-3Q&\cdot&2H-8P-4Q&\cdot\\
H-P-5Q&\cdot&2H-6P-6Q&\cdot&3H-11P-7Q\\
\cdot&2H-4P-8Q&\cdot&3H-9P-9Q&\cdot\\
2H-2P-10Q&\cdot&3H-7P-11Q&\cdot&4H-12P-12Q
\end {array}$$ With $H \sim 13Q$, we obtain the following multiplicities for the discrepencies at $(P,Q)$. $$\begin {array}{ccccc}
(0,0)&\cdot&(-5,12)&\cdot&(-10,24)\\
\cdot&(-3,10)&\cdot&(-8,22)&\cdot\\
(-1,8)&\cdot&(-6,20)&\cdot&(-11,32)\\
\cdot&(-4,18)&\cdot&(-9,30)&\cdot\\
(-2,16)&\cdot&(-7,28)&\cdot&(-12,40)
\end {array}$$ For the given Suzuki curve, Beelen [@Bee07FF] gives an example of a two-point code for which the floor bound exceeds the order bound. The example generalizes to any Suzuki curve. For both the Suzuki curve over ${{\mathbb F}}_8$ and over ${{\mathbb F}}_{32}$ (for which $q_0=4, q=32, g=124, N=1025, m=41=5^2+4^2$), the example is the only two-point code for which the floor bound exceeds the order bound.
Let $A = B = K-H, Z=2P+2Q.$ With $\dim L(H) - \dim L(H-2P-2Q) = 4$, it follows that $L(A+Z)=L(A)$ and $L(B+Z)=L(B).$ For the code $C_\Omega(D,G) = C_L(D,G)^\perp$ with $G=K+C=A+B+Z=2K-2H+2P+2Q$, the threshold divisor $C=K-2H+Z$. The floor bound gives minimum distance $d \geq \deg C + \deg Z = d^\ast + 4.$ This is one better than the order bound.
We give an example of the ABZ bound for codes that improves both the floor bound and the order bound.
\[E:abz\] Let $A = B = K-H, Z=(q_0+2)P+2Q$. For the code $C_\Omega(D,G) = C_L(D,G)^\perp$ with $G=K+C=A+B+Z=2K-2H+(q_0+2)P+2Q$, the threshold divisor $C=K-2H+Z$. For the ABZ bound we use $\dim L(A) - \dim L(A-C) + \dim L(B) - \dim L(B-C) =
2 (\dim L(K-H) - \dim L(H-Z)) = 2 \deg (K-H) - \deg K + 2 \dim L(H).$ The bound $d \geq 10$ for $q_0=2$ is one better than both the floor bound and the order bound.
We illustrate the use of tables $K^\pm_i(j,\ell)$ and $N^\pm_i(j,\ell)$ for the comparison of delta sets $\Delta_P(\ell Q,iP+jQ+Q)$ and $\Delta_P(\ell Q \mp Q,iP+jQ+Q)$ (Proposition \[P:tables\]). The functions $d_P=d_Q=d'_P=d'_Q$ all agree and we can omit the index. The functions are defined on residue classes modulo $m$. With Lemma \[L:q0\] and Theorem \[T:descsuz\], for $|a|+|b| \leq q_0$, $$d(k) = (q_0-a)(q-1), \quad \text{for $k=a(q_0+1)+bq_0-q_0(q_0+1) \pmod{m}$}$$
\[E:suz1\] For the Suzuki curve over ${{\mathbb F}}_{32}$, let $C=55P+31Q.$ In this case there is a unique choice $B=-5Q$ such that $\Delta_P(B,C) \geq 90.$ The improvement over the choice $B=0$ can be seen as follows. $$\begin{aligned}
&\begin{array}{rrrrrrrrrrrrr}
(i=55,j=30) &\ell &=0 &-1 &-2 &-3 &-4 &-5 \\[1ex] \hline
&d(\ell) &0 &31 &62 &93 &124 &\cdot \\
(K^+) &k^+ &0 &32 &64 &(96) &(128) &\cdot \\[1ex]
&d(k^+-i) &62 &93 &124 &0 &31 &\cdot \\
&d(\ell)-d(k^+-i) &-62 &-62 &-62 &93 &93 &\cdot \\
(N^+) &\leq i+j &1 &1 &1 &0 &0 &\cdot \\[1ex] \hline
\end{array} \end{aligned}$$ $$\begin{aligned}
&\begin{array}{rrrrrrrrrrrrr}
(i=55,j=30) &\ell &= 0 &-1 &-2 &-3 &-4 &-5 \\[1ex] \hline
&d(\ell-j) &\cdot &217 &124 &155 &186 &217 \\
(K^-) &k^- &\cdot &303 &(211) &243 &275 &307 \\[1ex]
&d(k^-) &\cdot &155 &217 &93 &124 &155 \\
&d(k^-)-d(\ell-j) &\cdot &-62 &93 &-62 &-62 &-62 \\
(N^-) &\leq i+j &\cdot &1 &0 &1 &1 &1 \\[1ex] \hline
\end{array}\end{aligned}$$ The tables use $k^+ = d(\ell) - \ell$ and $k^- = d(\ell-j)-\ell+i+j.$\
From the tables we obtain $$\begin{aligned}
I_P(0,C) \backslash I_P(-5Q,C) &= \{ k^+ \in I_P(-5Q,5Q) : k^+ P \in \Gamma_P \} \\
&= \{ 0, 32, 64, (96), (128) \}. \\
I_P(-5Q,C) \backslash I_P(0,C) &= \{ k^- \in I_P(-C,5Q) : -5Q+k^- P \not \in \Gamma_P \} \\
&= \{ 307, 275, 243, (211), 303 \}.\end{aligned}$$ The net gain is therefore $4-3=1.$ To reach this conclusion it is sufficient to consult the rows ($N^+$) and ($N^-$). $$\begin{aligned}
\Delta_P(55P+31Q) \supseteq &\{ A_1 = 36P-5Q, \ldots, A_{45} = 163P-5Q \} \\
& \cup \{ A_{46} = 180P-5Q, \ldots, A_{90} = 307P-5Q \}\end{aligned}$$
\[E:suz2\] We illustrate the improvemnt of the ABZ bound for cosets over the order bound. Both $\# \Delta_P(0,9P+9Q) = \# \Delta_P(9Q,9P+9Q) = 40.$ This is the optimum for the order bound. For $r \geq 0$, $$\# \Delta_P(0,C) = \# \Delta_P(\leq rP,C) + \# \Delta_P(\geq rP+P,C).$$ For $r, s \geq 0$ such that $$\# \Delta_P(\geq rP+P+sQ,C) > \# \Delta_P(\geq rP+P,C)$$ we obtain an improvement using the ABZ bound with choices $B=rP, Z=sQ$ (Theorem \[T:cbabz\]). As in the previous example we compare delta sets and find $$\begin{aligned}
&I_P(0,9P+9Q) \backslash I_P(9Q,9P+9Q) = \{ k^- \in I_P(-9P,9Q) : k^- P \in \Gamma_P \} \\
= &\{ 141, 109, 77, (45), 137, 105, 73, 41, (9) \}. \\[1ex]
&I_P(9Q,9P+9Q) \backslash I_P(0,9P+9Q) = \{ k^+ \in I_P(9Q,9Q) : 9Q+k^+ P \not \in \Gamma_P \} \\
= &\{ 115, 147, 179, (211), 119, 151, 183, 215, (247) \}.\end{aligned}$$ The information shows that although the delta sets $\Delta_P(0,9P+9Q)$ and $\Delta_P(9Q,9P+9Q)$ have the same size, the first contains more divisors of small degree and the latter more divisors of high degree. For $Z=9Q$ and for $141 \leq r \leq 146$ (or $109 \leq r \leq 114$) we see that $$\# \Delta_P(\geq rP+P+sQ,C) - \# \Delta_P(\geq rP+P,C) = 5.$$ The order bound gives minimum distance $d \geq 40$ for the AG code with $C=9P+9Q$ and $d \geq 50$ for the AG code with $C=10P+9Q.$ Thus we improve the minimum distance for $C=9P+9Q$ from $d \geq 40$ to $d \geq 45.$ $$\begin{aligned}
\Delta_P(9P+9Q) \supseteq &\{ A_1 = 0, \ldots, A_{18} = 109P \} \\
& \cup \{ A_{19} = 112P+9Q, \ldots, A_{45} = 256P+9Q \}\end{aligned}$$
The ABZ bound, while more general than the order bound, is still only a special case of the main theorem. The following choice of divisors in $\Delta_P(12P+12Q)$ gives $\gamma_P(12P+12Q) \geq 56.$ This improves both the order bound and the ABZ bound (for all possible choices of $A, B,$ and $Z$ as integer combinations of $P$ and $Q$). $$\begin{aligned}
\Delta_P(12P+12Q) \supseteq &\{ A_1 = 0, \ldots, A_{24} =116P \} \\
& \cup \{ A_{25} = 118P+6Q, \ldots, A_{32} = 141P+6Q \} \\
& \cup \{ A_{33} = 143P+12Q, \ldots, A_{56} = 259P+12Q \} \end{aligned}$$
Coset decoding {#S:CosetDecoding}
==============
For a given vector $y \in {{\mathbb F}}^n$, and for an extension of linear codes ${{\mathcal C}}' \subset {{\mathcal C}}\subset {{\mathbb F}}^n$, coset decoding determines the cosets of ${{\mathcal C}}'$ in ${{\mathcal C}}$ that are nearest to the vector $y.$ If $y$ is at distance $d(y,{{\mathcal C}}) \leq t$ from ${{\mathcal C}}$ and the minimum distance $d({{\mathcal C}}/{{\mathcal C}}')$ between distinct cosets is at least $w > 2t$ then there exists a unique nearest coset $c + {{\mathcal C}}'$ with $d(y,c+{{\mathcal C}}') \leq t$. We describe a coset decoding procedure that returns the unique coset when the estimate $d({{\mathcal C}}/{{\mathcal C}}') \geq w$ is obtained with Theorem \[T:cosetbound\]. The procedure follows the majority coset decoding procedure in [@Duu93Thesis], [@Duu93].\
Shift bound or Coset bound (Theorem \[T:cosetbound\]): Let ${{\mathcal C}}/{{\mathcal C}}_1$ be an extension of ${{\mathbb F}}$-linear codes with corresponding extension of dual codes ${{\mathcal D}}_1/{{\mathcal D}}$ such that $\dim {{\mathcal C}}/{{\mathcal C}}_1 = \dim {{\mathcal D}}_1/{{\mathcal D}}= 1$. If there exist vectors $a_1, \ldots, a_w$ and $b_1, \ldots, b_w$ such that $$\begin{cases}
a_i \ast b_j \in {{\mathcal D}}&\text{for $i+j \leq w$}, \\
a_i \ast b_j \in {{\mathcal D}}_1 \backslash {{\mathcal D}}&\text{for $i+j = w+1$},
\end{cases}$$ then $d({{\mathcal C}}/{{\mathcal C}}_1) \geq w.$\
For a given $x \in {{\mathcal D}}_1 \backslash {{\mathcal D}}$, we may assume, after rescaling if necessary, that $a_i \ast b_{w+1-i} \in x+{{\mathcal D}}$, for $i=1,\ldots,w.$ Define the following cosets of $a_i$ and $b_{w+1-i}$, for $i=1,\ldots,w,$ $$\begin{aligned}
A_i &= a_i + \langle a_1, \ldots, a_{i-1} \rangle, \\
B_{w+1-i} &= b_{w+1-i} + \langle b_1, \ldots, b_{w-i} \rangle.\end{aligned}$$ For $c \in {{\mathcal C}}$, the coset $c+{{\mathcal C}}_1$ is uniquely determined by $x \cdot c$. For a given $y \in {{\mathbb F}}^n$ such that $d(y,{{\mathcal C}}) \leq t$, the decoding procedure will look for a pair $a' \in A_i, b' \in B_{w+1-i}$ such that, for all $c \in {{\mathcal C}}$ with $d(y,c) \leq t$, $(a' \ast b') \cdot y = x \cdot c.$ The vector $a' \ast b'$ is defined as the Hadamard or coordinate-wise product of the vectors $a'$ and $b'$. We use $(a' \ast b') \cdot y = (a' \ast y) \cdot b'$.
(Decoding up to half the coset bound) \[T:decoding\] Let $2t < w \leq d({{\mathcal C}}/{{\mathcal C}}_1)$, for ${{\mathcal C}}/{{\mathcal C}}_1$ and $w$ as in Theorem \[T:cosetbound\]. For $y \in {{\mathbb F}}^n$ such that $d(y,{{\mathcal C}}) \leq t$, let $$\begin{aligned}
&I = \{ 1 \leq i \leq w : \; (\exists a'_i \in A_i) \; (a'_i \ast b_j) \perp y, 1 \leq j \leq w-i \;\}, \\
&I^\ast = \{ 1 \leq j \leq w : \; (\exists b'_j \in B_{w+1-j}) \; (a_i \ast b'_j) \perp y, 1 \leq i \leq j-1 \;\}. \end{aligned}$$ For every $c \in {{\mathcal C}}$ with $d(y,c) \leq t$, $x \cdot c = (a'_i \ast b'_i) \cdot y$, for a majority of $i \in I \cap I^\ast.$
For $c \in {{\mathcal C}}$, let $a'_i \in A_i$ be such that $a'_i \ast y = a'_i \ast c.$ The vector $a'_i$, if it exists, satisfies $(a'_i \ast b_j) \cdot y = 0$, for $j = \{1,\ldots,w-i \}.$ Moreover, for any $b' \in B_{w+1-i}$, $(a'_i \ast b') \cdot y =
(a'_i \ast b') \cdot c = (a_i \ast b_{w+1-i}) \cdot c = x \cdot c.$ Let $$\begin{array}{lcl}
\Gamma = \{ 1 \leq i \leq w : \; (\exists a'_i \in A_i) \; a'_i \ast y = a'_j \ast c \; \}, &
&\Delta = \{ 1 \leq i \leq w \} \backslash \Gamma, \\
\Gamma^\ast = \{ 1 \leq j \leq w : \; (\exists b'_j \in B_{w+1-j}) \; b'_j \ast y = b'_j \ast c \}, &
&\Delta^\ast = \{ 1 \leq j \leq w \} \backslash \Gamma^\ast.
\end{array}$$ We know a priori only the sets $I$ and $I^\ast$. Clearly, $\Gamma \subset I$ and $\Gamma^\ast \subset I^\ast.$ Moreover, for $c \in {{\mathcal C}}$ with $d(y,c) \leq t$, $|\Delta|, |\Delta^\ast| \leq t.$ For $i \in I \cap I^\ast$, $(a'_i \ast b'_i) \cdot y = x \cdot c$ if either $i \in \Gamma$ or $i \in \Gamma^\ast$. Regardless of the actual sets $I$ and $I^\ast$, this is certainly the case if $i \in \Gamma \cap \Gamma^\ast$ and it fails only when $i \in \Delta \cap \Delta^\ast$. Now $$|\Gamma \cap \Gamma^\ast| - |\Delta \cap \Delta^\ast| = w - |\Gamma \cap \Delta^\ast| - |\Gamma^\ast \cap \Delta|
\geq w - 2t > 0.$$ Thus, the majority of $i \in I \cap I^\ast$ will give a value $(a'_i \ast b'_i) \cdot y = x \cdot c.$
If $\dim {{\mathcal C}}/ {{\mathcal C}}' > 1$ then the procedure can be applied iteratively to a sequence of extensions ${{\mathcal C}}' = {{\mathcal C}}_r \subset {{\mathcal C}}_{r-1} \subset \cdots {{\mathcal C}}_1 \subset {{\mathcal C}}_0 = {{\mathcal C}}$ such that $\dim {{\mathcal C}}_{i} / {{\mathcal C}}_{i-1} = 1$, for $i=1,\ldots,r.$ For given $y_0 \in {{\mathbb F}}^n$ with $d(y_0,{{\mathcal C}}_0) \leq t$, the procedure returns the unique coset $c_0+{{\mathcal C}}_1$ such that $d(y_0,c_0+{{\mathcal C}}_1) \leq t$. At the next iteration, for $y_1 = y_0-c_0 \in {{\mathbb F}}^n$ with $d(y_1,{{\mathcal C}}_1) \leq t$, the procedure returns the unique coset $c_1+{{\mathcal C}}_2$ such that $d(y_1,c_1+{{\mathcal C}}_2) \leq t$, and so on.\
Let ${\cal A} = \{ A_1 \leq A_2 \leq \cdots \leq A_w \} \subset \Delta_P(C)$ be a sequence of divisors with $A_{i+1} \geq A_i + P$, for $i=1,\ldots,w-1.$ Theorem \[T:cbdiv\] (Main theorem) together with Lemma \[L:cosetgammas\] shows that $d(C_\Omega(D,G-P) / C_\Omega(D,G)) \geq w$, for $G$ such that $C=G-K-P,$ and for $D \cap (A_w-A_1) = \emptyset.$ We show how the coset decoding procedure applies to the given extension. For a divisor $A_i \in \Delta_P(C)$, also $K+C+P-A_i=G-A_i \in \Delta_P(C).$ Thus, there exist functions $f_i \in L(A_i) \backslash L(A_i-P)$ and $g_i \in L(G-A_i) \backslash L(G-A_i-P).$ Let $(a_i \ast b_{w+1-j}) = ((f_i g_j)(P_n), \ldots, (f_i g_j) (P_n)),$ for $i \leq j.$ Then $$\begin{cases}
a_i \ast b_j \in C_L(D,G) &\text{for $i+j \leq w$}, \\
a_i \ast b_j \in C_L(D,G) \backslash C_L(D,G-P) &\text{for $i+j = w+1$},
\end{cases}$$ Moreover, we have the following interpretation for the sets $\Gamma, \Gamma^\ast, \Delta, \Delta^\ast.$ $$\begin{array}{llcll}
i \in \Gamma &\Leftrightarrow~ A_i \in \Gamma_P(Q), & &i \in \Delta &\Leftrightarrow~ A_i \in \Delta_P(Q), \\
i \in \Gamma^\ast &\Leftrightarrow~ A_i \in \Delta_P(C-Q), & &i \in \Delta^\ast &\Leftrightarrow~ A_i \in \Gamma_P(C-Q).
\end{array}$$ The order bound (Theorem \[T:order\]) and the floor bound (Theorem \[T:floor\]) as well as their generalizations the ABZ bound for cosets (Theorem \[T:cbabz\]) and the ABZ bound for codes (Theorem \[T:Floor2\]) are all obtained in this paper as special cases of the main theorem. Thus, in each case coset decoding can be performed with Theorem \[T:decoding\].
\#1[0=]{}
[CMdST07]{}
Maria Bras-Amor[ó]{}s. Acute semigroups, the order bound on the minimum distance, and the [F]{}eng-[R]{}ao improvements. , 50(6):1282–1289, 2004.
Maria Bras-Amor[ó]{}s and Michael E. O’Sullivan. On semigroups generated by two consecutive integers and improved [H]{}ermitian codes. , 53(7):2560–2566, 2007.
Peter Beelen. The order bound for general algebraic geometric codes. , 13(3):665–680, 2007.
Peter Beelen and Nesrin Tuta[ş]{}. A generalization of the [W]{}eierstrass semigroup. , 207(2):243–260, 2006.
Hao Chen and Ronald Cramer. Algebraic geometric secret sharing schemes and secure multi-party computations over small fields. In [*Advances in cryptology—[CRYPTO]{} 2006*]{}, volume 4117 of [ *Lecture Notes in Comput. Sci.*]{}, pages 521–536. Springer, Berlin, 2006.
Ronald Cramer, Vanesa Daza, Ignacio Gracia, Jorge Jim[é]{}nez Urroz, Gregor Leander, Jaume Mart[í]{}-Farr[é]{}, and Carles Padr[ó]{}. On codes, matroids and secure multi-party computation from linear secret sharing schemes. In [*Advances in cryptology—CRYPTO 2005*]{}, volume 3621 of [ *Lecture Notes in Comput. Sci.*]{}, pages 327–343. Springer, Berlin, 2005.
Ronald Cramer, Ivan Damg[å]{}rd, and Ueli Maurer. General secure multi-party computation from any linear secret-sharing scheme. In [*Advances in cryptology—[EUROCRYPT]{} 2000 ([B]{}ruges)*]{}, volume 1807 of [*Lecture Notes in Comput. Sci.*]{}, pages 316–334. Springer, Berlin, 2000.
Antonio Campillo, Jos[é]{} Ignacio Farr[á]{}n, and Carlos Munuera. On the parameters of algebraic-geometry codes related to [A]{}rf semigroups. , 46(7):2634–2638, 2000.
C[í]{}cero Carvalho, Carlos Munuera, Ercilio da Silva, and Fernando Torres. Near orders and codes. , 53(5):1919–1924, 2007.
C[í]{}cero Carvalho and Fernando Torres. On [G]{}oppa codes and [W]{}eierstrass gaps at several points. , 35(2):211–225, 2005.
Iwan M. Duursma. . Dissertation, Technische Universiteit Eindhoven, Eindhoven, 1993.
Iwan M. Duursma. Majority coset decoding. , 39(3):1067–1070, 1993.
Iwan M. Duursma. Algebraic geometry codes: general theory. In C. Munuera E. Martinez-Moro and D. Ruano, editors, [*Advances in Algebraic Geometry Codes*]{}, Series on Coding Theory and Cryptography. World Scientific, to appear.
Gui Liang Feng and T. R. N. Rao. Decoding algebraic-geometric codes up to the designed minimum distance. , 39(1):37–45, 1993.
Arnaldo Garc[í]{}a, Seon Jeong Kim, and Robert F. Lax. Consecutive [W]{}eierstrass gaps and minimum distance of [G]{}oppa codes. , 84(2):199–207, 1993.
Masaaki Homma and Seon Jeong Kim. The complete determination of the minimum distance of two-point codes on a [H]{}ermitian curve. , 40(1):5–24, 2006.
W. Cary Huffman and Vera Pless. . Cambridge University Press, Cambridge, 2003.
Johan P. Hansen and Henning Stichtenoth. Group codes on certain algebraic curves with many rational points. , 1(1):67–77, 1990.
Seon Jeong Kim. On the index of the [W]{}eierstrass semigroup of a pair of points on a curve. , 62(1):73–82, 1994.
Christoph Kirfel and Ruud Pellikaan. The minimum distance of codes in an array coming from telescopic semigroups. , 41(6, part 1):1720–1732, 1995. Special issue on algebraic geometry codes.
Benjamin Lundell and Jason McCullough. A generalized floor bound for the minimum distance of geometric [G]{}oppa codes. , 207(1):155–164, 2006.
Gretchen L. Matthews. Weierstrass pairs and minimum distance of [G]{}oppa codes. , 22(2):107–121, 2001.
Gretchen L. Matthews. Codes from the [S]{}uzuki function field. , 50(12):3298–3302, 2004.
Hiren Maharaj and Gretchen L. Matthews. On the floor and the ceiling of a divisor. , 12(1):38–55, 2006.
Michael E. O’Sullivan. New codes for the [B]{}erlekamp-[M]{}assey-[S]{}akata algorithm. , 7(2):293–317, 2001.
Seungkook Park. . Dissertation, University of Illinois, Urbana, 2007.
Oliver Pretzel. , volume 8 of [*Oxford Lecture Series in Mathematics and its Applications*]{}. The Clarendon Press Oxford University Press, New York, 1998.
Serguei A. Stepanov. . Kluwer Academic/Plenum Publishers, New York, 1999.
Henning Stichtenoth. . Universitext. Springer-Verlag, Berlin, 1993.
Michael Tsfasman, Serge Vl[ă]{}du[ţ]{}, and Dmitry Nogin. , volume 139 of [ *Mathematical Surveys and Monographs*]{}. American Mathematical Society, Providence, RI, 2007.
J. H. van Lint. , volume 86 of [*Graduate Texts in Mathematics*]{}. Springer-Verlag, Berlin, third edition, 1999.
Jacobus H. van Lint and Richard M. Wilson. On the minimum distance of cyclic codes. , 32(1):23–40, 1986.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the recently proposed dark left-right gauge model (DLRM) of particle interactions, the usual left-handed lepton doublet $(\nu,e)_L$ transforming under $SU(2)_L$ is accompanied by the $unusual$ right-handed fermion doublet $(n,e)_R$ transforming under $SU(2)_R$, where $n_R$ is $not$ the Dirac mass partner of $\nu_L$. In this scenario, whereas $\nu_L$ is certainly a neutrino, $n_R$ should be considered a $scotino$, i.e. a dark-matter fermion. Variants of this basic idea are discussed, including its minimal $scotogenic$ realization.'
---
UCRHEP-T464\
April 2009
[**Variants of the Dark Left-Right Gauge Model:\
Neutrinos and Scotinos\
**]{}
: The gauge group $SU(3)_C \times
SU(2)_L \times U(1)_Y$ of the Standard Model (SM) of particle interactions treats left-handed and right-handed fermions differently, with the electric charge given by $Q = T_{3L} + Y$. To restore left-right symmetry, it is often proposed that the extension $SU(3)_C \times SU(2)_L \times SU(2)_R
\times U(1)_{B-L}$ be considered, where $Q = T_{3L} + T_{3R} + (B-L)/2$. In that case, the fermion content of the SM gains one extra particle, i.e. $\nu_R$ in the right-handed lepton doublet $(\nu,l)_R$. Connecting this with the usual left-handed lepton doublet $(\nu,l)_L$ through a Higgs bidoublet, $\nu_R$ pairs with $\nu_L$ to obtain a Dirac mass, just as $l_R$ does with $l_L$. Assuming $SU(2)_R \times U(1)_{B-L}$ is broken to $U(1)_Y$ through a Higgs triplet transforming as $(1,1,3,1)$, $\nu_R$ gets a large Majorana mass, thereby inducing a small seesaw mass for $\nu_L$. The above is a well-known scenario for what the addition of $\nu_R$ would do for understanding the existence of tiny neutrino masses. For a more general discussion of the $SU(2)_R$ breaking scale, see Ref. [@m04].
Suppose the mass connection between $\nu_R$ and $\nu_L$ is severed without affecting $l_R$ and $l_L$, then $\nu_L$ and $\nu_R$ can be different particles, with their own interactions. Whereas $\nu_L$ is clearly still the well-known neutrino, $\nu_R$ may become something else entirely. As shown in Ref. [@klm09], it may in fact be a $scotino$, i.e. a dark-matter fermion, and to avoid confusion, it is renamed $n_R$. This is accomplished in a nonsupersymmetric $SU(3)_C \times SU(2)_L \times
SU(2)_R \times U(1)$ model with the imposition of a global U(1) symmetry $S$, such that the breaking of $SU(2)_R \times S$ will leave the generalized lepton number $L = S - T_{3R}$ unbroken. It is called the dark left-right model (DLRM), to distinguish it from the alternative left-right model (ALRM) proposed 22 years ago [@m87-1; @bhm87] which has the same crucial property that $n_R$ is $not$ the mass partner of $\nu_L$.
: The fermion structure of the DLRM under $SU(3)_C \times SU(2)_L \times
SU(2)_R \times U(1) \times S$ is given by [@klm09] $$\begin{aligned}
&& \psi_L = \pmatrix{\nu \cr e}_L \sim (1,2,1,-1/2;1), ~~~
\psi_R = \pmatrix{n \cr e}_R \sim (1,1,2,-1/2;1/2), \\
&& Q_L = \pmatrix{u \cr d}_L \sim (3,2,1,1/6;0), ~~~ d_R \sim (3,1,1,-1/3;0),
\\
&& Q_R = \pmatrix{u \cr h}_R \sim (3,1,2,1/6;1/2), ~~~ h_L \sim (3,1,1,-1/3;1), \end{aligned}$$ where $h$ is a new heavy quark of charge $-1/3$. The above fermionic content was first studied in Ref. [@rr78] and also in Ref. [@bm88], without the identification of $n_R$ as a scotino.
To allow $e_L$ to pair with $e_R$ to form a Dirac fermion, the Higgs bidoublet $$\Phi = \pmatrix{\phi_1^0 & \phi_2^+ \cr \phi_1^- & \phi_2^0} \sim (1,2,2,0;1/2)$$ is added so that $m_e$ is obtained from $v_2 = \langle \phi_2^0 \rangle$. At the same time, $\nu_L$ is connected to $n_R$ through $\phi_1^0$. However, $\langle \phi_1^0 \rangle = 0$ will be maintained because $\phi_1^0$ has $S - T_{3R} = 1$, whereas that of $\phi_2^0$ is zero. As shown in Ref. [@klm09], the spontaneous breaking of $SU(2)_R \times S$ leaves the residual symmetry $L = S - T_{3R}$ unbroken, where $L$ is the conventional lepton number assigned to $\nu$ and $e$. Here $n$ has $L = S - T_{3R} = 0$, whereas $W_R^\pm$ has $L = S - T_{3R} = \mp 1$, and it does not mix with $W_L^\pm$, in contrast to the case of the conventional left-right model, where such mixing is unavoidable. Further, the bidoublet $$\tilde{\Phi} = \sigma_2 \Phi^* \sigma_2 = \pmatrix{\bar{\phi}_2^0 & -\phi_1^+
\cr -\phi_2^- & \bar{\phi}_1^0} \sim (1,2,2,0;-1/2)$$ is prevented by $S$ from coupling $\psi_L$ to $\psi_R$, thereby ensuring the absence of tree-level flavor-changing neutral currents, which was not possible in the conventional nonsupersymmetric left-right model.
In the quark sector, $Q_L$ couples to $Q_R$ through $\tilde{\Phi}$, but not $\Phi$. Hence $m_u$ is obtained from $v_2$, and there is no mixing between $d$ and $h$. The former has $L=0$, but the latter has $L = S - T_{3R} = 1$. For $d_L$ to pair with $d_R$, and $h_R$ to pair with $h_L$, the Higgs doublets $$\Phi_L = \pmatrix{\phi_L^+ \cr \phi_L^0} \sim (1,2,1,1/2;0), ~~~
\Phi_R = \pmatrix{\phi_R^+ \cr \phi_R^0} \sim (1,1,2,1/2;-1/2)$$ are needed. Note that $v_4 = \langle \phi_R^0 \rangle$ will break $SU(2)_R \times U(1)$ to $U(1)_Y$ as desired, and the leptoquark $h$ gets a heavy mass of order $v_4$.
: The fermion sector may be more exotic. For example, Eq. (1) may be replaced by $$\begin{aligned}
&& \psi_L = \pmatrix{\nu \cr e}_L \sim (1,2,1,-1/2;1), ~~~
e_R \sim (1,1,1,-1;1),\\
&& \psi_R = \pmatrix{n \cr E}_R \sim (1,1,2,-1/2;-1/2), ~~~
E_L \sim (1,1,1,-1;0).\end{aligned}$$ In this case, $E$ has $L=0$, $n$ has $L=-1$, $m_e$ comes from $v_3 = \langle
\phi_L^0 \rangle$, $m_E$ from $v_4$, and neither $\Phi$ nor $\tilde{\Phi}$ couples to $\bar{\psi}_L \psi_R$.
As another example, Eqs. (2) and (3) may be replaced by $$\begin{aligned}
&& Q_L = \pmatrix{u \cr d}_L \sim (3,2,1,1/6;0), ~~~ u_R \sim (3,1,1,2/3;0), \\
&& Q_R = \pmatrix{f \cr d}_L \sim (3,1,2,1/6;-1/2), ~~~
f_L \sim (3,1,1,2/3;-1).\end{aligned}$$ In this case, $f$ is an exotic quark of charge 2/3 and $L=-1$. Here, $Q_L$ couples to $Q_R$ through $\Phi$, but not $\tilde{\Phi}$, $m_d$ comes from $v_2$, $m_u$ from $v_3$, and $m_f$ from $v_4$.
: With the above Higgs content, $\nu_L$ and $n_R$ remain massless. Consider now the various ways that they acquire masses:
\(1) In the DLRM [@klm09], Higgs triplets under $SU(2)_L$ and $SU(2)_R$ are used separately for $\nu_L$ and $n_R$ masses.
\(2) In Ref. [@rr78], they are massless.
\(3) In Ref. [@bm88], they acquire radiative masses separately from the addition of two charged scalar singlets.
\(4) In the ALRM [@m87-1], the usual lepton doublet is actually part of a bidoublet: $$\pmatrix{\nu_e & E^c \cr e & N^c_E}_L \sim (1,2,2,0),$$ which means that $\nu_e$ and $e$ have $SU(2)_R$ interactions. In the original proposal, $\nu_L$ and $n_R$ are massless, but they can acquire seesaw masses separately through the many other fields available in the representation of $E_6$, as explained in Ref. [@m00]. One of the three $n_R$ copies in this supersymmetric model pairs with the neutral gaugino from the breaking of $SU(2)_R \times U(1) \to U(1)_Y$ to form a Dirac fermion. The other two are light and considered as sterile neutrinos which mix with $\nu_L$ through the soft term $\nu_e N^c_E - e E^c$ which breaks $R$ parity.
\(5) A simple variation of the DLRM also allows neutrino masses to be radiatively generated by dark matter (i.e. [*scotogenic*]{}) in one loop [@m06-1; @m06-2; @m06-3; @kms06; @ks06; @kko07; @hkmr07; @bm08; @m08-1; @m08-2; @m08-3; @m08-4; @akrsz09; @ms09; @m09; @sty09]. Instead of $\Delta_L$, a scalar singlet $\chi \sim (1,1,1,0;-1)$ is added, then the trilinear scalar term $Tr(\Phi
\tilde{\Phi}^\dagger) \chi$ is allowed.
(500,100)(120,45) (270,50)(310,50) (390,50)(350,50) (310,50)(350,50) (430,50)(390,50) (290,35)\[b\][$\nu_L$]{} (410,35)\[b\][$\nu_L$]{} (352,33)\[b\][$n_R$]{} (352,97)\[b\][$\chi$]{} (305,70)\[b\][$\phi^0_1$]{} (396,70)\[b\][$\phi^0_1$]{} (310,116)\[b\][$\phi^0_2$]{} (390,116)\[b\][$\phi^0_2$]{} (315,111)(330,85)[3]{} (385,111)(370,85)[3]{} (350,50)(40,180,120)[3]{} (350,50)(40,0,60)[3]{} (350,50)(40,90,120)[3]{} (350,50)(40,90,60)[3]{}
Using the soft term $\chi^2$ to break $L$ to $(-)^L$, a scotogenic neutrino mass is obtained as shown in Fig. 1. It is also possible to do this in two loops [@ms07; @m08-5] and three loops [@knt03; @cs04; @aks09].
\(6) Since the scotogenic mechanism of Fig. 1 does not care how $n_R$ acquires a Majorana mass, it may be accomplished with three neutral singlet fermions $n_L$ with $S=0$ instead of the Higgs triplet $\Delta_R$. Now the Yukawa coupling $\bar{n}_L (n_R \phi^0_R - e_R \phi_R^+)$ is allowed, as well as a Majorana mass for $n_L$. Hence $n_R$ gets an induced Majorana mass which is essential for Fig. 1. Note that $n_L$ does not couple to $(\nu_L \phi_L^0 - e_L \phi_L^+)$ because of $S$.
In this minimal variant, the $Z'$ mass comes entirely from the $\Phi_R$ doublet as in the ALRM, hence the prediction $$M^2_{W_R} = {(1-2x) \over (1-x)} M^2_{Z'} + {x^2 \over (1-x)^2} M^2_{W_L},$$ where $x \equiv \sin^2 \theta_W$ and zero $Z-Z'$ mixing has been assumed [@bhm87], i.e. $v_2^2/(v_2^2+v_3^2) = x/(1-x)$. Currently, the experimental bound on $M_{Z'}$ is 850 GeV [@klm09].
The diagram of Fig. 1 is exactly calculable [@m06-1]. The $\bar{n}_R \nu_L
\bar{\phi}_1^0$ coupling is given by $(m_\alpha/v_2) U_{\alpha i}$, where $\alpha = e,\mu,\tau$ and $i = 1$ to 6 refer to the mass eigenstates of the $6 \times 6$ $(n_R,n_L)$ mass matrix $${\cal M}_n = \pmatrix{0 & m_D \cr m_D & m_n}.$$ In the bases (Re$\phi_1^0$, Re$\chi$), (Im$\phi_1^0$, Im$\chi$), the respective mass-squared matrices are $$\pmatrix{m_\phi^2 & \mu v_2 \cr \mu v_2 & m_\chi^2 + \mu_\chi^2}, ~~~
\pmatrix{m_\phi^2 & -\mu v_2 \cr -\mu v_2 & m_\chi^2 - \mu_\chi^2}.$$ Let their mass eigenstates and mixing angles be $(m^2_{R1}, m^2_{R2}, \theta_R)$ and $(m^2_{I1}, m^2_{I2}, \theta_I)$, then $$\begin{aligned}
({\cal M}_\nu)_{\alpha \beta} &=& \sum_i {m_\alpha m_\beta U_{\alpha_i} U_{\beta i}
M_i \over 16 \pi^2 v_2^2} ~[~ \cos^2 \theta_R {m^2_{R1} \over m^2_{R1} - M_i^2}
\ln {m^2_{R1} \over M_i^2} - \cos^2 \theta_I {m^2_{I1} \over m^2_{I1} - M_i^2}
\ln {m^2_{I1} \over M_i^2} \nonumber \\ && + \sin^2 \theta_R {m^2_{R2} \over
m^2_{R2} - M_i^2} \ln {m^2_{R2} \over M_i^2} - \sin^2 \theta_I {m^2_{I2} \over
m^2_{I2} - M_i^2} \ln {m^2_{I2} \over M_i^2}~].\end{aligned}$$ In the limit $\mu_\chi^2 = 0$, ${\cal M}_\nu$ vanishes because $m_{R1} = m_{I1}$, $m_{R2} = m_{I2}$, and $\theta_R = -\theta_I$. In the limit $\mu = 0$, ${\cal M}_\nu$ also vanishes because $m_{R1} = m_{I1}$ and $\theta_R = \theta_I
= 0$. Furthermore, if $m_n = 0$ in Eq. (13), ${\cal M}_\nu$ is again zero because $n$ is a Dirac particle. This latter is an example of the inverse seesaw mechanism [@ww83; @mv86; @m87-2; @dv05; @kk07]. Hence neutrino masses are suppressed in this scenario by three possible limits, and the scale of $SU(2)_R$ breaking (associated with the masses of the dark-matter particles of Fig. 1) may well be as low as 1 TeV, as advocated.
: The neutral component $n$ of the $SU(2)_R$ doublet $(n,e)_R$ is proposed as a dark-matter fermion (scotino). Variants of this basic idea, the dark left-right model (DLRM), are discussed. A minimal version is considered, where neutrino masses are radiatively generated by dark matter (scotogenic) and naturally suppressed, allowing the $SU(2)_R$ breaking scale to be as low as 1 TeV.
: This work was supported in part by the U. S. Department of Energy under Grant No. DE-FG03-94ER40837.
[99]{} E. Ma, Phys. Rev. [**D69**]{}, 011301(R) (2004). S. Khalil, H.-S. Lee, and E. Ma, Phys. Rev. [**D79**]{}, 041701(R) (2009). E. Ma, Phys. Rev. [**D36**]{}, 274 (1987). K. S. Babu, X.-G. He, and E. Ma, Phys. Rev. [**D36**]{}, 878 (1987). P. Ramond and D. B. Reiss, Phys. Lett. [**80B**]{}, 87 (1978). K. S. Babu and V. S. Mathur, Phys. Rev. [**D38**]{}, 3550 (1988). E. Ma, Phys. Rev. [**D62**]{}, 093022 (2000). E. Ma, Phys. Rev. [**D73**]{}, 077301 (2006). E. Ma, Mod. Phys. Lett. [**A21**]{}, 1777 (2006). E. Ma, Ann. Fondation de Broglie [**31**]{}, 85 (2006). J. Kubo, E. Ma, and D. Suematsu, Phys. Lett. [**B642**]{}, 18 (2006). J. Kubo and D. Suematsu, Phys. Lett. [**B643**]{}, 336 (2006). Y. Kajiyama, J. Kubo, and H. Okada, Phys. Rev. [**D75**]{}, 033001 (2007). T. Hambye, K. Kannike, E. Ma, and M. Raidal, Phys. Rev. [**D75**]{}, 095003 (2007). K. S. Babu and E. Ma, Int. J. Mod. Phys. [**A23**]{}, 1813 (2008). E. Ma, Phys. Lett. [**B659**]{}, 885 (2008). E. Ma, Phys. Lett. [**B661**]{}, 273 (2008). E. Ma, Mod. Phys. Lett. [**A23**]{}, 647 (2008). E. Ma, Mod. Phys. Lett. [**A23**]{}, 721 (2008). D. Aristizabal Serra, J. Kubo, D. Restrepo, D. Suematsu, and O. Zapata, Phys. Rev. [**D79**]{}, 013011 (2009). E. Ma and D. Suematsu, Mod. Phys. Lett. [**A24**]{}, 583 (2009). E. Ma, Phys. Lett. [**B671**]{}, 366 (2009). D. Suematsu, T. Toma, and T. Yoshida, arXiv:0903.0287 \[hep-ph\]. E. Ma and U. Sarkar, Phys. Lett. [**B653**]{}, 288 (2007). E. Ma, Phys. Lett. [**B662**]{}, 49 (2008). L. M. Krauss, S. Nasri, and M. Trodden, Phys. Rev. [**D67**]{}, 085002 (2003). K. Cheung and O. Seto, Phys. Rev. [**D69**]{}, 113009 (2004). M. Aoki, S. Kanemura, and O. Seto, Phys. Rev. Lett. [**102**]{}, 051805 (2009). D. Wyler and L. Wolfenstein, Nucl. Phys. [**B218**]{}, 205 (1983). R. N. Mohapatra and J. W. F. Valle, Phys. Rev. [**D34**]{}, 1642 (1986). E. Ma, Phys. Lett. [**B191**]{}, 287 (1987). F. Deppisch and J. W. F. Valle, Phys. Rev. [**D72**]{}, 036001 (2005). S. K. Kang and C. S. Kim, Phys. Lett. [**B646**]{}, 248 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the nonlocal properties of states resulting from the mixture of an arbitrary entangled state $\rho$ of two $d$-dimensional systems and completely depolarized noise, with respective weights $p$ and $1-p$. We first construct a local model for the case in which $\rho$ is maximally entangled and $p$ at or below a certain bound. We then extend the model to arbitrary $\rho$. Our results provide bounds on the resistance to noise of the nonlocal correlations of entangled states. For projective measurements, the critical value of the noise parameter $p$ for which the state becomes local is at least asymptotically $\log(d)$ larger than the critical value for separability.'
author:
- 'Mafalda L. Almeida$^{1}$, Stefano Pironio$^{1}$, Jonathan Barrett$^2$, Géza Tóth $^{1,3}$, and Antonio Acín$^{1,4}$'
title: Noise robustness of the nonlocality of entangled quantum states
---
In 1964, Bell showed that some entangled states are nonlocal, in the sense that measurements on them yield outcome correlations that cannot be reproduced by a locally causal model [@Bell]. This nonlocal character of entangled states may be demonstrated through the violation of Bell inequalities. All pure entangled states violate such an inequality, hence are nonlocal [@Gisin]. For noisy states, the picture is much subtler. Werner constructed in 1989 a family of bipartite mixed states, which, while being entangled, return outcome correlations under projective measurements that can be described by a local model [@Werner]. This result has been extended to general measurements [@Barrett] and more parties [@TA]. Thus, while entanglement is necessary for a state to be nonlocal, in the case of mixed states it is not sufficient.
Beyond these exploratory results, little is known about the relation between noise, entanglement, and quantum nonlocality. Understanding this relation, apart from its fundamental interest, is important from the perspective of Quantum Information Science. In this context, entanglement is commonly viewed as a useful resource for various information-processing tasks. Not all entangled states, however, are useful for every task: for example, quantum computation with slightly entangled states can be efficiently simulated on a classical computer [@comp], and bound entangled states are useless for teleportation [@HHH]. For certain tasks, such as quantum communication complexity problems [@commcompl], or device-independent quantum key distribution [@AGM], entangled states are useful only to the extent that they exhibit nonlocal correlations. Indeed, in these scenarios two (or more) distant observers, Alice and Bob, directly exploit the correlations $$\label{qcorr}
P_{MN}(a,b)=\tr(\rho_{AB}\,M_a\otimes N_b)\,,$$ obtained by performing measurements $M$ and $N$ on a distributed entangled state $\rho_{AB}$ (in the above formula, $M_a$ and $N_b$ are the positive operators associated with the measurement outcomes $a$ and $b$). If the entangled state $\rho_{AB}$ can be simulated by a local model, these correlations can be written as $$\label{loccorr}
P_{MN}(a,b)=\int\!\! \mu(d\lambda)\, P_M(a|\lambda)P_N(b|\lambda)\, ,$$ where $\lambda$ denotes a shared classical variable distributed with probability measure $\mu$, and $P_M(a|\lambda)$ and $P_N(b|\lambda)$ are the local response functions of Alice and Bob. For all practical purposes then, the entangled state $\rho_{AB}$ can be replaced by classical correlations, and so does not provide any improvement over what is achievable using classical resources [@Lluis].
In this work, we estimate the resistance to noise of the nonlocal correlations of bipartite entangled states in $\compl^d\otimes
\compl^d$, where $d$ is the local Hilbert space dimension of each subspace. To do this, we analyze the nonlocal properties of states resulting from the mixture of an arbitrary state $\rho$ with completely depolarized noise, $$\label{noisy_states}
\rho(p)=p\,\rho+(1-p)\frac{\one}{d^2}\, .$$ Our goal is to find the minimal amount of noise that destroys the nonlocal correlations of any state $\rho$, i.e., the maximal value $p_L$ such that $\rho(p)$ is local for any $\rho$ when $p\leq p_L$. Clearly, for sufficiently small values of $p\leq p_S$, the state $\rho(p)$ becomes separable for any $\rho$ [@sep; @GB], thus local. We give here lower bounds on $p_L$ that are more constraining than the one obtained from the separability condition. If we restrict Alice and Bob to perform projective measurements only, the bound that we obtain for the locality limit is asymptotically $\log(d)$ larger than the separability limit.
A key step in the proof of our results is the construction of a local model for states of the form (\[noisy\_states\]) when $\rho=\ket{\phi_d}\bra{\phi_d}$ is maximally entangled, i.e. $\ket{\phi_d}=1/\sqrt d\sum_{i=1}^d \ket{ii}$. Thus we also provide a lower bound on $p_L^{\phi}$, defined as the maximal value of $p$ such that $$p\,\ket{\phi_d}\bra{\phi_d}+(1-p)\frac{\one}{d^2}$$ is local. This last result implies in particular the existence of entangled states whose nonlocal correlations are more robust than those of maximally entangled ones.
The results presented here concern mostly the simpler but physically relevant case in which Alice and Bob are restricted to projective measurements. Extensions to completely general measurements are discussed at the end of the paper. Our results also provide bounds for the notion of state steerability introduced in [@WJD].
As mentioned, we start by analyzing the case in which the state $\rho$ in (\[noisy\_states\]) is maximally entangled. Such states are called isotropic states and are the unique ones invariant under $U\otimes U^*$ transformations for all unitary operators $U$ on $\compl^d$ [@HH]. If Alice and Bob each make on these states a projective measurement, specified by a set of $d$ orthogonal projectors $Q=\{Q_a\}$ for Alice and $R=\{R_b\}$ for Bob, with $a,b=1,\ldots,d$, the resulting joint outcome probabilities are given by $$\label{quantum prob proj} \frac{p}{d}\tr\left(Q_a^T
R_b\right)+\frac{1-p}{d^2}\, .$$ Our first aim is to construct a local model for isotropic states, that is, to write the quantum probabilities in the form (\[loccorr\]) for some value of the noise parameter $p$.
Our construction is inspired by the model given in Ref. [@Werner] for Werner states, which are $U\otimes U$ invariant, and which we adapt to the $U\otimes U^*$ symmetry of isotropic states. The local classical variables $\lambda$ in our model are taken to be complex $d$-dimensional vectors which we can thus formally identify with $d$-dimensional quantum states $\ket{\lambda}$. The probability measure $\mu$ is the unique measure invariant under all unitary transformations $U$ on $\compl^d$. In analogy with the quantum formalism, Alice’s response function is defined as $$\label{aliceresp} P_Q(a|\lambda)=\bra{\lambda}Q_a^T\ket{\lambda}\,
.$$ Bob’s response function is suggested by the perfect correlations of maximally entangled states and taken to be $$\label{bobresp} P_R(b|\lambda)=\begin{cases}
1& \text{if } \bra{\lambda}R_b\ket{\lambda}=\max_{i}{\bra{\lambda}R_i\ket{\lambda}}\\
0& \text{otherwise}\,.
\end{cases}$$ It satisfies \[Ubob\] P\_[U\^R U]{}(b|)=P\_R(b|U). To obtain the joint probabilities predicted by this model, and to compare them with , it is necessary to compute the integral for our specific choice of measure $\mu$ and response functions. Following Werner (see [@Werner] for details), one can show that the $U$-invariance of $\mu$, the form of Alice’s response function, and the relation satisfied by Bob’s response function, imply $$\label{integral LHV}
\int\!\!\mu(d\lambda)\,P_Q(a|\lambda)P_R(b|\lambda)=\tr\left(Q_a^T\hat
B(b,R)\right)\,,$$ where $\hat B(b,R)$ is a positive operator depending on Bob’s response function. One can further show, exploiting the fact that the relation holds for all one-dimensional projectors $Q_a$ [@Werner], that $\hat B(b,R)=(p^\phi/d)\,
R_b+(1-p^\phi)/d^2\, \one$, for some $p^\phi\in \mathbb{R}$, and thus that $$\label{integral LHV 2}
\int\!\!\mu(d\lambda)P_Q(a|\lambda)P_R(b|\lambda)\!=\!\frac{p^\phi}{d}\tr\left(Q_a^T
R_b \right)+\frac{1-p^\phi}{d^2}\,.$$ These correlations are thus already of the prescribed form (\[quantum prob proj\]). To determine the value of $p^\phi$ for which holds, it is sufficient to compute the integral in the simplest case where $Q_a^T=R_b$, which gives $$\label{pcr}
p^\phi=\frac{1}{d-1}\left(-1+d^2\int\!\!
\mu(d\lambda)\,\bra{\lambda}R_b
\ket{\lambda}P_R(b|\lambda)\right)\, .$$ It now remains to evaluate this integral for the specific choice (\[bobresp\]) for $P_R(b|\lambda)$. After patient algebra, one obtains $$\label{boundmax}
p^\phi=\frac{1}{d-1}\left(-1+\sum_{k=1}^d\frac{1}{k}\right)
\,\xrightarrow[\mathrm{large}\; d]{}\,\frac{\log(d)}{d}\,.$$ For $d=2$, $p^\phi=1/2$ is equal to the critical value for two-dimensional Werner states, as expected since Werner and isotropic states are equivalent up to local unitary transformations when $d=2$. In the limit of large $d$, $p^\phi$ is asymptotically $\log(d)$ larger than the critical probability $p^\phi_S=1/(d+1)$ for the separability of isotropic states [@HH].
Our next goal is to generalize the local model for isotropic states to mixed states of the form $$\label{pure+noise}
\rho=p\,\proj{\psi}+(1-p)\frac{\one}{d^2}\, ,$$ where $\ket{\psi}$ is an arbitrary pure state in $\compl^d\otimes\compl^d$. This automatically also implies a model for the general states (\[noisy\_states\]), since any mixed state $\rho$ is a convex combination of pure states. To do this, we incorporate Nielsen’s protocol [@nielsen] for the conversion of bipartite pure states by local operations and classical communication (LOCC) into our model. Recall that a maximally entangled state $\ket{\phi_d}$ can be transformed by LOCC in a deterministic way into an arbitrary state $\ket{\psi}$ by a single measurement on Alice’s particle followed by a unitary operation on Bob’s side, depending on Alice’s measurement outcome. Indeed, consider an arbitrary pure entangled state written in its Schmidt form $\ket{\psi}=\sum_{j=0}^{d-1}\nu_j\ket{jj}$, and denote by $D_\nu$ the $d\times d$ diagonal matrix with entries $(D_\nu)_{jj}=\nu_j$. Taking the $d$ cyclic permutations $\Pi_i=\sum_{j=0}^{d-1} \ket{j}\bra{j+i \,(\mathrm{mod}\, d)}$, where $i=0,\ldots,d-1$, it is possible to write $$\label{nielsenrel} \ket{\psi}=\sqrt {d}(A_i\otimes \Pi_i)
\ket{\phi_d}\quad
\text{for all }
i=0,\ldots,d-1\, ,$$ with $A_i=D_\nu\Pi_i$. The operators $W_i=A_i^\dagger A_i$ define a measurement, since they are positive and sum to the identity, $\sum_i W_i=\one$. In order to convert $\ket{\phi_d}$ into $\ket{\psi}$, Alice first carries out this measurement, obtaining the outcome $i$ with probability $\bra{\phi_d}W_i\ket{\phi_d}=1/d$. She then communicates her result to Bob who applies the corresponding unitary operation $\Pi_i$, the resulting normalized state being $\ket{\psi}$, as implied by (\[nielsenrel\]).
The quantum-like properties of our local model, i.e., the fact that the hidden variable $\ket{\lambda}$ can be thought of as a quantum state and the quantum form of the response function (\[aliceresp\]), allow us to adapt Nielsen’s construction to it. The idea is that at the source, before sending the classical instructions $\ket{\lambda}$ to each party, a measurement defined by the operators $A^*_i$ is simulated on $\ket{\lambda}$, giving outcome $i$ with probability $q_i(\lambda)=\bra{\lambda}A_i^T
A_i^*\ket{\lambda}$. The classical description of the normalized hidden states $\ket{\lambda^A_i}=A^*_i\ket{\lambda}/\sqrt{q_i}$ and $\ket{\lambda^B_i}=\Pi_i\ket{\lambda}$ are then sent, respectively, to Alice and Bob, who use them in the response functions and instead of $\ket{\lambda}$. The joint probabilities $P_{QR}(a,b)$ predicted by the model for measurements $Q$ and $R$ are thus given by $$\begin{gathered}
\label{nielsenmodel}
\int\!\! \mu(d\lambda)\, \sum_{i=0}^{d-1} q_i(\lambda) P_Q(a|\lambda^A_i)P_R(b|\lambda^B_i) \\
=\sum_{i=0}^{d-1} \int\!\! \mu(d\lambda)\, \bra{\lambda}A_i^T
Q_a^T A_i^*\ket{\lambda} P_{\Pi^{\dagger}_i R_b \Pi_i}(b|\lambda)\end{gathered}$$ where we used property . Replacing the integral in the last expression by the right-hand side of , we obtain $$\sum_{i=0}^{d-1}\left(\frac{p^\phi}{d}\tr(A_i^T Q_a^TA_i^*\Pi_i^\dagger R_b \Pi_i)+
\frac{1-p^\phi}{d^2}\tr(A_i^T Q_a^TA_i^*)\right)\, .$$ Using Eqs. (\[nielsenrel\]) and , and the fact that $\sum A_iA_i^\dagger=d\sigma$, where $\sigma=\tr_B\proj{\psi}$, one can check that these probabilities are equal to the quantum probabilities $\tr\left(\tilde\rho\,Q_a\otimes R_b \right)$ for the state $$\label{noisy_psi}
\tilde\rho=p^\phi\proj{\psi}+(1-p^\phi)\sigma\otimes\frac{\one}{d}\, .$$ Not surprisingly, the measurement at the source modifies the local noise of Alice, which is no longer completely depolarized, and introduce some bias depending on $\ket{\psi}$.
This result can already be interpreted as a measure of the robustness of the nonlocal correlations of an arbitrary entangled state $\ket{\psi}$. By mixing a state-dependent local noise, with mixing probability $1-p^\phi$, it is always possible to wash out the nonlocal correlations of the state $\ket{\psi}$.
In order to extend this result to the case of completely depolarized noise, one can add some extra local noise to Alice such that the resulting state has the form , with the penalty that $p< p^\phi$. Writing the reduced density matrix $\sigma$ in its diagonal form $\sigma=\sum_j \mu_j^2
\proj{j}$, and defining $\sigma_k=\sum_j \mu_{j+k
\,(\mathrm{mod}\,d)}^2 \proj{j}$, it is clear that the state $$\label{mixtonoise}
q\tilde\rho+\frac{1-q}{d-1}\sum_{k=1}^{d-1}\sigma_k\otimes\frac{\one}{d}$$ has the form for $q(1-p_d)=(1-q)/(d-1)$, in which case the probability $p$ is given by $$\label{asympt pr}
p^\rho=\frac{p^\phi}{(1-p^\phi)(d-1)+1}\,\xrightarrow[\mathrm{large}\;
d]{}\,\frac{\log(d)}{d^2}\,.$$ The state (\[mixtonoise\]) is clearly local, since it is a convex combination of local states. We have thus shown that the noisy states (\[noisy\_states\]) have a local model for projective measurements whenever $p\leq p^\rho$. The probabilities $p^\phi$ and $p^\rho$ represent the main results of this work and provide lower bounds on $p_L^{\phi}$ and $p_L$. Several implications of our findings are discussed in what follows.
First of all, one may ask about the tightness of our bound. Actually, our model is based on Werner’s construction, and this model is known not to be tight in the case $d=2$ [@AGT]. Even if it is not tight, it would be interesting to understand whether the model predicts the right asymptotic dependence with the Hilbert-space dimension $d$. An upper bound on $p_L$ follows from the results of [@ADGL], where it was shown that a state of the form $\varrho_2=p \ket{\phi_2}\bra{\phi_2}+(1-p)\one/d^2$, where $\ket{\phi_2}=1/\sqrt{2}(\ket{00}+\ket{11})$ is a projector onto a two-qubit maximally entangled state, violates the Clauser-Horne-Shimony-Holt inequality [@CHSH] whenever $p>p^{\varrho_{2}}$, where $$\label{upprob}
p^{\varrho_2}=\frac{4(d-1)}{(\sqrt 2-1)d^2+4d-4}\,\xrightarrow[\mathrm{large}\; d]{}\,\frac{4}{(\sqrt 2-1)d}\,,$$ which tends to zero when $d\to\infty$. This result together with our previous model thus imply that $p^\rho\leq p_L\leq
p^{\varrho_2}$.
state separability locality (projective meas.) locality (general meas.)
------------------ ------------------------------------------------ ------------------------------------------------------------------------------------------------ ---------------------------------------------------------------------------------------
$\ket{\phi_d}$ $p^\phi_S=\frac{1}{d+1}$ $\Theta\left(\frac{\log d}{d}\right)\leq p^\phi_L\leq\frac{\pi^2}{16\,K}\simeq 0.67$ $\Theta\left(\frac{3}{e\,d}\right)\leq p^\phi_L\ \leq\frac{\pi^2}{16\,K}\simeq 0.67$
arbitrary $\rho$ $\frac{1}{d^2-1}\leq p_S \leq \frac{2}{d^2+2}$ $\Theta\left(\frac{\log d}{d^2}\right)\leq p_L\leq \Theta\left(\frac{4}{(\sqrt{2}-1)d}\right)$ $\Theta(\frac{3}{e\,d^2})\leq p_L\leq\Theta\left(\frac{4}{(\sqrt{2}-1)d}\right)$
Our results, when combined with (\[upprob\]), also provide a strict proof of the fact that the nonlocal correlations of maximally entangled states, under projective measurements, are not the most robust ones. Indeed, we have a local model for isotropic states whenever $p\leq p^\phi$, while there exist quantum states of the form (\[noisy\_states\]) violating a Bell inequality when $p>p^{\varrho_2}$. For sufficiently large dimension, $p^{\varrho_2}<p^\phi$, so we have a Bell inequality violation in a range of $p$ for which we have shown the existence of a local model for isotropic states.
It is also interesting to compare the bounds derived here for nonlocality with those known for entanglement. To our knowledge, the best upper and lower bound on the critical probability $p_S$ such that the states (\[noisy\_states\]) are guaranteed to be separable were obtained in Ref. [@GB]: $$\label{sep_bounds}
\frac{1}{d^2-1}\leq p_S\leq \frac{2}{d^2+2}\, .$$ Interestingly, the upper bound is obtained, as above, for the case in which the state $\rho$ in (\[noisy\_states\]) is equal to a projector onto $\ket{\phi_2}$. Comparing with Eq. (\[asympt pr\]), we see that the critical noise probability for nonlocality under projective measurements is, at least, asymptotically $\log(d)$ larger than the one for separability, as it is for isotropic states.
Finally, let us briefly mention how the above results can be extended to the case of general measurements. The idea is, as above, to start by constructing a model for isotropic states, adapting the one for Werner states of Ref. [@Barrett]. As noted in [@Barrett], it is sufficient to simulate measurements $M$ and $N$ defined by operators $M_a=c_a Q_a$ and $N_b=c_b R_b$ proportional to one-dimensional projectors $Q_a$ and $R_b$ to be able to simulate any measurement by Alice and Bob. In our corresponding model, the hidden states are again vectors $\ket{\lambda}$ in $\compl^{d}$ chosen with the Haar measure $\mu$. Alice’s response function is basically the same as before, $$\label{alicepovm}
P_M(a|\lambda)=\bra{\lambda}M_a^T\ket{\lambda}\,,$$ while Bob’s is, taking inspiration from [@Barrett], chosen as $$\begin{gathered}
\label{bobpovm}
P_N(b|\lambda) = \bra{\lambda}N_b\ket{\lambda}
\Theta\!\left(\bra{\lambda}R_b\ket{\lambda}-\frac{1}{d}\right)\\
+ \frac{c_b}{d}\left[1-\sum_k \bra{\lambda}N_k\ket{\lambda}\Theta\!\left(\bra{\lambda}R_k\ket{\lambda}-
\frac{1}{d}\right)\right]\,,
\end{gathered}$$ where $\Theta$ is the Heaviside step function. Evaluation of the integral with the definitions and can be done along the same steps as in [@Barrett] and yields the joint measurement outcome probabilities for an isotropic state with the critical value $$\label{probpovm}
\tilde p^\phi=\frac{(3d-1)(d-1)^{d-1}}{(d+1)d^d}\,\xrightarrow[\text{large }d]{}\,\frac{3}{e}\frac{1}{d}\, .$$
Since this model has the same quantum-like properties as the one for projective measurements, cf. definition (\[alicepovm\]), it can also be extended to arbitrary noisy states (\[noisy\_states\]) using Nielsen’s protocol. The corresponding critical probability is given by (\[pcr\]) with $p^\phi$ replaced by the above value of $\tilde p^\phi$.
In conclusion, we have obtained bounds on the robustness of the nonlocal correlations of arbitrary entangled states. Our results are summarized in Table I. In the particular but interesting case where the state is maximally entangled, we derived better bounds by exploiting the symmetry of isotropic states [@WJD]. Apart from their fundamental significance, our results are interesting from the point of view of the characterization of quantum information resources: if the noise affecting a state is larger than our bounds, its outcome correlations for local measurements can be reproduced by classical means alone.
*Note added.* While completing this work, we learned that our local model for isotropic states was independently derived in [@WJD] in the context of state steerability. We note that all our models imply the non-steerability of the corresponding quantum states because Alice’s response function is always quantum (see [@WJD] for details).
*Acknowledgements.* We acknowledge financial support from the EU Qubit Applications Project (QAP) Contract number 015848, the Spanish projects FIS2004-05639-C02-02, Consolider QOIT, the Spanish MEC for “Ramon y Cajal" and “Juan de la Cierva" grants, the Generalitat de Catalunya, the Fundação para a Ciência e a Tecnologia (Portugal) through the grant SFRH/BD/21915/2005, the National Research Fund of Hungary OTKA under Contract No. T049234, and the Hungarian Academy of Sciences (Bolyai Programme). Research at Perimeter Institute for Theoretical Physics is supported in part by the Government of Canada through NSERC and by the Province of Ontario through MRI.
[99]{}
J. S. Bell, Physics [**1**]{}, 195 (1964).
N. Gisin, Phys. Lett. A [**154**]{}, 201 (1991).
R. F. Werner, Phys. Rev. A [**40**]{}, 4277 (1989).
J. Barrett, Phys. Rev. A [**65**]{}, 042302 (2002).
G. Tóth and A. Acín, Phys. Rev. A [**74**]{}, 030306(R) (2006).
R. Jozsa and N. Linden, quant-ph/0201143; G. Vidal, Phys. Rev. Lett. [**91**]{}, 147902 (2003).
M. Horodecki, P. Horodecki and R. Horodecki, Phys. Rev. A [**60**]{}, 1888 (1999).
See for instance G. Brassard, quant-ph/0101005; C. Brukner, M. Zukowski, and A. Zeilinger, Phys. Rev. Lett. [**89**]{}, 197901 (2002).
A. Acín [*et al.*]{} quant-ph/0702152; A. Acín, N. Gisin, and L. Masanes, Phys. Rev. Lett. [**97**]{}, 120405 (2006); J. Barrett, L. Hardy, and A. Kent, [*ibid*]{} **95**, 010503 (2005).
This does not mean that the state can be replaced by shared randomness in any scenario, see Ll. Masanes, Phys. Rev. Lett. [**96**]{}, 150501 (2006).
K. Zyczkowski, P. Horodecki, A. Sanpera and M. Lewenstein, Phys. Rev. A [**58**]{}, 883 (1998); G. Vidal and R. Tarrach, Phys. Rev. A [**59**]{}, 141 (1999); S. L. Braunstein [*et al.*]{}, Phys. Rev. Lett. [**83**]{}, 1054 (1999).
L. Gurvits and H. Barnum, Phys. Rev. A [**66**]{}, 062311 (2002).
H. M. Wiseman, S. J. Jones and A. C. Doherty, Phys. Rev. Lett. [**98**]{}, 140402 (2007).
M. Horodecki and P. Horodecki, Phys. Rev. A [**59**]{}, 4206 (1999).
M. A. Nielsen, Phys. Rev. Lett. [**83**]{}, 436 (1999).
A. Acín, N. Gisin and B. Toner, Phys. Rev. A **73**, 062105 (2006).
A. Acín, T. Durt, N. Gisin and J. I. Latorre, Phys. Rev. A **65**, 052325 (2002).
J. F. Clauser, M. A. Horne, A. Shimony and R. A. Holt, Phys. Rev. Lett. [**23**]{}, 880 (1969).
D. Collins, N. Gisin, N. Linden, S. Massar and S. Popescu, Phys. Rev. Lett. **88**, 040404 (2002).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Intelligent systems sometimes need to infer the probable goals of people, cars, and robots, based on partial observations of their motion. This paper introduces a class of probabilistic programs for formulating and solving these problems. The formulation uses randomized path planning algorithms as the basis for probabilistic models of the process by which autonomous agents plan to achieve their goals. Because these path planning algorithms do not have tractable likelihood functions, new inference algorithms are needed. This paper proposes two Monte Carlo techniques for these “likelihood-free” models, one of which can use likelihood estimates from neural networks to accelerate inference. The paper demonstrates efficacy on three simple examples, each using under 50 lines of probabilistic code.'
author:
- |
[**Marco F. Cusumano-Towner**]{}$^1$, [**Alexey Radul**]{}$^1$, [**David Wingate**]{}$^2$, [**Vikash K. Mansinghka**]{}$^1$\
$^1$Probabilistic Computing Project, Massachusetts Institute of Technology\
$^2$Computer Science Department, Brigham Young University
bibliography:
- 'references.bib'
title: Probabilistic programs for inferring the goals of autonomous agents
---
INTRODUCTION
============
Intelligent systems sometimes need to infer the probable goals of people, cars, and robots, based on partial observations of their motion. These problems are central to autonomous driving and driver assistance [@franke1998autonomous; @urmson2008autonomous; @aufrere2003perception], but also arise in aerial robotics, reconnaissance, and security applications [@kumar2012opportunities; @liao2006location; @tran2008event]. In these settings, knowledge of the beliefs and goals of an agent makes it possible to infer their probable future actions.
Because the mental state of another agent is inherently unobservable and uncertain, it is natural to take a Bayesian approach to inferring it. Probabilistic models can be used to describe how an agent’s latent high-level goals and beliefs about the environment interact to yield its probable actions. Most existing work along these lines has focused on modeling goal-directed behavior using Markov decision processes and related approaches from stochastic control [@baker2007goal; @ziebart2009planning]. While promising, these approaches involve significant task-specific engineering. They also calculate policies that prescribe actions for every possible state of the world, sometimes in the inner loop of an inference algorithm. This leads to fundamental scaling challenges, even for simple environments and goal priors.
This paper introduces a class of probabilistic programs that formulate goal inference problems as approximate inference in generative models of goal-directed behavior. The proposed approach reflects three contributions: First, agents are assumed to follow paths generated by fast randomized path planning code that can incorporate heuristics drawn from video game engines and robotics. This can scale to larger environments than approaches based on optimal control. Second, hierarchical models for goals and paths are represented as probabilistic programs. This allows one to formulate a broad class of single- and multi-agent problems with common modeling and inference machinery. Ordinary probabilistic programming constructs can handle complex maps, hierarchical goal priors, and partially observed environments. Third, this paper proposes an approach to real-time approximate inference, using neural networks to learn proposals for the internal choices made by any path planners. Together, these contributions lead to a practical proposal for goal inference that has the potential to scale to a broad class of real-world problems and real-time applications. We demonstrate the efficacy of prototype implementations of these algorithms on three simple examples, each written in under 50 lines of probabilistic code.
{width="100.00000%"}
Note that this proposal does not require planning algorithms to be rewritten as probabilistic programs, but instead allows optimized, low-level, or legacy planning codes to be treated as black boxes. This avoids the implementation and performance cost of rewriting an existing path planner in a high-level probabilistic programming language, and exposing the thousands of random choices it might make to generic inference algorithms. One difficulty is that such optimized black-box planners may well make too many internal random choices to have tractable input-output likelihoods. This paper proposes two novel Monte Carlo techniques for these “likelihood-free” models, each extending Metropolis-Hastings: (i) a [*cascading resimulation*]{} algorithm that makes joint proposals to ensure cancellation of the unknown likelihoods, and (ii) a [*nested inference*]{} algorithm that uses estimated likelihoods derived from inference over the internal random choices of the planner. Cascading resimulation is simple to implement, but nested inference enables use of a broad class of Monte Carlo, variational, and neural network mechanisms to handle the intractable likelihoods.
MODELING GOAL-DIRECTED BEHAVIOR USING RANDOMIZED PATH PLANNERS {#sec:SP}
==============================================================
This paper defines probabilistic models of goal-directed behavior using randomized path-planning algorithms. Algorithm \[alg:planner\] describes one such planner, called . This planner can be applied to a broad class of environments with complex obstacles. The planner assumes a bounded two-dimensional space (e.g., the square $[0,1]^2$) and a world map $M$ that is a set of polygonal obstacles. The planner takes as input a start location $s \in [0, 1]^2$, a goal location $g \in
[0, 1]^2$, the map $M$, and a sequence of $T$ time points $\mathbf{t} = (t_1, \ldots, t_T)$, and returns either a sequence of locations $\mathbf{z} \in
[0, 1]^{2T}$ on a path from $s$ to $g$ at each time $t_i$, or ‘no-path-found’. The planner operates by growing a rapidly-exploring random tree (RRT) [@lavalle1998rapidly] from the start location $s$ to fill the space, searching for a clear line of sight between the tree and the goal. If a path is found, it is then refined to minimize its length using local optimization. Finally, the agent walks the path at a constant speed, producing the output locations $\mathbf{z}$. See Appendix \[sec:planner-details\] for more details.
Many variations of this planner are possible, including versions that take into account costs other than path length, and spaces encoding configurations other than geographic position (e.g., configuration spaces of an articulated robot). The planner parameters $N$ and $R$ trade off the cost of planning with the (probable) optimality of the paths (see Figure \[fig:planner\]). Figure \[fig:goal\_inference\] and Figure \[fig:fig\_real\_data\] show this planner being used as a modeling primitive in the Venture probabilistic programming platform [@mansinghka2014venture]. The planner was implemented in C and imported as a foreign modeling primitive into Venture. Venture supports likelihood-free primitives and design of custom inference strategies, including those of Section \[sec:lfbn\].
{width="100.00000%"}
$\left\{
\begin{array}{l}
\mbox{World map}\; M;\; \mbox{Start, goal}\; s,g \in [0,1]^2\\
\mbox{Time points}\; \mathbf{t} \in \mathbf{R}_{+}^{T}\\
\mbox{Refinement amount}\; N;\; \mbox{Restarts}\; R\\
\mbox{Max. \# tree nodes}\; J; \mbox{Min. \# tree nodes}\; S
\end{array}
\right.$ $V \gets \{s\}$ $a \sim \mbox{Uniform}([0,1]\times[0,1])$ $b \gets \textproc{nearest-vertex}(V, a)$ $\epsilon \sim \mbox{Uniform}([0, 1])$ $c \gets \epsilon a + (1-\epsilon) b$ $V.\textproc{add-edge}(b \to c)$ $V.\textproc{add-edge}(c \to g)$ $\mathbf{return}\; \textproc{path-in-tree}(V, s, g)$ $\mathbf{return}\;$ ‘no-path-found’ $\mathbf{p}^{(r)} \sim \textproc{rrt}(M, s, g)$ $\mathbf{p}^{(r)} \gets \textproc{simplify-path}(\mathbf{p}^{(r)})$ $\mathbf{p}^{(r)} \sim \textproc{refine-path}(M, s, g, \mathbf{p}^{(r)})$ $d^{(r)} \gets \textproc{path-length}(\mathbf{p}^{(r)}, s, g)$ $r^* \gets \textproc{argmin}(\mathbf{d})$ $\mathbf{return}\; \mathbf{p}^{(r^*)}$ $\mathbf{p} \sim \textproc{plan-path}(M, s, g; R, N)$ $\mathbf{z} \gets \textproc{walk-path}(\mathbf{p}, \mathbf{t})$ $\mathbf{return}\; \mathbf{z}$
INFERENCE IN PROBABILISTIC PROGRAMS WITH LIKELIHOOD-FREE PRIMITIVES {#sec:lfbn}
===================================================================
The path planner of Algorithm \[alg:planner\] can be used in a probabilistic program either by implementing the planner in a probabilistic programming language, or by treating the planner as a primitive random choice. We treat the planner as a random choice, as this allows use of an optimized C implementation of the planner. However, probabilistic programming languages such as Church, Stan, BLOG, and Figaro all require random choices to have tractable marginal likelihoods [@goodman2012church; @carpenter2016stan; @milch20071; @pfeffer2009figaro]. Computing the marginal likelihood of for outputs $\mathbf{z}$ and inputs $M$, $s$, $g$, and $\mathbf{t}$ would involve an intractable intregral over the (thousands of) internal random choices made in .
This section introduces two Monte Carlo strategies for inference in probabilistic programs that include random choices with intractable marginal likelihoods, referred to as “likelihood-free” primitives. The first strategy, shown in Algorithm \[alg:cascading-mh\], is called [*Cascading Resimulation Metropolis-Hastings*]{}; it makes block proposals to likelihood-free random choices, exploiting cancellation of the unknown likelihoods. The second, shown in Algorithm \[alg:nested-inference-mh\], is called [*Nested Inference Metropolis-Hastings*]{}; it uses Monte Carlo estimates of the unknown likelihoods in place of the likelihoods themselves. Although simple techniques like likelihood-weighting can also be used in the presence of likelihood-free primitives, they tend to work well only when a global proposal that is well-matched to the posterior is available. The algorithms we introduce do not have this limitation.
We first introduce notation. Let $\mathcal{T}$ be the set of primitive random choices available to a probabilistic program (e.g. $\{\textproc{flip}, \textproc{uniform\_continuous}, \textproc{agent-path}\}$). For $t \in \mathcal{T}$, let $\mathcal{X}_{t}$ denote the set of valid arguments for the primitive, let $\mathcal{Z}_t$ denote the set of possible outputs, and let $p_{t}(z; x)$ denote the marginal likelihood of output $z \in \mathcal{Z}_t$ given arguments $x \in \mathcal{X}_t$, where $\int p_t(z; x) dz = 1$ for all $x \in \mathcal{X}_t$. We do not require evaluation of $p_{t}(z; x)$ to be computationally tractable.
Following @wingate2011lightweight, for a probabilistic program ${\mathcal{P}}$, we assume there is a name $i \in \mathcal{I}$ assigned to every possible random choice, for some countable $\mathcal{I}$. We assume that distinct random choices are assigned unique names within every execution of ${\mathcal{P}}$. The set of names used in an execution is some finite set $I \subseteq \mathcal{I}$. We require that all random choices with name $i$ are of the same type $t_i \in \mathcal{T}$. Each unique completed execution of ${\mathcal{P}}$ can therefore be represented as the finite set of names $I \subseteq \mathcal{I}$ of those random choices made in the execution, together with the result values. We denote these results $z \in \bigtimes_{i \in I} \mathcal{Z}_{t_i}$, and denote this complete package $\rho = (I, z)$. The tuple $\rho$ is called an *execution trace* of the probabilistic program ${\mathcal{P}}$.
This paper focuses on probabilistic programs where $I$ is the same for all executions — that is, the set of random choices made is not affected by any of those choices. Relaxations of this are left for future work; more general formalizations of probabilistic programs can be found in [@wingate2011lightweight; @mansinghka2014venture].
We consider random choice $j$ to *depend on* random choice $i$ if changing the result $z_i$ of $i$ can lead to a change in the inputs $x_j$ of $j$, even if all other results $z_{I\setminus\{i\}}$ are held fixed. We assume that it is possible to construct a directed acyclic dependency graph $G = (I, E)$ among random choices $I$, where an edge $(i, j) \in E
\subset I \times I$ exists if and only if random choice $j$ depends on random choice $i$ in the above sense. The parents of a random choice $j$ are denoted by $\pi_G(j) := \{i \in I: (i, j) \in E\}$. The arguments $x_j$ of each random choice $j$ are then a (deterministic) function $f_j$ of the results of random choices in $\pi_G(j)$, which are denoted $z_{\pi_G(j)}$; we write $x_j =
f_j(\pi_G(j))$. Let $c_G(i) := \{ j \in I : (i, j) \in E\}$ denote the ‘children’ of choice $i$. Also, let $F \subseteq I$ denote the random choices with intractable likelihoods (the “likelihood-free” choices). Let $C \subseteq (I \setminus F)$ denote the random choices that are constrained based on data, which must have tractable likelihoods. Let $z_C \in \bigtimes_{i \in C}
\mathcal{Z}_i$ denote the values we are constraining those random choices to. The joint probability density of an execution trace $\rho = (I, z)$ is: $$p(z) := \prod_{i \in I} p_{t_i}(z_i; f_i(z_{\pi_G(i)}))$$ where we have omitted the dependence on $I$ because it is the same for all executions.
CASCADING RESIMULATION METROPOLIS-HASTINGS
------------------------------------------
How can a probabilistic program cope with complex, likelihood-free primitives? Our core insight is that if the proposal distribution $m(z'_i;\cdot)$ for a random choice $z_i$ is equal to the prior $p_{t_i}(z_i'; f_i(z_{\pi_G(i)}))$, then the likelihoods will cancel in a Metropolis-Hastings (MH) acceptance ratio and therefore do not need to be explicitly computed. Sampling from the prior is achieved simply by simulating the random choice. A (prototypical) acceptance ratio looks like this: $$\arraycolsep=0.1pt
\begin{array}{ccccccccc}
\alpha = \;&\frac{ p_{t_i}(z'_i; \cdot) }{p_{t_i}(z_i; \cdot)} & \cdot & \frac{m(z_i;\cdot)}{m(z'_i;\cdot)} & \;=\; & \;\frac{ p_{t_i}(z'_i; \cdot) }{p_{t_i}(z_i; \cdot)}&\cdot&
\frac{ p_{t_i}(z_i; \cdot) }{p_{t_i}(z'_i; \cdot)}&
\;= 1\\
& \mbox{{\small target }} && \mbox{{\small proposal }} && \mbox{{\small target }} && \mbox{{\small proposal }} &
\end{array}$$
We use blocked proposals in which a change to a likelihood-free choice is proposed from the prior whenever a proposal is made to any of its parents. A likelihood-free choice that is proposed may itself have likelihood-free choices as children, in which case these children are also proposed, generating a cascade of proposals. Algorithm \[alg:cascading-mh\] shows the Cascading Resimulation MH transition operator, which extends an initial custom proposal $m(z_i'; z)$ to random choice $i$ (which must not be likelihood-free) to also include any likelihood-free random choices $H$ in the cascade, such that the intractable likelihoods cancel.
$\left\{\begin{array}{l}
\mbox{Prob. program with dep. graph}\; G = (I,E)\\
\mbox{Likelihood-free random choices}\; F \subseteq I\\
\mbox{Proposed-to random-choice}\; i \in (I \setminus F)\\
\mbox{Custom proposal density}\; m(z_i'; z)\; \mbox{for choice $i$}\\
\mbox{Previous values $z$ for all random choices}\\
\end{array} \right.$ $z_{i}' \sim m(\cdot; z)$ $z_{I \setminus \{i\}}' \gets z_{I \setminus \{i\}}$ $\ell \gets 1$ $\ell' \gets 1$ $B \gets \{i\} \cup c_G(i)$ $H \gets \{\}$ $A \gets \{i\}$ $j \gets \textproc{pop}(B)$ $z_j' \sim p_{t_j}(\cdot; f_j(z_{\pi_G(j)}'))$ $\textproc{insert}(B, c_G(j))$ $H \gets H \cup \{j\}$ $\ell \gets \ell \cdot p_{t_j}(z_j; f_j(z_{\pi_G(j)}))$ $\ell' \gets \ell' \cdot p_{t_j}(z_j'; f_j(z_{\pi_G(j)}'))$ $A \gets A \cup \{j\}$ $\alpha \gets (\ell' / \ell) \cdot (m(z_i; z') / m(z_i'; z))$ $s \sim \mbox{Uniform}(0, 1)$ $z \gets z'$
Algorithm \[alg:cascading-mh\] is a Metropolis-Hastings transition over the random choices $\{i\} \cup H$ with target density equal to the local posterior $p\left(z_{\{i\} \cup H} | z_{I \setminus \{\{i\} \cup H\}}\right)$, and with proposal density: $$m(z_i'; z) \prod_{j \in H} p_{t_j}(z_j'; f_j(z_{\pi_G(j)}'))$$ The Metropolis-Hastings acceptance ratio is: $$\begin{aligned}
\alpha &= \left( \frac{\prod_{j \in H \cup A} p_{t_j}(z_j'; f_j(z_{\pi_G(j)}'))}
{\prod_{j \in H \cup A} p_{t_j}(z_j; f_j(z_{\pi_G(j)}))}\right. \notag \\
&\qquad\cdot
\left.\frac{m(z_i; z') \prod_{j \in H} p_{t_j}(z_j; f_j(z_{\pi_G(j)}))}
{m(z_i'; z) \prod_{j \in H} p_{t_j}(z_j'; f_j(z_{\pi_G(j)}'))}\right) \notag \\
&= \frac{m(z_i; z') \prod_{j \in A} p_{t_j}(z_j'; f_j(z_{\pi_G(j)}'))}
{m(z_i'; z) \prod_{j \in A} p_{t_j}(z_j; f_j(z_{\pi_G(j)}))}\end{aligned}$$ We illustrate Cascading Resimulation MH in Figure \[fig:goal\_inference\], on the task of inferring the goal of a simulated drone in an observed environment.
NESTED INFERENCE METROPOLIS-HASTINGS {#sec:nested-inference}
------------------------------------
In some problems, Cascading Resimulation MH will generate many expensive simulations of likelihood-free choices, most of which will be rejected. For these problems, and for real-time applications, we propose an alternative Metropolis-Hastings algorithm, called Nested Inference MH, that uses Monte Carlo estimates of the intractable likelihoods in the acceptance ratio. The likelihood estimates are obtained using auxiliary “nested inference” algorithms, which sample probable values for the internal random choices made by a likelihood-free choice (e.g. a randomized planning algorithm) given its inputs and outputs, and calculate a weight that can be used to form an importance sampling estimate of the unknown likelihood.
Nested Inference MH is based on an interpretation of likelihood-free random choices like as probabilistic programs in their own right. Let $u \in \mathcal{U}_t$ be an execution trace of a likelihood-free random choice of type $t$. We denote the joint density on execution traces $u$ and return values $z$ of the random choice, given input arguments $x$, by $p_t(u, z; x)$. The marginal likelihood of the random choice is given by the (intractable) integral $p_t(z; x) = \int p_t(u, z; x) du$. We denote the conditional trace density for arguments $x$ and output $z$ by $p_t(u | z; x) := p_t(u, z; x)/p_t(z;x)$.
Nested inference assumes the existence of a [*nested inference algorithm*]{} that samples execution traces $u$ according to some density $q_t(u; x, z)$ that approximates the conditional density on traces of the likelihood-free choice, i.e., $q_t(u; x, z) \approx p_t(u | z; x)$. We require that $q_t(u; x, z) > 0$ for all $u$ where $p_t(u | z; x) > 0$. Using the nested inference algorithm as an importance sampler, we produce an unbiased importance sampling estimate $\hat{p}_t(z;x)$ of the random choice’s intractable likelihood for arguments $x$ and output $z$ by sampling $K$ times $u_k \sim q_t(\cdot; x, z)$ from the inference algorithm, as follows: $$\hat{p}_t(z;x) := \frac{1}{K} \sum_{k=1}^K \frac{p_t(u_k, z; x)}{q_t(u_k; x, z)}\; \mbox{for} \;u_k \sim q_t(\cdot; x, z).$$
Nested inference also assumes that the ratio $p_t(u, z; x) / q_t(u; x, z)$ can be evaluated. While in principle the nested inference algorithm can be produced by recoding the likelihood-free primitive in a high-level probabilistic programming language, this is by no means required, nor do we expect it to be the common case. In this paper, we focus on nested inference algorithms that use learned neural networks.
The accuracy of the likelihood estimate is determined by the accuracy of the nested inference algorithm. Specifically, for $K = 1$ the variance of the estimate is: $$\begin{aligned}
&\mbox{Var}_{u \sim q(\cdot; x, z)} \left[ \frac{p_t(u, z; x)}{q_t(u;x,z)} \right] \notag \\
&\qquad\propto \mbox{D}_{\chi^2}\left( p_t(u | z; x) || q_t(u; x, z) \right),\end{aligned}$$ where $\mbox{D}_{\chi^2}$ denotes the chi-square divergence [@nielsen2014chi], and where $p_t(u|z;x)$ and $q_t(u;x,z)$ on the right-hand side represent density functions over $u$, not specific density values. Similarly, we can view $\log (p_t(u, z; x) / q_t(u; x, z))$ for $u \sim q(\cdot; x, z)$ as a (biased) estimator of $\log p_t(z;x)$, where the bias is: $$\begin{aligned}
&\mbox{E}_{u \sim q_t(\cdot; x, z)} \left[ \log \frac{p_t(u, z; x)}{q_t(u;x,z)} \right] - \log p_t(z;x) \notag \\
&\qquad= -\mbox{D}_{KL}( q_t(u; x, z) || p_t(u | z; x) ),\end{aligned}$$ where $\mbox{D}_{KL}$ denotes the Kullback-Leibler (KL) divergence [@kullback1951information].
### Nested Inference Metropolis-Hastings {#nested-inference-metropolis-hastings}
Algorithm \[alg:nested-inference-mh\] describes a Nested Inference MH transition in which a custom proposal is made to a likelihood-free random choice $i$ that uses estimated likelihoods produced using a nested inference algorithm. It assumes that all children of $i$ also have nested inference algorithms themselves. Heterogeneous configurations are also possible.
$\left\{\begin{array}{l}
\mbox{Prob. program with dep. graph}\; G = (I,E)\\
\mbox{Proposed-to random choice}\; i\\
\mbox{Custom proposal density}\; m(z_i'; z)\\
\mbox{Previous values $z$ for all random choices}\\
\mbox{Previous likelihood estimates $\ell$ for all choices}
\end{array} \right.$ $z_{i}' \sim m(\cdot; z)$ $z_{I \setminus \{i\}}' \gets z_{I \setminus \{i\}}$ $u_{i,k}' \sim q_{t_i}(\cdot; x_i, z_i')$ $\ell_i' \gets \frac{1}{K} \sum_{k=1}^K \frac{p_{t_i}(u_{i,k}', z_i'; x_i)}{q_{t_i}(u_{i,k}'; x_i, z_i')}$ $u_{j,k}' \sim q_{t_j}(\cdot; x_j', z_j)$ $\ell_j' \gets \frac{1}{K} \sum_{k=1}^K \frac{p_{t_j}(u_{j,k}', z_j; x_j')}{q_{t_j}(u_{j,k}'; x_j', z_j)}$ $\alpha \gets \left( \prod_{j \in \{i\} \cup c_G(i)} \frac{\ell_j'}{\ell_j}\right) \cdot \left( \frac{m(z_i; z')}{m(z_{i}'; z)} \right)$ $s \sim \mbox{Uniform}(0, 1)$ $z_i \gets z_i'$ $\ell_j \gets \ell_j'$
Although this transition uses Monte Carlo estimates of likelihoods in the acceptance ratio, it is a standard Metropolis-Hastings transition on an extended state space that includes the result $z_i$ of the proposed-to random choice $i$, $K$ traces $u_{i,k}$ of the proposed-to random choice, and $K$ traces $u_{j,k}$ of each child $j$ of the proposed-to random choice. The target density on the extended space is: $$\resizebox{\hsize}{!}{
$\displaystyle p(z_i | z_{I \setminus \{i\}}) \prod_{j \in \{i\} \cup c_G(i)} \frac{1}{K} \sum_{k=1}^K p_{t_j}(u_{j,k} | z_j; x_j) \prod_{\substack{r=1\\r \ne k}}^K q_{t_j}(u_{j,r}; x_j,z_j)$
}$$ The proposal density on the extended space is: $$m(z_i'; z) \prod_{j \in \{i\} \cup c_G(i)} \prod_{k=1}^K q_{t_j}(u_{j,k}'; x_j, z_j)$$ The values $z_j$ of other random choices $j \not \in \{i\} \cup c_G(i)$ are constant. See Appendix \[sec:nested-inference-derivations\] for derivation. The marginal density of $z_i$ in the extended target density is the local posterior $p(z_i | z_{I \setminus \{i\}})$ for the result of random choice $i$ given the values of all other random choices. Single-site Nested Inference MH transitions that propose to different random choices $i$ but use the same database of nested-inference likelihood estimates $\ell$ can be composed to form Markov chains that converge to the posterior $p(z_{I \setminus C} | z_C)$.
Our use of unbiased likelihood estimates in place of the true likelihoods when computing the Metropolis-Hastings acceptance ratio in Algorithm \[alg:nested-inference-mh\] is closely related to pseudo-marginal MCMC [@andrieu2009pseudo] and particle MCMC [@andrieu2010particle]. Indeed, each single-site Nested Inference MH transition can be seen as a compositional variant of a ‘grouped independence MH’ transition [@beaumont2003estimation] in which several pseudo-marginal likelihoods (one for each random choice $j \in \{i\} \cup c_G(i)$) are used in the same update. The database of nested-inference likelihood estimates $\ell$ stores the ‘recycled’ pseudo-marginal likelihood estimates from previous transitions.
The convergence rate of a Markov chain based on Nested Inference MH transition operators depends on the accuracy of the nested inference algorithm and $K$. In the limit of exact nested inference algorithm ($q_t(u;x,z) = p_t(u|z;x)$) the likelihood estimates are exact, and the algorithm is identical to standard Metropolis-Hastings. If the nested inference algorithm is very inaccurate, it may routinely propose traces $u$ that are incompatible with the output $z$ of the random choice, resulting in low acceptance rates. Better characterizing how the convergence rate depends on the accuracy of the nested inference algorithms and on $K$ is an important area for future work.
### Learning a nested inference algorithm {#sec:learning}
It is possible to learn a nested inference algorithm $q_t(u; x, z)$ that approximates $p_t(u|z;x)$. The idea of learned inference for probabilistic generative models goes back at least to @morris2001recognition and has also been used in @stuhlmuller2013learning and @kingma2013auto. We apply this idea to nested inference as follows. Let $q_{t,\theta}(u;x,z)$ denote a nested inference algorithm that is parameterized by $\theta$ — for example, $\theta$ might be the weights of a neural network used as part of the inference algorithm. We establish a training distribution $d_t(x)$ over the arguments to the primitive $t$, and approximately solve the following optimization problem: $$\begin{aligned}
\min_{\theta} \left\{ \mbox{E}_{\begin{subarray}{l}x \sim d_t(\cdot)\\z|x \sim p_t(\cdot; x)\end{subarray}} \left[ \mbox{D}_{KL}(p_t(u | z; x) || q_{t, \theta}(u; x, z)) \right]
\right\}\\
= \min_{\theta} \left\{\mbox{E}_{\begin{subarray}{l}x \sim d_t(\cdot)\\z|x \sim p_t(\cdot; x)\\u|x,z \sim p_t(\cdot | z; x)\end{subarray}} \left[ \log \frac{p_t(u | z; x)}{q_{t, \theta}(u; x, z)} \right]
\right\}\end{aligned}$$ The goal is for $q_{t,\theta}(u;x,z)$ to approximate $p_t(u|z;x)$ well (i.e., have small KL divergence) for typical input arguments $x \sim d_t(\cdot)$. We approximate this objective function by drawing $M$ independent sets of input arguments $x^{(i)}$ from the training distribution, and running a traced execution of the likelihood-free random choice (e.g. planner) on each set of arguments, recording[^1] the trace $u^{(i)}$ and output $z^{(i)}$: $$\begin{array}{rl}
x^{(i)} \sim d_t(\cdot)& \mbox{{Sample planner arguments}}\vspace{2mm}\\
u^{(i)}, z^{(i)} \sim p_t(\cdot, \cdot; x^{(i)})& \hspace{-2mm} \begin{array}{l}
\mbox{{Run likelihood-free planner,}}\\
\mbox{{record trace $u^{(i)}$, output $z^{(i)}$}}
\end{array}
\end{array}$$ We use the resulting dataset $D = \{(x^{(i)}, z^{(i)}, u^{(i)}) : i = 1\dots M\}$ to define an approximate objective function $J_D(\theta)$ that is an unbiased estimate of the original objective function: $$\begin{aligned}
J_D(\theta) &:= \frac{1}{M} \sum_{i=1}^M \log \frac{p_t(u^{(i)} | z^{(i)}, x^{(i)})}{q_{t,\theta}(u^{(i)}; x^{(i)}, z^{(i)})}\\
&= C - \frac{1}{M} \sum_{i=1}^M \log q_{t,\theta}(u^{(i)}; x^{(i)}, z^{(i)})\end{aligned}$$ where $C$ does not depend on $\theta$. Note that minimizing $J_D(\theta)$ over $\theta$ is equivalent to maximizing the log-likelihood of the data $D$. Because we use forward simulations to produce $u^{(i)}, z^{(i)}$ jointly from $p_t(\cdot, \cdot; x^{(i)})$, we have one exact conditional sample $u^{(i)} | z^{(i)} \sim p_t(\cdot | z^{(i)}; x^{(i)})$ for each training example.
EXAMPLE APPLICATIONS {#sec:applied}
====================
We have implemented four example applications, designed to illustrate the flexibility of our framework:
1. Inferring the probable goal of a simulated drone. This example shows that small changes to the environment, such as including an additional doorway, can yield large changes in the inferred goals.
2. Inferring the probable goal of a simulated drone with a more complex planner. Specifically, we model the drone as following a multi-part path produced by a planner that first chooses a waypoint uniformly at random and then recursively solves the two path planning problems induced by the choice of waypoint. This example shows (a) applicability of the framework to more complex models of goal-directed behavior, and (b) that Nested Inference MH with a learned neural network can outperform Cascading Resimulation MH.
3. Inferring whether or not two people walking around tables in a room are headed for the same goal or different goals. This example demonstrates applicability to simple hierarchical models for goals and also demonstrates applicability to real-world (as opposed to synthetic) data.
4. Jointly inferring a simulated agent’s goals and its beliefs about an obstacle in the map whose location, size, and orientation is unknown to the probabilistic program. This example is described in the appendix due to space constraints.
{width="1.00\linewidth"}
EXAMPLE 1: SENSITIVITY OF GOAL INFERENCE TO SMALL MAP CHANGES
-------------------------------------------------------------
Figure \[fig:goal\_inference\] shows a comparison of goal inference in two different maps given the same observations. The map for the scenario on the left has an enclosure with two openings, one on the top and one on the bottom, while the map for the scenario on the right has a single opening. In the map on the left, the inferred goal samples fall outside the enclosure, because if the drone intended to go inside the enclosure, it could have taken a much shorter path. In the map on the right, a significant fraction of goal samples fall inside the enclosure, as relatively efficient paths into the enclosure go through the partial trajectory that has been observed so far. Samples shown are the final states of 480 independent replicates of a Markov chain initialized from the prior, with $1000$ Cascading Resimulation MH transitions (Algorithm \[alg:cascading-mh\]) using the prior as the proposal. Planner parameters are $R = 10$, $N = 1000$, $\epsilon = 0.01$, $v = 0.5$, $J = 10000$, $S = 2000$.
EXAMPLE 2: HANDLING PATH PLANNERS WITH WAYPOINTS VIA NESTED INFERENCE
---------------------------------------------------------------------
Next, we used a model where the agent may choose a waypoint and separately plan a path to the waypoint and a path from the waypoint to the goal (, Algorithm \[alg:agent-waypoint-path\]). Unlike the simpler $\textproc{agent-path}$ model, which typically samples from a small number of modes concentrated at efficient routes from the start to the goal, yields paths that are unpredictable without knowledge of the waypoint. Parameters $R$ and $N$ of are omitted for simplicity. We consider the same goal inference task as in Example 1 but with the alternative planner. Cascading Resimulation MH performs poorly on this task, because the prior is a poor proposal for the internal random choices of .
$\left\{
\begin{array}{l}
\mbox{World map}\; M;\;\;\mbox{Start, goal}\; s,g \in [0,1]^2\\
\mbox{Time points}\; \mathbf{t} \in \mathbf{R}^{T}_{+}
\end{array}
\right.$ $g' \sim \mbox{Uniform}([0,1] \times [0,1])$ $w \sim \mbox{Bernoulli}(0.5)$ $\mathbf{p}_1 \sim \textproc{plan-path}(M, s, g')$ $\mathbf{p}_2 \sim \textproc{plan-path}(M, g', g)$ $\mathbf{p} = (\mathbf{p}_1, \mathbf{p}_2)$ $\mathbf{p} \sim \textproc{plan-path}(M, s, g)$ $\mathbf{\tilde{z}} \gets \textproc{walk-path}(\mathbf{p}, \mathbf{t})$ $\mathbf{z} \sim \textproc{add-noise}(\mathbf{\tilde{z}})$ $\mathbf{return}\; \mathbf{z}$
Algorithm \[alg:agent-waypoint-path-nested-inference\] shows a nested inference algorithm for that uses a neural network to propose the waypoint ($g'$) and whether the waypoint is used ($w$), given the goal and observations, and then executes the rest of the planner, conditioned on $w$ and $g'$. The network was trained on 10,000 runs of with random goal input $g \sim \mbox{Uniform}([0, 1]^2)$ and fixed world map $M$ and start $s$. The nested inference algorithm splits the trace $u$ of into $u_1 = (w, g')$ and $u_2$ (the random choices made within executions of ), so that $u = (u_1, u_2)$. The density of the nested inference algorithm is then $q_t(u; x, z) = q_{\theta}(u_1; x, z) p_t(u_2; x)$, and the density ratio $p_t(u, x, z) / q_t(u; x, z)$, which is used by Nested Inference MH when estimating the planner likelihoods, simplifies to $p_t(z | u; x) / q_{\theta}(u_1; x, z)$. To evaluate this ratio, we separately evaluate the density $p_t(z | u;, x)$ of and the density $q_{\theta}(u_1; x, z)$ of the neural network’s stochastic outputs.
$\left\{
\begin{array}{l}
\mbox{Arguments to planner}\; \mathbf{x} = (M, s, g, \mathbf{t})\\
\mbox{Hypothetical output of planner}\; \mathbf{z}
\end{array}
\right.$ $w, g' \sim q_{\theta}(\cdot; \mathbf{x}, \mathbf{z})$ $\mathbf{p}_1 \sim \textproc{plan-path}(M, s, g')$ $\mathbf{p}_2 \sim \textproc{plan-path}(M, g', g)$ $\mathbf{return}\; (\mathbf{p}_1, \mathbf{p}_2)$ $\mathbf{return}\; \textproc{plan-path}(M, s, g)$
We compared three strategies for goal inference: Nested Inference MH using Algorithm \[alg:agent-waypoint-path-nested-inference\] and $K=1$, Cascading Resimulation MH, and Nested Inference MH using a “resimulation” nested inference algorithm $(q_t(u;x,z) = p_t(u;x))$ and $K=2,10$. Figure \[fig:nested-combined\] shows that neural Nested Inference MH converges faster than the other strategies. Planner parameters were the same as in Example 1. All inference strategies were implemented using a custom Python inference library. Integration of Nested Inference MH with Venture is left for future work.
EXAMPLE 3: MODELING REAL-WORLD HUMAN MOTION {#sec:real-world-human-motion}
-------------------------------------------
The Venture program of Figure \[fig:fig\_real\_data\](c) defines a model with two agents whose destinations may or may not be the same. The environment ([`world`]{}) and the start locations of the agents are known. The [`is_common_goal`]{} flag determines whether the agents share the same goal destination. The paths of both agents are modeled using . The corresponding Bayesian network is shown in Figure \[fig:fig\_real\_data\](e). We collected video of two collaborators walking in a scene containing tables, for two conditions—one in which the they meet at a common location, and one where they diverge. For the common-goal condition we constructed short and extended sequences of observed locations (Figure \[fig:fig\_real\_data\](a) and (b)). We used Cascading Resimulation MH for inference, initialized from the prior, with a joint prior proposal over all latent variables. We ran 60 chains of $200$ transitions each, and rendered the final states in Figures \[fig:fig\_real\_data\](a-b). The speed for each individual was set to their average speed along the observed path. The estimated probabilities of [`is_common_goal=True`]{} for the short and extended sequences are 0.63 and 0.82 respectively. This trend qualitatively matches human judgments, shown in Figure \[fig:fig\_real\_data\](d) (the model was not calibrated to match human judgments). See Appendix \[sec:additional\_experiments\] for additional results.
{width="96.00000%"}
DISCUSSION
==========
This paper introduced a class of probabilistic programs for formulating goal inference as approximate inference in probabilistic generative models of goal-directed behavior. The technical contributions are (i) a probabilistic programming formulation that makes complex goal and map priors easy to specify; (ii) the use of randomized path planning algorithms as the backbone of generative models; and (iii) the introduction of Monte Carlo techniques that can handle the intractable likelihoods of these path planners. The experiments showed that it is possible for short probabilistic programs to make meaningful inferences about goal-directed behavior.
From the standpoint of robotics, autonomous driving, or reconnaissance, the examples in this paper are quite preliminary. More experiments are needed to explore the accuracy of approximate inference in these models, as well as the accuracy of the models themselves, especially on real-world problems. The probabilistic programming formulation makes it easy to explore variations of models, environments, and inference strategies.
The problem of inferring the mental states of autonomous agents is central to probabilistic artificial intelligence. It may also be a natural application for structured generative models and for probabilistic programming, but only if sufficiently fast and flexible inference schemes can be developed. We hope this paper helps to encourage the use of probabilistic programming for building intelligent software that can draw meaningful inferences about goal-directed behavior.
### Acknowledgements {#acknowledgements .unnumbered}
The authors would like to thank Feras Saad for obtaining human judgment data, and Leslie Kaelbling, Erin Bartuska, and Feras Saad for helpful conversations. Tree and car 3D models in figures are from <http://www.f-lohmueller.de/> [@models3d]. This research was supported by DARPA (PPAML program, contract number FA8750-14-2-0004), IARPA (under research contract 2015-15061000003), the Office of Naval Research (under research contract N000141310333), the Army Research Office (under agreement number W911NF-13-1-0212), and gifts from Analog Devices and Google. MCT is supported by the Department of Defense (DoD) through the National Defense Science & Engineering Graduate Fellowship (NDSEG) Program.
### References {#references .unnumbered}
PLANNER DETAILS {#sec:planner-details}
===============
We now describe details of the planner omitted from the main text, including the procedures , , and , which are defined in Algorithm \[alg:planner-details\]. Paths $\mathbf{p}$ are represented as sequences of points, with lines connecting the points. The path $\mathbf{p}$ begins with start $s$ and ends with goal $g$. To be a valid path with respect to map $M$, no point in the path may lie within an obstacle (polygon) of $M$ (i.e. $M.\textproc{is-valid}(\mathbf{p}_i)$), and no line between two adjacent path points may intersect an obstacle of $M$ (i.e. $M.\textproc{clear-line}(\mathbf{p}_i, \mathbf{p}_{i+1})$).
$\left\{
\begin{array}{l}
\mbox{World map}\; M;\; \mbox{Start, goal}\; s,g \in [0,1]^2\\
\mbox{Time points}\; \mathbf{t} \in \mathbf{R}_{+}^{T}\\
\mbox{Refinement amount}\; N;\; \mbox{Restarts}\; R\\
\mbox{Agent speed}\; v \in \mathbf{R}_{+}
\end{array}
\right.$ $\mathbf{p}'_1 \gets s$ $j \gets 2$ $\mathbf{p}'_j \gets \mathbf{p}_i$ $j \gets j + 1$ $\mathbf{pass}$ $\mathbf{p}'_j \gets g$ $\mathbf{return}\; \mathbf{p}'$ $d \gets \textproc{path-length}(\mathbf{p}, s, g)$ $\epsilon \sim \mathcal{N}(0,\sigma^2)$ ${\mathbf{p}}' \gets \mathbf{p} + \epsilon \cdot \mathbf{e}_l$ $d' \gets \textproc{path-length}({\mathbf{p}}', s, g)$ $(d, \mathbf{p}) \gets ({\mathbf{p}}',d')$ $\mathbf{return}\; \mathbf{p}$ $d \gets 0.0$ $d^* \gets t v$ $\delta \gets ||\mathbf{p}_{j} - \mathbf{p}_{j+1}||_2$ $e \gets d^* - d$ $\mathbf{return}\; \frac{\delta - e}{\delta} \mathbf{p}_j + \left(1 - \frac{\delta - e}{\delta}\right) \mathbf{p}_{j+1}$ $d \gets d + \delta$ $\mathbf{return}\; g$
$\mathbf{z}_i \gets \textproc{walk-to}(\mathbf{p}, t_i, v)$ $\mathbf{return}\; \mathbf{z}$
ADDITIONAL EXPERIMENTS {#sec:additional_experiments}
======================
JOINTLY INFERRING THE BELIEF AND GOAL OF AN AGENT
-------------------------------------------------
The Venture program of Figure \[fig:joint-map-goal-inference\](a) defines a model in which the belief of an agent about its environment, upon which the agent’s motion plan depends, is uncertain. The environment contains two, static objects (`known_objects`): a tree and a central divider wall that divides the $[0,1]\times[0,1]$ square into a left and right side. There are passageways between the left and right side that go above and below the divider. However, the agent has knowledge of (or belief in) an additional obstacle wall (`obstacle`), and the agent plans their path to the destination (`goal`) taking this additional obstacle into account. Figure \[fig:joint-map-goal-inference\](a) also shows a Bayesian network representation of this model. We seek to infer both the agent’s goal and the agent’s beliefs about the location, orientation, and size of the obstacle.
We used Cascading Resimulation Metropolis-Hastings (Algorithm \[alg:cascading-mh\]) with a single repeated transition operator based on an independent joint proposal to `goal` ($\mbox{Uniform}([0,1]^2)$) and to the unknown parameters of `obstacle` (start post location, orientation, and length, proposed from the prior). We initialized from the prior. Parameters of the planner were $R = 10$, $N = 1000$, $\epsilon = 0.01$, $v =0.5$, $J = 10000$, $S = 2000$. We ran several independent Markov chains of $1000$ iterations each, on a synthetic dataset in which the agent takes a path from the right to the left of the map by going below the divider. The final state of four such chains are visualized in Figure \[fig:joint-map-goal-inference\](b). For this dataset, the goal destination of the agent is revealed with certainty because the agent reaches and stops in the upper left corner. The obstacle inferences indicate that agent believes the upper route to its goal is blocked, because otherwise the agent would have taken the shorter, upper route, to its goal. However, the specific details of how the obstacle blocks the upper passageway remain uncertain.
GOAL INFERENCE IN A DRIVING SCENARIO
------------------------------------
Figure \[fig:fig3\] shows an application of the multi-agent common-goal model of Figure \[fig:fig\_real\_data\] to a driving scenario. We show 60 independent replicates of 3000 iterations of Cascading Resimulation Metropolis-Hastings each. The results illustrate that this model can be used with varied environments.
REAL-WORLD HUMAN MOTION, ALTERNATE SEQUENCE
-------------------------------------------
We extended the experiment described in Section \[sec:real-world-human-motion\] and shown in Figure \[fig:fig\_real\_data\] by running Cascading Resimulation Metropolis-Hastings on an alternate sequence of observed person locations in which the individals diverge to separate individual goal destination. The inferences, shown in Figure \[fig:extra\_real\], confirm the expectations, with all samples indicating `is_common_goal = False`. Samples were obtained from the final state of $120$ independent Markov chains, with initialization from the prior, followed by $1200$ iterations of Cascading Resimulation Metropolis-Hastings.
![Inferring whether or not two people are headed to the same destination, as in Figure \[fig:fig\_real\_data\], but for a different sequence of observed locations. The final frame shows approximate posterior inference samples obtained from cascading resimulation Metropolis-Hastings. Inference gives low probability of a common goal for this sequence (there were no `is_common_goal = True` samples).[]{data-label="fig:extra_real"}](images/fig_real_data_diverging.pdf){width="48.00000%"}
{width="100.00000%"}
{width="100.00000%"}
INFERENCE WITH WAYPOINT PLANNER {#sec:waypoint-additional}
-------------------------------
Figure \[fig:fig\_waypoint\_detail\] compares waypoints and paths proposed by Nested Inference MH with a neural nested inference algorithm with waypoints and paths proposed by Cascading Resimulation MH on an illustrative example data set. The poor quality of the prior as the proposal, as used by Cascading Resimulation, results in unecessary rejections, and slow convergence. The neural network proposes waypoints near the bend in the path.
The KL divergence estimates of Figure \[fig:nested-combined\](d) were obtained by binning 960 independent reference samples (30,000 transitions of Cascading Resimulation MH, initialized from the prior) and binning 960 independent approximate inference samples for each inference algorithm evaluated. The world unit square was binned into 25 squares (5-by-5), and a discrete distribution was estimated for each sampler by counting the number of samples falling into each bin, adding a pseudocount of $0.1$ to each bin, and normalizing. The KL divergence from the resulting reference sampler histogram was computed to each resulting approximate inference algorithm histogram. For each inference strategy (Cascading Resimulation MH, Neural Nested Inference MH with $K=1$, Resimulation Nested Inference MH with $K=2$ and Resimulation Nested Inference MH with $K=10$), the number of MH transitions was varied over several orders of magnitude, and the final state in each chain was recorded, to obtain samples for each inference algorithm evaluated. The number of MH transitions used to obtain samples shown in Figures \[fig:nested-combined\](a,b,c) are 10, 1, and 10, respectively. Figure \[fig:fig\_waypoint\_additional\] shows additional samples comparing Nested Inference Metropolis-Hastings with a neural nested inference algorithm with Cascading Resimulation Metropolis-Hastings.
![ Proposed waypoints and paths produced by Nested Inference MH with a neural nested inference algorithm, and Cascading Resimulation MH, when evaluating the MH acceptance ratio for a proposed goal $g'$ in the center of the enclosure that is the ground truth goal. Because the neural nested inference algorithm generates reasonable proposed waypoints, the proposed paths have a high probability of being consistent with the observed data. Because Cascading Resimulation proposes the waypoint and path from the prior for each proposed goal $g'$, the paths proposed are unlikely to be consistent with the observations, resulting in a high MH rejection rate, even when the proposed goal $g'$ is the ground truth goal, as is the case here. []{data-label="fig:fig_waypoint_detail"}](images/fig_nested_inference_detail.pdf){width="48.00000%"}
{width="100.00000%"}
NESTED INFERENCE DERIVATIONS {#sec:nested-inference-derivations}
============================
The variance of the likelihood estimate with $K = 1$ is: $$\begin{aligned}
&\mbox{Var}_{u \sim q(\cdot; x, z)} \left[ \frac{p_t(u, z; x)}{q_t(u; x, z)} \right]\\
&\qquad= p_t(x;z)^2 \cdot \int \left( \frac{p_t(u| z; x)}{q_t(u; x, z)} - 1 \right)^2 q(u; x, z) du\\
&\qquad\propto \int \left( \frac{p_t(u| z; x)}{q_t(u; x, z)} - 1 \right)^2 q(u; x, z) du\\
&\qquad= \mbox{D}_{\chi^2}(p_t(u|z;x) || q_t(u; x, z))\end{aligned}$$ The bias of the log likelihood estimate with $K = 1$ is: $$\begin{aligned}
& \mbox{E}_{u \sim q_t(\cdot; x, z)} \left[ \log \frac{p_t(u, z; x)}{q_t(u;x,z)} \right] - \log p_t(z;x)\\
&\qquad = \mbox{E}_{u \sim q_t(\cdot; x, z)} \left[ \log \frac{p_t(u| z; x)}{q_t(u;x,z)} \right]\\
&\qquad = -\mbox{D}_{KL}( q_t(u; x, z) || p_t(u | z; x) )\end{aligned}$$ Algorithm \[alg:nested-inference-mh\] can be intuitively understood as an approximation to a single-site MH update where a single value $z_i$ is being updated with proposal $m(z_i'; z)$ and target density $p(z_i | z_{I \setminus \{i\}})$, and where estimates of likelihoods are used in place of actual likelihoods when computing the MH acceptance ratio. However, Algorithm \[alg:nested-inference-mh\] is theoretically justified by recognizing that it is a standard joint MH transition on an extended state space that consists of $z_i$ (the value of the proposed-to random choice), $u_{i,k}$ for $k=1\ldots K$ (a set of $K$ traces for the proposed-to random choice), and $u_{j,k}$ for $j \in c_G(i)$ and $k=1\ldots K$ (a set of $K$ traces for each of the children of the proposed-to random choice). The extended target density is: $$\resizebox{\hsize}{!}{
$\displaystyle p(z_i | z_{I \setminus \{i\}}) \prod_{j \in \{i\} \cup c_G(i)} \frac{1}{K} \sum_{k=1}^K p_{t_j}(u_{j,k} | z_j; x_j) \prod_{\substack{r=1\\r \ne k}}^K q_{t_j}(u_{j,r}; x_j,z_j)$
}$$ Note that the marginal target density of $z_i$ is the original target density $p(z_i | z_{I \setminus \{i\}})$, which is proportional to $p_{t_i}(z_i; x_i) \prod_{j \in c_G(i)} p_{t_j}(z_j; x_j)$. Substituting $p_{t_i}(z_i; x_i) \prod_{j \in c_G(i)} p_{t_j}(z_j; x_j)$ for $p(z_i | z_{I \setminus \{i\}})$ in the extended target density expression and simplifying gives the following unnormalized extended target density: $$\begin{aligned}
\prod_{j \in \{i\} \cup c_G(i)} \frac{1}{K} \sum_{k=1}^K p_{t_j}(u_{j,k}, z_j; x_j) \prod_{\substack{r=1\\r \ne k}}^K q_{t_j}(u_{j,r}; x_j,z_j)\end{aligned}$$ The extended proposal density is: $$m(z_i'; z) \prod_{j \in \{i\} \cup c_G(i)} \prod_{k=1}^K q_{t_j}(u_{j,k}'; x_j, z_j)$$
The ratio of the unnormalized extended target density over the extended proposal density, for proposed values $z_i'$ and $u_{j,k}'$ for all $j \in \{i\} \cup c_G(i)$ and $k=1\ldots K$, with all other $z_{I \setminus \{i\} \cup c_G(i)}' = z_{I \setminus \{i\} \cup c_G(i)}$ is: $$\begin{aligned}
\frac{1}{m(z_i';z)} \prod_{j \in \{i\}\cup c_G(i)} \frac{1}{K} \sum_{k=1}^K \frac{p_{t_j}(u_{j,k}', z_j'; x_j')}{q_{t_j}(u_{j,k}';x_j',z_j')}\end{aligned}$$ Note that each factor $\frac{1}{K} \sum_{k=1}^K \frac{p_{t_j}(u_{j,k}, z_j; x_j)}{q_{t_j}(u_{j,k};x_j,z_j)}$ within this ratio takes the form of a nested inference likelihood estimate. Composing the full MH acceptance ratio for the extended target and proposal densities gives the acceptance ratio used in Algorithm \[alg:nested-inference-mh\]. Note that Algorithm \[alg:nested-inference-mh\] samples the proposed joint state $z_i$, $u_{i,k}$ for $k=1\ldots K$, and $u_{j,k}$ for $j \in c_G(i)$ and $k=1\ldots K$ precisely according to the extended proposal density. Finally, note that the nested-inference likelihood estimates $\ell$ for accepted proposals are retained between updates in Algorithm \[alg:nested-inference-mh\]. These estimate values serve as summaries of the previous iterates for the traces $u_{i,k}$ and $u_{j,k}$. Although the transition is an MH transition on the extended space including the traces, the previous estimates $\ell$ are sufficient for evaluating the extended MH acceptance ratio and retaining the previous trace iterates themselves is not necessary.
[^1]: This training regime cannot be applied to a true black-box path planner, since a recording of its internal randomness is now necessary. However, such recordings can be produced from a straightforwardly instrumented version of the algorithm. The likelihood estimator for the planner can still be treated as a black-box by the Nested Inference MH transition.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss the possible influence of a cosmic magnetic field on the macroscopic quantum tunneling process associated, in a cosmological context, to the decay of the “false vacuum." We find a close analogy with the effects of an external magnetic field applied to a Josephson junction in the context of low-temperature/high-temperature superconducting devices.'
author:
- 'A. Barone$^{1}$, M. Gasperini$^{2,3}$ and G. Rotoli$^{4}$'
title: 'Macroscopic quantum tunneling and the “cosmic" Josephson effect'
---
Introduction {#sec1}
============
It is well known that the interest in the Josephson effect [@1] lies in a variety of phenomena of paramount importance for the deep physics involved, and for the many – already realized or still potential – applications. In the field of superconductivity it represents a powerful probe of fundamental properties (such as, for instance, the unconventional symmetry of the order parameter in important classes of superconductors). However, the Josephson effect also offers an unparalleled tool to shed light on the intriguing underlying physics of many phenomena which go far beyond the realm of superconductivity.
Many years ago the scientific community working in the field of superconductivity was greatly impressed by two papers discussing the fate of the false vacuum [@cole1], namely the stability/unstability of a physical system in a state which is not an absolute minimum of its energy density, and which is separated from the minimum by an effective potential barrier. It was shown, in those papers, that even if the state of the early Universe is too cold to activate a “thermal" transition (via thermal fluctuations) to the lowest energy (i.e. “true vacuum") state, a quantum decay from the false vacuum to the true vacuum may still be possible through a barrier penetration via macroscopic quantum tunneling (MQT).
A few years later, the occurrence of the MQT in Josephson junctions [@5] became a subject of great interest. For an excellent description of the experiments in this context, the reader is refereed to the paper by Clarke et al [@6]. The macroscopic quantum nature of the superconductive state, the role of the Josephson effect as a probe of MQT and of other quantum and macroscopic effects (such as energy level quantization and macroscopic quantum coherence), produced a great development of both theoretical analyses and experimental confirmations. Unexpected results were also obtained, such as the first observation [@9] of MQT in high-temperature Josephson junctions (see Fig. 1), together with a clear observation of the energy-level-quantization effect [@10]. More recently, the influence of an external magnetic field applied to a Josephson junction was investigated, and it was found that the external field may play an important role in the activation of the MQT effect [@11].
Very recently, the decay of the false vacuum state in a cosmological context has attracted renewed interest, especially in view of its possible relevance in the process of tunneling among the many vacuum states of the string landscape [@13]. It seems therefore appropriate to point out that – just like in the case of the MQT effect occurring in the context of a superconducting Josephson junction – an external cosmic background field may have an important influence on the activation of the cosmological MQT effect, and on the effective decay rate of the false vacuum. The aim of this paper, in particular, is to estimate the effects of a cosmic magnetic field of a primordial origin, and to discuss its possible relevance for the false-vacuum $\to$ true-vacuum transition under the assumption that such field satisfies the required energy density bounds imposed by inflation and backreaction (see e.g. [@15]).
The paper is organized as follows. In Sec. \[Sec2\] we briefly recall the effect of an external magnetic field on the Josephson tunneling structures and, more generally, on the process of macroscopic quantum tunneling. In Sec. \[Sec3\] we generalize the original computation [@cole1] of the tunneling through the potential barrier for the decay of the false vacuum, including a cosmic background field of primordial origin. Our concluding remarks are finally presented in Sec. \[Sec4\].
External magnetic field and macroscopic quantum tunneling {#Sec2}
=========================================================
In the context of the Josephson effect, it is well known that the application of a magnetic-field $B$ to a Josephson junction produces a space modulation of the relative macroscopic phase [@1]. The effect of the applied magnetic-field on the related MQT process was also a subject of deep researches, and stimulated a variety of interesting papers. Quite recently, in particular, the influence of such an external field on the so-called “crossover" temperature $T_0$ has been explicitly computed [@11], and it has been shown to be nonmonotonically dependent on the flux of the applied magnetic field $B$.
![Details of full width half maximum of escape events distribution, as function of temperature, for a biepitaxial high temperature superconductivity Josephson junction in the experiment reported in [@9]. The so-called crossover temperature $T_0$ is indicated by arrow. The plateau below $T_0$ indicate the setting of the condition of MQT (dotted line is just a guide for eyes). In the inset we show $T_0$ as a function of the flux $f(B)$ of the applied magnetic-field (see [@11] for the choice of units and the normalization of the plotted variables).[]{data-label="fig1"}](f12.pdf){width="87mm"}
Let us recall that the crossover temperature $T_0$ is the (thermal bath) temperature at which quantum tunneling becomes the dominant activation mechanism of a superconducting Josephson junction: for $T<T_0$ the escape from the potential barrier is due to the MQT effect, while for $T>T_0$ it is mainly due to the effect of thermal fluctuations. An example of this is reported in Fig. \[fig1\] where we plot the details of MQT transition in terms of the full width half maximum of escape event distributions for the YBa$_2$Cu$_3$O$_{7-x}$ (YBCO) biepitaxial Josephson junction of Ref. [@9]. The crossover temperature $T_0$, indicated by an arrow, separate the MQT regime from thermal regime. In general the function $T_0=T_0(B)$ has a complicated analytic form [@11] which depends on the junction parameters, and which tends to be periodic in the flux of the applied magnetic-field. Its general behavior is illustrated in the inset of Fig. \[fig1\].
It may be noted that, in a particular range of values of the applied field, there is no crossover and the junction remains classical until the zero temperature limit. This means that the external field may have a dramatic effect in controlling the “shape" of the effective potential barrier and the overall efficiency of the MQT process – at least in the context of superconductive Josephson structures. What is important, for the purpose of this paper, is that the magnetic field produces a spatial variation of the phase whose main outcome is an increase in the tunneling rate [@11], despite a decrease in the crossover temperature.
Coming back from the laboratory to the early universe, we may ask whether the above aspects of Josephson MQT can also be retrieved in a cosmological context (after all, the process of false vacuum decay discussed in [@cole1] is fully controlled by the tunneling probability). In particular, it would be an intriguing possibility if we could figure out an analogous behavior of the crossover temperature. However, physical analogies deserve great interest but require, in general, also great caution. In our case it is probably inappropriate to look for a cosmological effect exactly analogous to the one produced by the magnetic field on a Josephson junction, which influences so drastically the thermal vs quantum activation of the transition process.
The reason is to be found in the substantial different nature of the scalar field appearing in Coleman’s model of vacuum decay and the (scalar) phase difference between two superconductors. Of course, it would be interesting to think of a “cosmological" Josephson effect in which different regions of the Universe have a phase difference in their (Wheeler-De Witt) wave functions, but this would represent a completely different problem from the one originally discussed in [@cole1].
To find a possible cosmological analogous of the magnetic-field effects on MQT in Josephson structures we should consider, instead, the scenario of false vacuum decay under the action of a cosmic background field nonminimally coupled to the scalar field of Coleman’s model.
We should recall, to this purpose, that the scalar dilaton field $\phi$ of superstring models is necessarily coupled to various antisymmetric tensor fields of second and higher ranks, including (in Type I and Heterotic superstrings) the electromagnetic field tensor $F_{\mu\nu}$ (see e.g. [@maurizio]). Such a coupling is typically of the form F\^F\_/(2M), \[21\] where $M$ is an appropriate mass scale setting the effective coupling constant. Hence, if we are interested into MQT-type transitions among string vacua, we can identify Coleman’s scalar with such a fundamental dilaton field, and revisit the original computation of the false vacuum decay taking into account the contribution of an external magnetic-field nonminimally coupled to $\phi$ according to Eq. (\[21\]). This will be the subject of the next section.
Magnetic-field perturbation in the semiclassical Coleman approach {#Sec3}
=================================================================
Let us consider the mechanism of false vacuum decay discussed in [@cole1], based on a scalar field $\phi$ with a symmetric self-interaction potential $U_+(\phi)=U_+(-\phi)$, and two minima at $\phi= \pm a$. Let us choose, in particular, the potential $$U_{+}(\phi)=U_0\left( \phi ^{2}-a^{2}\right) ^{2}/a^4,
\label{31}$$ where $U_0$ is the height of the effective barrier separating the two minima of the unperturbed potential. Including the electromagnetic interaction (\[21\]) the unperturbed scalar system is thus described by the following Lagrangian density: = \_\^/2- U\_[+]{}() -F\^F\_/(2M) \[32\] (we are using units in which $\hbar=c=1$, and the action is dimensionless). Note that, for simplicity, we are working in the context of a flat space-time geometry. However, all subsequent computations could be easily generalised to the case of a curved cosmological geometry without changing the main results presented here.
Following [@cole1], let us now introduce a small breaking of the potential symmetry, U\_+ U= U\_+ +( -a)/(2a), \[33\] parametrised by the energy density parameter $\vep>0$ (which also represents, in the Coleman model without external electromagnetic interactions, the energy-density difference between the true and the false vacuum). The total effective potential thus becomes, in our case, V() =U\_[+]{}()+ ( -a)/(2a) +B\^[2]{}/M. \[34\] Note that we have included only the contribute of the magnetic field $B$. Indeed, in a primordial cosmological context, one can safely neglect large-scale electric fields, assuming that they are rapidly dissipated by the high conductivity of the primeval plasma [@16].
In order to obtain the exponent of the tunneling rate between the false and the true vacuum we need now the expression of the Euclidean action $S_E$ associated to the total effective potential (\[34\]). Considering, as in [@cole1], only corrections to $U_{+}(\phi)$ of the first order in $\vep$ and $B^2$ one obtains, in the “thin-wall" approximation, $$S_{E}=\pi ^{2}R^{3}S_{1}-2\pi^{2}\left[ \varepsilon
+\left(B^{2}a/{M}\right)\right] \left(R^{4}/{4}\right).
\label{35}$$ Here $S_1=(8U_0)^{1/2}(a/3)$ is the “wall" action, whose expression depends only on $U_{+}$, and $R$ is the radius of the four-dimensional hyperspherical “bubble" of true vacuum emerging, via MQT, from the initial false vacuum state. By minimizing the Eulidean action with respect to $R$ we obtain, for our model, $
R={3S_{1}}/(\varepsilon +aB^{2}/M).
$
It should be noted that above results are valid, within our approximations, provided the energy-density difference between the false and true vacuum is small with respect to the height of the unperturbed potential barrier, namely, for $\varepsilon +aB^{2}/M \ll 8U_{0}$. Assuming that such condition is satisfied, we can finally obtain the exponent of the tunneling rate – also called “bounce" coefficient, according to [@cole1] – by evaluating the Euclidean action at the extremum value of $R$ given above. We find, as function of $U_{0}$: $$S_{E}=\frac{27\pi ^{2}S_{1}^{4}}{2\left( \varepsilon +aB^{2}/{M}\right)^{3}}=
\frac{32}{3}\frac{\pi ^{2}a^{4}U_{0}^{2}}{\left( \varepsilon +aB^{2}/{M}\right)^{3}}.
\label{38}$$ Using this equation, which represents the man result of this paper, we are now in the position of discussing the possible effects of a cosmic magnetic field on the decay of the false vacuum.
Let us suppose, to this purpose, that the tunneling process occurs at a cosmological epoch characterized by a value $H< \Mp$ of the Hubble parameter ($\Mp$ Planck mass), and localized around the end of the inflationary phase. Assuming that our scalar field is a dominant source of inflation we can estimate (using the Einstein equations at the given epoch) that $U_0\sim H^2M^2_P$; we can also give to the scalar field a typical value $\phi\sim a\sim \Mp$. A cosmic electromagnetic background field, on the other hand, can be naturally obtained by the inflationary amplification of the quantum vacuum fluctuations, and in this case we can evaluate its energy density by setting $B^2\sim H^4$ [@maurizio] (note that the energy bounds imposed by the process of quantum backreaction [@15] are satisfied, as $H^4 \ll \Mp^2 H^2$). Finally, a typical value of the photon-scalar coupling of Eq. (\[21\]), for a scalar component of the multiplet of fundamental interactions, can be estimated to be of the order of $M\sim \Mp$. In general, we can now consider two cases.
[*A. Dominant magnetic energy.*]{} Let us use the above values of $U_0, \phi, a, B^2, M$, and let us first assume that $\varepsilon \ll B^{2}a/M\sim H^4$, namely that the magnetic energy density dominates over the small potential-energy asymmetry $\vep$ used by Coleman to trigger the MQT transition. In that case we have a fully magnetic activation of the MQT effect, and the bounce coefficient (\[38\]) becomes $$S^{\rm mag}_E=\frac{32}{3}\frac{\pi ^{2}a^{4}U_{0}^{2}}{\left(aB^{2}/{M}\right)^{3}} \sim \frac{32\pi^2 }{3}\left(\frac{\Mp}{H}\right)^8.
\label{39}$$ The approximation condition (\[38\]) becomes $ H^2 \ll \Mp^2$, and is always satisfied as the inflation scale $H$ is always smaller than the Planck scale $\Mp$. This implies, according to Eq. (\[39\]), $S^{\rm mag}_E \gg 1$, namely a very small value of the Bounce coefficient.
We may recall, in particular, that in the standard inflationary scenario the typical inflation scale must satisfy the constraint $H \laq 10^{-5}\Mp$ (following from the inflationary production of a primordial background of gravitational radiation, see e.g. [@17]). One thus obtains, in this context, $S^{\rm mag}_E \gaq 10^{40}$. Such a lower bound can be somewhat relaxed in a string cosmology context, where the inflationary scale is allowed to be as high as the string mass scale $\Ms$, i.e. $H \laq \Ms \sim 10^{-1} \Mp$ (see e.g. [@maurizio]). In that case one obtains $S^{\rm mag}_E \gaq 10^{8}$.
Larger values of the bounce coefficient can be obtained, however, if the Coleman term proportional to $\vep$ dominates the potential-energy asymmetry. In that case the magnetic field $B$ may enhance the probability of tunneling transition, playing a role very similar to that played in the case of a microscopic Josephson junction.
[*B. Subdominant magnetic energy.*]{} Assuming that $\varepsilon \gg B^{2}a/M\sim H^4$ the bouncing coefficient (\[38\]) becomes $$S_E\simeq \frac{32}{3}\frac{\pi ^{2}a^{4}U_{0}^{2}}{\varepsilon ^{3}}\left( 1-
\frac{3B^{2}a}{\varepsilon M}\right) =S_{E}^0\left( 1-\frac{3B^{2}a}{
\varepsilon M}\right) ,
\label{310}$$ where $S_E^0$ denotes the Coleman result without magnetic-field contributions [@cole1]. Using the previous values for the parameters of our cosmological scenario we have $S_E^0 \sim 32\pi^2 \Mp^8 H^4/3\vep^3$, and we obtain that the variation of the bounce coefficient induced by the magnetic-field is given by - 3 1. \[312\] The corresponding enhancement of the escape rate is, in first approximation, $$\frac{\Delta\Gamma}{\Gamma _{0}}=e^{-\left( S_E-S_{E}^0\right)}-1\simeq 3 S_E^0 \frac {H^4}{\epsilon }
\sim\left(\frac{M^2_P H^2}{\varepsilon}\right)^4.
\label{313}$$
It is interesting to note that the condition for validity of our approximation, namely a small energy perturbation with respect to the height of the unperturbed barrier, implies in this case $\vep \ll 8 U_0 \sim 8\Mp^2 H^2$ (and is still compatible with the previous assumption $ \vep \gg H^4$). It follows that $\Delta\Gamma/\Gamma_0 \gg 1$. Hence, for $H^4 \ll \vep \ll \Mp^2 H^2$, even a small variation of the bounce, induced by the cosmic magnetic background according to Eq. (\[312\]), may result into a large variation of the escape rate for the decay of the false vacuum state.
Conclusion {#Sec4}
==========
The decay of the false vacuum discussed in two pioneer papers by Coleman and, Callan and Coleman [@cole1] is based on a process of MQT very similar to the one occurring in the context of superconductive Josephson structures [@1]. The behavior of such superconducting devices is strongly influenced by the presence of an external magnetic field. In this paper we have shown that also in a cosmological context the false vacuum decay may be affected by the contribution of a cosmic magnetic field.
Working in the thin-wall approximation, to first order in the magnetic energy density, we have found that a magnetic enhancement of the MQT escape rate is possible, and consistent with a realistic cosmological scenario satisfying the conventional bounds imposed by inflation and quantum backreation.
The results obtained in this paper can be directly applied, in particular, to the process of tunneling among the vacuum states of the string landscape [@13], and can be easily generalised to take into account the contribution of cosmic background fields of nonmagnetic origin. A more detailed investigation, taking into account other possible contributions to the tunneling rate in the Planckian regime, is postponed to a future paper.
ACKNOWLEDGMENTS {#acknowledgments .unnumbered}
===============
The authors warmly thank A. J. Leggett for fruitful suggestions. The work has been partially supported by the EC STREP Project MIDAS.
[999]{}
B. D. Josephson, Phys. Lett. [**1**]{},251 (1962); A. Barone and G. Paternò, [*Physics and Application of the Josephson Effect*]{} (John Wiley, New York, 1982).
S. Coleman, Phys. Rev. [**D 15**]{}, 2929 (1977); C. G. Callan and S. Coleman, Phys. Rev. [**D 16**]{}, 1762 (1977).
[5]{}A. J. Leggett, Suppl. Prog. Theor. Phys. [**69**]{}, 80 (1980).
J. Clarke, A. N. Cleland,, M. H. Devoret, D. Esteve and J. M. Martinis, Science [**239**]{}, 992 (1988) (and references cited therein).
T. Bauch, F. Lombardi, F. Tafuri, A. Barone, G. Rotoli, P. Delsing and T. Claeson, Phys. Rev. Lett. [**94**]{}, 087003 (2005).
T. Bauch, T. Lindstrom, F. Tafuri, G. Rotoli, P. Delsing, T. Claeson and F. Lombardi, Science [**311**]{}, 57 (2006).
Y. N. Ovchinnikov, A. Barone and A. A. Varlamov, Phys. Rev. Lett. [**99**]{}, 037004 (2007); Y. N. Ovchinnikov, A. Barone and A. A. Varlamov, Phys. Rev. [**B78**]{}, 054521 (2008).
[13]{}L. M. Krauss, J. Dent and G. D. Starkman, Int. J. Mod. Phys. [**D 17**]{}, 2501 (2008); Yung-Son Piao, [*Tunneling for large $N$*]{}, arXiv:0810.3654.
[15]{}V. Demozzi, V. Mukhanov and H. Rubinstein, JCAP [**08**]{}, 025 (2009).
M. Gasperini, [*Elements of String Cosmology*]{} (Cambridge University Press, Cambridge, 2007).
[16]{}M. S. Turner and L. M. Widrow, Phys. Rev. [**D 37**]{}, 2743 (1988).
[17]{}B. Allen, Phys. Rev. [**D 37**]{}, 2078 (1988).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'FFCam is a variable star near the North celestial pole with hydrogen lines in emission. Its optical variability of $\sim$0.3 mag was discovered by HIPPARCOS. The spectral type assigned to the star in SIMBAD is B9, but its position coincides with a ROSAT X–ray source. This suggests the presence of a high-temperature region in the system that could originate at or near a companion object. We undertook a spectroscopic monitoring of FFCam since the beginning of 2012 and found an extremely variable H$\alpha$ line profile as well as periodically variable radial velocities of numerous absorption lines. The main conclusion from our study is that FFCam is a binary system with an orbital period of 7.785 days, a B–type primary and a K–type secondary component. We discuss the spectral features, their variations, and the nature of FFCam.'
author:
- |
T. GARREL$^1$, A. S. MIROSHNICHENKO$^2$, S. DANFORD$^2$,\
S. CHARBONNEL$^3$, F. HOUPERT$^4$, K. N. GRANKIN$^5$, and A. V. KUSAKIN$^6$\
\
*$^1$ Observatoire de Foncaude, Montpellier, France\
*$^2$ University of North Carolina at Greensboro,\
*Department of Physics and Astronomy, Greensboro, NC, USA\
*$^3$ Durtal Observatory, Durtal, France\
*$^4$ Verny Observatory, Verny, France\
*$^5$ Crimean Astrophysical Observatory, Nauchny, Ukraine\
*$^6$ Fessenkov Astrophysical Institute, Almaty, Kazakhstan*******
title: Spectroscopy of the mysterious Be system FFCam
---
Introduction
============
FFCam = HD60062 is a fairly bright ($V \sim$ 7.5–8.0 mag) star not far from the North Celestial Pole (R.A. 7h 47m, Dec. +81$^{\circ} 40^{\prime}$, 2000). It was discovered as a variable star by the HIPPARCOS mission (ESA 1997) that obtained 146 measurements of its visual magnitude on 38 different days in 1989–1993 (see Fig. \[f3\]b). These data were first analyzed by Woitas (1997), who found no periodic variations. The object was included in the General Catalog of Variable Stars as FFCam (Kazarovets et al.1999) and classified as a Be star. We are still searching for the origin of this classification, although it does have an emission-line spectrum. An ROSAT X–ray source with a flux of 0.15$\pm$0.02 countss$^{-1}$ was recently found in 0$.\!\!^{\prime\prime}$1 from the visible/IR star position (Haakonsen & Rutledge 2009). FFCam is located far from the Galactic plane (b = 29$^{\circ}$) that implies a low reddening. Its HIPPARCOS parallax leads a distance of 580$^{+270}_{-150}$ pc (ESA 1997) which may be wrong, if the object is a binary system (see Sect. \[results\]). Both the distance and apparent brightness suggest that the system is not very luminous. The current presentation is the first study of the spectroscopic behavior and the spectral energy distribution of FFCam.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![The H$\alpha$ line profile variations in the spectrum of FFCam in January – February 2012. Intensity is normalized to the underlying continuum, radial velocity is heliocentric. []{data-label="f1"}](Garrel_f1a.eps "fig:"){width="5.4cm" height="4.9cm"} ![The H$\alpha$ line profile variations in the spectrum of FFCam in January – February 2012. Intensity is normalized to the underlying continuum, radial velocity is heliocentric. []{data-label="f1"}](Garrel_f1b.eps "fig:"){width="5.4cm" height="4.8cm"}
![The H$\alpha$ line profile variations in the spectrum of FFCam in January – February 2012. Intensity is normalized to the underlying continuum, radial velocity is heliocentric. []{data-label="f1"}](Garrel_f1c.eps "fig:"){width="5.4cm" height="4.8cm"} ![The H$\alpha$ line profile variations in the spectrum of FFCam in January – February 2012. Intensity is normalized to the underlying continuum, radial velocity is heliocentric. []{data-label="f1"}](Garrel_f1d.eps "fig:"){width="5.4cm" height="4.8cm"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Observations
============
Photometric $UBVRI$ observations were obtained on 2012 April 13 and 15 with the 1.25–m telescope of the Crimean Observatory in Ukraine. $BVR$ observations were obtained on 2012 April 11, 12 and May 24 with a 0.2–m Newtonian reflector at the Tien-Shan Observatory near Almaty, Kazakhstan. Medium-resolution spectra were obtained at three amateur sites in France (Montpellier, Durtal, and Verny) and in the USA at the Three College Observatory (TCO, near Greensboro, North Carolina). Two types of spectrographs were used: long-slit LHires III in the H$\alpha$ region with a spectral resolving power $R \sim$ 17,000 and échelle ($\lambda\lambda$ 4300–7200 Å, $R \sim$ 10,000). All the spectrographs were manufactured by the Shelyak company (www.shelyak.com). In total we obtained over 100 spectra in 2010–2012, mostly in January–April 2012. IRAF was used to reduce the TCO data, the amateurs data were reduced with software packages developed for amateur spectrographs (Audela[^1] and ISIS[^2].)
Results
=======
The main features of the FFCam optical spectrum are the following. The Balmer H$\alpha$ and H$\beta$ lines are seen in variable emission with mostly single– or double–peaked profiles (see Fig. \[f1\]), while almost no emission is observed in H$\gamma$. Fe [ii]{} lines show weak double-peaked emission profiles. He [i]{} lines are in absorption. Numerous absorption lines of neutral metals are weak ($\le$ 10% of the continuum, see Fig. \[f2\]a).
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![[**Panel a.**]{} Absorption lines in the spectrum of FFCam at different phases of the orbital cycle. Intensity is normalized to the underlying continuum. Radial velocity is measured by fitting the line profiles to a Gaussian. [**Panel b.**]{} Variations of the H$\alpha$ line equivalent width measured by intensity integration with the line profile in the continuum normalized spectra. Only the emission component of the line profile was used for the measurements.[]{data-label="f2"}](Garrel_f2a.eps "fig:"){width="5.7cm" height="5.8cm"} ![[**Panel a.**]{} Absorption lines in the spectrum of FFCam at different phases of the orbital cycle. Intensity is normalized to the underlying continuum. Radial velocity is measured by fitting the line profiles to a Gaussian. [**Panel b.**]{} Variations of the H$\alpha$ line equivalent width measured by intensity integration with the line profile in the continuum normalized spectra. Only the emission component of the line profile was used for the measurements.[]{data-label="f2"}](Garrel_f2b.eps "fig:"){width="5.7cm" height="5.7cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Strong absorption components of the H$\beta$ and H$\gamma$ lines and strong pure absorption He [i]{} lines along with weak absorption lines of neutral metals imply that FFCam has a composite spectrum. It is a binary system with a brighter B–type component (later than B2) and a cooler, fainter component probably of an early K–type. The temporal behaviour of the emission-line spectrum (Fig. \[f2\]b) implies that the amount of circumstellar gas in the system varies with time.
Analyzing radial velocity variations, we found them strictly periodic and sinusoidal (Fig. \[f3\]a). Therefore, the binary orbit is circular. The following orbital elements were determined. Radial Velocity Maxima = JD 2455941.594 + 7.785\*E (days). The radial velocity semi-amplitude for the cool component K$_{2}$ = 85.0 kms$^{-1}$. These elements lead to a mass function of f(M) = 0.5 M$_{\odot}$.
The emission line intensity variations do not correlate with the orbital period. Since the stars are very close to each other, the cool component most likely fills its Roche lobe. This can cause mass transfer into the Roche lobe of the hot component and formation of an accretion disk around it.
Absorption lines of the hot component move roughly in anti-phase with those of the cool component, but they seem to be contaminated by the accretion disk. Small variations of the cool component line intensities and chaotic brightness variations (Fig. \[f3\]b) suggest no eclipses in the system.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Panel a. Absorption-line radial velocity curve of FF Cam. Panel b. Visual brightness variations of FFCam plotted against the spectroscopic orbital phase. Filled circles are HIPPARCOS data, open circles are $V$–band magnitudes from Crimea.[]{data-label="f3"}](Garrel_f3a.eps "fig:"){width="5.5cm" height="5.0cm"} ![Panel a. Absorption-line radial velocity curve of FF Cam. Panel b. Visual brightness variations of FFCam plotted against the spectroscopic orbital phase. Filled circles are HIPPARCOS data, open circles are $V$–band magnitudes from Crimea.[]{data-label="f3"}](Garrel_f3b.eps "fig:"){width="5.5cm" height="5.0cm"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The low value of the mass function and the presence of a B–type star in the system imply that the system is viewed nearly pole-on. For a lower limit of 5 M$_{\odot}$ for the system mass, the orbital inclination angle is 27$^{\circ}$. Assuming the HIPPARCOS distance, no interstellar reddening, and a components brightness ratio of $\Delta V$ = 1 mag (Fig. \[f4\]), the cool component’s radius is 10 R$_{\odot}$. The X–ray flux is $\sim 10^{-5}$ of the hot star optical flux. This is over an order of magnitude larger than the expected photospheric flux from the B–type component. Our current data offer no explanation to this fact.
Conclusions
===========
We have found that FFCam is not a Be star, but rather a short-period (7.785 days) semi-detached binary system with mass transfer from the cool component to the hot component. The mass transfer is variable and results in fast variations of the emission-line profiles as well as in the observed photometric variations. High-resolution spectroscopy with a good phase coverage and high signal-to-noise spectra are needed to constrain the component spectral types and orbital parameters.
![Spectral energy distribution of FF Cam. Symbols: filled upward triangles – TD1 UV satellite, filled circles – Crimean $UBVRI$, filled squares – 2MASS $JHK$, filled downward triangles – WISE, open circle – AKARI, open square – IRAS. The thick line is a Kurucz (1993) model atmosphere for a B7–star (70% contribution to the observed flux), the thin line is a model atmosphere for a K0–star (30% contribution). No interstellar extinction is taken into account.[]{data-label="f4"}](Garrel_f4.eps){width="6.0cm" height="5.5cm"}
Acknowledgements {#acknowledgements .unnumbered}
================
A.M. acknowledges support from the American Astronomical Society International Travel Grant program and from the Department of Physics and Astronomy of the University of North Carolina at Greensboro. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France, data products from the Two Micron All Sky Survey (2MASS) and the Wide-Field Infrared Survey Explorer ([*WISE*]{}), and the BeSS database, operated at LESIA, Observatoire de Meudon, France (accessible at\
http://basebe.obspm.fr).
References {#references .unnumbered}
==========
- ESA 1997, European Space Agency SP–1200. The Hipparcos and Tycho Catalogues.
- Haakonsen, C. B., and Rutledge, R. E.: 2009, [**]{} [**184**]{}, 138.
- Kazarovets, E. V., Samus, N. N., Durlevich, O. V., et al.: 1999, IBVS, 4659.
- Kurucz, R.L.: 1993, [*CD-ROM No. 13, Smithsonian Astrophysical Observatory*]{}.
- Woitas, J.: 1997, Inform. Bull. Var. Stars, No 4444.
[^1]: http://www.audela.org/dokuwiki/doku.php/en/start
[^2]: http://www.astrosurf.com/buil/isis/isis$\_$en.htm
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'A.Narwid, Z.Ko[ł]{}aczkowski, A.Pigulski,'
- 'T.Ramza'
title: ' $\beta$ Cephei stars in the inner part of the Galaxy '
---
Introduction
============
The analysis of the OGLE-II observations by means of the image subtraction method resulted in a catalogue of over 200,000 variable star candidates published by @wozn02. The observations were carried out in the years 1997–2000 in 49 Galactic fields covering $\sim$11 square degrees in the sky. In the search for short-period pulsators in the Galaxy, we analyzed the data from this catalogue. The analysis yielded a lot of candidates for $\beta$ Cephei, $\delta$ Scuti, SPB and $\gamma$ Doradus stars. In this paper, we present global properties of a sample of $\sim$230 stars with shortest periods (exluding high-amplitude $\delta$ Scuti stars, see Pigulski et al., these proceedings). This sample consists mainly of a mixture of $\beta$ Cephei and low-amplitude $\delta$ Scuti stars.
Analysis
========
The analysis consisted of an automatic extraction of up to five periodic terms for all stars in the catalogue with consecutive prewhitening followed by an automatic classification based upon the periods, amplitudes, and Fourier coefficients (for stars with detected harmonics or subharmonic). Then, for stars selected in this way, a detailed analysis was performed in an interactive way. As a result, we selected about 230 short-period stars that are good candidates for $\beta$ Cephei and low-amplitude $\delta$ Scuti stars. For these stars, we searched for the photometry from the MACHO survey. It turned out that 94 stars from this sample have the MACHO photometry available. However, it is of low quality for some stars, so that we combined OGLE-II and MACHO photometry for only 67 stars in our sample. The analysis of the combined photometry results in a better resolution, lower detection threshold and practically removes the ambiguity in the frequencies of the extracted terms.
Results
=======
The main results of our study can be summarized as follows:
1. In the OGLE-II catalogue of variable star candidates [@wozn02] we found about 230 low-amplitude stars with periods shorter than 0.5 d that are good candidates for $\beta$ Cephei and low-amplitude $\delta$ Scuti stars. In the colour-magnitude diagram (Fig. \[cmd\]), they populate the branch of main-sequence stars. Their $I$ magnitudes range from the OGLE-II saturation limit of $I \sim$ 11.5 down to magnitude 16.5.
2. The periods of the bulk of stars in our sample range between 0.04 and 0.2 d (Fig. \[perampl\]). Since the periods of modes excited in both $\beta$ Cephei and $\delta$ Scuti stars cover this range, the two types of pulsators cannot be distinguished merely from their periods. We are going to carry out the $UBV$ photometry or/and low-resolution spectroscopy of stars from our sample. This will allow us to classify definitely these stars.
3. Almost all other variable stars with periods shorter than 0.5 d were excluded in the process of automatic classification. However, a small-amplitude RR Lyrae and W UMa stars may contaminate our sample, especially for periods longer than $\sim$0.25 d. We estimate, however, that this contribution does not exceed 10%.
4. Half of the stars from our sample (114) shows more than one period of variability. Up to eight modes in a single star were found in the combined OGLE-II and MACHO data.
5. In some stars, equidistant in frequency triplets, an indication of rotational splitting, were found. For many stars, the range of periods of the excited modes is very wide. This may be an indication of the larger-than-average metallicity expected in young population of stars in the inner regions of our Galaxy.
6. All the information given above shows clearly that our sample includes many $\beta$ Cephei stars located 3–6 kpc away.
This work has been supported by the KBN grant No. 1P03D01627.
Woźniak, P.R. et al. 2002, AcA, 52, 129
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we want to discuss how the question about the rationality of $L^2$-Betti numbers is related to the Isomorphism Conjecture in algebraic $K$-theory and why in this context noncommutative localization appears as an important tool.'
address: |
Fachbereich Mathematik\
Universität Münster\
Einsteinstr. 62\
48149 Münster\
Germany
author:
- Holger Reich
title: '$L^2$-Betti numbers, Isomorphism Conjectures and Noncommutative Localization'
---
[^1]
[*$L^2$-Betti numbers*]{} are invariants of spaces which are defined analogously to the ordinary Betti-numbers but they take information about the fundamental group into account and are a priori real valued.
The [*Isomorphism Conjecture in algebraic $K$-theory*]{} predicts that $K_0( {{\mathbb C}}\Gamma )$, the Grothendieck group of finitely generated projective ${{\mathbb C}}\Gamma$-modules, should be computable from the $K$-theory of the complex group rings of finite subgroups of $\Gamma$.
Given a commutative ring one can always invert the set of all non-zerodivisors. Elements in the resulting ring have a nice description in terms of fractions. For noncommutative rings like group rings this may no longer be the case and other concepts for a [*noncommutative localization*]{} can be more suitable for specific problems.
The question whether $L^2$-Betti numbers are always rational numbers was asked by Atiyah in [@Atiyah(1976)]. The question turns out to be a question about modules over the group ring of the fundamental group $\Gamma$. In [@Linnell(1993)] Linnell was able to answer the question affirmatively if $\Gamma$ belongs to a certain class of groups which contains free groups and is stable under extensions by elementary amenable groups (one also needs a bound on the orders of finite subgroups). In fact Linnell proves the stronger result that there exists a semisimple subring in ${\mathcal{U}}\Gamma$, the algebra of operators affiliated to the group von Neumann algebra, which contains the complex group ring.
The main purpose of this short survey is to give a conceptional framework for Linnell’s result, to explain how the question about the rationality of $L^2$-Betti numbers relates to the Isomorphism conjecture, and why this may involve studying noncommutative localizations of group rings. (The impatient reader should right away take a look at Proposition \[strategy\], Theorem \[main-linnell\] and Addendum \[addendum\].)
Since probably not every reader is familiar with all three circles of ideas – $L^2$-Betti numbers – Isomorphism Conjectures – Noncommutative Localization – the paper contains introductions to all of these.
After a brief introduction to group von Neumann algebras and the notion of $\Gamma$-dimension we proceed to explain the algebra ${\mathcal{U}}\Gamma$ of operators affiliated to a group von Neumann algebra and introduce $L^2$-Betti numbers in a very algebraic fashion. (Once ${\mathcal{U}}\Gamma$ has been defined there is no more need for Hilbert-spaces.) Section \[section-atiyah\] explains the Atiyah Conjecture and contains in particular Proposition \[strategy\] which is a kind of strategy for its proof. That Proposition says that if one can factorize the inclusion ${{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma$ over a ring ${\mathcal{S}}\Gamma$ with good ring-theoretical properties in such way that a certain $K$-theoretic condition is satisfied, then the Atiyah conjecture follows. In Section \[section-candidates\] we present a number of candidates for the ring ${\mathcal{S}}\Gamma$. To do this we first review a number of concepts from the theory of noncommutative localization in Section \[section-localization\]. One of the candidates is the universal localization of ${{\mathbb C}}\Gamma$ with respect to all matrices that become invertible over ${\mathcal{U}}\Gamma$. Section \[section-linnell\] contains Linnell’s result. We would like to emphasize that the intermediate rings Linnell constructs can also be viewed as universal localizations, see Addendum \[addendum\] (U). In Section \[section-isomorphism\] we discuss the Isomorphism Conjecture which seems to be closely related to the $K$-theoretical condition mentioned above. In the last Section we discuss to what extent the functor $- \otimes_{{{\mathbb C}}\Gamma} {\mathcal{U}}\Gamma$, which plays an important role when one studies $L^2$-Betti numbers, is exact.
The only new result in this paper is the following.
Let ${\mathcal{D}}\Gamma$ denote the division closure of ${{\mathbb C}}\Gamma$ inside ${\mathcal{U}}\Gamma$. The flat dimension of ${\mathcal{D}}\Gamma$ over ${{\mathbb C}}\Gamma$ is smaller than $1$ for groups in Linnell’s class which have a bound on the orders of finite subgroups.
For the division closure see Definition \[definition-division-rational\], for Linnell’s class of groups see Definition \[definition-linnells-class\]. The result is proven as Theorem \[application\] below. As an immediate corollary one obtains.
\[corollary-euler-intro\] If the infinite group $\Gamma$ belongs to Linnell’s class ${\mathcal{C}}$ and has a bound on the orders of finite subgroups then the $L^2$-Euler characteristic (and hence the ordinary one whenever defined) satisfies $$\chi^{(2)} ( \Gamma ) \leq 0.$$
We also would like to mention that the above theorem leads to interesting non-trivial examples of stably flat universal localizations which appear in [@Neeman-Ranicki(2002)].
A reader who is interested in more information about $L^2$-Betti numbers and the Atiyah Conjecture should consult the book [@Lueck(2002)]. Almost all topics discussed here are also treated there in detail. More information and further results about the Atiyah Conjecture can be found in [@Linnell(1998)], [@Grigorchuk-Linnell-Schick-Zuk(2000)], [@Schick(2000c)] and [@Schick(2002a)].
[**Acknowledgement**]{}
I would like to thank Wolfgang L[ü]{}ck who as my thesis advisor introduced me to the Atiyah Conjecture and the Isomorphism Conjecture. I would also like to thank Thomas Schick. In discussions with him the idea that the Corollary above should be true evolved. Furthermore I would like to thank Andrew Ranicki and the ICMS in Edinburgh for organizing the lively and interesting Workshop on Noncommutative Localization.
The von Neumann Dimension {#section-dimension}
=========================
In this section we want to introduce group von Neumann algebras and explain a notion of dimension for finitely generated projective modules over such algebras.
For a (discrete) group $\Gamma$ we denote by ${{\mathbb C}}\Gamma$ the complex group ring and by $l^2 \Gamma$ the complex Hilbert space with orthonormal basis $\Gamma$. Each group element operates from the left on $l^2 \Gamma$. Linearly extending this action we obtain an inclusion $${{\mathbb C}}\Gamma \to {\mathcal{B}}( l^2 \Gamma )$$ into the algebra ${\mathcal{B}}( l^2 \Gamma )$ of bounded linear operators on the Hilbert space $l^2 \Gamma$. The group von Neumann algebra ${\mathcal{N}}\Gamma$ is defined as the closure of ${{\mathbb C}}\Gamma$ inside ${\mathcal{B}}( l^2 \Gamma )$ with respect to the weak (or strong, it doesn’t matter) operator topology. This algebra is closed under taking the adjoint, i.e.$a^{\ast} \in {\mathcal{N}}\Gamma$ for every operator $a \in {\mathcal{N}}\Gamma$.
A von Neumann algebra is by definition a $\ast$-closed subalgebra of the algebra of bounded linear operators on some Hilbert-space which is closed with respect to the strong (or weak, it doesn’t matter) operator topology. Similarly a $C^{\ast}$-algebra can be defined as a $\ast$-closed subalgebra of the algebra of bounded operators on some Hilbert space which is closed with respect to the topology given by the operator-norm. Every von Neumann algebra is in particular a $C^{\ast}$-algebra. In the situation described above the operator-norm closure of ${{\mathbb C}}\Gamma$ inside ${\mathcal{B}}( l^2 \Gamma )$ defines the so called reduced $C^{\ast}$-algebra $C^{\ast}_r \Gamma$ and we have a natural inclusion of $C^{\ast}_r \Gamma$ in ${\mathcal{N}}\Gamma$.
The bicommutant theorem of von Neumann (see for example Theorem 5.3.1 in [@Kadison-Ringrose(1983)]) is a first hint that the definition of ${\mathcal{N}}\Gamma$ is very natural also from a purely algebraic point of view (at least if we agree to consider ${\mathcal{B}}( l^2 \Gamma )$ as something natural). It says that the von Neumann algebra is the double commutant of ${{\mathbb C}}\Gamma$, i.e.$${\mathcal{N}}\Gamma = {{\mathbb C}}\Gamma^{\prime \prime},$$ where for a subset $A \subset {\mathcal{B}}( l^2 \Gamma )$ we write $A^{\prime}= \{ b \in {\mathcal{B}}( l^2 \Gamma ) | ba = ab \mbox{ for all } a \in A \}$ for the commutant of $A$ in ${\mathcal{B}}( l^2 \Gamma )$.
The group von Neumann algebra comes equipped with a natural trace. This trace is given as follows: $$\begin{aligned}
{{\rm tr}}_{\Gamma} : {\mathcal{N}}\Gamma & \to & {{\mathbb C}}\\
a & \mapsto & \langle a (e ) , e \rangle.\end{aligned}$$ Here $\langle - , - \rangle$ denotes the inner product in $l^2 \Gamma$ and $e$ is the unit element of the group considered as a vector in $l^2 \Gamma$. Applied to an element $a=\sum a_g g$ in the group ring ${{\mathbb C}}\Gamma$ the trace yields the coefficient of the identity element $a_e$. Of course we have the trace property ${{\rm tr}}_{\Gamma}( ab ) = {{\rm tr}}_{\Gamma} ( ba )$. \[argument ???\] Once we have such a trace there is a standard procedure to assign a complex number to each finitely generated projective ${\mathcal{N}}\Gamma$-module: if $p=p^2=(p_{ij}) \in M_n ( {\mathcal{N}}\Gamma )$ is an idempotent matrix over ${\mathcal{N}}\Gamma$ which represents $P$, i.e. such that $P \cong {{\rm im}}( p: {\mathcal{N}}\Gamma^n \to {\mathcal{N}}\Gamma^n)$ then we set $$\begin{aligned}
\label{definition-dim}
\dim_{ \Gamma} P = \Sigma_{i = 1}^n {{\rm tr}}_{\Gamma} ( p_{ii} )\end{aligned}$$ and call it the $ \Gamma$-dimension of $P$. We have the following standard facts.
The $ \Gamma$-dimension has the following properties.
1. $\dim_{ \Gamma} P$ is a nonnegative real number.
2. $\dim_{ \Gamma} P$ depends only on the isomorphism class of $P$.
3. [**Normalization.**]{} We have $\dim_{ \Gamma} {\mathcal{N}}\Gamma = 1$.
4. [**Additivity.**]{} If $0 \to L \to M \to N \to 0$ is a short exact sequence of finitely generated projective modules then $$\dim_{ \Gamma} M = \dim_{ \Gamma} L + \dim_{ \Gamma} N.$$
5. [**Faithfulness.**]{} $\dim_{ \Gamma} P = 0 $ if and only if $P=0$.
\(i) follows since one can always arrange that the idempotent $p=p^2$ in (\[definition-dim\]) is a projection, i.e. $p=p^2=p^{\ast}$ (see for example Proposition 4.6.2 on p.23 in [@Blackadar(1986)]). (v) follows from the fact that the trace is faithful, i.e. ${{\rm tr}}( a^{\ast} a) =0$ implies $a = 0$. (ii)-(iv) are straightforward.
Let $K_0 ( {\mathcal{N}}\Gamma )$ denote the Grothendieck-group of finitely generated projective ${\mathcal{N}}\Gamma$-modules then because of (i)-(iv) above we obtain a homomorphism $$\xymatrix{
K_0 ( {\mathcal{N}}\Gamma ) \ar[rr]^-{\dim_{ \Gamma }} & & {{\mathbb R}}.
}$$ We recall some terminology, compare page 5 in [@Schofield(1985)].
A projective rank function $\rho$ on a ring $R$ is a homomorphism $\rho: K_0 ( R ) \to {{\mathbb R}}$ satisfying $\rho( [R^1] )=1$ and $\rho( [P] ) \geq 0$ for every finitely generated projective $R$-module $P$. It is called faithful if moreover $\rho( [P] ) =0$ implies $P=0$.
In this terminology we can summarize the content of the proposition above by saying that $\dim_{ \Gamma} : K_0 ( {\mathcal{N}}\Gamma ) \to {{\mathbb R}}$ is a faithful projective rank function.
Other natural examples of faithful projective rank functions occur as follows: Suppose the ring $R$ is embedded in a simple artinian ring $M_n ( D )$, where $D$ is a skew field. Then $P \mapsto \frac{1}{n} \dim_D P \otimes_R M_n ( D )$ defines a faithful projective rank function on $R$.
We would like to emphasize the following additional properties of the $\Gamma$-dimension for ${\mathcal{N}}\Gamma$-modules which are not true for arbitrary projective rank functions. They give further justification for the the use of the word “dimension” in this context.
The $\Gamma$-dimension satisfies:
1. [**Monotony.**]{} The ${\mathcal{N}}\Gamma$-dimension is monotone, i.e. $P \subset Q$ implies that $\dim_{ \Gamma} P \leq \dim_{ \Gamma} Q$.
2. [**Cofinality.**]{} If $P = \bigcup_{i \in I } P_i$ is a directed union of submodules then $$\dim_{ \Gamma} P = \sup_{i \in I } \dim_{ \Gamma } P_i .$$
Of course cofinality implies monotony. To convince the reader that these properties are not automatic for projective rank functions we would like to treat an example.
\[example-monotone-false\] Let $\Gamma$ be a free group on two generators $x$ and $y$. By work of Cohn [@Cohn(1964)] we know that ${{\mathbb C}}\Gamma$ is a free ideal ring. In particular every finitely generated projective module is free and taking its rank yields an isomorphism $$\xymatrix{
K_0 ( {{\mathbb C}}\Gamma ) \ar[r]^-{\cong} & {{\mathbb Z}}.
}$$ This is a faithful projective rank function with values in ${{\mathbb Z}}$. However there is an exact sequence $$\xymatrix{
0 \ar[r] & {{\mathbb C}}\Gamma^2 \ar[rr]^-{(x-1 , y-1 )} & & {{\mathbb C}}\Gamma \ar[r] & {{\mathbb C}}\ar[r] & 0
}$$ which shows that the rank function is not monotone. (Geometrically the above resolution of ${{\mathbb C}}$ is obtained as the cellular chain complex with complex coefficients of the universal cover $E \Gamma$ of the model for the classifying space $B \Gamma$ given by the wedge of two circles.)
In fact one can always compose $\dim_{\Gamma}$ with the natural map $K_0 ( {{\mathbb C}}\Gamma ) \to K_0 ( {\mathcal{N}}\Gamma )$. In this way we obtain naturally a faithful projective rank function on ${{\mathbb C}}\Gamma$ for every group $\Gamma$. One rediscovers the example above in the case where $\Gamma$ is the free group on two generators.
The Algebra of Operators affiliated to ${\mathcal{N}}\Gamma$. {#section-affiliated}
=============================================================
The category of finitely generated projective ${\mathcal{N}}\Gamma$-modules has one drawback: it is not abelian. In particular if we start out with a complex of finitely generated projective ${\mathcal{N}}\Gamma$-modules then the homology modules are not necessarily finitely generated projective and hence the ${\mathcal{N}}\Gamma$-dimension as explained above is a priori not available. But this is exactly what we would like to do in order to define $L^2$-Betti numbers, i.e. we want to consider $$C^{{\mathit{cell}}}_{\ast} ( \widetilde{ X } ) \otimes_{ {{\mathbb Z}}\Gamma } {\mathcal{N}}\Gamma$$ the cellular chain-complex of the universal covering of a CW-complex $X$ of finite type tensored up to ${\mathcal{N}}\Gamma$ and assign a dimension to the homology modules.
There are several ways to get around this problem. The traditional way to deal with it is to work with certain Hilbert spaces with an isometric $\Gamma$-operation instead of modules, e.g.with $l^2 \Gamma^n$ instead of ${\mathcal{N}}\Gamma^n$. These Hilbert spaces have a $\Gamma$-dimension and one (re-)defines the homology as the kernel of the differentials modulo the [*closure*]{} of their images. This is then again a Hilbert space with an isometric $\Gamma$-action and has a well defined $\Gamma$-dimension.
A different approach is taken in [@Lueck(1997a)]: [*finitely presented*]{} ${\mathcal{N}}\Gamma$-modules do form an abelian category (because ${\mathcal{N}}\Gamma$ is a semihereditary ring) and the ${\mathcal{N}}\Gamma$-dimension can be extended to these modules in such a way that the properties (i)-(vi) still hold. (In fact in [@Lueck(1998a)] the $\Gamma$-dimension is even extended to arbitrary ${\mathcal{N}}\Gamma$-modules.)
A third possible approach is to introduce the algebra ${\mathcal{U}}\Gamma$ of operators affiliated to ${\mathcal{N}}\Gamma$. This algebra has better ring-theoretic properties and indeed finitely generated projective ${\mathcal{U}}\Gamma$-modules do form an abelian category. Moreover the notion of $\Gamma$-dimension extends to that algebra. We want to explain this approach in some detail in this section.
Recall that an unbounded operator $a: {{\rm dom}}(a) \to H$ on a Hilbert space $H$ is a linear map which is defined on a linear subspace ${{\rm dom}}(a) \subset H$ called the domain of $a$. It is called densely defined if ${{\rm dom}}(a )$ is a dense subspace of $H$ and it is called closed if its graph considered as a subspace of $H \oplus H$ is closed. Each bounded operator is closed and densely defined. For unbounded operators $a$ and $b$ the symbol $a \subset b$ means that restricted to the possibly smaller domain of $a$ the two operators coincide. The following definition goes back to [@Murray-Neumann(1936)].
A closed and densely defined (possibly unbounded) operator $a: {{\rm dom}}(a) \to l^2 \Gamma$ is affiliated to ${\mathcal{N}}\Gamma$ if $ba \subset ab$ for all $b \in {\mathcal{N}}\Gamma' $. The set $${\mathcal{U}}\Gamma = \{ a: {{\rm dom}}(a ) \to l^2 \Gamma \; | \;
a \mbox{ is \begin{tabular}{l} closed, \\ densely defined \\ and affiliated to ${\mathcal{N}}\Gamma$
\end{tabular} } \}$$ is called the algebra of operators affiliated to ${\mathcal{N}}\Gamma$.
Each group element $\gamma \in \Gamma$ acts by right multiplication on $l^2 \Gamma$. This defines an element $r_{\gamma} \in {\mathcal{N}}\Gamma'$ (we had $\Gamma$ acting from the left when we defined ${\mathcal{N}}\Gamma$). In order to prove that a closed densely defined operator $a$ is affiliated it suffices to check that its domain ${{\rm dom}}( a )$ is $\Gamma$-invariant and that for all vectors $v \in {{\rm dom}}(a )$ we have $r_{\gamma} a (v) = a r_{\gamma} (v )$ for all $\gamma \in \Gamma$. In this sense the affiliated operators are precisely the $\Gamma$-equivariant unbounded operators.
Observe that the naive composition of two unbounded operators $c$ and $d$ yields an operator $dc$ which is only defined on $c^{-1}( {{\rm dom}}(d) )$. Similarly addition is only defined on the intersection of the domains. It is hence not obvious that ${\mathcal{U}}\Gamma$ is an algebra.
The set ${\mathcal{U}}\Gamma$ becomes a ${{\mathbb C}}$-algebra if we define addition and a product as the closure of the naive addition respectively composition of operators.
This is proven in Chapter XVI in [@Murray-Neumann(1936)]. A proof is reproduced in Appendix I in [@Reich(1999)] and also in Chapter 8 of [@Lueck(2002)].
The subalgebra of all bounded operators in ${\mathcal{U}}\Gamma$ is ${\mathcal{N}}\Gamma$. In contrast to ${\mathcal{N}}\Gamma$ there seems to be no useful topology on ${\mathcal{U}}\Gamma$. So we left the realm of $C^{\ast}$-algebras and $C^{\ast}$-algebraic methods. The reason ${\mathcal{U}}\Gamma$ is nevertheless very useful is that we have gained good ringtheoretical properties. Let us recall the definition of von Neumann regularity.
\[vNregular-definition\] A ring $R$ is called von Neumann regular if one of the following equivalent conditions is satisfied.
1. Every $R$-module $M$ is flat, i.e. for every module $M$ the functor $- \otimes_R M$ is exact.
2. Every finitely presented $R$-module is already finitely generated projective.
3. The category of finitely generated projective $R$-modules is abelian.
4. \[vNregular-definition-vier\] For all $x \in R$ there exists a $y \in R$ such that $xyx=x$.
For (i) $\Leftrightarrow$ (iv) see for example Theorem 4.2.9 in [@Weibel(1994)]. (i) $\Rightarrow$ (ii) follows since every finitely presented flat $R$-module is projective, see Theorem 3.2.7 in [@Weibel(1994)]. Since the tensor product is compatible with colimits, directed colimits are exact and every module is a directed colimit of finitely presented modules we obtain (ii) $\Rightarrow$ (i). For (ii) $\Rightarrow$ (iii) one needs to check that cokernels and kernels between finitely generated projectives are again finitely generated projective. But a cokernel is essentially a finitely presented module. The argument for the kernel and (iii) $\Rightarrow$ (ii) are elementary.
Note that in particular fields, skew fields, simple artinian rings and semisimple rings are von Neumann regular (every module is projective over such rings). The first condition says that von Neumann regular rings form a very natural class of rings from a homological algebra point of view: they constitute precisely the rings of weak homological dimension $0$. The last condition, which seems less conceptional to modern eyes, was von Neumann’s original definition [@vonNeumann(1936)] and has the advantage that one can explicitely verify it in the case we are interested in. More information about von Neumann regular rings can be found in [@Goodearl(1979)].
The algebra ${\mathcal{U}}\Gamma$ is a von Neumann regular ring.
Using the polar decomposition and functional calculus one can explicitely construct a $y$ as it is required in the characterization \[vNregular-definition\] (iii) of von Neumann regularity given above. Compare Proposition 2.1 (v) in [@Reich(2001)].
In order to define $L^2$-Betti numbers it remains to establish a notion of dimension for finitely generated projective ${\mathcal{U}}\Gamma$-modules. \[Ueberleitung O.K?\]
\[NGUG-properties\] We have the following facts about the inclusion ${\mathcal{N}}\Gamma \subset {\mathcal{U}}\Gamma$.
1. The natural map $K_0 ( {\mathcal{N}}\Gamma ) \to K_0 ( {\mathcal{U}}\Gamma )$ is an isomorphism. In particular there is a $\Gamma$-dimension for finitely generated projective ${\mathcal{U}}\Gamma$-modules which we simply define via the following diagram: $$\xymatrix{
K_0 ( {\mathcal{N}}\Gamma ) \ar[dr]^{\dim_{\Gamma}} \ar[rr]^-{\cong} & & K_0 ( {\mathcal{U}}\Gamma ) \ar[dl]_{\dim_{\Gamma}} \\
& {{\mathbb R}}&
}$$
2. The ring ${\mathcal{U}}\Gamma$ is the Ore-localization (compare Proposition \[proposition-ore-localization\]) of ${\mathcal{N}}\Gamma$ with respect to the multiplicative subset of all non-zerodivisors. In particular $- \otimes_{{\mathcal{N}}\Gamma} {\mathcal{U}}\Gamma$ is an exact functor.
See [@Reich(2001)] Proposition 6.1 (i) and Proposition 2.1 (iii).
If we now start with a finitely presented (as opposed to finitely generated projective) ${{\mathbb C}}\Gamma$-module $M$ then because of \[vNregular-definition\] (ii) we know that $M \otimes_{{{\mathbb C}}\Gamma} {\mathcal{U}}\Gamma$ is a finitely generated projective ${\mathcal{U}}\Gamma$-module and it makes sense to consider its $\Gamma$-dimension.
The assignment $M \mapsto \dim_{ \Gamma} ( M \otimes_{{{\mathbb C}}\Gamma} {\mathcal{U}}\Gamma )$ is a Sylvester module rank function for finitely presented ${{\mathbb C}}\Gamma$-modules in the sense of Chapter 7 in [@Schofield(1985)].
We are now prepared to give a definition of $L^2$-Betti numbers using the $\Gamma$-dimension for ${\mathcal{U}}\Gamma$-modules. Let $X$ be a CW-complex of finite type, i.e. there are only finitely many cells in each dimension. Let $\widetilde{ X }$ denote the universal covering. It carries a natural CW-structure and a cellular free $\Gamma= \pi_1 ( X )$-action. There is one $\Gamma$-orbit of cells in $\widetilde{X}$ for each cell in $X$ and in particular the cellular chain complex $C_{\ast}^{{\mathit{cell}}} ( \widetilde{X} )$ is a complex of finitely generated free ${{\mathbb Z}}\Gamma$-modules.
\[definition-l2-betti\] For a CW-complex $X$ of finite type we define its $L^2$-Betti numbers as $$b_p^{(2)} ( X ) = \dim_{{\mathcal{U}}\Gamma} H_p ( C_{\ast}^{{\mathit{cell}}} ( \widetilde{X} ) \otimes_{{{\mathbb Z}}\Gamma} {\mathcal{U}}\Gamma ).$$
Note that by \[vNregular-definition\] (iii) the homology modules are finitely generated projective ${\mathcal{U}}\Gamma$-modules and hence have a well defined ${\mathcal{U}}\Gamma$-dimension.
\[remark-dimension-for-arbitrary\] As already mentioned it is possible to extend the notion of $\Gamma$-dimension to arbitrary ${\mathcal{N}}\Gamma$-modules in such a way that one still has “additivity” and “cofinality” [@Lueck(1998a)]. Of course one has to allow the value $\infty$, and in cases where this value occurs one has to interpret “additivity” and “cofinality” suitably. In [@Reich(2001)] it is shown that analogously there is a $\Gamma$-dimension for arbitrary ${\mathcal{U}}\Gamma$-modules which is compatible with the one for ${\mathcal{N}}\Gamma$-modules in the sense that for an ${\mathcal{N}}\Gamma$-module $M$ we have $$\begin{aligned}
\label{dim-compatible}
\dim_{{\mathcal{U}}\Gamma} M \otimes_{{\mathcal{N}}\Gamma} {\mathcal{U}}\Gamma = \dim_{{\mathcal{N}}\Gamma } M.\end{aligned}$$ Both notions of extended dimension can be used to define $L^2$-Betti numbers for arbitrary spaces by working with the singular instead of the cellular chain complex. From \[NGUG-properties\] (ii) we conclude that for a complex $C_{\ast}$ of ${\mathcal{N}}\Gamma$-modules we have $$H_{\ast} ( C_{\ast} \otimes_{{\mathcal{N}}\Gamma} {\mathcal{U}}\Gamma ) = H_{\ast} ( C_{\ast} ) \otimes_{{\mathcal{N}}\Gamma} {\mathcal{U}}\Gamma.$$ If we combine this with (\[dim-compatible\]) we see that the two possible definitions of $L^2$-Betti numbers coincide. In the following we will not deal with $L^2$-Betti numbers in this generality. We restrict our attention to CW-complexes of finite type and hence to finitely generated projective ${\mathcal{U}}\Gamma$-modules.
In order to illustrate the notions defined so far we would like to go through two easy examples.
\[example-finite\] Suppose $\Gamma$ is a finite group of order $\# \Gamma$. In this case all the functional analysis is obsolete. We have ${{\mathbb C}}\Gamma = {\mathcal{N}}\Gamma = {\mathcal{U}}\Gamma$ and $l^2 \Gamma={{\mathbb C}}\Gamma$. A finitely generated projective module $P$ is just a finite dimensional complex $\Gamma$-representation. One can check that $$\dim_{\Gamma} P = \frac{1}{\# \Gamma} \dim_{{{\mathbb C}}} P.$$
\[example-infinite-cyclic\] Suppose $\Gamma = C$ is the infinite cyclic group written multiplicatively with generator $z \in C$. In this case (using Fourier transformation) the Hilbert-space $l^2 \Gamma$ can be identified with $L^2 ( S^1 )$, the square integrable functions on the unit circle equipped with the standard normalized measure $\mu=\frac{1}{2 \pi} dz$. Under this correspondence the group element $z$ corresponds to the function $z \mapsto z$, where we think of $S^1$ as embedded in the complex plane. The algebras ${{\mathbb C}}\Gamma$, ${\mathcal{N}}\Gamma$ and ${\mathcal{U}}\Gamma$ can be identified as follows: $$\begin{aligned}
{{\mathbb C}}\Gamma & \leftrightarrow & {{\mathbb C}}[ z^{\pm 1} ]
\mbox{ \begin{tabular}{l} Laurent-polynomials considered \\ as functions on $S^1$ \end{tabular} } \\
{\mathcal{N}}\Gamma & \leftrightarrow & L^{\infty} ( S^1 )
\mbox{ \begin{tabular}{l} essentially bounded \\ functions on $S^1$ \end{tabular} } \\
{\mathcal{U}}\Gamma & \leftrightarrow & L ( S^1 )
\mbox{ \begin{tabular}{l} measurable functions \\ on $S^1$ \end{tabular} } \end{aligned}$$ The action on $L^2 (S^1 )$ in each case is simply given by multiplication of functions. The trace on ${\mathcal{N}}\Gamma$ becomes the integral $f \mapsto \int_{S^1} f d \mu $. For a measurable subset $A \subset S^1$ let $\chi_A$ denote its characteristic function. Then $p=\chi_A$ is a projection and $P_A =p L^{\infty} ( S^1)$ is a typical finitely generated projective $L^{\infty} (S^1)$-module. We have $$\dim_{\Gamma} P_A = {{\rm tr}}_{\Gamma} ( p ) = \int_{S^1} \chi_A d \mu = \mu ( A ).$$ In particular we see that every nonnegative real number can occur as the $\Gamma$-dimension of a finitely generated projective ${\mathcal{N}}\Gamma$- or ${\mathcal{U}}\Gamma$-module. The module $L^{\infty} ( S^1 ) / (z-1) L^{\infty} ( S^1 )$ is an example of a module which becomes trivial (and hence projective) over $L(S^1)$, because $(z-1)$ becomes invertible. In fact one can show that the there is an isomorphism $$\xymatrix{
K_0( L^{\infty} (S^1) ) \cong K_0( L ( S^1 ) ) \ar[r]^-{\cong} & L^{\infty}(S^1 ; {{\mathbb Z}}),
}$$ where $L^{\infty}(S^1 ; {{\mathbb Z}})$ denotes the space of integer valued measurable bounded functions on $S^1$, compare Proposition 6.1 (iv) in [@Reich(2001)]. Every such function can be written in a unique way as a finite sum $f = \sum_{n=- \infty}^{\infty} n \cdot \chi_{A_n}$ with $A_n=f^{-1} ( \{ n \} ) \subset S^1$ and corresponds to $\sum_{n=-\infty}^{\infty} n [P_{A_n}]$ under the above isomorphism.
Once we have the notion of $L^2$-Betti numbers it is natural to define $$\chi^{(2)}( X ) = \sum (-1)^i b_i^{(2)} ( X ).$$ A standard argument shows that for a finite CW-complex this $L^2$-Euler characteristic coincides with the ordinary Euler-characteristic. But in fact since $L^2$-Betti numbers tend to vanish more often than the ordinary Betti-numbers the $L^2$-Euler characteristic is often defined in cases where the ordinary one is not. We also define $L^2$-Betti numbers and the $L^2$-Euler characteristic of a group as $$b_p^{(2)} ( \Gamma )= b_p^{(2)}( B \Gamma ) \mbox{ and } \chi^{(2)} ( \Gamma )= \chi^{(2)}( B \Gamma ).$$ As an example of an application we would like to mention the following result which is due to Cheeger and Gromov [@Cheeger-Gromov(1986)].
Let $\Gamma$ be a group which contains an infinite amenable normal subgroup, then $$b_p^{(2)}( \Gamma ) = 0 \mbox{ for all } p \mbox{, and hence } \chi^{(2)} ( \Gamma ) = 0.$$
The Atiyah Conjecture {#section-atiyah}
=====================
The question arises which real numbers do actually occur as values of $L^2$-Betti numbers. This question was asked by Atiyah in [@Atiyah(1976)] where he first introduced the notion of $L^2$-Betti numbers. (The definition of $L^2$-Betti numbers at that time only applied to manifolds and was given in terms of the Laplace operator on the universal covering.) It turns out that the question about the values can be phrased as a question about the passage from finitely presented ${{\mathbb Z}}\Gamma$- or ${{\mathbb Q}}\Gamma$-modules to ${\mathcal{U}}\Gamma$-modules.
Let $\Lambda$ be an additive subgroup of ${{\mathbb R}}$ which contains ${{\mathbb Z}}$. Let $\Gamma$ be a finitely presented group. The following two statements are equivalent.
1. For all CW-complexes $X$ of finite type with fundamental group $\Gamma$ and all $p \geq 0$ we have $$b_p^{(2)} ( X ) \in \Lambda.$$
2. For all finitely presented ${{\mathbb Z}}\Gamma$-modules $M$ we have $$\dim_{{\mathcal{U}}\Gamma} ( M \otimes_{{{\mathbb Z}}\Gamma } {\mathcal{U}}\Gamma ) \in \Lambda.$$
Using the additivity of the dimension and the fact that the finitely generated free modules of the complex $C_{\ast}^{{\mathit{cell}}} ( \widetilde{X} ) \otimes_{{{\mathbb Z}}\Gamma} {\mathcal{U}}\Gamma$ have integer dimensions (ii) $\Rightarrow$ (i) is straightforward. For the reverse direction one needs to construct a CW-complex $X$ with fundamental group $\Gamma$ such that the presentation matrix of $M$ appears as the, say $5$-th differential in $C_{\ast}^{{\mathit{cell}}}( \widetilde{X} )$ whereas the $4$-th differential is zero. This is possible by standard techniques. For details see Lemma 10.5 in [@Lueck(2002)].
More generally one can induce up finitely presented modules over $R \Gamma$ for every coefficient ring $R$ with ${{\mathbb Z}}\subset R \subset {{\mathbb C}}$ and ask about the values of the corresponding $\Gamma$-dimensions. Let $S \subset R$ be a multiplicatively closed subset. Since each finitely presented $(S^{-1}R) \Gamma$-module is induced from a finitely presented $R \Gamma$-module (clear denominators in a presentation matrix) we can without loss of generality assume that $R$ is a field. In the following we will work for simplicity with the maximal choice $R = {{\mathbb C}}$.
Let us describe a candidate for $\Lambda$. We denote by $\frac{1}{\# {\mathcal{F}}in \Gamma } {{\mathbb Z}}$ the additive subgroup of ${{\mathbb R}}$ which is generated by the set of numbers $\{ \frac{1}{H} \; | \; H \mbox{ a finite subgroup of } \Gamma \}$. If there is a bound on the orders of finite subgroups then $\frac{1}{\# {\mathcal{F}}in \Gamma} {{\mathbb Z}}= \frac{1}{l} {{\mathbb Z}}$ where $l$ is the least common multiple of the orders of finite subgroups. If $\Gamma$ is torsionfree then $\frac{1}{ \# {\mathcal{F}}in \Gamma } {{\mathbb Z}}= {{\mathbb Z}}$.
The following Conjecture turned out to be too optimistic in general (compare Remark \[counterexample\] below). But it still has a chance of being true if one additionally assumes a bound on the orders of finite subgroups.
\[strong-Atiyah\] Let $M$ be a finitely presented ${{\mathbb C}}\Gamma$-module then $$\dim_{{\mathcal{U}}\Gamma} M \otimes_{{{\mathbb C}}\Gamma} {\mathcal{U}}\Gamma \in \frac{1}{\# {\mathcal{F}}in \Gamma} {{\mathbb Z}}.$$
We will see below in \[atiyah-implies-zero\] that this Conjecture implies the Zero-Divisor Conjecture.
As explained above the conjecture makes sense with any field ${{\mathbb F}}$ such that ${{\mathbb Q}}\subset {{\mathbb F}}\subset {{\mathbb C}}$ as coefficients for the group ring. With ${{\mathbb F}}= {{\mathbb Q}}$ the conjecture is equivalent to the corresponding conjecture about the values of $L^2$-Betti numbers. The conjecture with ${{\mathbb F}}= {{\mathbb C}}$ clearly implies the conjecture formulated with smaller fields.
To get a first idea let us dicuss the Conjecture in the easy case where $\Gamma$ is the infinite cyclic group. We have already seen in Example \[example-infinite-cyclic\] that in this case the inclusion ${{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma$ can be identified with ${{\mathbb C}}[ z^{\pm 1} ] \subset L( S^1 )$, the Laurent polynomials considered as functions on $S^1$ inside the algebra of all measurable functions on $S^1$. Clearly ${{\mathbb C}}\Gamma$ corresponds to ${{\mathbb C}}[ z^{\pm 1} ]$. The crucial observation now is that in this case we find a field in between ${{\mathbb C}}\Gamma$ and ${\mathcal{U}}\Gamma$. Let ${{\mathbb C}}( z )$ denote the field of fractions of the polynomial ring ${{\mathbb C}}[ z ]$ then we have $${{\mathbb C}}[ z^{\pm 1} ] \subset {{\mathbb C}}( z ) \subset L ( S^1 ).$$ Now let $M$ be a finitely presented ${{\mathbb C}}[z^{\pm 1}]$-module then $M \otimes_{{{\mathbb C}}[z^{\pm 1} ] } {{\mathbb C}}( z )$ is a finitely generated free ${{\mathbb C}}(z)$-module because ${{\mathbb C}}(z)$ is a field and hence $M \otimes_{{{\mathbb C}}[z^{\pm 1} ] } L (S^1 )$ is a finitely generated free $L(S^1)$-module. In particular its $\Gamma$-dimension is an integer as predicted by Conjecture \[strong-Atiyah\].
Note that ${{\mathbb C}}( z )$ is not contained in the group von Neumann algebra $L^{\infty} (S^1 )$ because a rational function like for example $z \mapsto \frac{1}{z-1}$ which has a pole on $S^1$ can not be essentially bounded. It hence was crucial for this proof that we had the algebra of affiliated operators ${\mathcal{U}}\Gamma$, here $L( S^1 )$, available.
The following generalizes these simple ideas.
\[strategy\] Suppose the inclusion map ${{\mathbb C}}\Gamma \to {\mathcal{U}}\Gamma$ factorizes over a ring ${\mathcal{S}}\Gamma$ such that the following two conditions are fulfilled.
1. The composite map $$\xymatrix{
{{\rm colim }}_{H \in {\mathcal{F}}in \Gamma} K_0 ( {{\mathbb C}}H ) \ar[r] & K_0 ( {{\mathbb C}}\Gamma ) \ar[r] & K_0 ({\mathcal{S}}\Gamma )
}$$ is surjective.
2. The ring ${\mathcal{S}}\Gamma$ is von Neumann regular.
Then Conjecture \[strong-Atiyah\] holds for the group $\Gamma$.
In the source of the map in (K) the colimit is taken over the finite subgroups of $\Gamma$. The structure maps in the colimit are induced by inclusions $K \subset H$ and conjugation maps $c_g: H \to H^g$, $ h \mapsto ghg^{-1}$.
We will see below (compare Theorem \[main-linnell\]) that there is a reasonably large class of groups for which a factorization of the inclusion ${{\mathbb C}}\Gamma \to {\mathcal{U}}\Gamma$ as required above is known to exist. In order to prove Proposition \[strategy\] we need one more fact about $\Gamma$-dimensions.
\[compatible-with-induction\] The $\Gamma$-dimension is compatible with induction, i.e. if $G$ is a subgroup of $\Gamma$ then there is a natural inclusion ${\mathcal{U}}G \subset {\mathcal{U}}\Gamma$ and for a finitely generated projective ${\mathcal{U}}G$-module $P$ we have $$\dim_{{\mathcal{U}}\Gamma} P \otimes_{{\mathcal{U}}G} {\mathcal{U}}\Gamma = \dim_{{\mathcal{U}}G} P.$$
There exists a natural inclusion $i: {\mathcal{U}}G \to {\mathcal{U}}\Gamma$ which extends the inclusion $i: {\mathcal{N}}G \to {\mathcal{N}}\Gamma$ because ${\mathcal{U}}G$ is the Ore localization of ${\mathcal{N}}G$. The latter inclusion is compatible with the trace, i.e. ${{\rm tr}}_{\Gamma}( i(a) ) = {{\rm tr}}_G ( a )$ for $a \in {\mathcal{N}}\Gamma$, see Lemma 1.24 in [@Lueck(2002)]. The claim follows from these facts.
Let $M$ be a finitely presented ${{\mathbb C}}\Gamma$-module. Then also $M \otimes_{{{\mathbb C}}\Gamma} {\mathcal{S}}\Gamma$ is finitely presented and hence finitely generated projective by \[vNregular-definition\] (ii) because we assume that ${\mathcal{S}}\Gamma$ is von Neumann regular. In particular $M \otimes_{{{\mathbb C}}\Gamma} {\mathcal{S}}\Gamma$ defines a class in $K_0 ( {\mathcal{S}}\Gamma )$. Our second assumption implies that this class comes from ${{\rm colim }}_{H \in {\mathcal{F}}in \Gamma} K_0 ( {{\mathbb C}}H )$ via the natural map. It remains to check that the composition $$\xymatrix{
{{\rm colim }}_{H \in {\mathcal{F}}in \Gamma} K_0 ( {{\mathbb C}}H ) \ar[r] &
K_0 ( {{\mathbb C}}\Gamma ) \ar[r] & K_0 ( {\mathcal{S}}\Gamma ) \ar[r] &
K_0( {\mathcal{U}}\Gamma ) \ar[r]^-{\dim_{\Gamma}} & {{\mathbb R}}}$$ lands inside the subgroup $\frac{1}{\# {\mathcal{F}}in \Gamma } {{\mathbb Z}}$ of ${{\mathbb R}}$. But from Example \[example-finite\] together with Proposition \[compatible-with-induction\] we conclude that for a finite subgroup $H$ and a finitely generated projective ${{\mathbb C}}H$-module $P$ we have $$\dim_{{\mathcal{U}}\Gamma} P \otimes_{{{\mathbb C}}H} {\mathcal{U}}\Gamma = \dim_{{{\mathbb C}}H} P = \frac{1}{\# H} \dim_{{{\mathbb C}}} P.$$
From \[vNregular-definition\] \[vNregular-definition-vier\] it follows that the homomorphic image of a von Neumann regular ring is again von Neumann regular. In particular the image of ${\mathcal{S}}\Gamma$ in ${\mathcal{U}}\Gamma$ would be von Neumann regular if ${\mathcal{S}}\Gamma$ is. (But it is not clear that the induced map for $K_0$ is surjective, compare Question \[question-surjective\].)
\[note-vNr-bound-semisimple\] Suppose ${\mathcal{S}}\Gamma$ is a subring of ${\mathcal{U}}\Gamma$ which contains ${{\mathbb C}}\Gamma$. If we assume the properties (K) and (R) and additionally we assume that $\Gamma$ has a bound on the orders of finite subgroups, then ${\mathcal{S}}\Gamma$ is semisimple.
The assumptions imply that the projective rank function $P \mapsto \dim_{\Gamma} P \otimes_{{\mathcal{S}}\Gamma} {\mathcal{U}}\Gamma$ for finitely generated ${\mathcal{S}}\Gamma$ modules takes values in $\frac{1}{l} {{\mathbb Z}}$, where $l$ is the least common multiple of the orders of finite subgroups. Since each finitely generated projective ${\mathcal{S}}\Gamma$-module is a subset of a ${\mathcal{U}}\Gamma$-module it is easy to see that the projective rank function is faithful. In order to prove that a von Neumann regular ring is semisimple it suffices to show that there are no infinite chains of ideals, see page 21 in [@Goodearl(1979)]. Since each ideal is a direct summand of ${\mathcal{S}}\Gamma$ and each subideal of a given ideal is a direct summand this can be checked using the faithful projective rank function with values in $\frac{1}{l} {{\mathbb Z}}$.
\[counterexample\] The lamplighter group is the semidirect product of ${{\mathbb Z}}$ and $\bigoplus_{-\infty}^{\infty} {{\mathbb Z}}/ 2$ where ${{\mathbb Z}}$ acts via shift on $\bigoplus_{-\infty}^{\infty} {{\mathbb Z}}/ 2$. The orders of finite subgroups that occur are precisely all powers of $2$. Conjecture \[strong-Atiyah\] hence predicts ${{\mathbb Z}}[ \frac{1}{2} ]$ as the range for the dimensions. However in [@Grigorchuk-Linnell-Schick-Zuk(2000)] a finitely presented ${{\mathbb Q}}\Gamma$-module is constructed whose $\Gamma$-dimension is $\frac{1}{3}$. \[Further reading...???\]
Noncommutative Localization {#section-localization}
===========================
Our next aim is to present several candidates for the ring ${\mathcal{S}}\Gamma$ which appears in Proposition \[strategy\]. In order to do this we first want to fix some language and review a couple of concepts from the theory of localization for noncommutative rings. For more on this subject the reader should consult Chapter II in [@Stenstroem(1975)], Chapter 7 in [@Cohn(1985)] and Chapter 4 in [@Schofield(1985)].\
Ore Localization {#ore-localization .unnumbered}
----------------
Classically the starting point for the localization of rings is the wish that certain [*elements*]{} in the ring should become invertible. In mathematical terms we have the following universal property.
Let $T \subset R$ be a subset which does not contain any zero-divisors. A ring homomorphism $f:R \to S$ is called $T$-inverting if $f(t)$ is invertible for all $t \in T$. A $T$-inverting ringhomomorphism $i:R \to R_T$ is called universally $T$-inverting if it has the following universal property: given any $T$-inverting ring homomorphism $f:R \to S$ there exists a unique ring homomorphism $\Phi: R_T \to S$ such that $$\xymatrix{
& & R_T \ar[d]^-{\Phi} \\
R \ar[urr]^-i \ar[rr]^-f & & S
}$$ commutes.
A generator and relation construction shows that there always exists a universal $T$-inverting ring and as usual it is unique up to canonical isomorphism. Given a ring homomorphism $R \to S$ let us agree to write $$T(R \to S)$$ for the set of elements in $R$ which become invertible in $S$. If one replaces $T$ by $\overline{T}=T(R \to R_T)$ the universal inverting ring does not change. We can hence always assume that $T$ is multiplicatively closed. A natural maximal choice for $T$ is the set ${{\rm NZD}}( R )$ of all non-zerodivisors of $R$.
If the ring $R$ is commutative it is well known that there is a model for $R_T$ whose elements are “fractions” or more precisely equivalence classes of pairs $(a,t) \in R \times T$. For noncommutative rings the situation is more complicated. It goes back to Ore \[reference !!!!\] that under a suitable assumption such a calculus of fractions still exists.
A multiplicatively closed subset $T \subset R$ which does not contain zero-divisors or zero itself satisfies the right Ore-condition if for given $(a,s)\in R \times T$ there always exists a $(b,t) \in R \times T$ such that $at=sb$.
It is clear that this condition is necessary if a calculus of right fractions exists because we need to be able to write a given wrong way (left) fraction $s^{-1}a$ as $bt^{-1}$. It is a bit surprising that this is the only condition.
\[proposition-ore-localization\] Let $T \subset R$ be a multiplicatively closed subset without zero divisors which satisfies the right Ore condition, then there exists a ring $RT^{-1}$ and a universal $T$-inverting ring-homomorphism $i: R \to R T^{-1}$ such that every element of $RT^{-1}$ can be written as $i(a)i(t)^{-1}$ with $(a,t) \in R \times T$.
Elements in $RT^{-1}$ are equivalence classes of pairs $(a,t) \in R \times T$. The pair $(a,t)$ is equivalent to $(b,s)$ if there exist elements $u$, $v \in R$ such that $au=bv$, $su=tv$ and $su=tv \in S$. For more details see Chapter II in [@Stenstroem(1975)].
\[remark-ore-flat\] Ore-localization is an exact functor, i.e. $RT^{-1}$ is a flat $R$-module, see page 57 in [@Stenstroem(1975)].
\[example-ore-false\] Let $\Gamma$ be the free group on two generators $x$ and $y$. The group ring ${{\mathbb C}}\Gamma$ does not satisfy the Ore condition with respect to the set ${{\rm NZD}}( {{\mathbb C}}\Gamma )$ of all non-zerodivisors. Let $C \subset \Gamma$ be the infinite cyclic subgroup generated by $x$. Now $x-1$ is a non-zerodivisor since it becomes invertible in ${\mathcal{U}}C$ (compare Example \[example-infinite-cyclic\]) and therefore in the overring ${\mathcal{U}}\Gamma$. In fact every non-trivial element in ${{\mathbb C}}\Gamma$ is a non-zerodivisor since one can embed ${{\mathbb C}}\Gamma$ in a skew field. The Ore condition would imply the existence of $(b,t) \in {{\mathbb C}}\Gamma \times
{{\rm NZD}}( {{\mathbb C}}\Gamma )$ with $(y-1)t = (x-1) b$ alias $$(x-1)^{-1} (y-1) = bt^{-1} .$$ This implies that $(-b, t)^{tr}$ is in the kernel of the map $(x-1 , y-1 ): {{\mathbb C}}\Gamma^2 \to {{\mathbb C}}\Gamma$. But this map is injective, compare Example \[example-monotone-false\].
Localizing Matrices {#localizing-matrices .unnumbered}
-------------------
Instead of elements one can try to invert maps. Let $\Sigma$ be a set of homomorphisms between right $R$-modules. A ring homomorphism $R \to S$ is called $\Sigma$-inverting if for every map $\alpha \in \Sigma$ the induced map $\alpha \otimes_R {{\rm id}}_S$ is an isomorphism.
A $\Sigma$-inverting ring homomorphism $i: R \to R_{\Sigma}$ is called universal $\Sigma$-inverting if it has the following universal property. Given any $\Sigma$-inverting ring homomorphism $f:R \to S$ there exists a unique ring homomorphism $\Psi: R_{\Sigma} \to S$ such that the following diagram commutes. $$\xymatrix{
& & R_{\Sigma} \ar[d]^-{\Psi} \\
R \ar[rr]^-f \ar[urr]^-{i} & & S .
}$$
From now on let us assume that $\Sigma$ is a set of matrices over $R$. For a ring homomorphism $R \to S$ we will write $$\Sigma ( R \to S )$$ for the set of all matrices over $R$ which become invertible over $S$. One can always replace a given set of matrices $\Sigma$ by $\overline{\Sigma} = \Sigma ( R \to R_{\Sigma} )$ without changing the universal $\Sigma$-inverting ringhomomorphism. There are different constructions which prove the existence of a universal $\Sigma$-inverting ring homomorphism. One possibility is a generator and relation construction where one starts with the free ring on a set of symbols $\overline{a}_{i,j}$ where $(a_{i,j})$ runs through the matrices in $\Sigma$ and imposes the relations which are given in matrix form as $\overline{A} A = A \overline{ A } = 1$, compare Theorem 2.1 in [@Cohn(1985)]. For more information the reader should consult Chapter 7 in [@Cohn(1985)] and Chapter 4 in [@Schofield(1985)].
Another construction due to Malcolmson [@Malcolmson(1982)] (see also [@Beachy(1993)]), \[in this volume?\] a kind of calculus of fractions for matrices, allows a certain amount of control over the ring $R_{\Sigma}$.
As an easy example we would like to mention the following: A set of matrices is lower multiplicatively closed if $1 \in \Sigma$ and $a$, $b \in \Sigma$ implies that $$\left( \begin{array}{cc} a & 0 \\ c & b \end{array} \right) \in \Sigma$$ for arbitrary matrices $c$ of suitable size. Observe that $\Sigma( R \to S)$ is always lower multiplicatively closed.
\[cramers-rule\] Let $R$ be a ring and $\Sigma$ be a lower multiplicatively closed set of matrices over $R$ then every matrix $a$ over $R_{\Sigma}$ satisfies an equation of the form $$s \left( \begin{array}{cc} 1 & 0 \\ 0 & a \end{array} \right)
\left( \begin{array}{cc} 1 & x \\ 0 & 1 \end{array} \right) = b$$ with $s \in \Sigma$, $x \in M ( R_{\Sigma} )$ and $b \in M ( R )$.
See Theorem 4.3 on page 53 in [@Schofield(1985)].
In particular every matrix $a$ over $R_{\Sigma}$ is stably associated over $R_{\Sigma}$ to a matrix $b$ over $R$, i.e. there exist invertible matrices $c$, $d \in GL(R_{\Sigma})$ such that $$c \left( \begin{array}{cc} a & 0 \\ 0 & 1_n \end{array} \right) d^{-1} =
\left( \begin{array}{cc} b & 0 \\ 0 & 1_m \end{array} \right)$$ with suitable $m$ and $n$.\
Division Closure and Rational Closure {#division-closure-and-rational-closure .unnumbered}
-------------------------------------
Recall that for a given ring homomorphism $R \to S$ we denoted by $T( R \to S)$ the set of all elements in $R$ which become invertible in $S$ and by $\Sigma( R \to S)$ the set of all matrices over $R$ that become invertible over $S$. The universal localizations $R_{T(R \to S)}$ and $R_{\Sigma( R \to S )}$ come with a natural map to $S$. In the case where $R \to S$ is injective one may wonder whether these maps embed the universal localizations into $S$. The intermediate rings in the following definition serve as potential candidates for such embedded versions of the universal localizations.
\[definition-division-rational\] Let $S$ be a ring.
1. A subring $R \subset S$ is called division closed in $S$ if $T(R \subset S) = R^{\times}$, i.e. for every element $r \in R$ which is invertible in $S$ the inverse $r^{-1}$ lies already in $R$.
2. A subring $R \subset S$ is called rationally closed in $S$ if $\Sigma( R \subset S)= {{\rm GL}}(R)$, i.e.for every matrix $A$ over $R$ which is invertible over $S$ the entries of the inverse matrix $A^{-1}$ are all in $R$.
3. Given a subring $R \subset S$ the division closure of $R$ in $S$ denoted $${\mathcal{D}}( R \subset S)$$ is the smallest division closed subring of $S$ which contains $R$.
4. Given a subring $R \subset S$ the rational closure of $R$ in $S$ denoted by $${\mathcal{R}}( R \subset S)$$ is the smallest rationally closed subring of $S$ containing $R$.
Note that the intersection of division closed intermediate rings is again division closed and similarly for rationally closed rings. This proves the existence of the division and rational closure. Moreover we really have closure-operations, i.e. $$\begin{aligned}
{\mathcal{D}}( {\mathcal{D}}(R \subset S) \subset S) & = & {\mathcal{D}}( R \subset S) \mbox{ and } \\
{\mathcal{R}}( {\mathcal{R}}(R \subset S) \subset S) & = & {\mathcal{R}}( R \subset S).\end{aligned}$$ In [@Cohn(1985) Chapter 7, Theorem 1.2] it is shown that the set $$\begin{aligned}
\label{cohns-description}
\{ a_{i,j} \in S \; | \; ( a_{i,j} ) \mbox{ invertible over } S, \; ( a_{i,j} )^{-1} \mbox{ matrix over } R \}\end{aligned}$$ is a subring of $S$ and that it is rationally closed. Since this ring is contained in ${\mathcal{R}}( R \subset S)$ the two rings coincide. The following observation is very useful in our context.
\[universally-closed\] A von Neumann regular ring $R$ is division closed and rationally closed in every overring.
Suppose $a \in R$ is not invertible in $R$, then the corresponding multiplication map $l_a: R \to R$ is not an isomorphism. Therefore the kernel or the cokernel is non-trivial. Both split of as direct summands because the ring is von Neumann regular. The corresponding projection onto the kernel or cokernel is given by left multiplication with a suitable idempotent. This idempotent shows that $a$ must be a zerodivisor and hence can not become invertible in any overring. A matrix ring over a von Neumann regular ring is again von Neumann regular and the same reasoning applied to matrix rings over $R$ yields that $R$ is also rationally closed in every overring.
In particular note that once we know that the division closure ${\mathcal{D}}( R \subset S)$ is von Neumann regular then it coincides with the rational closure ${\mathcal{R}}( R \subset S )$. The following proposition relates the division respectively rational closure to the universal localizations $R_{T(R \subset S)}$ and $R_{\Sigma ( R \subset S )}$.
\[abstract-embedded\] Let $R \subset S$ be a ring extension.
1. The map $R_{T( R \subset S ) } \to S$ given by the universal property factorizes over the division closure. $$\xymatrix{
& R_{T ( R \subset S )} \ar[d]^-{\Phi} & \\
R \ar[r]^-{\subset} \ar[ur] & {\mathcal{D}}( R \subset S) \ar[r]^-{\subset} & S
}$$
2. If the pair $(R , T(R \subset S))$ satisfies the right Ore condition, then $\Phi$ is an isomorphism.
3. The map $R_{\Sigma ( R \subset S )} \to S$ given by the universal property factorizes over the rational closure. $$\xymatrix{
& R_{\Sigma ( R \subset S )} \ar@{->>}[d]^-{\Psi} & \\
R \ar[ur] \ar[r]^-{\subset} & {\mathcal{R}}( R \subset S ) \ar[r]^-{\subset} & S
}$$ The map $\Psi$ is always surjective.
\(i) This follows from the definitions. (ii) Note that $T( R \subset S)$ always consists of non-zerodivisors. Thus we can choose a ring of right fractions as a model for $R_{T(R \subset S )}$. Every element in ${{\rm im}}\Phi$ is of the form $at^{-1}$ with $t \in T( R \subset S)$. Such an element is invertible in $S$ if and only if $a \in T ( R \subset S)$. We see that the image of $\Phi$ is division closed and hence $\Phi$ is surjective. On the other hand the abstract fraction $at^{-1} \in
R T(R \subset T)^{-1}$ is zero if and only if $a =0$ because $T( R \subset S)$ contains no zerodivisors, so $\Phi$ is injecitve. (iii) Only the last statement is maybe not obvious. By Cohn’s description of the rational closure (compare (\[cohns-description\])) we need to find a preimage for $a_{i,j}$, where $(a_{i,j})$ is a matrix invertible over $S$ whose inverse lies over $R$. The generator and relation construction of the universal localization immediately gives such an element.
In general it is not true that the map $\Psi$ is injective.
Some Candidates for ${\mathcal{S}}\Gamma$ {#section-candidates}
=========================================
We are now prepared to describe the candidates for the ring ${\mathcal{S}}\Gamma$ which appears in Proposition \[strategy\]. We consider the ring extension ${{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma$ and define $$\begin{aligned}
{\mathcal{D}}\Gamma & = & {\mathcal{D}}( {{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma ), \\
{\mathcal{R}}\Gamma & = & {\mathcal{R}}( {{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma ), \\
{{\mathbb C}}\Gamma_T & = & {{\mathbb C}}\Gamma_{T ( {{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma )} \mbox{ and } \\
{{\mathbb C}}\Gamma_{\Sigma } & = & {{\mathbb C}}\Gamma_{ \Sigma( {{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma ) } .\end{aligned}$$
These rings are organized in the following diagram $$\begin{aligned}
\label{diagram-candidates}
\xymatrix{
& {{\mathbb C}}\Gamma_T \ar[r] \ar[d] & {{\mathbb C}}\Gamma_{\Sigma} \ar[d] & \\
{{\mathbb C}}\Gamma \ar[r]^-{\subset} \ar[ur] & {\mathcal{D}}\Gamma \ar[r]^-{\subset} & {\mathcal{R}}\Gamma \ar[r]^-{\subset} & {\mathcal{U}}\Gamma
}\end{aligned}$$
A first hint that the rational or division closure may be a good candidate for ${\mathcal{S}}\Gamma$ is the following result which is implicit in [@Linnell(1993)]. At the same time its proof illustrates the usefulness of Cramer’s rule \[cramers-rule\].
\[RG-skew\] If $\Gamma$ is a torsionfree group then the Strong Atiyah Conjecture \[strong-Atiyah\] implies that ${\mathcal{R}}\Gamma$ is a skew field.
For $x \in {\mathcal{U}}\Gamma$ let $l_x: {\mathcal{U}}\Gamma \to {\mathcal{U}}\Gamma$ denote left multiplication with $x$. From the additivity and faithfulness of the dimension it follows that $x$ is invertible if and only if $\dim {{\rm im}}( l_x ) = 1$ or equivalently $\dim \ker ( l_x )=0$ or equivalently $\dim {{\rm coker}}(l_x) = 0$. Now let $X$ be a matrix over ${\mathcal{R}}\Gamma$ then by \[abstract-embedded\] (iii) we know that we can lift it to a matrix over ${{\mathbb C}}\Gamma_{\Sigma}$. Using Cramer’s rule \[cramers-rule\] and projecting down again we see that we can find invertible matrices $A$ and $B \in GL ( {\mathcal{R}}\Gamma )$ such that $$C= A \left( \begin{array}{cc} 1_n & 0 \\ 0 & X \end{array} \right) B$$ is a matrix over ${{\mathbb C}}\Gamma$. In particular if $0 \neq x \in {\mathcal{R}}\Gamma $ then for $X=(x)$ we know that there exists an $n$ such that $$\dim {{\rm im}}(l_x) +n = \dim ( {{\rm im}}\left( \begin{array}{cc} 1_n & 0 \\ 0 & l_x \end{array} \right) ) = \dim ( {{\rm im}}C ) \in {{\mathbb Z}}$$ because we assume for the matrix $C$ over ${{\mathbb C}}\Gamma$ that the dimension of its image is an integer. It follows that $\dim {{\rm im}}(l_x) = 1$ and hence that $x$ is invertible in ${\mathcal{R}}\Gamma$.
\[note-DG-RG\] If one of the rings ${\mathcal{D}}\Gamma$ or ${\mathcal{R}}\Gamma$ is a skew field then so is the other and the two coincide.
If ${\mathcal{D}}\Gamma$ is a skew field then it is also rationally closed, see \[universally-closed\]. If ${\mathcal{R}}\Gamma$ is a skew field then ${\mathcal{D}}\Gamma$ is a division closed subring of a skew field and hence itself a skew field.
\[atiyah-implies-zero\] The Atiyah Conjecture \[strong-Atiyah\] implies the Zero-Divisor Conjecture, i.e. the conjecture that the complex group ring of a torsionfree group does not contain any zero-divisors.
One can show that for a torsionfree amenable group the Atiyah Conjecture \[strong-Atiyah\] is equivalent to the Zero-Divisor Conjecture, see Lemma 10.16 in [@Lueck(2002)].
Another natural question is in how far the rings discussed above depend functorially on the group. Since an arbitrary group homomorphism $G \to G^{\prime}$ does not induce a map from ${\mathcal{U}}G$ to ${\mathcal{U}}G^{\prime}$ we can not expect functoriality but at least we have the following.
An injective group homomorphism induces maps on the rings ${\mathcal{D}}\Gamma$, ${\mathcal{R}}\Gamma$, ${{\mathbb C}}\Gamma_{T}$ and ${{\mathbb C}}\Gamma_{\Sigma}$. These maps are compatible with the maps in diagram \[diagram-candidates\] above.
We already know that the inclusion ${{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma$ is functorial for injective group homomorphisms. Let $G$ be a subgroup of $\Gamma$. Since ${\mathcal{U}}G$ is von Neumann regular it is division closed and rationally closed in every overring, compare \[universally-closed\]. Therefore ${\mathcal{D}}\Gamma \cap {\mathcal{U}}G$ is division-closed in ${\mathcal{U}}\Gamma$ and ${\mathcal{D}}G \subset {\mathcal{D}}\Gamma \cap {\mathcal{U}}G \subset {\mathcal{D}}\Gamma$. Analogously one argues for the rational closure. One immediately checks that $T( {{\mathbb C}}G \subset {\mathcal{U}}G) \subset
T( {{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma)$ and $\Sigma ( {{\mathbb C}}G \subset {\mathcal{U}}G ) \subset \Sigma ( {{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma )$. The universal properties imply the statement for ${{\mathbb C}}\Gamma_T$ and ${{\mathbb C}}\Gamma_{\Sigma}$.
Linnell’s Result {#section-linnell}
================
Before we state Linnell’s result we would like to introduce the class of groups it applies to.
\[definition-linnells-class\] Let ${\mathcal{C}}$ be the smallest class of groups which has the following properties.
1. Free groups are contained in ${\mathcal{C}}$.
2. If $1 \to G \to \Gamma \to H \to 1$ is an exact sequence of groups such that $G$ lies in ${\mathcal{C}}$ and $H$ is finite or finitely generated abelian then $\Gamma$ lies in ${\mathcal{C}}$.
3. The class ${\mathcal{C}}$ is closed under directed unions, i.e. if a group $\Gamma= \bigcup_{i \in I} \Gamma_i$ is a directed union of subgroups $\Gamma_i$ which lie in ${\mathcal{C}}$ then $\Gamma$ also lies in ${\mathcal{C}}$.
To put this definition into perspective we would like to make a couple of remarks.
\(i) If one replaces (LC1) above by the requirement that the trivial group belongs to ${\mathcal{C}}$ one obtains the smaller class of elementary amenable groups. Compare [@Chou(1980)] and \[.precise.ref..\] [@Kropholler-Linnell-Moody(1988)]. Elementary amenable groups are in particular amenable (see [@Day(1957)]) but it is not easy to find amenable groups that are not elementary amenable [@Grigorchuk(1998)]. A group which contains a non-abelian free subgroup is not amenable.
\(ii) One can show that if $\Gamma$ lies in ${\mathcal{C}}$ and $A$ is an elementary amenable normal subgroup then $\Gamma /A$ also belongs to ${\mathcal{C}}$.
\(iii) The class ${\mathcal{C}}$ is closed under free products.
In [@Linnell(1993)] Linnell proves Conjecture \[strong-Atiyah\] for groups in the class ${\mathcal{C}}$ which additionally have a bound on the orders of finite subgroups. In fact by carefully investigating the proof given there one can obtain the following statements.
\[main-linnell\] Suppose the group $\Gamma$ lies in ${\mathcal{C}}$ and has a bound on the orders of finite subgroups then
1. The composition $${{\rm colim }}_{H \in {\mathcal{F}}in \Gamma} K_0 ( {{\mathbb C}}H ) \to K_0 ( {{\mathbb C}}\Gamma ) \to K_0 ( {\mathcal{D}}\Gamma )$$ is surjective.
2. The ring ${\mathcal{D}}\Gamma$ is semi-simple and hence ${\mathcal{D}}\Gamma= {\mathcal{R}}\Gamma$ by Proposition \[universally-closed\].
As already mentioned this result is essentially contained in [@Linnell(1993)]. In the above formulation it is proven in [@Reich(1999)]. The proof is published in Chapter 10 in [@Lueck(2002)]. Below we will only make a couple of remarks about the proof.
Since we formulated the theorem with the division closure ${\mathcal{D}}\Gamma$ the reader may get the impression that this is the best candidate for an intermediate ring ${\mathcal{S}}\Gamma$ as in Proposition \[strategy\]. But in fact the situation is not so clear. We already stated that ${\mathcal{D}}\Gamma = {\mathcal{R}}\Gamma$ when the theorem applies. Moreover one can show the following.
\[addendum\] $\mbox{}$
1. In the situation of Theorem \[main-linnell\] the natural map ${{\mathbb C}}\Gamma_{ \Sigma } \to {\mathcal{R}}\Gamma$ is an isomorphism and hence ${\mathcal{D}}\Gamma = {\mathcal{R}}\Gamma \cong {{\mathbb C}}\Gamma_{\Sigma}$.
2. If $\Gamma$ lies in the smaller class of elementary amenable groups and has a bound on the orders of finite subgroups then ${{\mathbb C}}\Gamma$ satisfies the the right Ore condition with respect to the set ${{\rm NZD}}( {{\mathbb C}}\Gamma )$ of all non-zerodivisors and this set coincides with $T( {{\mathbb C}}\Gamma \subset {\mathcal{U}}\Gamma)$. Hence ${{\mathbb C}}\Gamma_T$ can be realized as a ring of fractions and the natural map ${{\mathbb C}}\Gamma_T \to {\mathcal{D}}\Gamma$ is an isomorphism, compare \[abstract-embedded\] (iii).
The statement (O) about the Ore localization appears already in [@Linnell(1991)]
We will now make some comments about the proof of Theorem \[main-linnell\] and the Addendum \[addendum\]. As one might guess from the description of the class of groups to which the Theorem (and the Addendum) applies the proof proceeds via transfinite induction on the class of groups, i.e. one proves the following statements.
1. (K), (R) and (U) hold for free groups.
2. If $1 \to G \to \Gamma \to H \to 1$ is an extension of groups where $H$ is finite or infinite cyclic and (K), (R) and (U) hold for $G$ then they hold for $\Gamma$. Similar with (O) replacing (U).
3. If $\Gamma$ is the directed union of the subgroups $\Gamma_i$ and (K), (R) and (U) hold for all $\Gamma_i$ then they hold for $\Gamma$ if $\Gamma$ has a bound on the orders of finite subgroups. Similar with (O) replacing (U).
\[induction-principle\] (I) The Kadison Conjecture says that there are no non-trivial idempotents in the group $C^{\ast}$-algebra of a torsionfree group. Linnell observed that Connes conceptional proof of this Conjecture for the free group on two generators given in [@Connes(1985a)] (see also [@Julg-Valette(1984)]) can be used to verify the stronger Conjecture \[strong-Atiyah\] in this case. Combined with Proposition \[RG-skew\] and Note \[note-DG-RG\] one concludes that ${\mathcal{D}}\Gamma = {\mathcal{R}}\Gamma$ is a skew field. This yields (R) and also (K) since $K_0$ of a skew field is ${{\mathbb Z}}$. Every finitely generated free group is a subgroup of the free group on two generators and every free group is a directed union of finitely generated free subgroups. This is used to pass to arbitrary free groups.
Some rather non-trivial facts about group rings of free groups (see [@Hughes(1970)] and [@Lewin(1974)]) are used to verify that ${\mathcal{R}}\Gamma$ also coincides with the universal localization ${{\mathbb C}}\Gamma_{\Sigma}$ and hence to verify (U) in this case. Recall that we saw in Example \[example-ore-false\] that (O) is false for free groups.
\(II) For information about crossed products we refer the reader to [@Passman(1989)] and to Digression \[digression-crossed-product\] below. If $1 \to G \to \Gamma \to H \to 1$ is an extension of groups then every set-theoretical section $\mu$ of the quotient map $\Gamma \to H$ (we can always assume $\mu(e)=e$ and $\mu(g^{-1})=\mu(g)^{-1}$) allows to describe the group ring ${{\mathbb C}}\Gamma$ as a crossed product ${{\mathbb C}}G \ast H$ of the ring ${{\mathbb C}}G$ with the group $H$. Similarly crossed products ${\mathcal{D}}G \ast H $, ${\mathcal{R}}G \ast H$ and ${{\mathbb C}}G_{\Sigma} \ast H$ exist and can serve as intermediate steps when one tries to prove the statements (R), (K) and (U) for $\Gamma$. For example there are natural inclusions $${{\mathbb C}}G \ast H \to {\mathcal{D}}G \ast H \to {\mathcal{D}}\Gamma.$$ If $H$ is a finite group and ${\mathcal{D}}G$ is semisimple then ${\mathcal{D}}G \ast H$ is semisimple and coincides with ${\mathcal{D}}\Gamma$. It is relatively easy to verify that ${\mathcal{D}}G \ast H$ is noetherian and semiprime if $H$ is an infinite cyclic group and then Goldie’s theorem (a criterion for the existence of an Ore-localization, see Section 9.4 in [@Cohn(1991)]) together with results from [@Linnell(1992)] are used to verify that ${\mathcal{D}}\Gamma$ is the Ore-localization of ${\mathcal{D}}G \ast H$ with respect to the set of all non-zerodivisors. This is roughly the line of argument in order to verify that condition (R) survives extensions by finite or infinite cyclic groups. Once we know that ${\mathcal{D}}\Gamma$ is an Ore localization of ${\mathcal{D}}G \ast H$ we can combine this with the assumption (which implies that ${{\mathbb C}}G_{\Sigma} \ast H \to {\mathcal{D}}G \ast H$ is an isomorphism) in order to verify (U). Similarly iterating Ore localizations one obtains that (O) is stable under extensions with an infinite cyclic group. Moody’s Induction Theorem (see Theorem \[moodys-induction-theorem\] below) plays a crucial role in the argument for (K). Moreover one has to assume that the class of groups which appears in the induction hypothesis is already closed under extensions by finite subgroups. Hence one is forced to start the induction with virtually free groups and in particular one has to prove that (K) holds for such groups. For this purpose results of Waldhausen [@Waldhausen(1978a)] about generalized free products can be used. Moreover the map induced by ${{\mathbb C}}\Gamma \to {{\mathbb C}}\Gamma_{\Sigma}$ on $K_0$ needs to be studied, compare Question \[question-surjective\]. Here it is important to deal with the universal matrix localization.
\(III) If $\Gamma$ is the directed union of the subgroups $\Gamma_i$, $i \in I$ then ${\mathcal{D}}\Gamma$ is the directed union of the subrings ${\mathcal{D}}\Gamma_i$ and similar for ${\mathcal{R}}\Gamma$. A directed union of von Neumann regular rings is again von Neumann regular (use Definition \[vNregular-definition\] (iv)) so ${\mathcal{D}}\Gamma$ is at least von Neumann regular if all the ${\mathcal{D}}\Gamma_i$ are semisimple. The fact that $K$-theory is compatible with colimits yields that (K) holds for $\Gamma$ if it holds for all the $\Gamma_i$. Now the assumption on the bound of the orders of finite subgroups implies that ${\mathcal{D}}\Gamma$ is even semisimple by Note \[note-vNr-bound-semisimple\]. That (U) and (O) are stable under directed unions is straightforward.
\[digression-crossed-product\] A crossed product $R \ast G = ( S, \mu )$ of the ring $R$ with the group $G$ consists of a ring $S$ which contains $R$ as a subring together with an injective map $\mu: G \to S^{\times}$ such that the following holds.
1. The ring $S$ is a free $R$-module with basis $\mu(G)$.
2. For every $g \in G$ the conjugation map $c_{\mu(g)}: S \to S$, $\mu(g) s \mu(g)^{-1}$ can be restricted to $R$.
3. For all $g$, $g^{\prime} \in G$ the element $\tau( g,g^{\prime} )= \mu(g) \mu(g^{\prime}) \mu( gg^{\prime} )^{-1}$ lies in $R^{\times}$.
The Isomorphism Conjecture in algebraic $K$-theory {#section-isomorphism}
==================================================
Condition (K) in Proposition \[strategy\] requires that the composite map $$\xymatrix{
{{\rm colim }}_{H \in {\mathcal{F}}in \Gamma} K_0 ( {{\mathbb C}}H ) \ar[r] & K_0 ( {{\mathbb C}}\Gamma ) \ar[r] & K_0 ( {\mathcal{S}}\Gamma )
}$$ is surjective. About the first map in this composition there is the following conjecture.
\[special-isomorphism-conjecture\] For every group $\Gamma$ the map $$\xymatrix{
{{\rm colim }}_{H \in {\mathcal{F}}in \Gamma} K_0 ( {{\mathbb C}}H ) \ar[r] & K_0 ( {{\mathbb C}}\Gamma )
}$$ is an isomorphism. In particular for a torsionfree group $\Gamma$ we expect $$K_0 ( {{\mathbb C}}\Gamma ) \cong {{\mathbb Z}}.$$
To form the colimit we understand ${\mathcal{F}}in \Gamma$ as the category whose objects are the finite subgroups of $\Gamma$ and whose morphisms are generated by inclusion maps $K \subset H$ and conjugation maps $c_g: H \to H^g$, $h \mapsto ghg^{-1}$ with $g \in \Gamma$. Observe that in the torsionfree case the colimit reduces to $K_0( {{\mathbb C}}) \cong {{\mathbb Z}}$.
In fact Conjecture \[special-isomorphism-conjecture\] would be a consequence of a much more general conjecture which predicts the whole algebraic $K$-theory of a group ring $R \Gamma$ in terms of the $K$-theory of the coefficients and homological data about the group. This more general conjecture is known as the Farrell-Jones Isomorphism Conjecture for algebraic K-theory [@Farrell-Jones(1993a)]. A precise formulation would require a certain amount of preparation and we refer the reader to [@Davis-Lueck(1998)], [@BFJR] and in particular to [@Lueck-Reich(2003)] for more information.
Conjecture \[special-isomorphism-conjecture\] would have the following consequence.
For every finitely generated projective ${{\mathbb C}}\Gamma$-module $P$ we have $$\dim_{\Gamma } ( P \otimes_{{{\mathbb C}}\Gamma} {\mathcal{U}}\Gamma ) \in \frac{1}{\# {\mathcal{F}}in \Gamma } {{\mathbb Z}}.$$
Use Example \[example-finite\] and Proposition \[compatible-with-induction\].
So if all finitely presented ${{\mathbb C}}\Gamma$-modules were also finitely generated projective then the Isomorphism Conjecture would imply the Strong Atiyah Conjecture. But of course this is seldom the case.
Conjecture \[special-isomorphism-conjecture\] is true for infinite cyclic groups and products of such by the Bass-Heller-Swan formula [@Bass-Heller-Swan(1964)]. Cohn’s results in [@Cohn(1964)] imply the Conjecture for free groups. Work of Waldhausen [@Waldhausen(1978a)] deals with generalized free products and HNN-extensions. (The reader should consult [@Ranicki(2001)] and [@Neeman-Ranicki(2002)] for a “noncommutative localization”-perspective on these results.) A version of the following result plays also an important role in the proof of \[main-linnell\].
\[moodys-induction-theorem\] Let $\Gamma$ be a polycyclic-by-finite group then the map $${{\rm colim }}_{H \in {\mathcal{F}}in \Gamma} K_0 ( {{\mathbb C}}H ) \to K_0 ( {{\mathbb C}}\Gamma )$$ is surjective.
See [@Moody(1987)], [@Moody(1989)], [@Cliff-Weiss(1988)] and Chapter 8 in [@Passman(1989)].
What happens if we replace the complex coefficients in Conjecture \[special-isomorphism-conjecture\] by integral coefficients? Thinking about the situation for finite groups it is at first glance very surprising that for infinite groups there are a lot of cases where there are results about $K_0 ( {{\mathbb Z}}\Gamma )$ whereas nothing is known about $K_0 ( {{\mathbb C}}\Gamma )$. See for example [@Farrell-Jones(1993a)]. The reason is that the elements of algebraic $K$-groups of the [*integral*]{} group ring have a topological interpretation. They occur as obstruction groups in certain topological problems. Many people put a lot of effort into solving these topological problems and each time this is successful one obtains a result about the algebraic $K$-groups of ${{\mathbb Z}}\Gamma$.
However with integral coefficients one does not expect an isomorphism as in Conjecture \[special-isomorphism-conjecture\]. In the case where $\Gamma$ is torsionfree one would still expect $K_0( {{\mathbb Z}}\Gamma ) \cong {{\mathbb Z}}$, but in general so called Nil-groups and also negative $K$-groups should enter in a “computation” of $K_0( {{\mathbb Z}}\Gamma )$. Moreover by a result of Swan (see Theorem 8.1 in [@Swan(1960a)]) the map $K_0 ( {{\mathbb Z}}H ) \to K_0 ( {{\mathbb Q}}H )$ is almost the trivial map for a finite group $H$, i.e. the map on reduced $K$-groups $\tilde{K}_0 ( {{\mathbb Z}}H ) \to \tilde{K}_0 ( {{\mathbb Q}}H )$ is trivial. Summarizing: In general in the square $$\xymatrix{
{{\rm colim }}_{H \in {\mathcal{F}}in \Gamma} K_0 ( {{\mathbb Z}}H ) \ar[d] \ar[r] & K_0 ( {{\mathbb Z}}\Gamma ) \ar[d] \\
{{\rm colim }}_{H \in {\mathcal{F}}in \Gamma} K_0 ( {{\mathbb Q}}H ) \ar[r] & K_0 ( {{\mathbb Q}}\Gamma ).
}$$ neither the upper horizontal arrow nor the vertical arrows are surjective. We see that the comparison to the integral group ring is not very useful for the question we are interested in.
The main techniques to prove results about the $K$-theory of ${{\mathbb Z}}\Gamma$ stems from “controlled topology”. See [@Quinn(1987)], [@Quinn(1979)], [@Quinn(1982a)], [@Farrell(1996)], [@Farrell(Trieste)] and [@Jones(Trieste)]. The set-up has been adapted to a more algebraic setting [@Pedersen-Weibel(1989)] and this “controlled algebra” (see [@Pedersen-Weibel(1989)],[@Carlsson(1989)] and [@Pedersen(2000)]) was used successfully to obtain “lower bounds” for the $K$-theory of group rings with arbitrary coefficients under certain curvature conditions on the group [@Carlsson-Pedersen(1995a)].
A result about Conjecture \[special-isomorphism-conjecture\] which uses this “controlled algebra” is the following result from [@BFJR]. Recall that a ring is called (right)-regular if it is right noetherian and every finitely generated right $R$-module admits finite dimensional projective resolution.
Let $\Gamma$ be the fundamental group of a closed riemannian manifold with strictly negative sectional curvature. Let $R$ be a regular ring, e.g. $R = {{\mathbb C}}$ then $$K_0( R ) \cong K_0 ( R \Gamma ).$$ Moreover $K_{-n} ( R \Gamma ) = 0$ and $K_1 ( R \Gamma ) = \Gamma_{{\mathit{ab}}} \otimes_{{{\mathbb Z}}} K_0 ( R ) \oplus K_1 ( R )$, where $\Gamma_{{\mathit{ab}}}$ denotes the abelianized group.
The assumption about $\Gamma$ implies that $\Gamma$ is torsionfree so the above verifies Conjecture \[special-isomorphism-conjecture\].
The author is optimistic that in the near future techniques similar to the ones used in [@BFJR] will lead to further results about Conjecture \[special-isomorphism-conjecture\]. In view of condition (K) in Proposition \[strategy\] the following seems to be an important question.
\[question-surjective\] Are the maps $$\begin{aligned}
K_0 ( {{\mathbb C}}\Gamma ) & \to & K_0 ( {{\mathbb C}}\Gamma_{\Sigma} ), \\
K_0 ( {{\mathbb C}}\Gamma ) & \to & K_0 ( {\mathcal{R}}\Gamma ) \\
\mbox{ or } \quad
K_0 ( {{\mathbb C}}\Gamma ) & \to & K_0 ( {\mathcal{D}}\Gamma ) \end{aligned}$$ surjective?
Note that this is true for groups in Linnell’s class ${\mathcal{C}}$ with a bound on the orders of finite subgroups by Theorem \[main-linnell\] (K).
Exactness Properties
====================
In this section we want to investigate to what extent the functor $- \otimes_{{{\mathbb C}}\Gamma} {\mathcal{U}}\Gamma$ and related functors are exact. Recall that this functor is crucial for the definition of $L^2$-Betti numbers, compare Definition \[definition-l2-betti\].
\[note-elementary-amenable-flat\] If $\Gamma$ is elementary amenable and there is a bound on the orders of finite subgroups then $- \otimes_{{{\mathbb C}}\Gamma} {\mathcal{U}}\Gamma$ is exact.
From Addendum \[addendum\] (O) we know that for these groups ${\mathcal{D}}\Gamma$ is an Ore-localization of ${{\mathbb C}}\Gamma$. In particular in this case $- \otimes_{{{\mathbb C}}\Gamma} {\mathcal{D}}\Gamma$ is exact. Since by Theorem \[main-linnell\] (R) ${\mathcal{D}}\Gamma$ is also semisimple (and hence von Neumann regular) we know that every module is flat over ${\mathcal{D}}\Gamma$.
The following tells us that we cannot always have exactness.
Suppose for the infinite group $\Gamma$ the functor $- \otimes_{{{\mathbb C}}\Gamma} {\mathcal{U}}\Gamma$ is exact, then all $L^2$-Betti numbers and also the Euler-characteristic $\chi^{(2)}( \Gamma )$ of the group $\Gamma$ vanishes.
Flatness implies $$H_p ( C_{\ast} ( E \Gamma ) \otimes_{{{\mathbb Z}}\Gamma} {\mathcal{U}}\Gamma ) = Tor_p^{{{\mathbb Z}}\Gamma} ( {{\mathbb Z}}, {\mathcal{U}}\Gamma )
= Tor_p^{{{\mathbb Z}}\Gamma} ( {{\mathbb Z}}, {{\mathbb Z}}\Gamma ) \otimes_{{{\mathbb Z}}\Gamma} {\mathcal{U}}\Gamma =0$$ for $p >0$. Moreover $b_0^{(2)}(\Gamma)=0$ for every infinite group (see Theorem 6.54 (8) (b) in [@Lueck(2002)]).
In particular we see that for the free group on two generators we cannot have exactness. We saw this phenomenon already in Example \[example-monotone-false\] because exactness of $- \otimes_{{{\mathbb C}}\Gamma} {\mathcal{N}}\Gamma$ would contradict the monotony of the dimension. (Recall from Proposition \[NGUG-properties\] (ii) that $-\otimes_{{\mathcal{N}}\Gamma} {\mathcal{U}}\Gamma$ is always exact.)
More generally we have.
\[note-free-group-not-flat\] If $\Gamma$ contains a nonabelian free group, then neither ${\mathcal{D}}\Gamma$ nor ${\mathcal{R}}\Gamma$, ${{\mathbb C}}\Gamma_{\Sigma}$ or ${\mathcal{U}}\Gamma$ can be flat over ${{\mathbb C}}\Gamma$.
Every free group contains a free group on two generators. Let $G \subset \Gamma$ be a free subgroup on two generators. Let ${{\mathbb C}}G^2 \to {{\mathbb C}}G $ be the injective homomorphism from Example \[example-monotone-false\]. Since ${{\mathbb C}}\Gamma$ is flat over ${{\mathbb C}}G$ we obtain an injective map ${{\mathbb C}}\Gamma^2 \to {{\mathbb C}}\Gamma$. On the other hand since ${\mathcal{D}}G$ is a skew-field we know that the non-trivial kernel of the corresponding map ${\mathcal{D}}G^2 \to {\mathcal{D}}G$ (which must appear for dimension reasons since $-\otimes_{{\mathcal{D}}G} {\mathcal{U}}G$ is exact and the $\Gamma$-dimension is faithful) is a one-dimensional free module which splits off ${\mathcal{D}}G^2$ as a direct summand. The same remains true for every overring of ${\mathcal{D}}G$. In particular for ${\mathcal{D}}\Gamma$, ${\mathcal{R}}\Gamma$ and ${\mathcal{U}}\Gamma$. But also for ${{\mathbb C}}\Gamma_{\Sigma}$ because ${\mathcal{D}}G = {\mathcal{R}}G \cong {{\mathbb C}}G_{\Sigma}$ by Addendum \[addendum\] (O) and since there is a natural map ${{\mathbb C}}G_{\Sigma} \to {{\mathbb C}}\Gamma_{\Sigma}$.
In this context we would also like to mention the following result from [@Lueck(1998b)].
If $\Gamma$ is amenable then ${\mathcal{N}}\Gamma$ (and hence ${\mathcal{U}}\Gamma$) is dimension-flat over ${{\mathbb C}}\Gamma$, i.e. for $p >0$ and every ${{\mathbb C}}\Gamma$-module $M$ we have $$\dim_{{\mathcal{N}}\Gamma} Tor_p^{{{\mathbb C}}\Gamma} ( M , {\mathcal{N}}\Gamma ) = \dim_{{\mathcal{U}}\Gamma} Tor_p^{{{\mathbb C}}\Gamma} ( M , {\mathcal{U}}\Gamma) = 0.$$
See [@Lueck(1998b)] or Theorem 6.37 on page 259 in [@Lueck(2002)] and recall that ${\mathcal{U}}\Gamma$ is flat over ${\mathcal{N}}\Gamma$ and the ${\mathcal{U}}\Gamma$-dimension and the ${\mathcal{N}}\Gamma$-dimension are compatible.
Given these facts it is tempting to conjecture that $- \otimes_{{{\mathbb C}}\Gamma} {\mathcal{U}}\Gamma$ is exact if and only if $\Gamma$ is amenable. However in [@Linnell-Lueck-Schick(2002)] it is shown that the condition about the bound on the orders of finite subgroups in Note \[note-elementary-amenable-flat\] is necessary.
Let $H$ be a nontrivial finite group and let $H \wr {{\mathbb Z}}$ denote the semidirect product $\bigoplus_{- \infty}^{\infty} H {\rtimes}{{\mathbb Z}}$, where ${{\mathbb Z}}$ is acting via shift on the $\bigoplus_{-\infty}^{\infty} H$. Then neither ${\mathcal{D}}\Gamma$ nor ${\mathcal{U}}\Gamma$ is flat over ${{\mathbb C}}\Gamma$ (see Theorem 1 in [@Linnell-Lueck-Schick(2002)]).
The main purpose of this section is to prove the following result which measures the deviation from exactness for groups in Linnell’s class.
\[application\] Let $\Gamma$ be in the class ${\mathcal{C}}$ with a bound on the orders of finite subgroups, then $${{\rm Tor}}_p^{{{\mathbb C}}\Gamma } ( - ; {\mathcal{D}}\Gamma) =0 \hspace{3em}
\mbox{for all } p \geq 2.$$
Note that for these groups ${\mathcal{D}}\Gamma={\mathcal{R}}\Gamma \cong {{\mathbb C}}\Gamma_{\Sigma}$ is semisimple and therefore the functor $ - \otimes_{{\mathcal{D}}\Gamma} {\mathcal{U}}\Gamma$ is exact. The functor $-\otimes_{{{\mathbb Z}}\Gamma } {{\mathbb C}}\Gamma$ is always exact. Therefore we obtain the corresponding statements for ${{\rm Tor}}_p^{{{\mathbb C}}\Gamma}( - ; {\mathcal{U}}\Gamma)$, ${{\rm Tor}}_p^{{{\mathbb Z}}\Gamma} ( - ; {\mathcal{D}}\Gamma)$ and ${{\rm Tor}}_p^{{{\mathbb Z}}\Gamma} (- ; {\mathcal{U}}\Gamma)$.
As an immediate consequence we obtain interesting examples of stably flat universal localizations.
If $\Gamma$ lies in Linnell’s class ${\mathcal{C}}$ and has a bound on the orders of finite subgroups then ${\mathcal{D}}\Gamma \cong {{\mathbb C}}\Gamma_{\Sigma}$ is stably flat over ${{\mathbb C}}\Gamma$, i.e. we have $$Tor_p^{{{\mathbb C}}\Gamma} ( {\mathcal{D}}\Gamma , {\mathcal{D}}\Gamma ) = 0 \hspace{3em} \mbox{ for all } p \geq 1.$$
We know that ${\mathcal{D}}\Gamma \cong {{\mathbb C}}\Gamma_{\Sigma}$ is a universal localization of ${{\mathbb C}}\Gamma$ and hence ${{\mathbb C}}\Gamma \to {{\mathbb C}}\Gamma_{\Sigma}$ is an epimorphism in the category of rings, see page 56 in [@Schofield(1985)]. By Theorem 4.8 b) in [@Schofield(1985)] we know that $Tor_1^{{{\mathbb C}}\Gamma} ( {\mathcal{D}}\Gamma , {\mathcal{D}}\Gamma ) = 0$. For $p \geq 2$ the result follows from Theorem \[application\].
Recent work of Neeman and Ranicki [@Neeman-Ranicki(2002)] shows that for universal localizations which are stably flat there exists a long exact localization sequence which extends Schofield’s localization sequence for universal localizations (see Theorem 4.12 in [@Schofield(1985)]) to the left. In the case of Ore-localizations the corresponding sequence was known for a long time, see [@Gersten(1974)], [@Grayson(1980)], [@Weibel-Yao(1992)] and [@Thomason-Trobaugh(1990)]. Observe that because of Note \[note-free-group-not-flat\] we know that whenever $\Gamma$ contains a free group ${{\mathbb C}}\Gamma_{\Sigma}$ cannot be an Ore-localization.
Here is another consequence of Theorem \[application\].
\[euler char\] If the infinite group $\Gamma$ belongs to ${\mathcal{C}}$ and has a bound on the orders of finite subgroups, then $$\chi^{(2)}(\Gamma) \leq 0.$$
Since the group is infinite we have $b_0^{(2)}( \Gamma )= 0$. Because of $$H_p( \Gamma ; {\mathcal{U}}\Gamma )={{\rm Tor}}_p^{{{\mathbb C}}\Gamma} ( {{\mathbb C}}; {\mathcal{U}}\Gamma)=0 \hspace{3em} \mbox{ for all } p \geq 0$$ we know that $b_1^{(2)}(\Gamma)$ is the only $L^2$-Betti number which could possibly be nonzero.
The $L^{(2)}$-Euler characteristic coincides with the usual Euler-characteristic and the rational Euler-Characteristic of [@Wall(1961)] whenever these are defined.
Before we proceed to the proof of Theorem \[application\] we would also like to mention the following consequences for $L^2$-homology.
\[UCT\] Let $\Gamma$ be in ${\mathcal{C}}$ with a bound on the orders of finite subgroups. Then there is a universal coefficient theorem for $L^2$-homology: Let $X$ be a $\Gamma$-space whose isotropy groups are all finite, then there is an exact sequence $$0 \to H_n (X ; {{\mathbb Z}}) \otimes_{ {{\mathbb Z}}\Gamma } {\mathcal{U}}\Gamma \to
H_n^{\Gamma} ( X ; {\mathcal{U}}\Gamma ) \to {{\rm Tor}}_1( H_{n-1} (X ; {{\mathbb Z}}) ; {\mathcal{U}}\Gamma )
\to 0.$$
We freely use the dimension theory for arbitrary ${\mathcal{U}}\Gamma$-modules, compare Remark \[remark-dimension-for-arbitrary\]. If $X$ has finite isotropy, then the set of singular simplices also has only finite isotropy groups. If $H$ is a finite subgroup of $\Gamma$, then ${{\mathbb C}}\left[ \Gamma / H \right] \cong {{\mathbb C}}\Gamma \otimes_{{{\mathbb C}}H} {{\mathbb C}}$ is induced from the projective ${{\mathbb C}}H$-module ${{\mathbb C}}$ and therefore projective. We see that the singular chain complex with complex coefficients $C_{\ast}= C_{\ast}^{sing} ( X ; {{\mathbb C}})$ is a complex of projective ${{\mathbb C}}\Gamma$-modules. The $E^2$-term of the Künneth spectral sequence (compare Theorem 5.6.4 on page 143 in [@Weibel(1994)]) $$E^2_{pq} = {{\rm Tor}}_p^{{{\mathbb C}}\Gamma} ( H_q( C_{\ast} ) ; {\mathcal{D}}\Gamma ) \Rightarrow
H_{p+q}(C_{\ast} \otimes {\mathcal{D}}\Gamma ) = H_{p+q} ( X ; {\mathcal{D}}\Gamma)$$ is concentrated in two columns. The spectral sequence collapses, and we get exact sequences $$0 \to H_n(X; {{\mathbb C}}) \otimes_{{{\mathbb C}}\Gamma} {\mathcal{D}}\Gamma \to H_n(X ; {\mathcal{D}}\Gamma ) \to
{{\rm Tor}}_1^{{{\mathbb C}}\Gamma } ( H_{n-1} (X ; {{\mathbb C}}) ; {\mathcal{D}}\Gamma ) \to 0.$$ Applying the exact functor $- \otimes_{{\mathcal{D}}\Gamma } {\mathcal{U}}\Gamma$ yields the result.
The proof of Theorem \[application\] depends on the following Lemma.
\[Tor lemma\]
1. \[Tl1\] Let $R \ast G \subset S \ast G $ be compatible with the crossed product structure. Let $M$ be an $R \ast G$-module. There is a natural isomorphism of right $S$-modules $${{\rm Tor}}_p^{R \ast G} ( M ; S \ast G) \cong {{\rm Tor}}_p^{R} ( {{\rm res}}^{R \ast G}_R M ; S )$$ for all $p \geq 0$.
2. \[Tl2\] Suppose $R \subset S$ is a ring extension and $R = \bigcup_{i \in I} R_i$ is the directed union of the subrings $R_i$. Let $M$ be an $R$-module. Then there is a natural isomorphism of right $S$-modules $${{\rm Tor}}_p^{R}( M ; S ) \cong {{\rm colim }}_{i \in I }{{\rm Tor}}_p^{R_i}( {{\rm res}}^{R}_{R_i} M ; S_i ) \otimes_{S_i} S$$ for all $p\geq 0$.
\[Tl1\] We start with the case $p=0$. We denote the crossed product structure map by $\mu$, compare Digression \[digression-crossed-product\]. Define a map $$h_M: {{\rm res}}^{R \ast G}_R M \otimes_R S \to M \otimes_{R \ast G} S \ast G$$ by $m \otimes s \mapsto m \otimes s$. Obviously $h$ is a natural transformation from the functor ${{\rm res}}^{R \ast G}_R (-) \otimes_R S$ to $- \otimes_{R \ast G} S \ast G$. If $M = R \ast G $ the map $h_{R \ast G}^{-1}: R \ast G \otimes_{R \ast G} S \ast G
\cong S \ast G \to {{\rm res}}^{R \ast G}_R R \ast G \otimes_R S $ given by $s\mu(g) \mapsto g \otimes c_g^{-1}(s)$ is a well-defined inverse. Since $h$ is compatible with direct sums we see that $h_{F}$ is an isomorphism for all free modules $F$. Now if $M$ is an arbitrary module choose a free resolution $F_{\ast} \to M$ of $M$ and apply both functors to $$F_1 \to F_0 \to M \to 0 \to 0.$$ Both functors are right exact, therefore an application of the five lemma yields the result for $p=0$. Now let $P_{\ast} \to M$ be a projective resolution of $M$, then $$\begin{aligned}
{{\rm Tor}}_p^{R}( {{\rm res}}_R^{R \ast G} M ; S ) & = & H_p( {{\rm res}}_R^{R \ast G} P_{\ast} \otimes_R S ) \\
& \stackrel{\cong}{\to} & H_p( P_{\ast} \otimes_{R \ast G} S \ast G ) \\
& = & {{\rm Tor}}_p^R ( M ; S \ast G ) .\end{aligned}$$ \[Tl2\] Again we start with the case $p=0$. The natural surjections ${{\rm res}}^R_{R_i} M \otimes_{R_i} S \to M \otimes_R S$ induce a surjective map $$h_M:{{\rm colim }}_{i \in I} {{\rm res}}^R_{R_i} M \otimes_{R_i} S \to M \otimes_R S$$ which is natural in $M$. Suppose the element of the colimit represented by $\sum_k m_k \otimes s_k \in
{{\rm res}}^R_{R_i} M \otimes_{R_i} S$ is mapped to zero in $M \otimes_R S$. By construction the tensor product $M \otimes_R S$ is the quotient of the free module on the set $M \times S$ by a relation submodule. But every relation involves only finitely many elements of $R$, so we can find a $j \in I$ such that $\sum_k m_k \otimes s_k=0$ already in ${{\rm res}}^R_{R_j} M \otimes_{R_j} S$. We see that $h_M$ is an isomorphism. Now let $P_{\ast} \to M$ be a projective resolution. Since the colimit is an exact functor it commutes with homology and we get $$\begin{aligned}
{{\rm colim }}_{i \in I} {{\rm Tor}}_p^{R_i} ( M ; S) & = &
{{\rm colim }}_{i \in I} H_p ( {{\rm res}}^R_{R_i} P_{\ast} \otimes_{R_i} S ) \\
& = & H_p( {{\rm colim }}_{i \in I} ( {{\rm res}}_{R_i}^R P_{\ast} \otimes_{R_i} S )) \\
& \stackrel{\cong}{\to } & H_p( P_{\ast} \otimes_{R} S ) \\
& = & {{\rm Tor}}_p^R( M ; S ) .\end{aligned}$$
The proof works via transfinite induction over the group as for the proof of Linnell’s Theorem \[main-linnell\] itself, compare (I), (II) and (III) on page .
[**(I)**]{} The statement for free groups is well known: let $\Gamma $ be the free group generated by the set $S$. The cellular chain complex of the universal covering of the obvious $1$-dimensional classifying space gives a projective resolution of the trivial module of length one $$0 \to \bigoplus_{S} {{\mathbb C}}\Gamma \to {{\mathbb C}}\Gamma \to {{\mathbb C}}\to 0 .$$ Now if $M$ is an arbitrary ${{\mathbb C}}\Gamma$-module we apply $-\otimes_{{{\mathbb C}}} M$ to the above complex and get a projective resolution of length $1$ for $M$ (diagonal action). (Use that for $P$ a projective ${{\mathbb C}}\Gamma$-module $P \otimes_{{{\mathbb C}}} M$ with the diagonal respectively the left $\Gamma$-action are noncanonically isomorphic ${{\mathbb C}}\Gamma$-modules.)
[**(II)**]{} The next step is to prove that the statement remains true under extensions by finite groups. So let $1 \to G \to \Gamma \to H \to 1$ be an exact sequence with $H$ finite. We know that ${\mathcal{D}}\Gamma = {\mathcal{D}}G \ast H$, see Lemma 10.59 on page 399 in [@Lueck(2002)] or Proposition 8.13 in [@Reich(1999)]. Let $M$ be a ${{\mathbb C}}\Gamma$-module, then with Lemma \[Tor lemma\] and the induction hypothesis we conclude $$\begin{aligned}
{{\rm Tor}}_p^{{{\mathbb C}}\Gamma}(M; {\mathcal{D}}\Gamma) & = & {{\rm Tor}}_p^{{{\mathbb C}}G \ast H} ( M ; {\mathcal{D}}G \ast H) \\
& \cong &
{{\rm Tor}}_p^{{{\mathbb C}}G} ( {{\rm res}}^{{{\mathbb C}}G \ast H}_{{{\mathbb C}}G} M; {\mathcal{D}}G ) \\
& = & 0 \quad \mbox{ for } p>1.\end{aligned}$$ The case $H$ infinite cyclic is only slightly more complicated. This time we know from Lemma 10.69 in [@Lueck(2002)] or Proposition 8.18 in [@Reich(1999)] that ${\mathcal{D}}\Gamma=({\mathcal{D}}G \ast H) T^{-1}$ is an Ore localization, where $T= T( {\mathcal{D}}G \ast H \subset {\mathcal{U}}\Gamma)$, i.e. the set of all elements in ${\mathcal{D}}G \ast H$ which become invertible in ${\mathcal{U}}\Gamma$. Since Ore localization is an exact functor we get $$\begin{aligned}
{{\rm Tor}}_p^{{{\mathbb C}}\Gamma}(M ; {\mathcal{D}}\Gamma) & = & {{\rm Tor}}_p^{{{\mathbb C}}G \ast H} (M ; ({\mathcal{D}}G \ast H)T^{-1} ) \\
& \cong & {{\rm Tor}}_p^{{{\mathbb C}}G \ast H} (M; {\mathcal{D}}G \ast H ) \otimes_{{\mathcal{D}}G \ast H } {\mathcal{D}}\Gamma \end{aligned}$$ and conclude again with Lemma \[Tor lemma\] that this module vanishes if $p>1$.
[**(III)**]{} The behaviour under directed unions remains to be checked. Let $\Gamma = \bigcup_{i \in I}
\Gamma_i $ be a directed union, then using Definition \[vNregular-definition\] we see that $\bigcup_{i \in I} {\mathcal{D}}\Gamma_i$ is von Neumann regular and it is easy to check that it coincides with the division closure ${\mathcal{D}}\Gamma$. Now Lemma \[Tor lemma\] gives $$\begin{aligned}
{{\rm Tor}}_p^{{{\mathbb C}}\Gamma} ( M ; {\mathcal{D}}\Gamma ) & \cong &
{{\rm colim }}_{i \in I} {{\rm Tor}}_p^{{{\mathbb C}}\Gamma_i}( {{\rm res}}^{{{\mathbb C}}\Gamma}_{{{\mathbb C}}\Gamma_i} M ; {\mathcal{D}}\Gamma ) \\
& = & {{\rm colim }}_{i \in I} {{\rm Tor}}_p^{{{\mathbb C}}\Gamma_i}( {{\rm res}}^{{{\mathbb C}}\Gamma}_{{{\mathbb C}}\Gamma_i} M ; {\mathcal{D}}\Gamma_i ) \otimes_{{\mathcal{D}}\Gamma_i} {\mathcal{D}}\Gamma \\
& = & 0 \quad \mbox{ for } p>1.\end{aligned}$$
\#1[0=]{}
[10]{}
M. F. Atiyah. Elliptic operators, discrete groups and von [N]{}eumann algebras. , 32-33:43–72, 1976.
A. Bartels, F. T. Farrell, L. E. Jones, and H. Reich. On the isomorphism conjecture in algebraic ${K}$-theory. Preprintreihe SFB 478 — Geometrische Strukturen in der Mathematik, Heft 175, Münster, 2001.
H. Bass, A. Heller, and R. G. Swan. The [W]{}hitehead group of a polynomial extension. , 22:61–79, 1964.
J. A. Beachy. On universal localization at semiprime [G]{}oldie ideals. In [*Ring theory (Granville, OH, 1992)*]{}, pages 41–57. World Sci. Publishing, River Edge, NJ, 1993.
B. Blackadar. . Springer-Verlag, New York, 1986.
G. Carlsson. Homotopy fixed points in the algebraic [$K$]{}-theory of certain infinite discrete groups. In [*Advances in homotopy theory (Cortona, 1988)*]{}, volume 139 of [*London Math. Soc. Lecture Note Ser.*]{}, pages 5–10. Cambridge Univ. Press, Cambridge, 1989.
G. Carlsson and E. K. Pedersen. Controlled algebra and the [N]{}ovikov conjectures for ${K}$- and ${L}$-theory. , 34(3):731–758, 1995.
J. Cheeger and M. Gromov. ${L}\sb 2$-cohomology and group cohomology. , 25(2):189–215, 1986.
C. Chou. Elementary amenable groups. , 24(3):396–407, 1980.
G. Cliff and A. Weiss. Moody’s induction theorem. , 32(3):489–500, 1988.
P. M. Cohn. Free ideal rings. , 1:47–69, 1964.
P. M. Cohn. . Academic Press Inc. \[Harcourt Brace Jovanovich Publishers\], London, second edition, 1985.
P. M. Cohn. . John Wiley & Sons Ltd., Chichester, second edition, 1991.
A. Connes. Noncommutative differential geometry. , (62):257–360, 1985.
J. F. Davis and W. L[ü]{}ck. Spaces over a category and assembly maps in isomorphism conjectures in ${K}$- and ${L}$-theory. , 15(3):201–252, 1998.
M. M. Day. Amenable semigroups. , 1:509–544, 1957.
F. T. Farrell. . Published for the Tata Institute of Fundamental Research, Bombay, 1996.
F. T. Farrell and L. E. Jones. Isomorphism conjectures in algebraic ${K}$-theory. , 6(2):249–297, 1993.
T. Farrell. The [B]{}orel [C]{}onjecture. In [*Topology of high-dimensional manifolds*]{}, volume 9(1) of [ *ICTP Lecture Notes*]{}, pages 225–298. ICTP, Trieste, 2002.
S. M. Gersten. The localization theorem for projective modules. , 2:317–350, 1974.
K. R. Goodearl. . Pitman (Advanced Publishing Program), Boston, Mass., 1979.
D. R. Grayson. -theory and localization of noncommutative rings. , 18(2):125–127, 1980.
R. I. Grigorchuk. An example of a finitely presented amenable group that does not belong to the class [E]{}[G]{}. , 189(1):79–100, 1998.
R. I. Grigorchuk, P. A. Linnell, T. Schick, and A. [Ż]{}uk. On a question of [A]{}tiyah. , 331(9):663–668, 2000.
I. Hughes. Division rings of fractions for group rings. , 23:181–188, 1970.
L. Jones. Foliated control theory and its applications. In [*Topology of high-dimensional manifolds*]{}, volume 9(2) of [ *ICTP Lecture Notes*]{}, pages 405–460. ICTP, Trieste, 2002.
P. Julg and A. Valette. -theoretic amenability for [${\rm SL}\sb{2}({\bf Q}\sb{p})$]{}, and the action on the associated tree. , 58(2):194–215, 1984.
R. V. Kadison and J. R. Ringrose. . Academic Press Inc. \[Harcourt Brace Jovanovich Publishers\], New York, 1983. Elementary theory.
P. H. Kropholler, P. A. Linnell, and J. A. Moody. Applications of a new ${K}$-theoretic theorem to soluble group rings. , 104(3):675–684, 1988.
J. Lewin. Fields of fractions for group algebras of free groups. , 192:339–346, 1974.
P. A. Linnell. Zero divisors and group von [N]{}eumann algebras. , 149(2):349–363, 1991.
P. A. Linnell. Zero divisors and ${L}\sp 2({G})$. , 315(1):49–53, 1992.
P. A. Linnell. Division rings and group von [N]{}eumann algebras. , 5(6):561–576, 1993.
P. A. Linnell. Analytic versions of the zero divisor conjecture. In [*Geometry and cohomology in group theory (Durham, 1994)*]{}, pages 209–248. Cambridge Univ. Press, Cambridge, 1998.
P. A. Linnell, W. Lück, and T. Schick. The [O]{}re condition, affiliated operators, and the lamplighter group. Preprintreihe SFB 478 — Geometrische Strukturen in der Mathematik, Heft 205 Münster. To appear in the Proceedings of the school/conference on “High-dimensional Manifold Topology” in Trieste, May/June 2001, 2002.
W. L[ü]{}ck. Hilbert modules and modules over finite von [N]{}eumann algebras and applications to ${L}\sp 2$-invariants. , 309(2):247–285, 1997.
W. L[ü]{}ck. Dimension theory of arbitrary modules over finite von [N]{}eumann algebras and ${L}\sp 2$-[B]{}etti numbers. [I]{}. [F]{}oundations. , 495:135–162, 1998.
W. L[ü]{}ck. Dimension theory of arbitrary modules over finite von [N]{}eumann algebras and ${L}\sp 2$-[B]{}etti numbers. [I]{}[I]{}. [A]{}pplications to [G]{}rothendieck groups, ${L}\sp 2$-[E]{}uler characteristics and [B]{}urnside groups. , 496:213–236, 1998.
W. Lück. . Springer-Verlag, Berlin, 2002. Ergebnisse der Mathematik und ihrer Grenzgebiete, 3. Folge, Band 44.
W. L[ü]{}ck and H. Reich. The [B]{}aum-[C]{}onnes and the [F]{}arrell-[J]{}ones conjectures in ${K}$- and ${L}$-theory. in preparation, 2003.
P. Malcolmson. Construction of universal matrix localization. , 951, 1982.
J. A. Moody. Induction theorems for infinite groups. , 17(1):113–116, 1987.
J. A. Moody. Brauer induction for ${G}\sb 0$ of certain infinite groups. , 122(1):1–14, 1989.
F. Murray and J. [von Neumann]{}. On rings of operators. , 37:116–229, 1936.
A. Neeman and A. Ranicki. Noncommutative localization and chain complexes i. e-print AT.0109118, 2002.
D. S. Passman. . Academic Press Inc., Boston, MA, 1989.
E. K. Pedersen. Controlled algebraic [$K$]{}-theory, a survey. In [*Geometry and topology: Aarhus (1998)*]{}, volume 258 of [ *Contemp. Math.*]{}, pages 351–368. Amer. Math. Soc., Providence, RI, 2000.
E. K. Pedersen and C. A. Weibel. ${K}$-theory homology of spaces. In [*Algebraic topology (Arcata, CA, 1986)*]{}, pages 346–361. Springer-Verlag, Berlin, 1989.
F. Quinn. Ends of maps. [I]{}. , 110(2):275–331, 1979.
F. Quinn. Ends of maps. [I]{}[I]{}. , 68(3):353–424, 1982.
F. Quinn. Applications of topology with control. In [*Proceedings of the International Congress of Mathematicians, Vol. 1, 2 (Berkeley, Calif., 1986)*]{}, pages 598–606, Providence, RI, 1987. Amer. Math. Soc.
A. Ranicki. Noncommutative localization in topology. unpublished notes, 2001.
H. Reich. Group von [N]{}eumann algebras and related algebras. Dissertation Universität Göttingen, http://www.math.uni-muenster.de/u/lueck/publ/reich/reich.dvi, 1999.
H. Reich. On the [$K$]{}- and [$L$]{}-theory of the algebra of operators affiliated to a finite von [N]{}eumann algebra. , 24(4):303–326, 2001.
T. Schick. Integrality of ${L}\sp 2$-[B]{}etti numbers. , 317(4):727–750, 2000.
T. Schick. Erratum: “[I]{}ntegrality of ${L}\sp 2$-[B]{}etti numbers” \[[M]{}ath. [A]{}nn. [**317**]{} (2000), no. 4, 727–750; 1777117\]. , 322(2):421–422, 2002.
A. H. Schofield. . Cambridge University Press, Cambridge, 1985.
B. Stenstr[ö]{}m. . Springer-Verlag, New York, 1975. Die Grundlehren der Mathematischen Wissenschaften, Band 217, An introduction to methods of ring theory.
R. G. Swan. Induced representations and projective modules. , 71:552–578, 1960.
R. W. Thomason and T. Trobaugh. Higher algebraic [$K$]{}-theory of schemes and of derived categories. In [*The Grothendieck Festschrift, Vol. III*]{}, volume 88 of [ *Progr. Math.*]{}, pages 247–435. Birkhäuser Boston, Boston, MA, 1990.
J. von Neumann. , volume 27 of [*Princeton Mathematical Series*]{}. Princeton University Press, Princeton, New Jersey, 1960.
F. Waldhausen. Algebraic ${K}$-theory of generalized free products. [I]{}, [I]{}[I]{}. , 108(1):135–256, 1978.
C. T. C. Wall. Rational [E]{}uler characteristics. , 57:182–184, 1961.
C. Weibel and D. Yao. Localization for the [$K$]{}-theory of noncommutative rings. In [*Algebraic $K$-theory, commutative algebra, and algebraic geometry (Santa Margherita Ligure, 1989)*]{}, volume 126 of [*Contemp. Math.*]{}, pages 219–230. Amer. Math. Soc., Providence, RI, 1992.
C. A. Weibel. . Cambridge University Press, Cambridge, 1994.
[^1]: Research supported by the SFB “Geometrische Strukturen in der Mathematik” in Münster, Germany.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Temperature and applied magnetic field dependent magnetization measurements on 34 single crystalline samples of (R, R’,R”...)Ni$_{2}$Ge$_{2}$ compounds (R, R’, R”, etc. being primarily Gd$-$Lu, Y), were made. These measurements reveal that, despite extremes in local moment anisotropy, the average de Gennes parameter is a remarkably good predictor of the paramagnetic to antiferromagnetic ordering temperature. In addition, the pronounced metamagnetic phase transitions seen in the low temperature phase of TbNi$_{2}$Ge$_{2}$ are found to be remarkably robust to high substitution levels of Gd and 25% substitutions of other heavy rare earths.'
address: 'Ames Laboratory and Department of Physics and Astronomy, Iowa State University, Ames, IA 50011, USA'
author:
- 'S. A. Law, S. L. Bud’ko, P. C. Canfield'
title: 'Effects of mixed rare earth occupancy on the low temperature properties of (R, R’,R”...)Ni$_{2}$Ge$_{2}$ single crystals'
---
magnetic order ,de Gennes scaling ,metamagnetism
75.20.Hr ,75.30.Kz ,75.50.Ee
Introduction
============
The RNi$_{2}$Ge$_{2}$ series is one of the model series for the study of how the 4f-shell electrons manifest themselves in intermetallic compounds. Single crystals can be grown for all stable R isotopes (R = La$-$Nd, Sm$-$Lu, and Y) and very likely could be made for R = Pm as well.[@1] This range of R allows for the formation of non-magnetic (R = La, Lu, and Y), hybridizing (R = Ce and Yb), isotropic local moment (R = Eu and Gd), extremely axial local moment (R = Tb), and extremely planar local moment systems (R = Er and Tm) as well as a range of less extreme, but still anisotropic examples. All of the local moment members of the RNi$_{2}$Ge$_{2}$ series manifest antiferromagnetic order below *T$_{N}$*, and the *T$_{N}$* values for the heavy rare earths scale well with the de Gennes parameter (dG = (g$_{J}$-1)$^{2}$J(J+1)). In addition to an extensive study of the anisotropic thermodynamic and transport properties of the pure RNi$_{2}$Ge$_{2}$ series,[@1] studies of the magnetic structures in GdNi$_{2}$Ge$_{2}$ and TbNi$_{2}$Ge$_{2}$,[@2; @3; @4] the evolution of an Ising spin glass state in (Y$_{1-x}$Tb$_{x}$)Ni$_{2}$Ge$_{2}$,[@5] and the effects of band filling on the ordering wave vector in EuNi$_{2}$Ge$_{2}$ and GdNi$_{2}$Ge$_{2}$ have been made.[@6; @7]
Rare earth intermetallic compounds offer the possibility of utilizing the intelligible complexity of the 4f-shell to tune magnetic properties. By changing the rare earth element, the size and anisotropy of the local moment as well as the value of the transition temperature can be changed. The ability to modify the low temperature properties of this series by partial substitution of two or more R species onto the one unique crystallographic site offers the possibility of either tuning the magnetic properties with greater control or perhaps even creating new properties. In this paper, we present the results of a broad study of the properties of (R, R’,R”...)Ni$_{2}$Ge$_{2}$ single crystals, specifically focused on their low temperature local moment magnetism.
Experimental methods
====================
Single crystals of (R, R’,R”...)Ni$_{2}$Ge$_{2}$ were grown out of a self-flux rich in Ni and Ge.[@1; @5] Generally the ratio of starting elements was R$_{0.07}$Ni$_{0.465}$Ge$_{0.465}$ with the 7% atomic for the rare earth split appropriately between R, R’, R”, etc. (e.g., the (Tb$_{0.75}$Y$_{0.25}$)Ni$_{2}$Ge$_{2}$ growth contained Tb$_{0.525}$Y$_{0.175}$Ni$_{0.465}$Ge$_{0.465}$). The high purity elements were placed into Al$_{2}$O$_{3}$ crucibles and sealed in quartz ampoules under a partial atmosphere of Ar gas. The quartz ampoules were then put into box furnaces, heated to just under 1200$^\circ$C, and slowly cooled (generally over 75$-$100 hours) to 1000$^\circ$C, at which point they were removed from the furnace and the excess liquid was decanted.[@8],[@9] The single crystals had a plate$-$like morphology and could have dimensions as large as 1 x 1 x 0.2 cm$^{3}$, although more typical dimensions were 0.3 x 0.3 x 0.1 cm$^{3}$. The crystallographic *c*-axis was found to be perpendicular to the plates and well-defined facets allowed for the easy identification of the in-plane orientation as well.
Temperature and field dependent magnetization measurements were taken on Quantum Design MPMS units for *T* $\geq$ 1.8 K and *H* $\leq$ 7 T, and resistivity and specific heat measurements taken on the mélange sample (see below) were performed on Quantum Design PPMS units.
The antiferromagnetic ordering temperature, *T$_{N}$*, was determined from low field, temperature dependent magnetization data by invoking the similarity in temperature dependence between *C$_{p}$(T)* and ${d(MT/H)}/{dT}$[@10] near *T$_{N}$*. Features in *M(H)* curves were associated with critical magnetic fields by determining *H$_{c}$* to be the field at which ${dM}/{dH}$ was maximal. A given sample’s de Gennes value was the weighted average of the constituent rare earth ions’ elemental de Gennes values.
Experimental results
====================
(Pr$_{0.11}$Nd$_{0.11}$Sm$_{0.11}$Gd$_{0.11}$Tb$_{0.11}$Dy$_{0.11}$Ho$_{0.11}$Er$_{0.11}$Tm$_{0.11}$)Ni$_{2}$Ge$_{2}$ mélange sample
------------------------------------------------------------------------------------------------------------------------------------
{width="6.8cm"}
{width="6.8cm"}
{width="6.8cm"}
{width="6.8cm"}
Single crystals of (Pr$_{0.11}$Nd$_{0.11}$Sm$_{0.11}$Gd$_{0.11}$Tb$_{0.11}$Dy$_{0.11}$Ho$_{0.11}$Er$_{0.11}$Tm$_{0.11}$)Ni$_{2}$Ge$_{2}$ (henceforth referred to as the mélange sample) were grown to see if a thorough mixture of the single site local moment anisotropy would result in a partial suppression of long-range order, or perhaps even a complete loss of long range order resulting in a spin glass state akin to that associated with random site anisotropy. Figure 1 presents the magnetic susceptibility, specific heat data, and electrical resistivity for this sample. Figure 2 presents the anisotropic magnetization as a function of applied field for *T* = 2 K for *H $||$ c* plane and *H $||$ ab* plane. Both temperature dependent thermodynamic measurements reveal a phase transition at 11.4 $\pm$ 0.5 K. The transport data also reveal a somewhat weaker feature near this temperature. This phase transition is consistent with long range antiferromagnetic order, and magnetic X-ray diffraction measurements reveal that there is an ordering wave vector (0, 0, 0.777) in the low temperature state. [@11]
The mélange sample’s transition temperature falls onto the de Gennes scaling line associated with the heavy rare earth members of the RNi$_{2}$Ge$_{2}$ series \[see ref. 1 and figure 9 below\] indicating that there is little or no suppression of magnetic ordering in the sample even though there is a wide range of local moment anisotropy. This opens up the possibility of selectively mixing rare earths on this one unique crystallographic site to tune *T$_{N}$* systematically. In the rest of this paper, we will mix a variety of heavy rare earths to test how well de Gennes scaling holds and to also investigate the effects of this mixing on the details of *M(T)* and *M(H)* data.
Mixed heavy rare earth samples
------------------------------
Table 1 presents a summary of the magnetic data on the 34 mixed rare earth samples studied in this work. These samples can be divided and discussed in a number of ways. The first sub-group is a (Tb$_{1-x}$Gd$_{x}$)Ni$_{2}$Ge$_{2}$ series with samples ranging from the Ising-like TbNi$_{2}$Ge$_{2}$ to the Heisenberg-like GdNi$_{2}$Ge$_{2}$. Figure 3 presents anisotropic *M(T)/H* data for representative Gd substitutions: x = 0.15 and x = 0.90; the axial nature of Tb is clearly observed across the whole range of finite *x* values. Figure 4 presents a *T* vs. *x* diagram for the low field ordering of the (Tb$_{1-x}$Gd$_{x}$)Ni$_{2}$Ge$_{2}$ series. The highest points which plot the paramagnetic to antiferromagnetic transition temperature vary with *x* in a roughly linear manner. The next highest transition varies in a non-monotonic fashion, reaching a minimum near *x* $\sim$ 0.33. The third, lowest temperature phase transition was tentatively identified for pure Gd, [@1] but is much more clearly seen and separated for light Tb doping, finally disappearing totally as pure TbNi$_{2}$Ge$_{2}$ is approached.
R-site composition de Gennes factor T$_{m}$ (K) $\theta_{c}$(K) $\theta_{ab}$(K) $\theta_{ave}$(K) p$_{eff}$($\mu_{B}$)
---------------------------------------------- ------------------ ------------------ ----------------- ------------------ ------------------- ----------------------
Tb$_{0.85}$Gd$_{0.15}$ 11.29 18.3, 7.3 8.4 -39.5 -18.2 9.3
Tb$_{0.75}$Gd$_{0.25}$ 11.81 18.8, 7.3, 4.1 12.4 -36.3 -13.8 9.3
Tb$_{0.66}$Gd$_{0.33}$ 12.13 20.3, 6.1 8.6 -39.5 -14.8 9.3
Tb$_{0.5}$Gd$_{0.5}$ 13.13 21.1, 7.6, 3.6 2.1 -31.7 -17.8 9.2
Tb$_{0.33}$Gd$_{0.66}$ 13.86 22.1, 10.1, 4.3 -13 -30 -21.2 8.8
Tb$_{0.15}$Gd$_{0.85}$ 14.96 24, 16 -13 -21.1 -18 8.6
Tb$_{0.1}$Gd$_{0.9}$ 15.23 25.4, 17.6, 8.3 -12.1 -19.7 -16.9 8.3
Tb$_{0.05}$Gd$_{0.95}$ 15.49 25.4, 17.1, 3.1 -18.1 -18.1 -18.1 8.3
Y$_{0.25}$Tb$_{0.75}$ 7.88 11.3, 4.3 20.3 -41.4 -11.6 8.8
Gd$_{0.25}$Tb$_{0.75}$ 11.81 18.8, 7.3, 4.1 1 2.4 -36.3 -13.3 9.3
Dy$_{0.25}$Tb$_{0.75}$ 9.65 13.8, 7.8 11.5 -35.3 -14.5 10.1
Ho$_{0.25}$Tb$_{0.75}$ 9.01 11.8, 4.9 10.8 -29.1 -12 10.3
Er$_{0.25}$Tb$_{0.75}$ 8.52 12.3 8.8 -39 -18.4 10
Tm$_{0.25}$Tb$_{0.75}$ 8.17 12.6 13.2 -41.5 -12.8 9.6
Yb$_{0.25}$Tb$_{0.75}$ 7.96 10.8, 5.3 16.2 -40.9 -13.7 9.3
Lu$_{0.25}$Tb$_{0.75}$ 7.88 12.3, 7.4, 3.3 13.1 -43.5 -13.3 9.4
Gd$_{0.75}$Er$_{0.25}$ 12.45 21.8, 7.1 -16.5 -9.3 -9.3 8.4
Gd$_{0.5}$Er$_{0.5}$ 9.15 17.3, 5.6 -13.3 -7 -10.5 8.8
Gd$_{0.25}$Er$_{0.75}$ 5.85 8.8, 4.6 -21.6 0.05 -5.2 9.3
Gd$_{0.5}$Tb$_{0.5}$ 13.13 21.1, 7.6, 3.6 2.1 -31.7 -17.8 9.2
Gd$_{0.5}$Er$_{0.5}$ 9.15 17.3, 5.6 -13.3 -7 -10.5 8.8
Gd$_{0.5}$Tm$_{0.5}$ 8.46 15.8, 3.3 -22.6 -3.4 -8.2 8.2
Tb$_{0.5}$Er$_{0.5}$ 6.53 7.5 6.8 -19.8 -8.8 9.6
Tb$_{0.5}$Tm$_{0.5}$ 5.84 6.3 5.6 -22.7 -10.4 9
Dy$_{0.5}$Ho$_{0.5}$ 5.79 4.8 6.4 -15.5 -6.5 10.7
Dy$_{0.5}$Tm$_{0.5}$ 4.13 3.6 0.78 -10.7 -5.8 9.6
Ho$_{0.5}$Er$_{0.5}$ 3.53 3.1 -4.8 -0.94 -2.6 10.2
Er$_{0.5}$Tm$_{0.5}$ 1.86 2.1 -30.1 6.8 -1.8 8.8
Tb$_{0.9}$Lu$_{0.1}$ 9.45 14.8, 8.8 14.1 -45.7 -14.1 9.8
Gd$_{0.33}$Tb$_{0.33}$Er$_{0.33}$ 9.50 14.5 -0.35 -15 -8.9 9.2
Gd$_{0.33}$Tb$_{0.33}$Tm$_{0.33}$ 9.05 15.7 1.7 -21.4 -11.9 8.9
Gd$_{0.33}$Er$_{0.33}$Tm$_{0.33}$ 6.43 11.75, 4.0 -20 0.93 -5.4 8
Gd$_{0.25}$Tb$_{0.25}$Dy$_{0.25}$Ho$_{0.25}$ 9.46 13.2, 2.2 2.8 -19.2 -10.5 10
Mélange 5.35 11.8 3.8 -15.9 -6.18 8.2
: Summary of magnetization measurements on (R,R’,R”,...)Ni$_{2}$Ge$_{2}$ single crystals
\
{width="6.8cm"}
{width="6.8cm"}
{width="6.8cm"}
Figure 5 presents the anisotropic low temperature *M(H)* isotherms for the (Tb$_{1-x}$Gd$_{x}$)Ni$_{2}$Ge$_{2}$ series and figure 6 presents the critical fields as a function of *x*. As can be seen in figures 5a and 6a, when Tb is added to GdNi$_{2}$Ge$_{2}$, the metamagnetic transition seen for field applied in the basal plane drops and disappears for *x* $<$ 0.66. A somewhat more complex behavior is found for the field applied along the *c*-axis, figures 5b and 6b. Whereas for *x* $\leq$ 0.5, there are clear transitions near 2 T and 4-5 T, for *x* $>$ 0.85 the upper transition is lost and even the lower transition drops to zero. Detailed neutron or magnetic X$-$ray diffraction measurements are needed to identify specific regions of figures 4 and 6 with specific ordering wave vectors.
{width="6.8cm"}
{width="6.8cm"}
{width="6.8cm"}
{width="6.8cm"}
Although the ordering temperatures for the (Tb$_{1-x}$Gd$_{x}$)Ni$_{2}$Ge$_{2}$ series scale well with the de Gennes factor (see figure 9 below), this series can only probe de Gennes values between those of pure TbNi$_{2}$Ge$_{2}$ and pure GdNi$_{2}$Ge$_{2}$ (i.e. 10.50 $<$ *dG* $<$ 15.75). In order to probe a wider range of de Gennes values and to further explore the response of the TbNi$_{2}$Ge$_{2}$ system to a range of perturbations, a related series, (Tb$_{0.75}$R$_{0.25}$)Ni$_{2}$Ge$_{2}$, was studied. The *M(T)* data for field applied along the *c*-axis for these compounds are shown in figure 7. Their de Gennes factor ranges from $\sim$7.8 for R = Lu, Y to $\sim$11.8 for R = Gd. As shown in figure 9 below, de Gennes scaling of *T$_{N}$* continues to hold. Figure 8 presents the anisotropic, low temperature magnetization for the members of this series. Regardless of which heavy R is substituted for Tb, the (Tb$_{0.75}$R$_{0.25}$)Ni$_{2}$Ge$_{2}$ samples manifest clear metamagnetism for *H $||$ c* near 4-5 T. Other transitions are seen near 3 T and/or 1 T. For *H $||$ ab* there are no clear metamagnetic transitions for *H* $<$ 5.5 T. The data for R = Tm and Er appear to be clearly offset from the data for R = Y, Lu at high fields for in-plane magnetization. This raises the question of whether the planar moments associated with Er and Tm are somehow decoupled from the ordering Tb ions. In an attempt to see if the minority rare earth is ordered or acting as a paramagnetic impurity that is simply undergoing Brillouin saturation, the *H $||$ ab* magnetization was decomposed into a part associated with the Tb sublattice and an additional part, presumably associated with the minority rare earth. The inset to figure 8b presents the *M(H)* data for (Tb$_{0.75}$Er$_{0.25}$)Ni$_{2}$Ge$_{2}$ and (Tb$_{0.75}$Tm$_{0.25}$)Ni$_{2}$Ge$_{2}$ with the *M(H)* data for (Tb$_{0.75}$Y$_{0.25}$)Ni$_{2}$Ge$_{2}$ subtracted. Although the data approach saturation for *H*$>$2.5 T, the saturated moment is half that associated with the paramagnetic planar moments. [@1] These results imply that the Er and Tm are indeed taking part in the ordered state, a fact that is supported by magnetic X-ray diffraction studies carried out on other R-mixtures in this family. [@11],[@12]
{width="6.8cm"}
{width="6.8cm"}
{width="6.8cm"}
Lower values of the de Gennes parameter were obtained with the (Gd$_{1-x}$Er$_{x}$)Ni$_{2}$Ge$_{2}$, (R$_{0.5}$Er$_{0.5}$)Ni$_{2}$Ge$_{2}$ and (R$_{0.5}$Tm$_{0.5}$)Ni$_{2}$Ge$_{2}$ series. The *T$_{N}$* values for these samples, the series discussed above, and those for pure RNi$_{2}$Ge$_{2}$ and (Y$_{1-x}$Tb$_{x}$)Ni$_{2}$Ge$_{2}$ samples from refs. 1 and 5 are presented in figure 9. The *T$_{N}$* values for three ternary R-mixtures, one quaternary R-mixture, and the mélange sample are also shown. All of these data scale quite well, indicating that sample averaged de Gennes value is a good predictor of *T$_{N}$* for the heavy R members of the (R,R’,R”...)Ni$_{2}$Ge$_{2}$ series.
{width="15cm"}
Discussion
==========
Although the RNi$_{2}$Ge$_{2}$ series manifests extremes in local moment anisotropy (ranging from axial through isotropic to planar), the magnetic ordering temperature does not appear to depend too strongly on this anisotropy. (This makes it different from, for example, the RRh$_{4}$B$_{4}$ series in which extreme axial anisotropy for R = Tb is thought to be responsible for a higher ordering temperature than that found for R = Gd. [@13]) Although this effect can be seen by observing that the *T$_{N}$* values for the pure, heavy rare earth RNi$_{2}$Ge$_{2}$ compounds do not significantly deviate from de Gennes scaling (i.e. *T$_{N}$* for the axial R = Tb is not higher than that for the isotropic R = Gd), this relative insensitivity of *T$_{N}$* to local moment anisotropy can be further illustrated by examining mixtures which have similar average de Gennes values and dramatically different net anisotropies. A particularly clear example of this can be found by examining data for the (Tb$_{0.9}$Lu$_{0.1}$)Ni$_{2}$Ge$_{2}$, (Gd$_{0.25}$Tb$_{0.25}$Dy$_{0.25}$Ho$_{0.25}$)Ni$_{2}$Ge$_{2}$, and (Gd$_{0.5}$Tm$_{0.5}$)Ni$_{2}$Ge$_{2}$ samples. These samples are part of the cluster of data points in figure 9 centered near *dG* $\sim$ 9 and *T$_{N}$* $\sim$ 15 K. Figure 10 presents the anisotropic, low temperature magnetic susceptibility data for each of these samples. (Tb$_{0.9}$Lu$_{0.1}$)Ni$_{2}$Ge$_{2}$ is extremely axial (which is expected, given that it is a slight non-magnetic dilution away from the pure, Ising-like TbNi$_{2}$Ge$_{2}$), (Gd$_{0.25}$Tb$_{0.25}$Dy$_{0.25}$Ho$_{0.25}$)Ni$_{2}$Ge$_{2}$ is only weakly axial (actually approaching isotropy in the paramagnetic state), and (Gd$_{0.5}$Tm$_{0.5}$)Ni$_{2}$Ge$_{2}$ is manifesting the planar anisotropy associated with pure TmNi$_{2}$Ge$_{2}$. Despite these clear and large differences in local moment anisotropy, the de Gennes factor appears to be the primary agent in determining the antiferromagnetic ordering temperature.
{width="6.8cm"}
{width="6.8cm"}
{width="6.8cm"}
One feature that contributes to the RNi$_{2}$Ge$_{2}$ series’ compliance with de Gennes scaling is that so many of the heavy R members order with a similar type of wave vector: (0, 0, *l*) with *l* ranging from 0.75 to 0.81. This trend continues for the samples containing multiple rare earths with (Gd$_{0.25}$Tb$_{0.25}$Dy$_{0.25}$Ho$_{0.25}$)Ni$_{2}$Ge$_{2}$ and (Gd$_{0.33}$Er$_{0.33}$Tm$_{0.33}$)Ni$_{2}$Ge$_{2}$ having *l* = 0.752 and 0.759, respectively, [@12] and the mélange sample with its nine moment-bearing rare earths (including several light ones) having *l* = 0.755 [@11].
Acknowledgements
================
We would like to acknowledge Nate Kelso for his seminal contributions at the inception of this work. We would also like to thank H. S. Thompson and H. Humbert for their moral support. Ames Laboratory is operated for the U.S. Department of Energy by Iowa State University under Contract No. W-7405-Eng.-82. This work was supported by the Director for Energy Research, Office of Basic Energy Sciences.
\[\]
[00]{}
S.L. Bud’ko, Z. Islam, T.A. Wiener, I.R. Fisher, A.H. Lacerda, P.C. Canfield, J. Magn. Magn. Mater. 205(1) (1999) 53.
J.W. Kim, L. Tan, D. Wermeille, S.L. Bud’ko, P.C. Canfield, A.I. Goldman, J. Phys.: Condens Matter 17(46) (2005) L493.
J.W. Kim, A. Kreyssig, L. Tan, D. Wermeille, S.L. Bud’ko, P.C. Canfield, A.I. Goldman, App. Phys. Lett. 87(20) (2005) 202505/1.
A. Islam, C. Detlefs, A.I. Goldman, S.L. Bud’ko, P.C. Canfield, J.P. Hill, D. Gibbs, T. Vogt, A. Zheludev, Phys. Rev. B 58(13) (1998) 8522.
T.A. Wiener, I.R. Fisher, S.L. Bud’ko, A. Kracher, P.C. Canfield, Phys. Rev. B 62(22) (2000) 15056.
Z. Islam, C. Detlefs, C. Song, A.I. Goldman, V. Antropov, B. Harmon, S.L. Bud’ko, T. Wiener, P.C. Canfield, D. Wermeille, K.D. Finkelstein, Phys. Rev. Lett. 83(14) (1999) 2817.
J. Park, D.P. Brammeier, C.G. Olson, P.C. Canfield, D.W. Lynch, Phys. Rev. B 70(7) (2004) 075105/1.
Z. Fisk, J.P. Remeika, in: K.A. Gscheidner Jr., L. Eyring (Eds.), Handbook on the Physics and Chemistry of Rare Earths, vol. 12, Elsevier, Amsterdam, 1989.
P.C. Canfield, Z. Fisk, Philos. Mag. B 65 (1992) 1117.
M.E. Fisher, Mil. Mag. 7 (1962) 1731.
A. Goldman private communication.
J.W. Kim, Y. Lee, D. Wermeille, B. Sieve, L. Tan, S.L. Bud’ko, S. Law, P.C. Canfield, B.N. Harmon, A.I. Goldman, Phys. Rev. B 72(6) (2005) 064403/1.
D.R. Noakes, G.K. Shenoy, Phys. Lett. A, 91 (1982) 35.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a parametric family of Riordan arrays, which are obtained by multiplying any Riordan array with a generalized Pascal array. In particular, we focus on some interesting properties of one-parameter Catalan triangles. We obtain several combinatorial identities that involve two special Catalan matrices, the Chebyshev polynomials of the second kind, some periodic sequences, and the Fibonacci numbers.'
bibliography:
- 'catalan-bibarray2.bib'
---
\[theorem\][Corollary]{} \[theorem\][Lemma]{} \[theorem\][Proposition]{}
\[theorem\][Definition]{} \[theorem\][Example]{} \[theorem\][Conjecture]{}
\[theorem\][Remark]{}
José Agapito [^1], Ângela Mestre [^2], and Maria M. Torres\
Centro de Análise Funcional, Estruturas Lineares e Aplicações\
Grupo de Estruturas Algébricas, Lineares e Combinatórias\
Departamento de Matemática\
Faculdade de Ciências, Universidade de Lisboa\
1749-016 Lisboa\
Portugal\
<[email protected]>\
<[email protected]>\
<[email protected]>\
\
Pasquale Petrullo\
Dipartimento di Matematica e Informatica\
Università degli Studi della Basilicata\
Potenza\
Italy\
<[email protected]>
Introduction {#seIntro}
============
There are two infinite lower triangular matrices in the mathematical literature, that are both called Catalan triangles. Let us denote them by ${\boldsymbol{C}}$ and ${\boldsymbol{B}}$, respectively. By way of illustration, their first rows are shown below: $$\label{eqballot}
{\boldsymbol{C}}=\begin{pmatrix}
1 & & & & & & &\\
1 & 1 & & & & & &\\
2 & 2 & 1 & & & & &\\
5 & 5 & 3 & 1 & & & &\\
14 & 14 & 9 & 4 & 1& & &\\
42 & 42 & 28 & 14 & 5 & 1& &\\
132 & 132 & 90 & 48 & 20 & 6 & 1&\\
\vdots & \vdots & \vdots & \vdots & \vdots &\vdots &\vdots &\ddots
\end{pmatrix} \phantom{A}\text{,}\phantom{A}
{\boldsymbol{B}}=\begin{pmatrix}
1 & & & & & & &\\
2 & 1 & & & & & &\\
5 & 4 & 1 & & & & &\\
14 & 14 & 6 & 1 & & & &\\
42 & 48 & 27 & 8 & 1& & &\\
132 & 165 & 110 & 44 & 10 & 1 & &\\
429 & 572 & 429 & 208 & 65 & 12 & 1 &\\
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \ddots
\end{pmatrix}.$$
Note that the $0^\mathrm{th}$ column of ${\boldsymbol{C}}$ is given by the sequence of Catalan numbers $(C_n)_{n\ge 0}$, where $C_n=\frac{1}{n+1}{{2n}\choose{n}}$. In turn, the $0^\mathrm{th}$ column of ${\boldsymbol{B}}$ is given by $(C_{n+1})_{n\ge 0}$. This similarity explains why they are called Catalan triangles.
According to the online encyclopedia of integer sequences (OEIS) [@OEIS], the non-zero entries of ${\boldsymbol{C}}$ (sequence [[](http://oeis.org/A033184)]{}) are given by the formula ${\boldsymbol{C}}_{n,k}=\frac{k+1}{n+1} {{2n-k}\choose{n}}$, for $n\ge k\ge 0$. They are known as the (ordinary) ballot numbers. Aigner [@Aig2008] used these numbers (and generalizations of them) to enumerate various combinatorial instances. Furthermore, Ferrari and Pinzani [@FPCatalan] gave an interpretation of ${\boldsymbol{C}}$, using the ECO method, and a suitable change of basis in the vector space of one-variable polynomials. On the other hand, the non-zero entries of ${\boldsymbol{B}}$ ([[](http://oeis.org/A039598)]{}) are given by the formula ${\boldsymbol{B}}_{n,k}=\frac{k+1}{n+1} {{2n+2}\choose{n-k}}$, for $n\ge k\ge 0$. The numbers ${\boldsymbol{B}}_{n,k}$ first appeared in the work of Shapiro [@ShaCatalan; @ShaRuns], in problems connected with non-intersecting paths, random walks, Eulerian numbers, runs, slides, and moments.[^3] We will refer from now on to ${\boldsymbol{C}}$ as the Aigner array, and to ${\boldsymbol{B}}$ as the Shapiro array.
The aim of this paper is to discuss some interesting features regarding the product matrix ${\boldsymbol{R}}( r )={\boldsymbol{R}}{\boldsymbol{P}}( r )$, where ${\boldsymbol{R}}$ is a given Riordan array, and the parameter $r$ is any real or complex number. The notation ${\boldsymbol{P}}( r )$ stands for a generalized Pascal array, such that ${\boldsymbol{P}}(0)={\boldsymbol{I}}$ is the identity matrix, and ${\boldsymbol{P}}(1)={\boldsymbol{P}}$ is the classical Pascal array of binomial numbers, that comes arranged as a lower triangular matrix. We elaborate, in particular, on the product matrix ${\boldsymbol{C}}( r )={\boldsymbol{C}}{\boldsymbol{P}}( r )$. Clearly, we have ${\boldsymbol{C}}(0)={\boldsymbol{C}}$. It turns out that ${\boldsymbol{C}}(1)={\boldsymbol{B}}$. The factorization ${\boldsymbol{B}}={\boldsymbol{C}}{\boldsymbol{P}}$ resembles Barry’s definition [@BarryThesis see Sections 6.3 and 6.5] of the generalized ballot array given by $\mathbf{Bal}=\mathbf{Cat}\cdot\mathbf{Bin}$; where $\mathbf{Bin}={\boldsymbol{P}}$, and $\mathbf{Cat}=\operatorname{diag}(1,{\boldsymbol{C}})$.
Recently, Yang [@Yang] introduced a generalized Catalan matrix, given by $$C[a,b;r]=\left(\Big(\frac{1-\sqrt{1-4rz}}{2rz}\Big)^a,z\Big(\frac{1-\sqrt{1-4rz}}{2rz}\Big)^b\right),$$ where $a,b$ are integer numbers, and $r$ is an arbitrary parameter. Such a generalization provides a unified way of presenting the matrices ${\boldsymbol{C}}$ and ${\boldsymbol{B}}$, along with many other matrices related to Catalan numbers. One can check that the $0^{\mathrm{th}}$ column of $C[a,b;r]$ forms a sequence of degree $n$ monomials in $r$, whose coefficients are polynomials in $a$ and $b$, that have rational coefficients. The one-parameter Catalan triangles that we will study are instead given by $${\boldsymbol{C}}( r )=\left(\frac{1-2rz-\sqrt{1-4z}}{2z(1-r+r^2z)},\frac{1-2rz-\sqrt{1-4z}}{2(1-r+r^2z)}\right).$$ We shall show explicitly that the $0^\mathrm{th}$ column of ${\boldsymbol{C}}( r )$ forms a sequence $({\boldsymbol{C}}( r )_{n,0})_{n\ge 0}$ of degree $n$ monic polynomials in $r$. These polynomials have positive integer coefficients. In particular, the constant term of ${\boldsymbol{C}}( r )_{n,0}$ is the Catalan number $C_n$. For this reason, we shall refer to ${\boldsymbol{C}}( r )$ as a Catalan triangle too.
It is also worth mentioning that He [@HeCatalan] introduced recently, a family of Catalan triangles which depend on two parameters. His $(c,r)$-Catalan triangles are based on the sequence characterization of Bell-type Riordan arrays. We emphasize that they are also different from the Catalan triangles ${\boldsymbol{C}}( r )$. One way to see this difference is to compare, once again, the $0^\mathrm{th}$ column of both types of matrices. On the one hand, one can check that the $0^\mathrm{th}$ column of a $(c,r)$-Catalan triangle forms a sequence of degree $n$ homogeneous polynomials in $c$ and $r$, whose coefficients are the well-known Narayana numbers. On the other hand, the $0^\mathrm{th}$ column of a Catalan triangle ${\boldsymbol{C}}( r )$ forms a sequence of one-variable polynomials of degree $n$, whose coefficients are the ordinary ballot numbers. Recall that both the Narayana and the ballot numbers are known to refine the Catalan numbers.
The fundamental theorem of Riordan arrays, and the isomorphism between Riordan arrays and Sheffer polynomials (the former being the coefficient matrices for the latter), allow us to interpret ${\boldsymbol{C}}( r )$ as a change of basis matrix in the space of polynomials in one variable. Thus, for instance, the Shapiro array ${\boldsymbol{B}}$ is seen as the change of basis between the Chebyshev polynomials of the second kind, and the standard polynomial sequence of powers of one variable. In this way, by specifying values for the one-variable of the corresponding polynomials, several formulas related to periodic sequences, the sequence of natural numbers, and the sequence of Fibonacci numbers, can be stated and generalized.
General setting {#seGeneralsetting}
===============
The literature on Riordan arrays is vast and constantly growing (see, for instance, [@AMPT; @Bri; @Chen; @dellaRicciaRiordan; @HeCatalan; @HHS; @LMMSidentities; @LMrecurrence; @MRSV; @MS11; @Rog; @ShGeWoWoRgroup; @SprugnoliRioarrayCombsum; @SprugnoliRiordanarraysAbelGould; @WW]). In this section we recall the fundamental theorem of Riordan arrays, and discuss some general facts regarding the generalized Sheffer polynomial sequence associated with any Riordan array. Then we define a one-parameter family of Riordan arrays, and focus our attention on Catalan arrays in Section \[serCatalan\].
Riordan Arrays {#sseRiordanarrays}
--------------
Let $d(z)=d_0+d_1z+d_2z^2+\cdots$, and let $h(z)=h_0+h_1z+h_2z^2+\cdots$ be two formal power series in $z$ with coefficients in a given integral domain $R$, with unit $1$. Assume that $d_0=d(0)\neq0$, and $h_0=h(0)=0$. A *Riordan array* ${\boldsymbol{R}}$ is an infinite lower triangular matrix, whose entries are given by $$\label{eqRiodefi}
{\boldsymbol{R}}_{n,k}=[z^n] d(z) h(z)^k.$$ The operator $[z^n]$ acts on a formal power series $f(z)=\sum_{\sst k\ge0}f_kz^k$ by extracting its $n^{\mathrm{th}}$ coefficient; that is, we have $[z^n]f(z)=f_n$. It is customary to denote ${\boldsymbol{R}}=(d(z),h(z))$. In addition, the set of Riordan arrays for which $h'(0)\neq 0$, and that is equipped with the multiplication given by $$\label{eqRiomult}
(d_1(z),h_1(z)) (d_2(z),h_2(z)) = (d_1(z) d_2(h_1(z)),h_2(h_1(z))),$$ forms a group. We denote this group by ${\mathfrak{R}\mathfrak{i}\mathfrak{o}}$. The group identity is the usual identity matrix ${\boldsymbol{I}}=(1,z)$. Furthermore, for any ${\boldsymbol{R}}\in{\mathfrak{R}\mathfrak{i}\mathfrak{o}}$, its inverse is given by ${\boldsymbol{R}}^{{{^{\sst{-1}}}}}=(1/d(\bar{h}(z)),\bar{h}(z))$, where $\bar{h}(z)$ denotes the compositional inverse of $h(z)$.
The *order* of a formal power series $g(z)=\sum_{k\ge0} g_k z^k$ is the minimal index $k$ such that $g_k\neq0$. Since $R[[z]]$ denotes the ring of formal power series in $z$ with coefficients in $R$, we will write $R_r[[z]]$ to denote the set of formal power series of order $r$. Therefore, if ${\boldsymbol{R}}=(d(z),h(z))\in{\mathfrak{R}\mathfrak{i}\mathfrak{o}}$, then $d(z)\in R_0[[z]]$, and $h(z)\in R_1[[z]]$. For simplicity, and without loss of generality, we further assume that $d_0=1$, and $h_1=h'(0)=1$, so that ${\boldsymbol{R}}_{n,n}=h_1^n=1$, for all $n\ge 0$.
Now, let us define an action $\ast$ of the group ${\mathfrak{R}\mathfrak{i}\mathfrak{o}}$ on $R_0[[z]]$ by $$\label{eqFTRA}
(d(z),h(z))\ast g(z) = d(z) g(h(z)).$$ Formula represents the traditional matrix-column multiplication. Therefore, it is also equivalent to saying that, for any $n\ge 0$, the following identity holds $$\label{eqFTRA2}
{\displaystyle\sum}_{k=0}^n {\boldsymbol{R}}_{n,k} g_k = [z^n] d(z) g(h(z)).$$ Equation is known as the fundamental theorem of Riordan arrays. It is usually used to deal with sums and combinatorial identities. In particular, let $g(z)=1+xz+x^2z^2+\cdots=\frac{1}{1-xz}$. We write $$\label{eqzngf}
p^{{\boldsymbol{R}}}_n(x)=\sum_{k=0}^n {\boldsymbol{R}}_{n,k} x^k = [z^n]\dfrac{d(z)}{1-xh(z)} .$$ We call $(p^{{\boldsymbol{R}}}_n(x))_{n\ge 0}$ the *Sheffer polynomial sequence* associated with ${\boldsymbol{R}}$. Formula implies that the generating function of $p^{{\boldsymbol{R}}}_n(x)$ is given by $$\label{eqGfpnx}
{\displaystyle\sum}_{n\ge 0} p^{{\boldsymbol{R}}}_n(x) z^n = \dfrac{d(z)}{1-xh(z)}.$$ Likewise, the generating function for the Sheffer polynomial associated with the inverse array ${\boldsymbol{R}}^{{^{\sst{-1}}}}=(1/d(\bar{h}(z)),\bar{h}(z))$ is $$\label{eqGfpnxinv}
{\displaystyle\sum}_{n\ge 0} p^{{\boldsymbol{R}}^{{{^{\sst{-1}}}}}}_n\!\!(x) z^n = \dfrac{1}{d(\bar{h}(z))}\dfrac{1}{1-x\bar{h}(z)}.$$
Since ${\boldsymbol{R}}$ is invertible, and the sequence $(x^n)_{n\ge 0}$ is a basis for the linear space $\C[x]$, the sequence $(p^{{\boldsymbol{R}}}_n(x))_{n\ge 0}$ is also a basis for $\C[x]$. Therefore, any ${\boldsymbol{R}}\in{\mathfrak{R}\mathfrak{i}\mathfrak{o}}$ can be interpreted as the change of basis matrix between the standard basis $(x^n)_{n\ge 0}$, and the basis of polynomials $(p^{{\boldsymbol{R}}}_n(x))_{n\ge 0}$. Hence, we have $$\label{eqFTRApninv}
{\displaystyle\sum}_{k=0}^n {\boldsymbol{R}}_{n,k} p^{{\boldsymbol{R}}^{{^{\sst{-1}}}}}_k(x) = x^n.$$
We will use Equation in Section \[seIdentities\] to deduce some interesting combinatorial identities involving the Riordan arrays ${\boldsymbol{C}}$ and ${\boldsymbol{B}}$.
One-parameter Riordan arrays {#sseParametricRiordan}
----------------------------
The classical Pascal array ${\boldsymbol{P}}$ of binomial numbers ${{n}\choose{k}}$ is a Riordan array, since we can write ${\boldsymbol{P}}=(p(z),zp(z))$, where $p(z)=\frac{1}{1-z}$. Given any real number $r$, the Riordan array ${\boldsymbol{P}}( r )=\big(p(rz),zp(rz)\big) = \big(\frac{1}{1-rz},\frac{z}{1-rz}\big)$ is known as the *generalized Pascal array of parameter* $r$. Its entries are given by ${\boldsymbol{P}}( r )_{n,k}={{n}\choose{k}}r^{n-k}$.
\[de1parameterRiordan\] Let ${\boldsymbol{R}}$ be in ${\mathfrak{R}\mathfrak{i}\mathfrak{o}}$. The *$r$-Riordan array* ${\boldsymbol{R}}( r )$ is defined by ${\boldsymbol{R}}( r ) = {\boldsymbol{R}}{\boldsymbol{P}}( r )$.
Set ${\boldsymbol{R}}=(d(z),h(z))$. Then, by Equation , we have ${\boldsymbol{R}}( r ) = \left(\frac{d(z)}{1-rh(z)},\frac{h(z)}{1-rh(z)}\right)$. Using Equation , it is easy to check that the entries of ${\boldsymbol{R}}( r )$ are given by $${\boldsymbol{R}}( r )_{n,k} = \big[z^n\big] \dfrac{d(z)}{1-rh(z)}\left(\dfrac{h(z)}{1-rh(z)}\right)^k = \big[z^{n}\big] \dfrac{d(z) h(z)^k}{\big(1-rh(z)\big)^{k+1}}.$$
For any real number $r$, the Sheffer polynomial sequence $(p_n^{{\boldsymbol{R}}( r )}(x))_{n\ge 0}$ satisfies the identity $$p_n^{{\boldsymbol{R}}( r )}(x) = p_n^{{\boldsymbol{R}}}(x+r).$$
By Equation , we have $$p_n^{{\boldsymbol{R}}( r )}(x) = {\displaystyle\sum}_{k=0}^n {\boldsymbol{R}}( r )_{n,k} x^k = \big[z^n\big] \dfrac{\dfrac{d(z)}{1-rh(z)}}{1-x\dfrac{h(z)}{1-rh(z)}} = \big[z^n\big] \dfrac{d(z)}{1-(x+r)h(z)} = p^{{\boldsymbol{R}}}_n(x+r).$$
Let $n$ be a non-negative integer. The constant term and the sum of coefficients of $p_n^{{\boldsymbol{R}}( r )}(x)$ are given by $$\begin{aligned}
p_n^{{\boldsymbol{R}}( r )}(0) &= {\boldsymbol{R}}( r )_{n,0} = {\displaystyle\sum}_{k=0}^n {\boldsymbol{R}}_{n,k} r^k. \label{eqColumn0Rr} \\[1em]
p_n^{{\boldsymbol{R}}( r )}(1) &= {\displaystyle\sum}_{k=0}^n {\boldsymbol{R}}( r )_{n,k} = {\boldsymbol{R}}(r+1)_{n,0}. \label{eqRowsumsRr}\end{aligned}$$
*r*-Catalan triangles {#serCatalan}
=====================
Let us recall that the generating function for the well-known Catalan numbers $C_n=\frac{1}{n+1}{{2n}\choose{n}}$ is $c(z)=\frac{1-\sqrt{1-4z}}{2z}$. We apply Definition \[de1parameterRiordan\] to the Riordan array ${\boldsymbol{C}}=\big(c(z),zc(z)\big)$.
\[derCatalan\] The *$r$-Catalan triangle* ${\boldsymbol{C}}( r )$ is given by ${\boldsymbol{C}}( r ) = {\boldsymbol{C}}{\boldsymbol{P}}( r )$. Then$$\label{eqrCatalan}
{\boldsymbol{C}}( r ) = \left(\dfrac{c(z)}{1-rzc(z)},\dfrac{zc(z)}{1-rzc(z)}\right) = \left(\dfrac{1-2rz-\sqrt{1-4z}}{2z(1-r+r^2z)},\dfrac{1-2rz-\sqrt{1-4z}}{2(1-r+r^2z)}\right).$$
In particular, we have ${\boldsymbol{C}}(0)={\boldsymbol{C}}$. Moreover, since the identity $c(z)=1+zc(z)^2$ holds, it follows that ${\boldsymbol{C}}(1)={\boldsymbol{C}}{\boldsymbol{P}}(1)=\big(\frac{c(z)}{1-zc(z)},\frac{zc(z)}{1-zc(z)}\big)=\big(c(z)^2,zc(z)^2\big)={\boldsymbol{B}}$.
The next formula follows from Definition \[derCatalan\].
\[leCr\] The entries of ${\boldsymbol{C}}( r )$ are given by $$\label{eqrCatalanentriesRiodef}
{\boldsymbol{C}}( r )_{n,k} = \big[z^{n-k}\big] \left(\dfrac{c(z)}{1-rzc(z)}\right)^{k+1} = {\displaystyle\sum}_{i=0}^{n-k} \frac{i+k+1}{n+1}{{2n-i-k}\choose{n}}{{i+k}\choose{k}} r^{i}.$$
The matrix ${\boldsymbol{C}}( r )$ is an array of polynomials in $r$. By way of illustration, the first entries of ${\boldsymbol{C}}( r )$ are $${\boldsymbol{C}}( r ) =\begin{pmatrix}
1 & & & & & \\
1+r & 1 & & & & \\
2+2r+r^2 & 2+2r & 1 & & & \\
5+5r+3r^2+r^3 & 5+6r+3r^2 & 3+3r & 1 & & \\
14+14r+9r^2+4r^3+r^4& 14+18r+12r^2+4r^3& 9+12r+6r^2& 4+4r& 1 & \\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots\\
\end{pmatrix}.$$ For every $k\ge 1$, column $k$ of ${\boldsymbol{C}}( r )$ is the $(k+1)$-fold convolution of column $0$. Thus, for example, the first columns of ${\boldsymbol{C}}(2)$ are given by the following sequences: $${\boldsymbol{C}}( 2 ) =\begin{pmatrix}
{\href{http://oeis.org/A001700}{\underline{A001700}}} & {\href{http://oeis.org/A008549}{\underline{A008549}}} & {\href{http://oeis.org/A045720}{\underline{A045720}}} &
{\href{http://oeis.org/A045894}{\underline{A045894}}} & {\href{http://oeis.org/A035330}{\underline{A035330}}} & \cdots \\
\end{pmatrix}.$$
The $0^\mathrm{th}$ column of ${\boldsymbol{C}}( r )$ forms a sequence $({\boldsymbol{C}}( r )_{n,0})_{n\ge 0}$ of monic polynomials, which are given by the matrix-column product ${\boldsymbol{C}}[1,r,r^2,\ldots]^{\sst\mathrm{T}}$, as observed in Equation . We can obtain some known sequences by assigning values to the parameter $r$. For instance, we have $({\boldsymbol{C}}(3)_{n,0})_{n\ge 0}={\href{http://oeis.org/A049027}{\underline{A049027}}}$ and $({\boldsymbol{C}}(4)_{n,0})_{n\ge 0}={\href{http://oeis.org/A076025}{\underline{A076025}}}$. Observe that, by Equation , the constant term of each polynomial in the sequence $({\boldsymbol{C}}( r )_{n,0})_{n\ge 0}$ is equal to the Catalan number $C_n$. Moreover, we also have $$C_{n+1}=\frac{1}{n+2}{{2n+2}\choose{n+1}}=\frac{1}{n+1}{{2n+2}\choose{n}}={\boldsymbol{B}}_{n,0}={\boldsymbol{C}}(1)_{n,0}=\sum_{i=0}^{n} \frac{i+1}{n+1}{{2n-i}\choose{n}}=\sum_{k=0}^n{\boldsymbol{C}}_{n,k}.$$ This is a well-known result. In this sense, we say that the polynomial sequence $({\boldsymbol{C}}( r )_{n,0})_{n\ge 0}$ is a refinement of the shifted Catalan sequence $(C_{n+1})_{n\ge 0}$.
Now, set $h(z)=zc(z)$. We have $z=\bar{h}(z)c(\bar{h}(z))$. It is easy to check that $\bar{h}(z)=z-z^2$. Hence, we have ${\boldsymbol{C}}^{{^{\sst{-1}}}}=\big(\nicefrac{1}{c(\overline{zc(z)})},\overline{zc(z)}\big) = (1-z,z-z^2)$. Since ${\boldsymbol{C}}( r )^{{^{\sst{-1}}}}= {\boldsymbol{P}}( r )^{{^{\sst{-1}}}}{\boldsymbol{C}}^{{^{\sst{-1}}}}$, it follows that $$\label{eqrCatalaninverse}
{\boldsymbol{C}}( r )^{{^{\sst{-1}}}}= \left(\dfrac{1}{1+rz},\dfrac{z}{1+rz}\right) \left(1-z,z-z^2\right) = \left(\dfrac{1+(r-1)z}{(1+rz)^2}, \dfrac{z+(r-1)z^2}{(1+rz)^2}\right).$$
\[leCrinv\] The entries of ${\boldsymbol{C}}( r )^{{^{\sst{-1}}}}$ are given by $$\label{eqCrinvpoly}
{\boldsymbol{C}}( r )^{{^{\sst{-1}}}}_{n,k} = \big[z^{n-k}\big] \left(\dfrac{1+(r-1)z}{(1+rz)^2}\right)^{k+1} = (-1)^{n-k}\sum_{i=0}^{n-k}{{i+k+1}\choose{n-i-k}}{{i+k}\choose{k}} r^{i}.$$
Lemma \[leCr\] and Lemma \[leCrinv\] show that both ${\boldsymbol{C}}( r )_{n,k}$ and $(-1)^{n-k}{\boldsymbol{C}}( r )^{{^{\sst{-1}}}}_{n,k}$ are polynomials in $r$ with positive integer coefficients. In addition, note that the polynomial ${\boldsymbol{C}}( r )_{n+1,1}$ is of degree $n$, has constant term $C_{n+1}$, and its leading coefficient is $n+1$. Both lemmas are useful to deduce several formulas and combinatorial identities. For instance, letting $r=0$ in Equation , we obtain $${\boldsymbol{C}}^{{^{\sst{-1}}}}_{n,k} = \big[z^{n-k}\big](1-z)^{k+1} = {{n-2k-2}\choose{n-k}}=(-1)^{n-k}{{k+1}\choose{n-k}}.$$ Similarly, letting $r=1$ in Equation , we get $$\label{eqShapiroinv}
\begin{array}{lcl}
{\boldsymbol{B}}_{n,k}^{{^{\sst{-1}}}}= \big[z^{n-k}\big]\left(\frac{1}{(1+z)^2}\right)^{k+1} &=& (-1)^{n-k}{\displaystyle{{n+k+1}\choose{n-k}}} \\[2ex] &=& (-1)^{n-k}{\displaystyle\sum}_{i=0}^{n-k}{{i+k+1}\choose{n-i-k}}{{i+k}\choose{k}}.
\end{array}$$ Thus, the following formula is immediate.
$$\label{eqCombident1}
{{n+k+1}\choose{n-k}}=\sum_{i=0}^{n-k}{{i+k+1}\choose{n-i-k}}{{i+k}\choose{k}}.$$
The next non-trivial combinatorial identity follows directly from Equation and Equation .
$$\label{eqCombident2}
{\displaystyle\sum}_{k=0}^n {\displaystyle\sum}_{i=0}^{n-k}\frac{i+k+1}{n+1}{{2n-i-k}\choose{n}}{{i+k}\choose{k}}r^i =
{\displaystyle\sum}_{i=0}^n\frac{i+1}{n+1}{{2n-i}\choose{n}}(r+1)^i.$$
As for the generating functions for the Sheffer polynomials associated with ${\boldsymbol{C}}( r )$ and ${\boldsymbol{C}}( r )^{{^{\sst{-1}}}}$, the next formulas are straightforward consequences of Equation and Equation , using Formula and Formula , respectively.
\[propGfShefferCrCrinv\] $$\begin{aligned}
{\displaystyle\sum}_{n\ge 0} p_n^{{\boldsymbol{C}}( r )}(x) z^n &= \dfrac{c(z)}{1-(x+r)zc(z)} \notag \\[1em]
{\displaystyle\sum}_{n\ge 0} p_n^{{\boldsymbol{C}}( r )^{{^{\sst{-1}}}}}\!\!(x) z^n &= \dfrac{1+(r-1)z}{1+(2r-x)z+(r^2-(r-1)x)z^2} \label{eqGfShefferCrinv}\end{aligned}$$
In Section \[seIdentities\], we will use Proposition \[propGfShefferCrCrinv\] to obtain more combinatorial identities.
Special identities {#seIdentities}
==================
Identities involving the Aigner array
-------------------------------------
Set $r=0$ and $x=-1$ in Equation . We obtain $$\begin{aligned}
1+{\displaystyle\sum}_{n\geq 1 }p_n^{{\boldsymbol{C}}^{{^{\sst{-1}}}}}\!\!(-1) (-z)^n &= \dfrac{1+z}{1-z-z^2} \notag \\[2ex]
&= 1+2z+3z^2+5z^3+8z^4+13z^5+21z^6+34z^7+55z^8+\cdots \notag \\[2ex]
&= F_2 + F_3z + F_4z^2 + F_5z^3 + \cdots, \end{aligned}$$ where $F_n$ is the $n^\mathrm{th}$ Fibonacci number, for $n\ge 2$. Hence, for $n\ge 0$, it holds that $$(-1)^np_n^{{\boldsymbol{C}}^{{^{\sst{-1}}}}}(-1)=F_{n+2}.$$ Since $p_n^{{\boldsymbol{C}}^{{^{\sst{-1}}}}}\!\!(x)=\sum_{k=0}^n {\boldsymbol{C}}^{{^{\sst{-1}}}}_{n,k} x^k$, we have $${\boldsymbol{C}}^{{^{\sst{-1}}}}\begin{pmatrix}1\\-1\\1\\-1\\\vdots\end{pmatrix} = \begin{pmatrix}F_2\\-F_3\\ F_4\\-F_5\\\vdots
\end{pmatrix}, \phantom{A}\text{or equivalently},\phantom{A} {\boldsymbol{C}}\begin{pmatrix}F_2\\-F_3\\ F_4\\ -F_5\\\vdots
\end{pmatrix}=\begin{pmatrix}1\\-1\\1\\-1\\\vdots
\end{pmatrix}.$$ Then $$\sum_{k=0}^n {{k+1}\choose{n-k}}=F_{n+2}, \phantom{A}\text{and}\phantom{A} \sum_{k=0}^n \dfrac{k+1}{n+1} {{2n-k}\choose{n}}(-1)^{n-k} F_{k+2}=1.$$
Identities involving the Shapiro array
--------------------------------------
Set $r=1$ in Equation . Shapiro, Woan, and Getu [@ShaRuns] observed that the Catalan triangle ${\boldsymbol{B}}$ verified the following interesting identity: $$\label{eqShaidentity}
\boldsymbol{B}
\left(
\begin{array}{c}
1 \\
2 \\
3 \\
4 \\
5 \\
\vdots
\end{array}
\right)=\left(
\begin{array}{c}
1 \\
4\\
4^2\\
4^3\\
4^4\\
\vdots
\end{array}
\right).$$ The problem of finding a matrix identity that extends the sequence $(1,4,4^2,4^3,\ldots)$ to the sequence $(1,k,k^2,k^3,\ldots)$ in Equation , was studied by Chen, Li, Shapiro and Yan [@Chen]. They generalized the following known recurrence relation for ${\boldsymbol{B}}$; namely, $${\boldsymbol{B}}_{n,k}=\boldsymbol{B}_{n-1,k-1}+2{\boldsymbol{B}}_{n-1,k}+{\boldsymbol{B}}_{n-1,k+1},$$ by constructing a matrix $\boldsymbol{M}$ that satisfies the identity $$\boldsymbol{M}
\left(
\begin{array}{c}
1 \\
1+t \\
1+t+t^2 \\
1+t+t^2+t^3 \\
\vdots
\end{array}
\right)=\left(
\begin{array}{c}
1 \\
k\\
k^2\\
k^3\\
\vdots
\end{array}
\right).$$ The first column of $\boldsymbol{M}$ (that is, its $0^\mathrm{th}$ column) was then interpreted in terms of weighted partial Motzkin paths. Here, we approach the problem of extending the sequence $(1,4,4^2,4^3,\ldots)$ to the sequence $(1,k,k^2,k^3,\ldots)$, by keeping ${\boldsymbol{C}}( 1 )={\boldsymbol{B}}$ fixed, and using Sheffer sequences. More explicitly, by Equation , we have $$1+{\displaystyle\sum}_{n\geq 1} p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x)z^n = \dfrac{1}{1-(x-2)z+z^2}.$$ Furthermore, by Equation , we also have $$\label{eqShaxk}
{\boldsymbol{B}}\begin{pmatrix}
1 \\
p_1^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x) \\
p_2^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x) \\
\vdots \\
\end{pmatrix}
=\begin{pmatrix}
1 \\
x \\
x^2 \\
\vdots \\
\end{pmatrix}, \phantom{A}\text{or equivalently, }\phantom{A}
{\boldsymbol{B}}^{{^{\sst{-1}}}}\begin{pmatrix}
1 \\
x \\
x^2 \\
\vdots \\
\end{pmatrix}
=\begin{pmatrix}
1 \\
p_1^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x) \\
p_2^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x) \\
\vdots \\
\end{pmatrix}.$$ By Identity , the polynomial $p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x)$ is given by $$\label{eqCheb}
p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x) = \sum_{k=0}^n {\boldsymbol{B}}^{{^{\sst{-1}}}}_{n,k} x^k = \sum_{k=0}^n (-1)^{n-k} {{n+k+1}\choose{n-k}} x^k.$$
Note that $p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x)=U_n\left(\frac{x-2}{2}\right)$, where $(U_n(x))_{n\ge 0}$ denotes the sequence of Chebyshev polynomials of the second kind [@Comtet]. Moreover, the sequence $(p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x))_{n\ge 0}$ satisfies the three-term recursion $$p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x)=(x-2)p_{n-1}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(x)-p_{n-2}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(x),\quad n\geq 2,$$ with initial values $p_0^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x)=1$ and $p_1^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x)=x-2$. Whenever $x\in\Z$, the polynomial sequence $(p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(x))_{n\ge 0}$ reduces to an integer sequence. By specifying the values of $x$ in Equation , we obtain many interesting relations.
When $x=2$, the recursion $p_{n}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(2)= -p_{n-2}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(2)$, with initial values $p_{0}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(2)=1$ and $p_{1}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(2)=0$, is solved by the period-$4$ sequence $1,0,-1,0$. Hence, we have $$\left(
\begin{array}{cccccc}
1 & & & & &\\
2 & 1 & & & &\\
5 & 4 & 1 & & &\\
14 & 14 & 6 & 1 & &\\
42 & 48 & 27 & 8 & 1&\\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots
\end{array}
\right)\left(
\begin{array}{c}
1 \\
0 \\
-1 \\
0 \\
1 \\
\vdots
\end{array}
\right)=\left(
\begin{array}{c}
1 \\
2\\
2^2\\
2^3\\
2^4\\
\vdots
\end{array}
\right).$$ When $x=3$, the recursion $p_{n}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(3)=p_{n-1}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(3)-p_{n-2}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(3)$, satisfying the initial conditions $p_{0}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(3)=p_{1}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(3)=1$, is solved by the period-$6$ sequence $1,1,0,-1,-1,0$. Thus, we have $$\left(
\begin{array}{cccccc}
1 & & & & &\\
2 & 1 & & & &\\
5 & 4 & 1 & & &\\
14 & 14 & 6 & 1 & &\\
42 & 48 & 27 & 8 & 1&\\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots
\end{array}
\right)\left(
\begin{array}{c}
1 \\
1 \\
0 \\
-1 \\
-1 \\
\vdots
\end{array}
\right)=\left(
\begin{array}{c}
1 \\
3\\
3^2\\
3^3\\
3^4\\
\vdots
\end{array}
\right).$$ These two cases are better understood by recalling that $U_n(\cos y)=\frac{\sin(n+1)y}{\sin y}$. Hence, we have $p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(2)=U_n\left(\cos\frac{\pi}{2}\right)$ and $p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!\!(3)=U_n\left(\cos\frac{\pi}{3}\right)$. These identities explain why we obtained periodic sequences.
\[ex:Nat\]
When $x=4$, the recursion $p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(4)=2p_{n-1}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(4)-p_{n-2}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(4)$, with initial values $p_0^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(4)=1$, $p_1^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(4)=2$, is solved by the sequence $(p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(4))_{n\ge 0}=(n+1)_{n\ge 0}$. Thus, we obtain the Identity : $$\left(
\begin{array}{cccccc}
1 & & & & &\\
2 & 1 & & & &\\
5 & 4 & 1 & & &\\
14 & 14 & 6 & 1 & &\\
42 & 48 & 27 & 8 & 1&\\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots
\end{array}
\right)\left(
\begin{array}{c}
1 \\
2 \\
3 \\
4 \\
5 \\
\vdots
\end{array}
\right)=\left(
\begin{array}{c}
1 \\
4\\
4^2\\
4^3\\
4^4\\
\vdots
\end{array}
\right).$$ The previous matrix identity is equivalent to Chen’s combinatorial formula [@Chen]: $$\label{eqMotivationChen}
\sum_{k=0}^n\frac{(k+1)^2}{n+1} {{2n+2}\choose{n-k}} = 4^n.$$
\[ex:Fib\]
When $x=5$, the recursion $p_n^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(5)=3p_{n-1}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(5)-p_{n-2}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(5)$, with initial conditions $p_0^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(5)=1$ and $p_1^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(5)=3$, is solved by the subsequence of Fibonacci numbers $(p_{n}^{{\boldsymbol{B}}^{{^{\sst{-1}}}}}\!(5))_{n\ge 0}=(F_{2n+2})_{n\ge 0}$. Thus, Shapiro’s triangle ${\boldsymbol{B}}$, and the Fibonacci numbers of even index, are related by the following identity: $$\left(
\begin{array}{cccccc}
1 & & & & &\\
2 & 1 & & & &\\
5 & 4 & 1 & & &\\
14 & 14 & 6 & 1 & &\\
42 & 48 & 27 & 8 & 1&\\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots
\end{array}
\right)\left(
\begin{array}{c}
1 \\
3 \\
8 \\
21 \\
55 \\
\vdots
\end{array}
\right)=\left(
\begin{array}{c}
1 \\
5\\
5^2\\
5^3\\
5^4\\
\vdots
\end{array}
\right).$$ Equivalently, we have $$\label{eqFib}
\sum_{k=0}^n\frac{k+1}{n+1} {{2n+2}\choose{n-k}}F_{2k+2} = 5^n.$$
Using the well-known binomial recurrence relation ${{n}\choose{k}} = {{n-1}\choose{k-1}}+{{n-1}\choose{k}}$ in Equation , and then inverting Formula , we obtain $$F_{2n+2}=(-1)^n\sum_{k=0}^n\left({{n+k+2}\choose{n-k}}-{{n+k+1}\choose{n-k-1}}\right)(-5)^k.$$ Moreover, seeing as $F_{2n+3}=F_{2(n+1)+2}-F_{2n+2}$, we can deduce an analogous formula for the Fibonacci numbers of odd index; that is $$F_{2n+1}=(-1)^n\sum_{k=0}^{n}\left({{n+k+2}\choose{n-k}}-{{n+k}\choose{n-k-2}}\right)(-5)^k.$$
Acknowledgements {#se:acknow}
================
The authors wish to thank the editor and the referee for their helpful comments.
------------------------------------------------------------------------
2010 *Mathematics Subject Classification*: Primary 15B; Secondary 05A19, 11B37, 11B83, 11C.
*Keywords*: Riordan array, Pascal array, Catalan triangle, Sheffer polynomial, combinatorial identity, recurrence, Chebyshev polynomial, Fibonacci number.
------------------------------------------------------------------------
(Concerned with sequence [[](http://oeis.org/A001700)]{}, [[](http://oeis.org/A008549)]{}, [[](http://oeis.org/A033184)]{}, [[](http://oeis.org/A035330)]{}, [[](http://oeis.org/A039598)]{}, [[](http://oeis.org/A045720)]{}, [[](http://oeis.org/A045894)]{}, [[](http://oeis.org/A049027)]{}, and [[](http://oeis.org/A076025)]{}.)
------------------------------------------------------------------------
[^1]: Corresponding author.
[^2]: Work performed within the activities of Centro de Análise Funcional, Estruturas Lineares e Aplicações (Faculdade de Ciências, Universidade de Lisboa) and supported by the fellowship SFRH/BPD/48223/2008 provided by the Portuguese Foundation for Science and Technology (FCT).
[^3]: Actually, Shapiro used the formula ${\boldsymbol{B}}_{n,k}=\frac{k}{n}{{2n}\choose{n+k}}$, for $n,k\ge 1$. His formula is just a shifted version of the formula shown previously for ${\boldsymbol{B}}_{n,k}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Our aim is to set up the cornerstones of Koszul duality in general operadic categories introduced in [@duodel]. In particular, we will prove that operads (in our generalized sense) governing the most important operad- and/or PROP-like structures as classical operads, their variants as cyclic, modular or wheeled operads, and also diverse versions of PROPs such as properads, dioperads, , and still more exotic stuff as permutads and pre-permutads are quadratic, and describe their Koszul duals.
To this end we single out some additional properties of operadic categories ensuring that free operads admit a nice explicit description, and investigate how these properties interact with discrete operadic (op)fibrations which we use as a mighty tool to construct new operadic categories from the old ones. Particular attention is payed to the operadic category of graphs and to its clones, but several other examples are given as well.
Our present work provides an answer to the questions “What does encode a type of operads?” and “How to construct Koszul duals to these objects?” formulated in the last Loday’s 2012 talk [@loday:talk].
address:
- 'Macquarie University, NSW 2109, Australia'
- 'Mathematical Institute of the Academy, [Ž]{}itn[á]{} 25, 115 67 Prague 1, The Czech Republic'
author:
- Michael Batanin
- Martin Markl
title: Koszul duality in operadic categories
---
Introduction {#introduction .unnumbered}
============
The present work sets up the basic stones of our general theory of “operad-like” structures. By them we mean, besides the classical operads in the sense of Boardman-Vogt and May [@boardman-vogt:73; @may:1972] and their more recent variants as cyclic, modular or wheeled operads [@getzler-kapranov:CPLNGT95; @getzler-kapranov:CompM98; @mms], also diverse versions of PROPs such as properads [@vallette:TAMS07], dioperads [@gan], [@mv], and still more exotic stuff as permutads and pre-permutads [@loday11:_permut] or protoperads [@leray]. Also Batanin’s $n$-operads [@batanin:conf; @batanin:AM08] appear in our scope. One may vaguely characterize operad and/or PROP-like structures as those generalizing compositions of multivalued functions.
[**History.**]{} To our knowledge, the first attempt to systematize this kind of objects was made by the second author in 2008 [@markl:handbook]. He considered structures with operations modeled by contractions along edges of graphs (called ‘pasting schemes’ in this context) of the type particular to the concrete situation. These schemes were required to satisfy an important property of [*hereditarity*]{}, which is a specific stability under contractions of subgraphs. This property was later redressed into categorical garment in the notion of a [*Feynman category*]{} [@kaufmann-ward:Fey]. Hereditarity however played an important rôle already in [@borisov-manin]. Let us close this brief fly over history by mentioning [@getzler:operads-revisited] predating Feynman categories, see also the follow-ups [@BKW; @comprehensive]. Finally, in [@BB] an approach to general operad-like structures through the use of polynomial monads was developed. We are commenting more on the connections between these approaches with ours, as well as on their advantages and disadvantages, at the end of the Introduction.
[**The setup.**]{} The approach of this work differs from the one of [@markl:handbook] or [@BB]. It is based on the notion of an [*operadic category*]{}. The idea goes back to the first author’s work on higher category theory based on a higher generalization of non-symmetric (non-$\Sigma$) operads [@batanin:en]. In this formalism, a higher version of Eckman-Hilton argument was described by reformulating the classical notion of a (symmetric) operad and Batanin’s notion of an $n$-operad in such a way that a comparison of the two notions became possible [@batanin:conf; @batanin:AM08]. The fruitfulness of this idea was then confirmed in [@batanin:br].
In our work on duoidal Deligne’s conjecture we came to understanding that the same categorical scheme is very useful and, indeed, necessary for the study of many other standard and nonstandard operad-like structures. Thus the concept of operadic categories was introduced by the authors in [@duodel].
Intuitively, morphisms in operadic categories poses [*fibers*]{} whose properties are modeled by the preimages of maps between finite sets. Unlike in Barwick’s operator categories [@barwick], the fibers need not be pullbacks. Each operadic category $\ttO$ has its [*operads*]{} and each $\ttO$-operad $P$ has its category of [*$P$-algebras*]{}.
An archetypal operadic category is the skeletal category $\Fin$ of finite sets. Also hereditary categories of graphs are operadic. Examples of different scent are Batanin’s $n$-trees and $n$-ordinals, or the operadic category supporting permutads. For convenience of the reader we recall definitions of operadic categories and related notions in the opening Section \[a0\].
The background scheme of our approach is the triad in Figure \[triad\],
$$\begin{aligned}
&\boxed{\hbox {level 1: $\ttO^+$-operads}}&
\\
&\Downarrow&
\\
&\boxed{\hbox {level 0: $\ttO$-operads = ${\sf 1_{\ttO^+}}$-algebras}}&
\\
&\Downarrow&
\\
&\boxed{\hbox {level -1: algebras of $\ttO$-operads}}&\end{aligned}$$
in which ‘$\Downarrow$’ means ‘is governed by.’ At level $0$ one sees operads in an operadic category $\ttO$. We consider algebras of these operads as objects at level $-1$. It turns out that $\ttO$-operads are algebras over the constant operad ${\sf 1}_{\ttO^+}$ in a certain operadic category $\ttO^+$ called the [*$+$-construction*]{} of $\ttO$, which we place at level $1$. The triad can in fact be continued upwards to infinity.
An example is the [*classical triad*]{} in which $\ttO$ is the operadic category $\Fin$ of finite sets. $\Fin$-operads are the classical operads, which simultaneously appear as algebras over the constant operad ${\sf 1}_\RTr$ in the operadic category $\RTr$ of rooted trees, which is $\Fin^+$. At level $-1$ we find algebras over the classical operads.
Strong inspiration of our setup was the seminal paper [@getzler-kapranov:CompM98] whose authors realized that modular operads are algebras over a certain (hyper)operad. They thus constructed levels $0$ and $-1$ of the triad for the operadic category $\ggGrc$ of connected genus-graded directed graphs. It turns out $\ggGrc^+$ at level $1$ is the category of graphs from $\ggGrc$ with a hierarchy of nested subgraphs. We will call the resulting scheme the [*Getzler-Kapranov triad*]{}.
The novelty of our approach is that we systematically put the structures we want to study at level $-1$ so that they appear as algebras over a certain operad. For instance, cyclic operads in our setup are algebras over the constant operad ${\sf 1}_{\Tr}$ in the operadic category $\Tr$ of trees, though they themselves are [*not*]{} operads in any operadic category.
[**The results and perspectives.**]{} The interaction between levels $0$ and $-1$ captures the bar-cobar duality. It is well-known today that the correct place for the bar construction ${\mathbb B}(A)$ of an algebra $A$ over the classical quadratic operad $\oP$ is the category of $\oP^!$-(co)algebras, where $\oP^!$ denotes the Koszul dual of $\oP$ [@markl-shnider-stasheff:book Definition II.3.7]. One thus should expect e.g. the bar-cobar duality between commutative associative and Lie algebras, because the Koszul dual of the operad $\Com$ governing commutative associative algebras is the operad $\Lie$ for Lie algebras. And, indeed, the bar construction of a commutative associative algebra is its Harrison complex, which is a dg-Lie algebra.
More recent appearance of this phenomenon is [@getzler-kapranov:CompM98] which shows that the bar construction of a modular operad (called the Feynman transform there) is [*not*]{} a modular operad, but a certain twisted version of it, later called an odd modular operad [@KWZ; @markl:odd]. Our approach provides the following explanation. Modular operads are algebras over the constant operad $\term_\ggGrc$ in the operadic category $\ggGrc$ of genus-graded connected directed graphs. This operad is quadratic, and its suitably defined Koszul dual $\oddGr
:=\term_\ggGrc^!$ governs odd modular operads. Our theory offers the following generalization.
If $P$ is a quadratic operad in an operadic category $\ttO$ and $A$ its algebra, then there exists a natural bar construction $\BAR(A)$ of $A$ which lives in the category of $P^!$-algebras over a suitably defined Koszul dual $P^!$. The related bar-cobar duality reflects the derived categories of $P$ and $P^!$-algebras.
This is of course true also for the interaction between levels $+1$ and $0$, but the crucial nice additional feature is that the constant $\tt0^+$-operad ${\sf 1}_{\ttO^+}$ is Koszul self-dual. In this sense, the $+$-constructions improves properties, analogously to the similar property of other in higher category theory [@baez-dolan; @KJBM] or of the suspension in topology. Since, as the triad in Figure \[triad\] teaches us, each $\ttO$-operad $P$ is simultaneously also a ${\sf 1}_{\ttO^+}$-algebra, one has its bar construction $\BAR(P)$, which is an ${\sf 1}_{\ttO^+}$-algebra by the self-duality, i.e. an $\ttO$-operad, again.
If $P$ is quadratic, one moreover has a natural map $$\label{Pozitri_letim_na_Vivat_tour.}
P \longleftarrow {\mathbb B}(P^!)$$ of dg-$\ttO$-operads. In parallel to the case of classical operads we say that $P$ is [ *Koszul*]{} if this map is a homology isomorphism. Algebras over ${\mathbb B}(P^!)$ are then strongly homotopy $P$-algebras. For the classical triad $$%\label{classical}
\boxed{\hbox {$\RTr$-operads}}\Rightarrow
\boxed{\hbox {classiscal operads = ${\sf 1_{\RTr}}$-algebras}}\Rightarrow
\boxed{\hbox {classical operad algebras}}$$ with $\ttO = \Fin$ we get the standard theory of Koszul duality for operads [@ginzburg-kapranov:DMJ94]. If $\oP$ is a classical quadratic Koszul operad, such as $\Com$, $\Ass$ or $\Lie$, $\BAR(\oP^!)$-algebras are classical strongly homotopy algebras such as $L_\infty$-, $A_\infty$- or $C_\infty$-algebras.
As less a standard example, consider the triad $$\boxed{\hbox {$\RTr^+$-operads}}\Rightarrow
\boxed{\hbox {$\RTr$-operads = ${\sf 1_{\RTr^+}}$-algebras}}\Rightarrow
\boxed{\hbox {algebras over $\RTr$-operads}}$$ related to the operadic category $\ttO = \RTr$ of directed rooted trees, in which $\RTr^+$ consists of directed rooted trees with a hierarchy of nested subtrees (see [@KJBM] how to iterate this construction). As we already know, classical operads are algebras over the constant $\RTr$-operad ${\sf 1}_\RTr$, which is quadratic self-dual. Its self-duality follows from the fact that $\RTr = \Fin^+$ and is also established explicitly in the present work. The canonical map (\[Pozitri\_letim\_na\_Vivat\_tour.\]) in this case equals $${\sf 1}_\RTr \longleftarrow {\mathbb B}({\sf 1}_\RTr).$$ We conjecture that ${\sf 1}_\RTr$ is Koszul. If it is so, then ${\mathbb B}({\sf
1}_\RTr)$-algebras would represent a canonical version of classical operads up to homotopy. A similar analysis can be made e.g. for the Getzler-Kapranov triad $$\boxed{\hbox {$\ggGrc^+$-operads}}\Rightarrow
\boxed{\hbox {$\ggGrc$-operads = ${\sf 1_{\ggGrc^+}}$-algebras}}\Rightarrow
\boxed{\hbox {algebras over $\ggGrc$-operads}}\ .$$
[**Aims of the present work.**]{} We focus on the interaction between levels $-1$ and $0$ of the triad in Figure \[triad\]. We describe free operads in operadic categories, and introduce quadratic operads and their Koszul duals. We then explicitly analyze operads whose algebras are the most common PROP-like structures, including the description of their Koszul duals. The remaining issues mentioned in the previous paragraphs will be addressed in a future work.
Our theory shall offer a framework for the study of Koszulity of operads in general operadic categories. We conjecture that most if not all operads governing the structures mentioned above are Koszul. An immediate gain would be canonical constructions of ‘up to coherent homotopies’ versions of these structures. So far it has been established, besides the classical examples, in [@markl:perm] for the operad governing permutads. This conjectural Koszulity might also provide interpretation of some numerical hypotheses formulated in the last pages of Loday’s slides [@loday:talk].
[**The plan.**]{} In Section \[a0\] we recall operadic categories and related notions, using almost verbatim the material of [@duodel]. In Section \[Pojedu\_vecer\_nebo\_ted?\] we single out some finer additional properties of operadic categories ensuring that free operads in these categories are of a particularly nice form. Section \[Ceka\_mne\_Psenicka.\] is devoted to our construction of an important operadic category of graphs which, as we show in Section \[Poletim\_letos\_do\_Sydney?\], satisfies all these additional requirements. We will also see that several subtle properties of graphs may be conveniently expressed in the language of our theory. In Section \[Minulou\_sobotu\_jsem\_odletal\_vlekarskou\_osnovu.\] we recall from [@duodel] discrete (op)fibrations and the related Grothendieck’s construction, and use it as a mighty tool that produces new operadic categories from old ones. One more application of this technique is given in Section \[+construction\] where we introduce a construction over an operadic category $\ttO$ as the Grothendieck construction $\int_\ttO\PO$ of the colored operad $\PO$ for free $\ttO$-operads.
Classical operads in the spirit of Peter May [@may:1972] are collections $\{P(n)\}_{n\geq 1}$ of $\Sigma_n$-modules[^1] with structure operations $$\gamma : P(n) \ot P(n_1) \ot \cdots \ot P(n_k) \to P(n_1 + \cdots + n_k)$$ given for any $n, \Rada n1k \geq 1$ satisfying appropriate associativity and equivariance axioms. An alternative definition based on structure operations of the form $$\label{letel_jsem_ve_snehovych_prehankach}
\circ_i : P(m) \ot P(n) \to P(m+n), \ m,n \geq 1,$$ was given much later in [@markl:zebrulka Definition 1.1]. It turns out that under some quite standard assumptions, for instance in the presence of units, augmentations or connectivity, both definitions agree, see e.g. [@markl:zebrulka Observation 1.2] or [@markl:handbook Proposition 13], though there are structures possessing $\gamma$-operations only [@markl:handbook Example 19]. Operad-like structures based on ‘partial compositions’ in (\[letel\_jsem\_ve\_snehovych\_prehankach\]) were later called Markl’s operads.
Also operads in general operadic categories exist in two disguises which are, under favorable conditions, equivalent – in a form where the compositions in all inputs are made simultaneously; this is how they were introduced in [@duodel] – and in Markl’s form where they are performed one after one. The crucial advantage of Markl’s form is, as in the classical case, that free Markl’s operads are naturally graded by the length of the chain of compositions.
Markl’s operads in the context of general operadic categories are introduced in Section \[section-markl\]. Sections \[letam\_205\]–\[Zitra\_bude\_foukat\_a\_ja\_musim\_preletet\_205.\] then contain material needed for the definition of free Markl’s operads. While the underlying structure of a classical operad is a collection of spaces equipped with actions of symmetric groups, for general operadic categories the situation is subtler. It turns out that the rôle of underlying collections is played by presheaves on a category $\QV(e)$, constructed using virtual isomorphisms in Section \[letam\_205\]. The precise relation of Markl’s operads to the category $\QV(e)$ is explained in Section \[Zitra\_bude\_foukat\_a\_ja\_musim\_preletet\_205.\]. Free Markl’s operads are then explicitly described in Section \[zitra\_letim\_do\_Pragy\]. Having free operads available, we define, in Section \[pairing\], quadratic operads and their Koszul duals.
The remaining sections are devoted to explicit calculations. In Section \[zitra\_budu\_pit\_na\_zal\] we study the constant operad $\term_\ggGrc$ whose algebras are Getzler-Kapranov’s modular operads. We show that this operad is quadratic and that its Koszul dual describes odd (twisted) modular operads. In Section \[Asi\_pojedu\_vecer.\] we make the similar analysis for operads describing ordinary and cyclic operads, and pre-permutads. In Section \[Ta\_moje\_lenost\_je\_strasna.\] we continue this analysis for wheeled properads, dioperads, $\frac12$-PROPs and permutads.
In Appendix \[a1\] we recall modular operads and their odd version, and classical Markl’s operads. To help the reader navigating, we included an index of terminology.
[**Comparison with other approaches.**]{} In their recent monograph [@kaufmann-ward:Fey], M. Kaufman and B. Ward propose a different theory of multivariable structures, based on the notion of a Feynman category. It is a symmetric monoidal functor ${\mathcal F}:\mathbb{S}{\tt C}\to {\tt M}$, where ${\tt M}$ is a symmetric monoidal category and $\mathbb{S}{\tt C}$ a free symmetric monoidal category on a groupoid ${\tt C}$. This functor should satisfy certain conditions which we do not specify here.
An operad in their sense is a symmetric monoidal functor from ${\tt M}$ to a symmetric monoidal category $\ttV$. They constructed Feynman categories which codify symmetric operads, cyclic operads, modular operads, and many other structures. In fact, it was shown in [@BKW] that Feynman categories codify exactly the same structures as colored operads.
Not all of the above types of operads posses the underlying operadic categories in our sense. Namely, there are no operadic categories for cyclic and modular operads. These objects in our set-up however appear as algebras over the constant operad in a appropriate operadic category, i.e. as the $(-1)$-parts of the corresponding triads.
Given an operadic category $\ttO$ one can, as we did in Section \[+construction\], construct a $\Sigma$-free colored operad $\PO$ whose algebras are $\ttO$-operads, and then convert $\PO$ into a Feynman category using [@BKW Theorem 5.16]. The resulting Feynman category ${\mathcal F}(\ttO)$ is a slightly more sophisticated version of the usual PROP associated to the operad $\PO$. The category of ${\mathcal F}(\tt O)$-operads in Kaufmann-Ward sense is equivalent to the category of $\PO$-algebras , hence to the category of $\ttO$-operads in our sense.
The passage from operadic categories to Feynman categories described above may create the impression that Feynman categories encompass more structures than operadic categories do. It is not the case, though. There is an inverse process which goes as follows.
By a coherence result [@BKW Proposition 1.6] we may assume without loss of generality that the Feynman category $\mathcal{F} :\mathbb{S}{\tt C}\to {\tt M}$ we start from is strict, meaning that all symmetric monoidal data involved are as strict as possible. Then we associate to $\mathcal{F}$ the category $$\ttO(\mathcal{F}) := \coprod_{c\in {\tt C}} {\tt M}/c.$$ which can be equipped with a natural structure of an operadic category. Moreover, the category of algebras of the terminal $\ttO(\mathcal{F})$-operad is equivalent to the category of $\mathcal{F}$-operads in the Kaufmann-Ward sense. The functors $$\xymatrix@C=4em@1{\OpCat \ar@/^{1.3em}/[r]^{{\mathcal F}(-)} &
\ar@/^{1.3em}/[l]_{\ttO(-)} {\tt FeyCat}
}$$ from operadic categories to Feynman categories and back are not adjoint to each other but are related in an interesting manner. Namely, the composite $:\ttO \circ {\mathcal F} (-):\OpCat \to \OpCat$ equals the construction described in Section \[+construction\] while the composite ${\mathcal F} \circ \ttO (-) : {\tt FeyCat} \to {\tt FeyCat}$ associates to a Feynman category $\mathcal F$ the Feynman category for $\mathcal{F}$-hyperoperads introduced in [@kaufmann-ward:Fey Section 4] as an analogue of the Baez-Dolan construction for symmetric operads.
The construction of $\ttO(\mathcal{F})$ admits the following explanation in terms of classifiers of internal algebras of polynomial monads. First, convert $\mathcal{F}$ to a colored operad $A$ in $\tt Set$ using . Then take its canonical $\Sigma$-cofibrant replacement $\mathbf A$ as a $\tt Cat$-polynomial monad [@Weber:polynomial Lemmas 3.13–3.17]. Its absolute classifier $\mathbf{A}^{\mathbf{A}}$ is a certain codescent object [@weber]. It is an operadic category isomorphic to $\ttO(\mathcal{F})$.
The discussion above demonstrates that, if we focus on operads with values in symmetric monoidal categories as it is necessary for the Feynman category approach, we are able to describe exactly the same structures using either operadic categories or Feynman categories. Feynman categories have an advantage of expressing all structures of interest in the same relatively simple language of symmetric monoidal functors. Our language, on the other hands, allows us to keep an important distinction between operads and their algebras expressed in the concept of a triad, as explained above.
Yet another difference is that Feynman categories tend to be a lot more complicated and combinatorially involved objects than operadic categories. For example, the skeletal category $\Fin$ of finite sets is the category whose operads are the ’classical’ symmetric operads of P. May while the Feynman category for these operads is the category of forests of rooted trees. Operadic categories are essentially the most distilled algebraic structures which contain all information determining operad-like structures of a given type along with their algebras.
Our approach has one more important feature: operadic categories can be used to define operads and their algebras with values in more general structures than in symmetric monoidal categories. This was one of our motivations for introducing them in [@duodel], where we needed operads with values in a duoidal category. This led to a proof of a very general form of Deligne’s conjecture.
Another crucial feature of our approach is the rôle of discrete operadic fibrations and/or opfibrations between operadic categories, and their relations to operads resp. cooperads via an analog of Grothendieck’s construction. We use it as our main tool for constructions of new operadic categories from the old ones. Our discrete operadic fibrations also explain some constructions in [@BW Section 7]. While operadic fibrations have their analogs in [@kaufmann-ward:Fey] as decorations of Feynman categories, we are not aware of a similar analog for opfibrations in the Kaufman-Ward language.
We are planning to write the details of comparison of our approach to the approaches of [@kaufmann-ward:Fey; @borisov-manin; @comprehensive; @getzler:operads-revisited; @BKW] and [@BB] in a separate paper. We however want to stress that in our opinion there can not be a unique best theory of operad-like structures. We expect that other approaches will emerge as well, and that these new approaches and already existing ones will be complementary to each other and all will be useful in the study of the multifaceted word of multivariable structures and their applications.
[**Conventions.**]{} Operadic categories and related notions were introduced in [@duodel]; some basic concepts of that paper are recalled in Section \[a0\]. We will freely use the terminology and notation from there. All operadic categories in this work will be strict and constant-free, see Definition \[constant-free\] below. Chosen local terminal objects of an operadic category $\ttO$ will be denoted by $U$ with various decorations such as $U',U'', U_c, U_a$, &c. We will call these chosen local terminal objects the [*trivial*]{} ones. Likewise, local terminal (not necessarily chosen) objects will be denoted by $u',u'', u_c, u_a$, &c.
A [**]{} is a morphism in $\ttO$ whose all fibers are trivial.[^2] will be indicated by $\sim$, by $\cong$; a preferred notation for both of them will be something resembling permutations, like $\sigma$, $\omega$, $\pi$, &c.
We will denote by $\QO\subset \ttO$ the subcategory of , and by $\DO\subset \ttO$ the subcategory of morphisms for which $|f|$ is order-preserving. By $\Fin $ we denote the operadic category of finite ordinals $\bar{n} =\{1,\ldots,n\} ,
n\in \bbN$, and their set-theoretic maps.
Operadic categories and their operads {#a0}
=====================================
In this introductory section we recall, for the convenience of the reader, some basic definition from [@duodel]. The material is taken almost verbatim. The reader may also wish to look at S. Lack’s paper [@lack] for a characterization of operadic categories in the context of skew monoidal categories, or at [@GKW] by R. Garner, J. Kock and M. Weber for yet another point of view.
Let $\Fin$ be the skeletal category of finite sets.[^3] The objects of this category are linearly ordered sets $\bar{n} =\{1,\ldots,n\} ,
n\in \bbN.$ Morphisms are arbitrary maps between these sets. We define the $i$th fiber $f^{-1}(i)$ of a morphism $f: T \to S$, $i\in S$, as the pullback of $f$ along the map $\bar{1}\to S$ which picks up the element $i$, so this is an object $f^{-1}(i)=\bar{n}_i \in \Fin$ which is isomorphic as a linearly ordered set to the preimage $\big\{j\in T \ | \ f(j) =i\big\}$. Any commutative diagram in $\Fin$ $$\xymatrix@C = +1em@R = +1em{
T \ar[rr]^f \ar[dr]_h & & S \ar[dl]^g
\\
&R&
}$$ then induces a map $f_i:h^{-1}(i)\to g^{-1}(i)$ for any $i\in R$. This assignment is a functor $\Fib_i:\Fin/R\to \Fin.$ Moreover, for any $j\in S$ we have the equality $f^{-1}(j) =
f^{-1}_{g(j)}(j).$ The above structure on the category $\Fin$ motivates the following abstract definition.
A [*strict operadic category*]{} is a category $\ttO$ equipped with a ‘cardinality’ functor $|\dash|:\ttO\to \Fin$ having the following properties. We require that each connected component of $\ttO$ has a chosen local terminal object $U_c$, $c\in \pi_0(\ttO)$. We also assume that for every $f:T\to S$ in $\ttO$ and every element $i\in |S|$ there is given an object $f^{-1}(i) \in \ttO,$ which we will call [*the $i$th fiber*]{} of $f$, such that $|f^{-1}(i)| = |f|^{-1}(i).$ We also require that
- For any $c\in \pi_0(\ttO)$, $|U_c| = 1$.
A [*trivial*]{} morphism $f:T\to S$ in $\ttO$ is a morphism such that, for each $i\in
|S|$, $f^{-1}(i) = U_{d_i}$ for some $d_i\in \pi_0(\ttO).$
The remaining axioms for a strict operadic category are:
- The identity morphism $id:T\to T$ is trivial for any $T\in \ttO;$
- For any commutative diagram in $\ttO$ $$\label{Kure}
\xymatrix@C = +1em@R = +1em{
T \ar[rr]^f \ar[dr]_h & & S \ar[dl]^g
\\
&R&
}$$ and every $i\in |R|$ one is given a map $$f_i: h^{-1}(i)\to g^{-1}(i)$$ such that $|f_i|: |h^{-1}(i)|\to |g^{-1}(i)|$ is the map $|h|^{-1}(i)\to |g|^{-1}(i)$ of sets induced by $$\xymatrix@C = +1em@R = +1em{ |T| \ar[rr]^{|f|} \ar[dr]_{|h|} & & |S|
\ar[dl]^{|g|}
\\
&|R|& }.$$ We moreover require that this assignment forms a functor ${\Fib}_i: \ttO/R \to \ttO$. If $R=U_c$, the functor ${\Fib}_1$ is required to be the domain functor $\ttO/R \to \ttO.$
- In the situation of (iii), for any $j\in |S|$, one has the equality $$\label{karneval_u_retardacku}
f^{-1}(j) = f_{|g|(j)}^{-1}(j).$$
- Let $$\xymatrix@C = +2.5em@R = +1em{ & S \ar[dd]^(.3){g} \ar[dr]^a &
\\
T \ar[ur]^f \ar@{-}[r]^(.7){b}\ar[dr]_h & \ar[r] & Q \ar[dl]^c
\\
&R&
}$$ be a commutative diagram in $\ttO$ and let $j\in |Q|, i = |c|(j).$ Then by axiom (iii) the diagram $$\xymatrix@C = +1em@R = +1em{
h^{-1}(i) \ar[rr]^{f_i} \ar[dr]_{b_i} & & g^{-1}(i) \ar[dl]^{a_i}
\\
&c^{-1}(i)&
}$$ commutes, so it induces a morphism $(f_i)_j: b_i^{-1}(j)\to a_i^{-1}(j).$ By axiom (iv) we have $$a^{-1}(j)=a_i^{-1}(j) \ \mbox{and} \ b^{-1}(j)=b_i^{-1}(j).$$ We then require the equality $$f_j = (f_i)_j$$.
We will also assume that the set $\pi_0(\ttO)$ of connected components is [*small*]{} with respect to a sufficiently big ambient universe.
It follows from axiom (iii) that the unique fiber of the canonical morphism $!_T: T\to
U_c$ is $T$.
A [*strict operadic functor*]{} between strict operadic categories is a functor $F: \ttO \to \ttP$ over $\Fin$ which preserves fibers in the sense that $F\big(f^{-1}(i)\big) = F(f)^{-1}(i)$, for any $f : T
\to S \in \ttO$ and $i \in |S| = |F(S)|$. We also require that $F$ preserves the chosen terminal objects, and that $F(f_i) = F(f)_i$ for $f$ as in (\[Kure\]). This gives the category ${\OpCat}$ of strict operadic categories and their strict operadic functors. For a family $E = \{E(T)\}_{T
\in \ttO}$ of objects of $\ttV$ and a morphism $f:T\to S$ let $$E(f) = \bigotimes_{i \in |S|} E({T_i})$$ In the following definition we tacitly use equalities (\[karneval\_u\_retardacku\]).
\[Jarca\_u\_mne\_prespala!\] An $\ttO$-operad is family $\calP = \{\calP(T)\}_{T
\in \ttO}$ of objects of $\ttV$ together with units $$I\to \calP(U_c),\ c \in \pi_0(\ttO),$$ and structure maps $$\gamma_f: \calP(f) \otimes \calP(S)\to \calP(T),\ f:T\to S,$$ satisfying the following axioms.
- Let $T \stackrel f\to S \stackrel g\to R$ be morphisms in $\ttO$ and $h := gf : T \to R$ as in (\[Kure\]). Then the following diagram of structure maps of $\calP$ combined with the canonical isomorphisms of products in $\ttV$ commutes: $$\xymatrix@C = 2em@R = .4em{
\ar@/^2.5ex/[rrd]^(.56){\hskip .5em\bigotimes_{i}\gamma_{f_i} \ot \id}
\ar[dd]_(.45){\id \ot \gamma_g}
\displaystyle\bigotimes_{i
\in |R|}
\calP(f_i) \ot \calP(g) \ot \calP(R) & &
\\ & & \ar@/^/[dl]_{\gamma_h}
\calP(h) \ot \calP(R)\ .
\\
\ar[r]^(.77){\gamma_f}{\rule{0pt}{2em}}
\displaystyle\bigotimes_{i \in |R|}
\calP(f_i) \ot \calP(S) \cong \calP(f) \ot \calP(S)&
\calP(T)&
}$$
- The composition $$\xymatrix@1@C = +2em{ \calP(T) \ar[r]& \rule{0pt}{2em} \displaystyle
\bigotimes_{i\in |T|} I\! \ot\! \calP(T) \ar[r]&\rule{0pt}{2em}
\displaystyle \bigotimes_{i\in |T|} \calP(U_{c_i})\! \ot\! \calP(T)\ar[r]^(.56)= &
\calP(\id_T)\! \ot\! \calP(T)
\ar[r]^(.65){\gamma_{\id}}&\calP(T) }$$ is the identity for each $T \in \ttO$, as well as the identity is
- the composition $$\xymatrix@1@C = +2.2em{ \calP(T)\! \ot \! I \ \ar[r]&\ \calP(T)\! \ot\!
\calP(U_c) \ \ar[r]^=
&\ \calP(!_T)\! \ot\! \calP(U_c)\
\ar[r]^{\hskip 1.8em \gamma_{!_T}}&\ \calP(T)},\ c:= \pi_0(T).$$
A [*morphism*]{} $\calP' \to \calP''$ of $\ttO$-operads in $\ttV$ is a collection of $\ttV$-morphisms , $T \in \ttO$, commuting with the structure operations. We denote by $\Oper\ttO(\ttV)$ or simply by $\Oper\ttO$ if $\ttV$ is understood the category of $\ttO$-operads in $\ttV$. Each operadic functor $F:\ttO\to \ttP$ induces the restriction $F^*: \Oper\ttP\to \Oper\ttO$.
A primary example of an operadic category is the category $\Fin$, while the cardinality functor is an example of a strict operadic functor. Thus $\Fin$ is the terminal object in the category of operadic categories and strict operadic functors. The category of $\Fin$-operads is isomorphic to the category of classical one-coloured (symmetric) operads.
\[Dnes\_na\_muslicky.\] The category of vines $\Vines$ [@lavers; @weber] is another example of an operadic category. It has the same objects as $\Fin$ but a morphism $\bar{n}\to \bar{m}$ is an isotopy class of merging descending strings in $\mathbb{R}^3$ (called [*vines*]{}) like in the following picture:
(5,-1.504111)(-0,1.7) (0.1,1.39) (1.88,1.39) (3.52,1.39) (1.28,-1.35) (3.74,-1.35) (4.9,1.39) (0.108106375,1.3233984)(0.108106375,0.5233984)(0.8481064,1.1633984)(1.3281064,0.3833984)(1.8081064,-0.3966016)(1.9081063,-0.43660158)(3.0281065,-0.63660157) (1.8881063,1.3433985)(1.9281064,0.5633984)(1.8481064,0.8433984)(1.8281064,0.4633984)(1.8081064,0.08339841)(2.0081065,-0.3566016)(3.0481064,-0.63660157) (4.8881063,1.3433985)(4.9281063,0.5633984)(4.3281064,-0.116601594)(3.9281063,-0.4566016)(3.5281065,-0.7966016)(2.1081064,-0.6766016)(1.2681063,-1.3766016) (3.5281065,1.3433985)(3.5681064,0.5633984)(3.1481063,0.5833984)(3.2281063,0.24339841)(3.3081064,-0.09660159)(3.7281063,-0.3566016)(3.7481065,-0.3966016) (3.9681063,-0.57660156)(4.1881065,-0.69660157)(4.1681066,-0.8966016)(3.9281063,-1.0966016)(3.6881063,-1.2966015)(3.7281063,-1.3346504)(3.7411063,-1.4166015) (3.0481064,-0.63660157)(3.4481063,-0.7366016)(3.6281064,-0.8766016)(3.6681063,-0.9966016)(3.7081063,-1.1166016)(3.6881063,-1.0166016)(3.7281063,-1.3166016)
There is a canonical identity-on-object functor $|-|:\Vines\to\Fin$ which sends a vine to the function assigning to the top endpoint of a string its bottom endpoint. A fiber of a vine $v:\bar{n}\to\bar{m}$ is equal to the fiber of $|v|:\bar{n}\to\bar{m}.$ The rest of the operadic category structure on $\Vines$ is quite obvious. The category of $\Vines$-operads is isomorphic to the category of braided operads [@Fiedorowicz]. This fact can be easily proved using the equivalent definition of braided operad given in [@batanin:br].
In fact, using Weber’s theory [@weber] one can associate an operadic category $\ttO(G)$ to each group operad (see [@yoshida] for its definition) $G$ is such a way that $\ttO(G)$-operads are exactly $G$-operads. The operadic categories $\Fin$ and $\Vines$ are special cases $\ttO(\Sigma)$ and $\ttO({\it Braid})$ of this construction for the symmetric group and braid group operads, respectively. We will provide the details elsewhere.
\[puget\] Let $\frC$ be a set. Recall from [@duodel Example 1.7] (see also [@lack Example 10.2]) that a [*$\frC$-bouquet*]{} is a map $b:
X\!+\!1\to \frC,$ where $X\in \Fin.$ In other words, a $\frC$-bouquet is an ordered $(k+1)$-tuple $(i_1,\ldots,i_k;i),\ X
= \bar{k}$, of elements of $\frC$. It can also be thought of as a planar corolla whose all edges (including the root) are colored. The extra color $b(1) \in \frC$ is called the [*root color*]{}. The finite set $X$ is the [*underlying set*]{} of the bouquet $b$.
A map of $\frC$-bouquets $b \to c$ whose root colors coincide is an arbitrary map $f: X\to Y$ of their underlying sets. Otherwise there is no map between $\frC$-bouquets. We denote the resulting category of $\frC$-bouquets by $\Bq(\frC)$.
The cardinality functor $|\dash|:\Bq(\frC)\to \Fin$ assigns to a bouquet $b :X+1\to \frC$ its underlying set $X$. The fiber of a map $b \to c$ given by $f : X \to Y$ over an element $y\in Y$ is a $\frC$-bouquet whose underlying set is $f^{-1}(y),$ the root color coincides with the color of $y$ and the colors of the elements are inherited from the colors of the elements of $X$.
\[end\] Operads over the category $\Bq(\frC)$ of $\frC$-bouquets introduced in Example \[puget\] are ordinary $\frC$-colored operads. Therefore, for each $\frC$-colored collection $E = \{E_c\}_{c\in \frC}$ of objects of $\ttV$ one has the [*endomorphism $\Bq(\frC)$-operad*]{} $\End^{\Bq(\frC)}_E$, namely the ordinary colored endomorphism operad [@berger-moerdijk:CM07 §1.2].
\[finset+deltaalg\] The category $\Deltaalg$ of finite ordinals (including the empty one) has an obvious structure of an operadic category. Operads over $\Deltaalg$ are ordinary non-$\Sigma$ operads [@batanin:AM08 Sec. 3, Prop. 3.1].
For any operadic category $\ttO$ the product category in the category of operadic categories exists, and -operads are coloured $\ttO$-operads [@duodel page 1637]. Likewise, the product with the operadic category of vines of Example \[Dnes\_na\_muslicky.\] describes braided versions of $\ttO$-operads. The product equals the subcategory $\DO\subset \ttO$ of morphisms for which $|f|$ is order-preserving.
Another important example is the operadic category $\Ordn$ of $n$-ordinals, $n \in \bbN$, see [@batanin:conf Sec. II]. $\Ordn$-operads are Batanin’s pruned $n$-operads which are allowed to take values not only in ordinary symmetric monoidal categories, but in more general [*globular*]{} monoidal $n$-categories and as such are therefore not covered by Feynman categories. Although $\Ordn$ does not fulfill the additional properties required for some constructions in this work, it was a crucial motivating example for our definition of operadic categories.
For each operadic category $\ttO$ with $\pi_0(\ttO)=\frC$, there is a canonical operadic ‘arity’ functor $$\label{Jarca}
\Ar: \ttO\to \Bq(\frC)$$ giving rise to the factorization $$\label{eq:factor}
\xymatrix@R=1em{&\ttO \ar@/_/[ld]_{\Ar} \ar@/^/[rd]^{|\dash|} &
\\
\Bq(\frC) \ar[rr]^{|\dash|} &&\Fin
}$$ of the cardinality functor $|\dash| : \ttO \to \Fin$. It is constructed as follows. Let the [*source*]{} $s(T)$ of $T \in \ttO$ be the set of fibers of the identity $\id : T \to T$. The bouquet $\Ar(T) \in \Bq(\frC)$ is defined as $b: s(T)+1 \to \frC$, where $b$ associates to each fiber $U_c
\in s(T)$ the corresponding connected component $c \in \frC$, and $b(1)
:= \pi_0(T)$. The assignment $T \mapsto \Ar(T)$ extends into an operadic functor.
\[Zitra\_do\_Osnabrucku\] For a $\frC$-colored collection $E = \{E_c\}_{c\in \frC}$ in $\ttV$ and an operadic category $\ttO$ with $\pi_0(\ttO) = \frC$, one defines the [ *endomorphism $\ttO$-operad*]{} $\End^\ttO_E$ as the restriction $$\End^\ttO_E := \Ar^*\big(\End^{\Bq(\frC)}_E\big)$$ of the $\Bq(\frC)$-endomorphism operad of Example \[end\] along the arity functor $\Ar$ of (\[Jarca\]).
\[Jaruska\_ma\_chripecku\] An [*algebra*]{} over an $\ttO$-operad $\calP$ in $\ttV$ is a collection $A = \{A_c\}_{c\in \pi_0(\ttO)}$, $A_c \in \ttV$, equipped with an $\ttO$-operad map $\alpha : \calP \to \End^\ttO_A$.
An algebra is thus given by suitable structure maps $$\alpha_T : \bigotimes_{c \in \pi_0(s(T))} A_c \ot \calP(T) \to
A_{\pi_0(T)}, \ T \in \ttO,$$ where $s(T)$ denotes, as before, the set of fibers of the identity $\id: T\to T$. We denote by $\Alg\calP(\ttV)$ or simply by $\Alg\calP$ when $\ttV$ is clear from the contect the category of $\calP$-algebras and their morphisms.
Sundry facts about operadic categories {#Pojedu_vecer_nebo_ted?}
======================================
The aim of this section is to some finer properties of operadic categories and formulate some additional axioms required for our constructions.
General facts
-------------
We are going to prove some consequences of the axioms for operadic categories needed later in this work.
\[l1\] Consider the commutative diagram $$\xymatrix@C=4em{S \ar[rd]^{f''} \ar[d]_{f'}
\\
T' \ar[r]^(0.4)\sigma & T''.
}$$ Let $j\in |T''|$ and $|\sigma|^{-1}(j) = \{i\}$ for some $i\in |T'|$. Then the unique fiber of the induced map equals $\inv{f'}(i)$. If $\inv{\sigma}(j)$ is trivial, in particular, if $\sigma$ is a , then $$\label{druhy_tyden_v_Sydney}
\inv{f'}(i) = \inv{f''}(j).$$
By Axiom (iv) of an operadic category,[^4] $\inv{f'}(i) = \inv{f'_j}(i)$ which readily gives the first part of the lemma. If $\inv{\sigma}(j)$ is trivial, then the fiber of $f'_j$ equals $
\inv{f''}(j)$ by Axiom (iii). This proves the second part of the lemma.
\[l2\] Consider the commutative diagram $$\xymatrix{S' \ar[rr]^\pi_\sim \ar[dr]_{f'} && S'' \ar[ld]^{f''}
\\
&T&
}$$ where $\pi$ is a . Then all $\pi_i: \inv{f'}(i) \to
\inv{f''}(i)$, $i \in |T|$, are , too.
Immediate from Axiom (iv).
\[l3\] Consider the commutative diagram $$\label{s1}
\xymatrix@C=4em{S' \ar[d]_{f'}\ar[dr]^{f} \ar[r]^\pi & S'' \ar[d]^{f''}
\\
T' \ar[r]^\sigma &\ T''.
}$$ Let $j\in |T''|$ and $|\sigma|^{-1}(j) = \{i\}$ for some $i\in |T'|$. Diagram (\[s1\]) determines:
- the map $f'_j: \inv{f}(j) \to
\inv{\sigma}(j)$ whose unique fiber equals $\inv{f'}(i)$, and
- [the induced map $\pi_j :
\inv{f}(j) \to
\inv{f''}(j)$.]{}
If $\inv{\sigma}(j)$ is trivial, in particular, if $\sigma$ is a , then $\pi$ induces a map $$\label{zitra_na_prohlidku_k_doktoru_Reichovi}
\pi_{(i,j)} : \inv{f'}(i) \to \inv{f''}(j)$$ which is a if $\pi$ is.
The first part immediately follows from Lemma \[l1\] and Axiom (iii). Under the assumption of the second part, one has equality (\[druhy\_tyden\_v\_Sydney\]) and $\pi_{(i,j)}$ defined as the composite $$\pi_{(i,j)} : \inv{f'}(i) = \inv{f}(j) \stackrel{\pi_j}\longrightarrow
\inv{f''}(j).$$ The rest follows from Lemma \[l2\].
Thus, in the situation of Lemma \[l3\] with $\sigma$ a , one has the [ *derived sequence*]{} $$\label{e1}
\left\{\pi_{(i,j)} :
\inv{f'}(i) \to \inv{f''}(j), j = |\sigma|(i)\right\}_{i \in |T'|}$$ consisting of if $\pi$ is a . Central constructions of this work will require the following:
\[bu\] Consider the corner $$\label{c1}
\xymatrix@C=3.5em{S' \ar[d]_{f'} &
\\
T' \ar[r]^\sigma_\sim &T''
}$$ in which $\sigma$ is a and $f' \in \DO$. Assume we are given objects $F''_j$, $j \in |T''|$ together with a collection of maps $$\label{e2}
\big\{\pi_{(i,j)} : \inv{f'}(i) \to F''_j,\ j = |\sigma|(i)\big\}_{i \in |T'|}.$$ Then the corner (\[c1\]) can be completed uniquely into the commutative square $$\label{eq:5}
\xymatrix@C=3.5em{S' \ar[d]_{f'} \ar[r]^\pi & S'' \ar[d]^{f''}
\\
T' \ar[r]^\sigma_\sim &T''
}$$ in which $f'' \in \DO$, $\inv{f''}(j) = F''_j$ for $j \in |T''|$, and such that derived sequence (\[e1\]) induced by $f''$ coincides with (\[e2\]).
The requirement that $f',f'' \in \DO$ is crucial, otherwise the factorization would not be unique even in ‘simple’ operadic categories as $\Fin$. It will sometimes suffice to assume the blow up for $\sigma = \id$ only, i.e. to assume
\[wbu\] For any $f' :S' \to T$ in $\DO$ and morphisms $\pi_i : \inv{f'}(i) \to F''_i$ in $\ttO$, $i \in |T|$, there exists a unique factorization of $f'$ $$\xymatrix@C=2em@R=1.2em{S' \ar[rr]^\omega \ar[dr]_{f'} && S'' \ar[ld]^{f''}
\\
&T&
}$$ such that $f'' \in \DO$ and $\omega_i = \pi_i$ for all $i \in |T|$.
Notice that $\omega \in \DO$ (resp. $\omega \in \QO$) if and only if $\pi_i \in \DO$ (resp. $\pi_i \in \QO$) for all $i \in |T|$. A concise formulation of the weak blow up is that the fiber functor $$\DO/T \to \ttO^{\times |T|}$$ is a discrete opfibration.
\[zase\_mam\_nejaky\_tik\] If the weak blow up axiom is satisfied in $\ttO$, then $$\QO \cap \DO = \ttO_{\rm disc},$$ the discrete category with the same objects as $\ttO$. In particular, the only in $\DO$ are the identities.
It is clear that each identity belongs to $\QO \cap \DO$. On the other hand, assume that $\phi : S \to T \in \QO \cap \DO$. Since it is a , all its fibers are trivial, $\inv{\phi}(i) = U_i$ for $i\in
|T|$. Consider now two factorizations of $\phi$, $$\label{posledni_den_v_Srni}
\xymatrix@R=1em{&\ar@{=}[ld]_{\id_S} \ar[dd]^(.35)\phi \ar[dr]^\phi S&
\\
S\ar[dr]^\phi &&T\ar@{=}[ld]_{\id_T}
\\
&T&
}$$ In the left triangle we have, for $i\in |T|$, $(\id_T)_i :U_i= \inv{\phi}(i) \to
\inv{\phi}(i) = U_i$, therefore $(\id_T)_i = \id_{U_i}$ by the terminality of $U_i$. Let us turn our attention to the right triangle.
By Axiom (ii) of an operadic category, all fibers of an identity are trivial, thus $$(\phi)_i :U_i= \inv{\phi}(i) \to
\inv{{\id_T}}(i) = U_c$$ for some chosen local terminal $U_c$. Since $U_c$ and $U_i$ are in the same component of $\ttO$, $U_c = U_i$ and $(\phi)_i$ must be the identity. We see that both factorizations in (\[posledni\_den\_v\_Srni\]) are determined by the collection $\id_{U_i} : \inv{\phi}(i) \to U_i$, $i \in |T|$, so, by the uniqueness in the blow up axiom, they are the same.
Corollary \[zase\_mam\_nejaky\_tik\] shows the power of the blow up axiom and illustrates how it determines the nature of an operadic category. While it is satisfied in operadic categories underlying ‘classical’ examples of operads, it is violated e.g. in Batanin’s category of $n$-trees [@batanin:AM08] which possesses non-invertible in $ \QO \cap \DO$. The blow up axiom has another simple implication which we formulate as
\[Zitra\_ma\_byt\_-18\_stupnu.\] Suppose we have two diagrams $$\xymatrix@C=3.5em{S' \ar[d]_{f'} \ar[r]^{\pi_1} & S_1'' \ar[d]^{f''_1}
\\
T' \ar[r]^\sigma_\sim &T''
}
\ \and \
\xymatrix@C=3.5em{S' \ar[d]_{f'} \ar[r]^{\pi_2} & S_2'' \ar[d]^{f''_2}
\\
T' \ar[r]^\sigma_\sim &T''
}$$ as in (\[s1\]) such that the fibers of $f_1''$ and $f_2''$ are the same and the derived sequences of $\pi_1$ and $\pi_2$ coincide. If the blow up axiom is satisfied, the diagrams are the same, i.e. $S''_1=S_2''$, $f_1'' = f_2''$ and $\pi_1=\pi_2$.
We will need in Section \[pairing\] the concept of quadraticity of operads in an operadic category $\ttO$, requiring an additional structure introduced in:
\[grad\] A [*grading*]{} on an operadic category $\ttO$ is a map $e :
{\rm Objects}(\ttO) \to \bbN$ of sets with the property that $$\label{Parramatta}
e(T) + e(F_1) + \cdots + e(F_k) = e(S)$$ for each $f : S \to T$ with fibers $\Rada F1k$. In this situation we define the [*grade*]{} $e(f)$ of $f$ by $e(f) := e(S)- e(T)$.
The identity endomorphism $\id: U \to U$ of a trivial object has its unique fiber $U$, hence $2e(U) = e(U)$ by (\[Parramatta\]), thus $e(U) = 0$. It is easy to see that a grading on $\ttO$ is the same as an $\ttO$-operad [@duodel Definition 1.1] in the discrete symmetric monoidal category $\bbN$ with the monoidal unit $0$ and the product given by the addition. A typical example of grading is the number of internal edges of ordered graphs in the operadic category $\Gr$ introduced in Definition \[Radeji\_bych\_sel\_s\_Jaruskou\_na\_vyhlidku\_sam.\], whence the notation. Notice finally that if $F : \ttO \to \ttP$ is a strict operadic functor and $\ttP$ is graded, then $\ttO$ is graded as well by the formula $$\label{Dnes_plynari_ale_ja_jsem_v_Australii.}
e(T) := e(F(T)),\ T \in \ttO.$$
\[constant-free\] An operadic category $\ttO$ is [*constant free*]{} if the cardinality functor $\ttO\to \Fin$ factorizes through the operadic category $\Surj$ of nonempty finite sets and surjections. Equivalently, $\ttO$ is constant free if $|f|$ is surjective for each $f \in \ttO$.
Notice that each constant-free operadic\[canon\] category $\ttO$ with no elements of cardinality $0$ bears the [*canonical grading*]{} given by $e(T) := |T| -1$.
\[plysacci\_postacci\] A morphism $\phi
: T \to S \in \DO$ in a graded operadic category $\ttO$ is [*elementary*]{} if all its fibers are trivial (= chosen local terminal) except precisely one whose grade is $\geq 1$. If $\inv{\phi}(i)$ is, for $i \in |S|$, the unique nontrivial fiber, we will sometimes write $\phi$ as the couple $(\phi,i)$. If we want to name the unique nontrivial fiber $F := \inv\phi( i)$ explicitly, we will write $F \fib_i T
\stackrel\phi\to S$ or $F \fib T
\stackrel\phi\to S$ when the concrete $i \in |S|$ is not important.
[**Notation.**]{} Assume that, in the set-up of Lemma \[l3\] with $\sigma$ a , the morphisms $f',f''$ are elementary, $\inv{f'}(a)$ is the only nontrivial fiber of $f'$, and $\inv{f''}(b)$ with $b:= |\sigma|(a)$ the only nontrivial fiber of $f''$. In this situation we denote by $$\label{n1}
\overline \pi : = \pi_{(a,b)} :\inv{f'}(a) \to \inv{f''}(b)$$ the only nontrivial part of the derived sequence (\[e1\]).
\[V\_Srni\_pred\_Borovou\_Ladou\] If $\pi$ is a , the only nontrivial fiber of $f''$ [*must be*]{} $\inv{f''}(b)$ with $b:= |\sigma|(a)$. Indeed, the maps in (\[zitra\_na\_prohlidku\_k\_doktoru\_Reichovi\]) are , so their fibers are, by definition, the chosen local terminal objects. When $\inv{f''}(j)$ is the chosen local terminal object, then the (unique) fiber of $\pi_{(i,j)}$ is $\inv{f'}(i)$, so it must be, by Axiom (iii) of an operadic category, a chosen local terminal object, too.
\[sec:sundry-facts-about-1\] Assume the blow up axiom and suppose that in (\[c1\]) the map $f'$ is elementary, with the unique fiber over $a \in |T'|$. Let $b := |\sigma|(a)$ and assume we are given a map $\overline \pi : \inv{f'}(a) \to
F$. Then (\[c1\]) can be uniquely completed into (\[eq:5\]) in which $f''$ is elementary with the unique nontrivial fiber $\inv{f''}(b) = F$ such that $\overline \pi$ is the map (\[n1\]).
By the blow up axiom, (\[eq:5\]) is uniquely determined by the maps between the fibers. The only map between nontrivial fibers is $\overline \pi$ while all maps between trivial ones are unique by the terminality of trivial objects, thus there is no room for choices of the induced maps between fibers.
Chains of morphisms
-------------------
This subsection contains an auxiliary material for the construction of free Markl’s operads in Section \[zitra\_letim\_do\_Pragy\].
\[d3\] Let $T\stackrel{(\phi,j)}{\longrightarrow} S
\stackrel{(\psi,i)}{\longrightarrow} P$ be elementary morphisms. If $|\psi|(j) = i$ we say that the fibers of $\phi$ and $\psi$ are [*joint*]{}. If $|\psi|(j)\ne i$ we say that $\phi$ and $\psi$ have [*disjoint fibers*]{} or, more specifically, that the fibers of $\phi$ and $\psi$ are [ *$(i,j)$-disjoint*]{}. The situation of disjoint fibers is captured in Figure \[Jarka\_ma\_novy\_pocitac.\].
(0,-2.8625)(6.24,2.8625) (0.31,-2.3375)(5.71,-2.3375) (0.31,0.0625)(5.71,0.0625) (5.71,2.4625)(5.71,2.4625)(0.31,2.4625) (0.71,2.4625)(1.71,0.0625)(2.71,2.4625)(2.71,2.4625) (1.71,0.0625)(1.71,-2.3375) (3.31,0.0625)(3.31,0.0625)(4.31,-2.3375)(5.31,0.0625)(5.31,2.4625)(5.31,2.4625) (3.31,0.0625)(3.31,2.4625) (1.71,2.4625)(1.0,0.4) (1.71,2.5)[$\xi^{-1}(k)$]{} (4.31,2.4625)(1.0,0.4) (4.31,0.0625)(1.0,0.4) (4.3,2.5)[$\xi^{-1}(i)$]{} (4.31,0.0625)[$\psi^{-1}(i)$]{} (1.71,0.0625) (1.71,-2.3375) (4.31,-2.3375) (0.51,2.4625)(0.51,0.0625) (0.51,0.0625)(0.51,-2.3375) (4.11,2.0625)(4.11,0.4625) (6.11,2.4625)[$T$]{} (6.11,0.0625)[$S$]{} (6.11,-2.3375)[$P$]{} (0.11,1.4625)[$\phi$]{} (0.11,-0.9375)[$\psi$]{} (4.51,1.2625)[$\phi_i$]{} (2.11,0.3)[$j$]{} (1.71,-2.7375)[$k$]{} (4.31,-2.7375)[$i$]{} (3.71,1.2625)[$\sim$]{}
\[l7\] If the fibers of $\phi$ and $\psi$ of elementary morphisms in Definition \[d3\] are joint, then the composite $\xi = \psi(\phi)$ is an elementary as well, with the nontrivial fiber over $i$, and the induced morphism $\phi_i:\xi^{-1}(i)\to \psi^{-1}(i) $ is elementary with the nontrivial fiber over $j$ that equals $\phi^{-1}(j)$. For $l\ne i$ the morphism $\phi_l$ equals the identity $U_c\to U_c$ of trivial objects.
If the fibers of $\phi$ and $\psi$ are [ $(i,j)$-disjoint]{} then the morphism $\xi = \psi(\phi)$ has exactly two nontrivial fibers and these are fibers over $i$ and $k: =|\psi|(j)$. Moreover, there is a canonical induced quasibijection:
$$\label{harmonika}
\phi_i:\xi^{-1}(i)\to \psi^{-1}(i) \in \DO$$
and the equality $$\label{pisu_opet_v_Sydney}
\xi^{-1}(k) = \phi^{-1}(j).$$
By Axiom (iv) of an operadic category, $\phi_i^{-1}(j) = \phi^{-1}(j)$, thus $e(\phi_i^{-1}(k))\geq 1$. If $k \in |\psi|^{-1}(i)$ is such that $k\ne j$, $\phi_i^{-1}(k) = \phi^{-1}(k) = U_c$. Therefore $\phi_i$ is an elementary morphism.
Let us prove that $\xi$ is elementary as well. For $i =k\in |P|$, we have $\phi_i: \xi^{-1}(i)\to \psi^{-1}(i)$, hence the grade of $\xi^{-1}(i)$ must be greater or equal to the grade of $\phi_i^{-1}(j) = \phi^{-1}(j)$, which is greater or equal $1$. For $k\ne i$, $\phi_k:\xi^{-1}(k)\to \psi^{-1}(k) = U'$ has the unique fiber equal to $\xi^{-1}(k).$ On the other hand for the unique $l$ such that $|\psi|(l) = k$, $$\phi_k^{-1}(l) = \phi^{-1}(l) = U'',$$ hence $\xi^{-1}(k) = U''$, so $\xi$ is elementary.
Let us prove the second part of the lemma. If $l\ne i,k$ then $\phi_l:\xi^{-1}(l)\to \psi^{-1}(l) =
U'$, where $U'$ is a trivial object. So $\phi_l$ has as its unique fiber equal to $\xi^{-1}(l)$. Since $|\psi|$ is surjective, there exists $l'\in |S|$ such that $|\psi|(l') = l$ and such an $l'$ is unique because $\psi$ is elementary. Hence $\phi_l^{-1}(l') = \phi^{-1}(l') = U'$ and so $\xi^{-1}(l) = U'$. This proves that the only nontrivial fibers of $\xi$ might be those over $i$ and $k$. Their grades are clearly $\geq 1$.
Let us prove that $\phi_i$ is a quasibijection. If $l\in
|\psi^{-1}(i)|$ then $\phi_i^{-1}(l) = \phi^{-1}(l).$ But $l\ne j$ since $|\psi|(l) = i \ne |\psi|(j)$ hence, $\phi^{-1}(l)$ is trivial. To prove that $\phi_i \in \DO$, notice that by Axiom (iii), $|\phi_i|$ is the map of sets $|\xi|^{-1} \to |\psi|^{-1}$ induced by the diagram $$\xymatrix@C=2em@R=1.2em{|T| \ar[rr]^{|\phi|}
\ar[dr]_{|\xi|} && |S| \ar[ld]^{|\psi|}
\\
&|P|&
}$$
Regarding (\[pisu\_opet\_v\_Sydney\]), we have by Axiom (iv) $\phi^{-1}(j) = \phi^{-1}_k(j)$. But $\phi_k:\xi^{-1}(k)\to \psi^{-1}(k) = U''$ and hence its unique fiber equals to $\xi^{-1}(k)$. So, $\phi^{-1}(j) = \xi^{-1}(k)$.
\[har\] We will call the pair $T\stackrel{(\phi,j)}{\longrightarrow} S
\stackrel{(\psi,i)}{\longrightarrow} P$ of morphisms in Definition \[d3\] with disjoint fibers [*harmonic*]{} if $\inv{\xi}(i) = \inv \psi
(i)$ and the map $\phi_i$ in (\[harmonika\]) is the identity.
\[zitra\_vylet\_do\_Sydney\] If the blow up axiom is satisfied then all pairs with disjoint fibers are harmonic.
The map $\phi_i$ in (\[harmonika\]) is a in $\DO$, so it is the identity by Corollary \[zase\_mam\_nejaky\_tik\].
\[move\] Assume that $$\label{eq:3}
\xymatrix@R = 1em@C=4em{& {P'} \ar[dr]^{(\psi',\, i)} &
\\
T\ar[dr]^{(\phi'',\, l)} \ar[ur]^{(\phi',\, j)} && S
\\
&{P''}\ar[ur]^{(\psi'',\, k)}&
}$$ is a commutative diagram of elementary morphisms. Assume that $\psi''|l| = i$ and $\psi'|j| = k$. Let $F', F'',G',G''$ be the only nontrivial fibers of $\phi',\phi'',\psi',\psi''$, respectively. Then one has canonical quasibijections $$\label{za_tyden_poletim_do_Prahy}
\sigma': F' \longrightarrow G'' \
\mbox { and }\ \sigma'': F'' \longrightarrow G'.$$ If both pairs in (\[eq:3\]) are harmonic, then $F' = G''$, $F'' =
G'$ and $\sigma',\sigma''$ are the identities.
Let $\xi : T \to S$ be the composition $\psi'\phi' = \psi''\phi''$. One has $G' =
\inv{\psi'}(j)$, $G'' = \inv{\psi''}(k)$ and, by Lemma \[l7\], $F' = \inv{\phi'}(j) = \inv{\xi}(k)$ and $F'' = \inv{\phi''}(l) = \inv{\xi}(i)$. We define $$\sigma' : F' = \inv{\xi}(k) \stackrel{\phi''_k}\longrightarrow
\inv{\psi''}(k) = G''
\ \and \
\sigma'' : F'' = \inv{\xi}(i) \stackrel{\phi'_i}\longrightarrow
\inv{\psi''}(i) = G'.$$ These maps are by Lemma \[l7\]. The second part of the corollary follows directly from the definition of the harmonicity.
\[bude\_pekna\_zima\] Consider a diagram $$\label{grand_mall}
\xymatrix@R=1.6em@C=3em{T'\ar[r]^{\sigma_T}_\sim\ar[d]_{(\phi',j')} &
T''\ar[d]^{(\phi'',j'')}
\\
P'\ar[r]^{\sigma_P}_\sim \ar[d]_{(\psi',i')} & P''\ar[d]^{(\psi'',i'')}
\\
S'\ar[r]^{\sigma_S}_\sim & S''
}$$ whose vertical maps are elementary with disjoint fibers as indicated, and the horizontal maps are . Denoting $k' := |\psi'|(j')$, $k'' := |\psi''|(j'')$, one has $$\label{zase_podleham}
|\sigma_S|(i') = i'' \ \and \ |\sigma_S|(k') = k'' .$$
Assume the blow up axiom. If we are given a subdiagram of (\[grand\_mall\]) consisting only of the morphisms $\phi',\phi'',\psi',\psi'',\sigma_T$ and $\sigma_S$, i.e. $$\label{Pojedu_za_Jarkou_na_chalupu?}
\xymatrix@R=1.6em@C=3em{T'\ar[r]^{\sigma_T}_\sim\ar[d]_{(\phi',j')} &
T''\ar[d]^{(\phi'',j'')}
\\
P' \ar[d]_{(\psi',i')} & P''\ar[d]^{(\psi'',i'')}
\\
S'\ar[r]^{\sigma_S}_\sim & S'',
}$$ then the conditions (\[zase\_podleham\]) are also sufficient for the existence of a unique $\sigma_P$ as in (\[grand\_mall\]).
The only nontrivial fiber of $\psi'$ is $\inv{\psi'}(i')$ and the only nontrivial fiber of $\psi''$ is $\inv{\psi''}(i'')$ so, by Remark \[V\_Srni\_pred\_Borovou\_Ladou\], $|\sigma_S|(i') = i''$. By the same argument, $|\sigma_P|(j') = j''$. Since $|-|$ is a functor, we have $$k'' = |\psi''||\sigma_P|(j') = |\sigma_S| |\psi'|(j') = |\sigma_S| (k')$$ proving the first part of the proposition.
To prove the second part, denote by $\xi'$ resp. by $\xi''$ the composition of the maps in the left resp. right column of (\[Pojedu\_za\_Jarkou\_na\_chalupu?\]). Since the left column of (\[Pojedu\_za\_Jarkou\_na\_chalupu?\]) is harmonic by Corollary \[zitra\_vylet\_do\_Sydney\], we may define a map $(\sigma_P)_{(i',i'')}$ by the commutativity of the diagram $$\label{streda_v_Srni}
\xymatrix@C=4em{\inv{\xi'}(i') \ar[r]^{(\sigma_T)_{(i',i'')}}
\ar@{=}[d]_{\phi'_{i'} = \id}
& \inv{\xi''}(i'')\ar[d]_{\phi''_{i''}}
\\
\inv{\psi'}(i') \ar[r]^{(\sigma_P)_{(i',i'')}} & \inv{\psi''}(i'').
}$$ The blow up axiom produces a commutative diagram $$\xymatrix@R=1.6em@C=3em{
P'\ar[r]^{\sigma_P}_\sim \ar[d]_{\psi'} & \tilde P''\ar[d]^{\tilde \psi''}
\\
S'\ar[r]^{\sigma_S}_\sim & S''
}$$ in which, by construction, $\tilde \psi''$ is elementary with the only nontrivial fiber ${\psi''}^{-1}(i'')$ over $i''$, and the map between nontrivial fibers induced by $\sigma_P$ is $(\sigma_P)_{(i',i'')}$. Consider now two commutative diagrams $$\label{Necham_auto_na_Rokyte.}
\xymatrix@R=1.6em@C=3em{
T'\ar[r]^{\sigma_P\circ \phi'}_\sim \ar[d]_{\xi'} & \tilde P''\ar[d]^{\tilde \psi''}
\\
S'\ar[r]^{\sigma_S}_\sim & S''
}
\ \and \
\xymatrix@R=1.6em@C=3em{
T'\ar[r]^{\phi''\circ \sigma_T}_\sim \ar[d]_{\xi'} & P''\ar[d]^{\tilde \psi''}
\\
S'\ar[r]^{\sigma_S}_\sim & S''.
}$$ In both diagrams, the right vertical map is elementary, with the only nontrivial fiber ${\psi''}^{-1}(i'')$. We will show that both $\sigma_P\circ \phi'$ and $\phi'' \circ \sigma_T$ induce the same maps between nontrivial fibers. One has $$(\sigma_P\circ \phi')_{(i',i'')} = (\sigma_P)_{(i',i'')} \circ\phi'_{i'}$$ while $$(\phi''\circ \sigma_T)_{(i',i'')} = \phi''_{i''}\circ (\sigma_T)_{(i',i'')}.$$ By the defining diagram (\[streda\_v\_Srni\]), the right hand sides of both equations coincide. By Corollary \[Zitra\_ma\_byt\_-18\_stupnu.\], the diagrams in (\[Necham\_auto\_na\_Rokyte.\]) are the same, therefore both squares in (\[grand\_mall\]) with $\sigma_P$ constructed above commute. This finishes the proof.
\[spousta-prace\] Assume the blow up axiom. Let $\rho: S \to T \in \DO$ be elementary with the unique fiber $F$ over $a \in |T|$. Suppose that we are given a chain of elementary morphisms $$\label{eq:4}
F \stackrel {\varphi_1} \longrightarrow F_1 \stackrel {\varphi_2} \longrightarrow
F_2 \stackrel {\varphi_3} \longrightarrow
F_3 \stackrel {\varphi_4} \longrightarrow \cdots \stackrel
{\varphi_{l-1}} \longrightarrow F_{l-1}.$$ Then there exists a unique factorization $$\label{eq:2}
S \stackrel {\rho_1} \longrightarrow S_1 \stackrel {\rho_2} \longrightarrow
S_2 \stackrel {\rho_3} \longrightarrow
S_3 \stackrel {\rho_4} \longrightarrow \cdots \stackrel
{\rho_{l-1}} \longrightarrow S_{l-1} \stackrel {\rho_l} \longrightarrow T$$ of $\rho$ into elementary morphisms such that $(\rho_l \circ \cdots \circ
\rho_s)^{-1}(a) = F_{s-1}$ for each $2 \leq s \leq l$, and $(\rho_s)_a = \varphi_s$ for each $1 \leq s < l$.
We will inductively construct maps in the commutative diagram $$\label{eq:10}
\xymatrix{S \ar[r]^{\rho_1}\ar[d]_{\rho} &
S_1\ar[r]^{\rho_2} \ar@/^.1pc/[dl]_{\eta_1} & S_2 \ar@/^.3pc/[dll]_{\eta_2}
\ar[r]^{\rho_3} & S_3 \ar[r]^{\rho_4}\ar@/^.6pc/[dlll]_{\eta_3} &
\cdots \ar[r]^{\rho_{l-1}} & S_{l-1}. \ar@/^.9pc/[dlllll]_{\eta_l}
\\
T
}$$ The blow up axiom implies that the maps $$\varphi_1 : F = \inv{\rho}(a) \to F_1, \ \id : \inv{\rho}(i) = U_i \to U_i
\mbox { for } i \ne a,$$ uniquely determine a decomposition $\rho = \eta_1 \circ
\rho_1$. Clearly, $\eta_1$ is elementary with the unique fiber $F_1$ and we may apply the same reasoning to $\eta_1$ in place of $\rho$. The result will be a unique decomposition $\eta_1 = \eta_2 \circ
\rho_2$. Repeating this process $(l-1)$-times and defining $\rho_l : =
\eta_l$ finishes the proof.
Assume that, in (\[eq:4\]), the pair $(\varphi_t,\varphi_{t+1})$ has, for some $1 \leq t \leq l-2$, $(i,j)$-disjoint fibers. Then the corresponding pair $(\rho_t,\rho_{t+1})$ in (\[eq:2\]) has $(i+a-1,j+a-1)$-disjoint fibers. This is an immediate consequence of Axiom (iv) of an operadic category.
\[spousta-prace-1\] Assumptions: blow up. Suppose that, in the notation of Lemma \[spousta-prace\], we are given two chains of elementary morphisms as in (\[eq:4\]) of the form
$$\label{eq:8}
F \stackrel {\varphi_1} \longrightarrow F_1 \stackrel{\varphi_2} \longrightarrow
\cdots \stackrel{\varphi_{u-1}} \longrightarrow {F_{u-1}}
\stackrel{\varphi'_{u}}\longrightarrow F'_{u} \stackrel
{\varphi'_{u+1}}\longrightarrow F_{u+1} \stackrel{\varphi_{u+2}}
\longrightarrow \cdots \stackrel {\varphi_{l-1}} \longrightarrow
F_{l-1}$$
and $$\label{eq:9}
F \stackrel {\varphi_1} \longrightarrow F_1 \stackrel{\varphi_2} \longrightarrow
\cdots
\stackrel{\varphi_{u-1}} \longrightarrow {F_{u-1}}
\stackrel{\varphi''_{u}}\longrightarrow F''_{u}
\stackrel {\varphi''_{u+1}}\longrightarrow F_{u+1}
\stackrel{\varphi_{u+2}} \longrightarrow \cdots
\stackrel {\varphi_{l-1}} \longrightarrow F_{l-1}$$
such that the diagram $$\xymatrix@R = 1em{& {F'_u} \ar[dr]^{\varphi'_{u+1}} &
\\
F_{u-1}\ar[dr]^{\varphi_u''} \ar[ur]^{\varphi_u'} && F_{u+1}
\\
&{F''_u}\ar[ur]^{\varphi''_{u+1}}&
}$$ commutes. Then the corresponding decompositions (\[eq:2\]) are of the form
$$\label{eq:6}
S \stackrel {\rho_1} \longrightarrow S_1 \stackrel{\rho_2} \longrightarrow
\cdots \stackrel{\rho_{u-1}} \longrightarrow {S_{u-1}}
\stackrel{\rho'_{u}}\longrightarrow S'_{u} \stackrel
{\rho'_{u+1}}\longrightarrow S_{u+1} \stackrel{\rho_{u+2}}
\longrightarrow \cdots \stackrel {\rho_{l-1}} \longrightarrow
S_{l-1} \stackrel {\rho_{l}} \longrightarrow T$$
respectively $$\label{eq:7}
S \stackrel {\rho_1} \longrightarrow S_1 \stackrel{\rho_2} \longrightarrow
\cdots
\stackrel{\rho_{u-1}} \longrightarrow {S_{u-1}}
\stackrel{\rho''_{u}}\longrightarrow S''_{u}
\stackrel {\rho''_{u+1}}\longrightarrow S_{u+1}
\stackrel{\rho_{u+2}} \longrightarrow \cdots
\stackrel {\rho_{l-1}} \longrightarrow S_{l-1}
\stackrel {\rho_{l}} \longrightarrow T,$$
and the diagram $$\label{eq:11}
\xymatrix@R = 1em{& {S'_u} \ar[dr]^{\rho'_{u+1}} &
\\
S_{u-1}\ar[dr]^{\rho_u''} \ar[ur]^{\rho_u'} && S_{u+1}
\\
&{S''_u}\ar[ur]^{\rho''_{u+1}}&
}$$ commutes.
We will rely on the notation used in the proof of Lemma \[spousta-prace\]. It is clear from the inductive construction described there that the initial parts of the canonical decompositions corresponding to (\[eq:8\]) resp. (\[eq:9\]) coincide and are equal to $$S \stackrel {\rho_1} \longrightarrow S_1 \stackrel{\rho_2} \longrightarrow
\cdots
\stackrel{\rho_{u-1}} \longrightarrow {S_{u-1}}.$$ Consider the following two stages of the inductive construction in the proof of Lemma \[spousta-prace\]: $$\xymatrix{S \ar[r]^{\rho_1}\ar[d]_{\rho} & \cdots
\ar[r]^{\rho_{u-1}} & S_{u-1} \ar[r]^{\rho'_u}\ar@/^.6pc/[dll]_{\eta_{u-1}} &
S'_u \ar@/^1.1pc/[dlll]_(.3){\eta'_u}
\ar[r]^{\rho'_{u+1}} & S'_{u+1} \ar@/^1.3pc/[dllll]_(.25){\eta'_{u+1}}
\\
T
}$$ and $$\xymatrix{S \ar[r]^{\rho_1}\ar[d]_{\rho} & \cdots
\ar[r]^{\rho_{u-1}} & S_{u-1} \ar[r]^{\rho''_u}\ar@/^.6pc/[dll]_{\eta_{u-1}} &
S''_u \ar@/^1.1pc/[dlll]_(.3){\eta''_u}
\ar[r]^{\rho''_{u+1}} & S''_{u+1}. \ar@/^1.3pc/[dllll]_(.25){\eta''_{u+1}}
\\
T
}$$ The maps $\eta_{u-1}$, $\eta'_{u+1}$ and $\eta''_{u+1}$ are elementary, with the nontrivial fibers $F_{u-1}$ resp. $F_{u+1}$. By construction, the horizontal maps in the factorizations $$\xymatrix@C=4em{S_{u-1} \ar[r]^{\rho'_{u+1} \circ \rho'_{u}} \ar[d]^{\eta_{u-1}} & S'_{u+1}
\ar[ld]^{\eta_{u+1}}
\\
T
}
\ \and \
\xymatrix@C=4em{S_{u-1} \ar[r]^{\rho''_{u+1} \circ \rho''_{u}} \ar[d]^{\eta_{u-1}} & S''_{u+1}
\ar[ld]^{\eta_{u+1}}
\\
T
}$$ induce the same maps between these nontrivial fibers, namely $\varphi'_{u+1} \circ \varphi'_{u} = \varphi''_{u+1} \circ
\varphi''_{u}$. By the uniqueness of the blow up, the diagrams in the above display coincide, so diagram (\[eq:11\]) with $S_{u+1} = S'_{u+1} = S''_{u+1}$ commutes. The fact that the remaining parts of (\[eq:6\]) resp. (\[eq:7\]) are the same is obvious.
Operadic category of graphs {#Ceka_mne_Psenicka.}
===========================
In this section we introduce an operadic category of graphs. This category and its modifications will play the fundamental rôle in this work. Recall that $\Fin$ denotes the category of finite ordinals and their set-theoretic (not necessarily order-preserving) maps.
Preordered graphs
-----------------
We start by an ordered version of the standard concept of graphs as recalled e.g. in [@markl-shnider-stasheff:book Definition II.5.23].
\[pre\] A [*preodered graph*]{} $\Gamma$ is a pair $(g,\sigma)$ consisting of an order-preserving map $g:F\to V$ in the category $\Fin$ of finite sets together with an involution $\sigma$ on $F$.
Elements of $F$ are the [*flags*]{}[^5] of $\Gamma$ and elements of $V$ its [*vertices*]{}. The fixed points of $\sigma$ are called the [*legs*]{} of $\Gamma$ while nontrivial orbits of $\sigma$ are its [*edges*]{}. The [*endpoints*]{} of an edge $e = \{h_1,h_2\}$ are $g(h_1)$ and $g(h_2)$.
For any $v\in V$, the set $g^{-1}(v)$ of flags adjacent to $v$ inherits a linear order from $F$ which we call the [*local order*]{} at $v$. We may thus equivalently define a preordered graph as a map $g :F\to V$ from a finite set $F$ into a linearly ordered set $V$ with the additional data consisting of linear orders of each $\inv g(v)$, $v \in V$. The lexicographic order combining the order of $V$ with the local orders makes $F$ a finite ordinal, and both definitions coincide.
A [*morphism*]{} of preodered graphs $\Phi:\Gamma'\to \Gamma''$ is a pair $(\psi,\phi)$ of morphisms of finite sets such that the diagram $$\label{graphsmorphism}
\xymatrix@C=3.5em{F' \ar[d]_{g'} & F''\ar@{_{(}->}[l]_{\psi}%)
\ar[d]^{g''}
\\
V' \ar@{->>}[r]^{\phi} &V''
}$$ commutes. We moreover require $\phi$ to be a surjection and $\psi$ an equivariant injection which is a bijection on fixed points of $\sigma$[^6] and which satisfies the following condition: If $\phi(i) \ne \phi(j)$ and $e'$ is an edge with endpoints $i$ and $j$ then there exists and edge $e''$ in $\Gamma''$ with endpoints $\phi(i)$ and $\phi(j)$ such that $e'=
\psi(e'')$.[^7] Preodered graphs and their morphisms form a category $\ttprGr$.
The [*fiber*]{} $\inv{\Phi}(i)$ of a map $\Phi = (\psi,\phi): \Gamma' \to \Gamma''$ in (\[graphsmorphism\]) over $i\in V''$ is a preodered graph whose set of vertices is $\phi^{-1}(i)$ and whose set of flags is $g^{-1}\big(\phi^{-1}(i)\big)$. The involution $\tau$ of $\inv{\Phi}(i)$ is defined as $$\tau(h) :=
\begin{cases} h& \hbox {if $h\in \Im(\psi)$ and}
\\
\sigma'(h)& \hbox {if $h\notin \Im(\psi)$,}
\end{cases}$$ where $\sigma'$ is the involution of $\Gamma'$. Observe that $h\notin \Im(\psi)$ if and only if $\sigma(h)\notin \Im(\psi)$.
\[pisu\_jednou\_rukou\] A [*local reordering*]{} morphism of graphs is a morphism (\[graphsmorphism\]) for which $\phi = \id$ and $\psi$ is an isomorphism. The map (\[graphsmorphism\]) is a [*local isomorphism*]{} of graphs if $\phi$ is a bijection and $\psi$ restricts to an order preserving isomorphism $\inv{g''}(j)\cong
\inv{g'}(i)$ for each $i \in V''$, $j = \phi(i)$. An [*order preserving*]{} morphism of graphs is a morphism (\[graphsmorphism\]) such that $\phi$ is order preserving.
\[contractions\] Let $\Gamma': F'\stackrel{g'}{\epi} V'$ be a preodered graph. Let $\phi: V'\epi V''$ be an order preserving surjection and let $F_i := {g'}^{-1}\big(\phi^{-1}(i)\big)$, $i\in V''$. Suppose that we are given subsets $E_i$, $i\in V''$, of edges of $\Gamma'$ formed by the flags in $F_i$. Denote by $\Gamma_i$ the graph given by the restriction $F_i\stackrel{g_i}{\longrightarrow} V_i := \phi^{-1}(i)$ of $g'$ whose involution is trivial everywhere except for the flags forming edges in $E_i$, in which case it coincides with the involution of $\Gamma'$.
Then there is a unique preodered graph $\Gamma'': F''\stackrel{g''}{\epi} V''$ and a morphism $\Phi:\Gamma'\to \Gamma''$ as in (\[graphsmorphism\]) whose fibers $\inv{\Phi}(i)$ are equal to $\Gamma_i$, $i\in V''$. Such a morphism is unique if $\psi$ is required to be order preserving.
We construct $\Gamma''$ as the graph whose set of vertices is $V''$ and whose set of flags $F''$ is $F'':=
F' \setminus \bigcup_{i\in V''}E_i$. The defining map $g'':F''\to V''$ is the restriction of the composite $\phi \circ g'$, as shown in $$\xymatrix@C=3.5em{F' \ar[d]_{g'} & \ F'':=
F' \setminus \bigcup_{i\in V''}E_i \ar@{_{(}->}[l]_(.7){\psi}%)
\ar[d]^{g''}
\\
V' \ar@{->>}[r]^{\phi} &V''.
}$$ We finally define $\Phi:\Gamma'\to \Gamma''$ as the couple $(\psi,\phi)$ with $\psi:F''\hookrightarrow F'$ the inclusion. It is easy to see that $(\psi,\phi)$ is a morphism of graphs with the required properties. It is also quite clear that $(\psi,\phi)$ is the unique such map with $\psi$ order preserving.
\[Ron\_v\_Praze\] We call the morphism $\Phi = (\psi,\phi): \Gamma' \to \Gamma''$ constructed in Lemma \[contractions\] a [*contraction*]{}. If $\psi$ is order preserving we call it a [*pure contraction*]{}.
Indeed, $\Gamma''$ is, roughly speaking, obtained from $\Gamma'$ by contracting edges belonging to $E_i$, $i \in V''$. The following claim is obvious.
Every morphism $\Phi = (\psi,\phi)$ of preodered graphs such that both $\phi$ and $\psi$ are order preserving is a pure contraction. The composite of two pure contractions is a pure contraction.
Let us prove a version of the weak blow up axiom for pure contractions.
\[bupforcontractions\] Let $\Phi = (\psi,\phi):\Gamma'\to \Gamma''$ be a pure contraction with fibers $\Gamma_i = (F_i,V_i)$, $i \in V''$. Given pure contractions $\Xi_i:\Gamma_i\to
\Lambda_i$ for each $i\in V''$, there exists a unique factorization of $\Phi$ as a composite of pure contractions $$\label{po_dlouhe_dobe_opet_pisu}
\xymatrix@C=2em@R=1.2em{\Gamma' \ar[rr]^a \ar[dr]_{\Phi} && \Gamma \ar[ld]^{b}
\\
&\Gamma''&
}$$ such that the induced map $a_i$ of the fibers equals $\Xi_i$, $i \in V''$.
Assume that the pure contraction $\Phi$ is given, as in Lemma \[contractions\], by an order preserving map $\phi : V' \to V''$ and subsets $E_i$, $i\in V''$, of edges. Suppose also that the pure contractions $\Xi_i$ are given by order-preserving maps $\phi_i : \inv \phi (i) \epi V_i$, $i \in V''$, and subsets $E_{ij} \subset E_i$ of edges of $\Xi_i^{-1}(j)$ for each $j \in V_i$. We then use Lemma \[contractions\] to build $\Gamma$ with the set of vertices $V$, and a pure contraction $a$ as follows. As $V$ we take the ordinal sum $\bigcup_{i \in V''_i} V_i$ and $$\label{vcera_do_mne_zase_sili}
\phi_a := \bigcup_{i \in V''_i} \phi_i : V' = \bigcup_{i \in V''_i}
\phi^{-1}(i) \longrightarrow V.$$ The pure contraction $a$ is then determined by $\phi_a : V' \to V$ and the subsets of edges $E_{ij}$, $j \in V_i$, $i \in V''$. It is easy to check that $\Gamma''$ is a result of a further pure contraction $b$. The uniqueness of the construction is clear.
Another version of the weak blow up axiom is described in
\[bupforisos\] Let $\Phi:\Gamma'\to \Gamma''$ be a pure contraction with fibers $\Gamma_i$, $i \in V''$. Given local isomorphisms $\Xi_i:\Gamma_i\to \Lambda_i$ for each $i\in V''$, there exists a unique factorization of $\Phi$ as in (\[po\_dlouhe\_dobe\_opet\_pisu\]) in which $a$ is a local isomorphism inducing the prescribed maps $\Xi_i$ on the fibers, and $b$ a pure contraction.
Let $\Lambda_i = (V_i,F_i)$. We construct $\Gamma$ in (\[po\_dlouhe\_dobe\_opet\_pisu\]) as the graph whose set of vertices $V$ equals the ordinal sum $\bigcup_{i \in V''_i} V_i$ and the set $F$ of flags the ordinal sum $\bigcup_{i \in V''_i} F_i$. There is an obvious isomorphisms $\psi_a$ between the set $F$ of flags of $\Gamma$ and the set $F'$ of flags of $\Gamma'$ induced by the local isomorphism between the fibers. We transport the involution of $\Gamma'$ to the flags of $\Gamma$ along this isomorphism. Then $a := (\psi_a,\phi_a)$ with $\phi_a$ as in (\[vcera\_do\_mne\_zase\_sili\]) is the requisite local isomorphism. It is easy to check as in the proof of Lemma \[bupforcontractions\] that $\Gamma''$ is a result of a further pure contraction and that the factorization is unique.
The last version of the weak blow up axiom which we will need is
\[local\] Let $\Phi:\Gamma'\to \Gamma''$ be a map with fibers $\Gamma_i$, $i \in V''$. Given local reorderings $\Xi_i:\Gamma_i\to \Lambda_i$ for each $i\in V''$, there exists a unique factorization of $\Phi$ as in (\[po\_dlouhe\_dobe\_opet\_pisu\]) in which $a$ is a local reordering that induces the prescribed maps on the fibers.
Each vertex $v$ of $\Gamma'$ belongs to a unique fiber of $\Phi$. So the prescribed reorderings of the fibers determine a reordering at each vertex of $\Gamma'$. We thus construct $\Gamma$ as the graph with the same vertices as $\Gamma'$ but with the local orders modified according to the above reorderings. The map $a: \Gamma' \to \Gamma$ is then the related local reordering map. Since it is an isomorphism, it determines the map $b : \Gamma \to \Gamma''$ uniquely.
In Lemma \[local\] we did not require $\Phi$ to be a pure contraction. When $\Phi$ is a pure contraction, $b : \Gamma \to
\Gamma''$ need not be pure, but it is still a contraction in the sense of Definition \[Ron\_v\_Praze\].
\[fact\] Any morphism $\Phi$ in $\ttprGr$ can be factorized as a local isomorphism followed by a pure contraction followed by a local reordering. Symbolically $$\label{mam_prohlidky}
\Phi = {\it Reo}\circ {\it Cont}\circ {\it Li}.$$
Assume the notation in (\[graphsmorphism\]). We first factorize $\phi$ as a bijection $\pi:V'\to V$ followed by an order preserving map $\xi: V \to V''$ such that $\pi$ restricts to order-preserving isomorphism $\phi^{-1}(i) \cong \xi^{-1}(i)$ for each $i\in V''$, cf. the bottom row of $$\label{vcera_s_Jarkou_na_Korabe}
\xymatrix@C=4em{F' \ar[r]^\eta_\cong \ar[d]_{g'} & \ar[d]_{g} F
& F'' \ar@{^{(}->}[l]_{\eta \circ \psi} \ar[d]_{g''}
\ar@/_2em/[ll]_{\psi}
\\
\ar@{->>}@/_2em/[rr]^{\phi} V' \ar[r]^\pi_\cong& \ar@{->>}[r]^{\xi} V &\ V''.
}$$ We then factorize $\pi \circ g'$ into the composite $F'\stackrel{\eta}{\longrightarrow} F\stackrel{g}{\longrightarrow} V$ where $\eta$ induces an order-preserving isomorphism $\eta^{-1}\big(g^{-1}(j)\big)
\cong g^{-1}(j)$ for each $j \in V$, cf. the left square in (\[vcera\_s\_Jarkou\_na\_Korabe\]). We induce an involution on $F$ from $F'$ via the isomorphism $\eta$. The pair $(\eta^{-1},\pi)$ is the required local isomorphism $ {\it Li}$ in (\[mam\_prohlidky\]).
The pair $( \eta \circ\psi, \xi)$ in the right square of (\[vcera\_s\_Jarkou\_na\_Korabe\]) is a morphism of graphs as well. We factorize $\eta \circ \psi$ as a bijection $\mu:F''\to F'''$ followed by an order preserving monomorphism $\lambda:F'''\to F'$ as in $$\xymatrix@C=3em{F \ar[d]_{g} & \ar@{_{(}->}[l]_\lambda
F''' \ar[dr]^(.6){g'''} &
\ar[l]_\mu^\cong F''\ar@/_2em/[ll]_{\eta \circ \psi}
\ar[d]^{g''}
\\
V \ar@{->>}[rr]^\xi && \ V''.
}$$ We finally define $g''':F'''\to V''$ as $\xi\circ g\circ
\lambda$. Since $\xi\circ g\circ \lambda \circ \mu = g''$, the diagram $$\xymatrix@C=3.5em{F''' \ar[d]_{g'''} & F''\ar[l]^\cong_{\mu} \ar[d]^{g''}
\\
V'' \ar[r]^{\id} & V''
}$$ commutes. It is a reordering morphism playing the rôle of ${\it Reo}$ in (\[mam\_prohlidky\]). The pair $(\lambda, \xi)$, which is clearly a pure contraction, is ${\it Cont}$ in (\[mam\_prohlidky\]).
Any isomorphism of preodered graphs can be factorized into a local isomorphism followed by a reordering morphism, symbolically ${\it Iso} = {\it Reo}\circ {\it Li}$.
The statement follows from Proposition \[fact\] combined with the obvious fact that the only pure contractions that are isomorphisms are the identity maps.
Any morphism $\Phi = (\psi,\phi)$ such that $\phi: V' \to V''$ is order-preserving is a composite of a pure contraction followed by a local reordering.
Another consequence of Proposition \[fact\]. Notice that the decomposition of $\phi$ in the bottom row of (\[vcera\_s\_Jarkou\_na\_Korabe\]) was specified so that if $\phi$ is order-preserving, $\pi$ must be the identity, thus $\eta = \id$ as well, so ${\it Li}$ in (\[mam\_prohlidky\]) is the identity morphism.
For each natural number $n \geq 0$ let $1_n$ (the [*corolla*]{}) be the graph $\bar{n}\to \bar{1}$ with the trivial involution. The corollas are not the local terminal objects in $\ttprGr$ since there are exactly $n!$ morphisms from any graph $\Gamma$ with $n$ legs to $1_n.$ Any such a morphism is completely determined by a linear order of the legs of $\Gamma.$
\[Radeji\_bych\_sel\_s\_Jaruskou\_na\_vyhlidku\_sam.\] The operadic category of [*ordered graphs $\Gr$*]{} is the coproduct of the categories ${\ttprGr}/1_n$. The cardinality functor assigns to a graph the (linearly ordered) set of its vertices.
A morphism $\Phi : \Gamma' \to \Gamma''$ of ordered graphs, i.e. a diagram $$\label{Ruda_je_rasista.}
\xymatrix@R=.3em@C=2.5em{
F' \ar[dddd]_{g'} && \ar@{_{(}->}[ll]_\psi \ar[dddd]^{g''} F''
\\
&\bar{n} \ar@{^{(}->}[lu] \ar[dd] \ar@{_{(}->}[ru] &
\\
{}
\\
&\bar{1} &
\\
V' \ar@{->>}[ru] \ar[rr]^\phi &&V'' \ar@{->>}[lu]
}$$ induces for each $i \in V''$ a commutative diagram $$\label{Jana_taky.}
\xymatrix@C=3.5em{g'^{-1}\big(\phi^{-1}(i)\big) \ar[d]_{g'}
&
g''^{-1}(i)\ar@{_{(}->}[l]_(.4){\psi} \ar[d]^{g''}
\\
\phi^{-1}(i) \ar[r]^{\phi} & \bar{1}
}$$ in $\Fin$ in which the morphisms $g',\phi,g''$ and $\psi$ are the restrictions of the corresponding morphisms from (\[Ruda\_je\_rasista.\]). We interpret the right vertical morphism as a corolla by imposing the trivial involution on $g''^{-1}(i)$. Due to the definition of fibers of maps of preodered graphs, the diagram above represents a map of the fiber of $\Phi$ over $i$ to a corolla, which makes it an ordered graph. We take it as the definition of the [*fiber in $\Gr$*]{}. In other words, the fiber gets a linear order on its legs from the ordinal $g''^{-1}(i)$. Finally, the chosen local terminal objects in $\Gr$ are $c_n =
\id:1_n\to 1_n$, that is corollas whose global order of legs coincides with the local order at this unique vertex.
It follows from the commutativity of the upper triangle in (\[Ruda\_je\_rasista.\]) that the map $\psi$ preserves the global orders of legs, therefore morphisms of ordered graphs induce order-preserving bijections of the legs of graphs. The category ${\tt \Delta}\Gr$ then consists of morphisms (\[Ruda\_je\_rasista.\]) in which, moreover, $\phi$ is order-preserving, that is, the order of vertices is preserved.
A $\Phi : \Gamma' \to \Gamma''$ in $\Gr$ is, by definition, a morphism (\[Ruda\_je\_rasista.\]) whose each fiber is the chosen local terminal object $c_n$ for some $n \geq 0$. It is clear that in this case both $\phi$ and $\psi$ must be bijections and, moreover, the local orders on $g''^{-1}(i) $ and $g'^{-1}\big(\phi(i)\big)$ coincide for each $i \in V''$. In other words, are local isomorphisms over $1_n$. So $\Gamma''$ as an ordered graph is obtained from $\Gamma'$ by reordering its vertices. In particular
All quasibijections in $\Gr$ are invertible.
Observe that the morphisms from ${\tt \Delta}\Gr$ are morphisms (\[graphsmorphism\]) of preodered graphs over $1_n$ such that $\phi$ is order-preserving.
\[dnes\_prednaska\_na\_Macquarie\] An operadic category $\ttO$ is [*factorizable*]{} if each morphism $f \in \ttO$ decomposes, not necessarily uniquely, as $\phi \circ \sigma$ for some $\phi \in \DO$ and $\sigma \in \QO$ or, symbolically, $\ttO = \DO \circ \QO$.
\[Clemens zitra odleta.\] The operadic category $\Gr$ is factorizable.
Given a morphism $\Gamma'\to \Gamma''$ over $1_n$ we use Proposition \[fact\] to factorize it as a local isomorphism $\Gamma'
\to \Gamma$ followed by a local reordering composed with a contraction, which is an order preserving morphism $\Gamma \to
\Gamma''$. We have a commutative diagram $$\xymatrix@C=3.5em{F \ar[d]_\cong & F''\ar@{_{(}->}[l]_{}
\\
F' & \bar{n}\ar@{_{(}->}[l]_{} \ar@{_{(}->}[u] \ar@{_{(}-->}[lu]}$$ of flags of the corresponding graphs. All maps in this diagram induce isomorphisms of the sets of legs. Thus there is a unique monomorphism $\bar{n} \hookrightarrow F$ which makes the diagram commutative and which, moreover, induces an isomorphism of the sets of legs. Therefore the factorization described above is the factorization over $1_n$ as required.
Lemma \[Dnes\_s\_Clemensem\_do\_Matu\] below involves a local reordering morphism $\rho : \Upsilon \to \Gamma''$ of directed graphs. By definition, such $\rho$ induces the identity between the vertices of the graphs $\Upsilon$ and $\Gamma''$, i.e. their vertices are ‘the same.’ This observation is important for the formulation of:
\[Dnes\_s\_Clemensem\_do\_Matu\] Consider a commutative diagram $$\xymatrix@C=1em@R=1em{
\Gamma' \ar[rrrr]^f \ar[rd]_g
%_a
%\ar@/_2em/[ddrr]_\Phi
&&&& \Gamma \ar[llld]
%_(.6)\alpha
\ar[ddll]
%_(.6)\beta
\\
&\Upsilon \ar[rd]^\rho&&&
\\
&&\Gamma''&&
}$$ of directed graphs in which $\rho : \Upsilon \to \Gamma''$ is a local reordering. Then the fiber of the map $g$ over $i \in \Vert(\Upsilon)$ is obtained from the fiber of $g$ over the same $i$ but considered as belonging to $|\Gamma''|$ by changing the global order of its legs according to the local reordering at the corresponding vertex over which the fibers are taken. Also the maps between the fibers induced by $f$ over $i\in |\Upsilon|$ resp. $i\in
|\Gamma''|$ are related by the same kind of global relabeling.
Direct verification.
The operadic category $\Gr$ satisfies the weak blow up axiom.
Let $\Phi : \Gamma' \to \Gamma''$ be an order-preserving map with fibers $\Gamma_i$, $i
\in |\Gamma''|$. Assume we are given a morphism $\Xi_i : \Gamma_i \to
\Lambda_i$ for each $i$. To prove the lemma, we must embed $\Phi$ into diagram (\[po\_dlouhe\_dobe\_opet\_pisu\]) in which $a$ induces the prescribed maps of fibers, and prove its uniqueness.
Let us first ignore the global orders of graphs involved, i.e. work in the category ${\ttprGr}$ of preordered graphs. We first factorize, using Proposition \[fact\], $\Phi$ into a pure contraction $c$ followed by a local reordering $\rho$ as in the bottom of $$\label{az_spony_mostu_sepnou_reku_noc_a_ty_jsi_jeden}
\xymatrix@C=3em{
A \ar[r]^\beta \ar[rd]^u &\ar[r]^\gamma\ar[d]^v B &
\Gamma\ar[d]^b \ar[ld]_w
\\
\Gamma' \ar[r]^c \ar[u]^\alpha \ar@/_2em/[rr]^\Phi & \Upsilon \ar[r]^\rho
&\ \Gamma''.
}$$
Let $\widehat \Gamma_i$ be the graph obtained from $\Gamma_i$ by modifying its global order according to the action of the local reordering $\rho$ as in Lemma \[Dnes\_s\_Clemensem\_do\_Matu\]. Notice that $\widehat \Gamma_i$ is the fiber of $c$ over $i \in |\Upsilon|$. Let $\widehat \Lambda_i$ be the graph $\Lambda_i$ with the global order modified in the same manner, and $\widehat \Xi_i : \widehat \Gamma_i \to
\widehat \Lambda_i$ the induced map. We factorize $\widehat \Xi_i$ as a quasibijection followed by a pure contraction and a local reordering, as in $$\widehat \Xi_i : \widehat \Gamma_i \stackrel{\alpha_i}\longrightarrow
A_i \stackrel{\beta_i}\longrightarrow B_i \stackrel{\gamma_i}\longrightarrow
\widehat \Lambda_i, \ i \in |\Upsilon|.$$ We then realize these families of maps as the induced maps between fibers step by step using Lemmas \[bupforcontractions\], \[bupforisos\] and \[local\] giving rise to the maps $\alpha$, $\beta$ and $\gamma$ in (\[az\_spony\_mostu\_sepnou\_reku\_noc\_a\_ty\_jsi\_jeden\]). It is clear that diagram (\[po\_dlouhe\_dobe\_opet\_pisu\]) with $a : = \gamma \circ
\beta \circ \alpha$ and $b : = \rho \circ w$ commutes. By Lemma \[Dnes\_s\_Clemensem\_do\_Matu\], $a$ induces the requisite maps between the fibers in the category of preordered graphs. Since the forgetful functor $\Gr \to {\ttprGr}$ is faithful, the same is true also in the category of ordered graphs.
We must prove that the graph $\Gamma$ in (\[po\_dlouhe\_dobe\_opet\_pisu\]) thus constructed carries a compatible global order. Since morphisms in $\ttprGr$ map legs to legs bijectively, the unique dashed arrow in $$\xymatrix@R=1em@C=1em{\Leg(\Gamma') && \Leg(\Gamma) \ar[ll]_\cong
\\
&\bar n \ar[ul]_(.3)\cong \ar@{-->}[ur] \ar[d]_(.3)\cong&
\\
&\Leg(\Gamma'') \ar[uul]^\cong\ar[uur]_\cong &
}$$ provides the requisite global order of $\Gamma$.
We need to prove that the factorization (\[po\_dlouhe\_dobe\_opet\_pisu\]) is unique. Let $\Gamma' = (V',F')$, $\Gamma'' = (V'',F'')$ and $\Lambda_i = (V_i,F_i)$ for $i \in V''$. Since the map $b: \Gamma \to
\Gamma'' \in {\tt \Delta}\Gr$ is order-preserving, the set $V$ of vertices of $\Gamma$ must be the ordinal sum $\bigcup_{i
\in V''}V_i$ of the sets of vertices of the fibers and, likewise, the set of flags $F$ of $\Gamma$ equals the ordinal sum $\bigcup_{i
\in V''} F_i$. It is not difficult to show that also the involution on $F$ is determined by the involutions on $F''$ and $F_i$, $i \in V''$. Thus the graph $\Gamma$ is uniquely determined by the input data, namely by $\Gamma$ and the fibers $\Lambda_i$, $i\in V''$.
Let us discuss the uniqueness of the maps in (\[po\_dlouhe\_dobe\_opet\_pisu\]). As each vertex of $\Gamma'$ belongs to a unique fiber of $\Phi$, the horizontal arrow in the diagram $$\xymatrix{
V' \ar@{->>}[rr] \ar@{->>}[rd] &&V \ar@{->>}[ld]
\\
&V''&
}$$ of the induced maps of vertices is uniquely determined by the maps $\Vert(\Gamma_i) \to \Vert(\Lambda_i)$, $i\in V''$, induced by the prescribed maps $\Xi_i$ of the fibers. Since both down-going maps are order-preserving by assumption, the right down-going map is uniquely determined by the remaining two. By the similar argument, the horizontal inclusion in the diagram $$\xymatrix{F'&& \ar@{_{(}->}[ll] F
\\
&F'' \ar@{_{(}->}[lu] \ar@{^{(}->}[ur] &
}$$ of the induced maps of flags is uniquely determined by the maps $\Flag(\Lambda_i) \to \Flag(\Gamma_i)$, $i \in V''$, induced by the prescribed maps of the fibers, so the right up-going inclusion is unique as well. This finishes the proof.
Further properties of operadic categories {#Poletim_letos_do_Sydney?}
=========================================
In Section \[Ceka\_mne\_Psenicka.\] we introduced the operadic category $\Gr$ of ordered graphs. We proved that it is a factorizable operadic category satisfying the weak blow up axiom in which all are invertible. It is moreover obvious that $\Gr$ is graded by the number of edges of graphs and, since we assumed $\phi$ in (\[graphsmorphism\]) to be surjective, constant free.
In this section we show how the (strong) blow up axiom for $\Gr$ follows from the above properties. We also prove that $\Gr$ fulfills a couple of other useful axioms. Let us start by formulating the following variant of Definition \[dnes\_prednaska\_na\_Macquarie\].
\[zase\_jsem\_podlehl\] An operadic category $\ttO$ is [*strongly factorizable*]{} if each morphism $f : T\to S $ decomposes [*uniquely*]{} as $\phi \circ \sigma$ for some $\phi \in \DO$ and $\sigma \in \QO$ such that the induced map between the fibers $$\sigma_i:f^{-1}(i)\to \phi^{-1}(i)$$ is the identity for each $i\in |S|$.
\[I\_laboratorni\_vysetreni\_mne\_ceka.\] Assume that in $\ttO$ all are invertible, $\ttO$ is factorizable and satisfies the weak blow up axiom. Then $\ttO$ is strongly factorizable and satisfies the (strong) blow up axiom. Schematically $$\hbox {\rm {\QBI} \& {\Fac} \& {\WBU}}
\Longrightarrow \hbox {\rm {\SBU} \& \SFac},$$ with the obvious meaning of the abbreviations.
Let $f:T\to S$ be a morphism in $\ttO$. We factorize it into a $\omega$ followed by an order preserving $\eta:T'\to
S$ as in the left upper triangle of $$\label{Nechce_se_mi_byt_pres_vikend_v_Praze.}
\xymatrix@C=3.5em{T \ar[r]^f \ar[d]^\omega_\sim &S
\\
T' \ar[ur]^\eta \ar[r]^\pi_\sim &\ Q. \ar[u]^\phi
}$$ Let $\pi_i : \eta^{-1}(i) \to f^{-1}(i)$ be, for $i\in |S|$, the inverse to $\omega_i:
f^{-1}(i)\to \eta^{-1}(i)$. Using the weak blow up axiom we factorize uniquely $\eta$ into $\phi \circ \pi$ such that $\pi$ on fibers induces the morphisms $\pi_i$, $i \in |S|$, see the lower right triangle of (\[Nechce\_se\_mi\_byt\_pres\_vikend\_v\_Praze.\]). Notice that $\pi$ is a quasibijection as well. We thus have a factorization of $f$ into a $\sigma := \pi \circ \omega$ followed by $\phi\in \DO$. Clearly, $\sigma$ induces identities of the fibers.
Suppose we have two such factorizations of $f$: $$\xymatrix@R=1em@C=3.5em{
&Q'\ar[rd]^{\phi'} \ar@{-->}[dd]^p_\sim &
\\
T \ar[ru]^{\sigma'}_\sim\ar[rd]^{\sigma''}_\sim &&S
\\
&Q''\ar[ru]^{\phi''}&
}$$ Since are invertible, we have a unique $p:Q'\to Q''$ which induces identities of the fibers over $S$. It follows from the uniqueness part of the weak blow up axiom that $p =
\id$. So the decomposition $f= \phi \circ \sigma$ is unique, thus $\ttO$ is strongly factorizable.
It remains to prove the general version of the blow up axiom. Let $$\xymatrix@C=3.5em{S' \ar[d]_{f'} &
\\
T' \ar[r]^\sigma_\sim &T''
}$$ be a corner for the blow up axiom as in (\[c1\]). By the weak blow up we have a unique factorization $S'\stackrel{\gamma}{\to} S''\stackrel{g}{\to} T'$ of $f$ as in $$\xymatrix@C=3.5em{S' \ar[d]_{f'} \ar[r]^\gamma
& S'' \ar[r]^\pi_\sim \ar[dl]_g & Q\ar[dl]_\eta
\\
T' \ar[r]^\sigma_\sim &T''&
}$$ such that $\gamma$ induces the prescribed morphisms of fibers. We then apply the strong factorization axiom to $\sigma \circ g$ and get a factorization $S''\stackrel{\pi}{\to} Q\stackrel{\eta}{\to} T''$ where $\pi \in \QO$ and $\eta \in\DO$. Clearly $(\pi \circ \gamma)\circ
\eta$ is the factorization required by the general blow up axiom.
The operadic category $\Gr$ satisfies the strong form of the blow up axiom.
The assumptions of Lemma \[I\_laboratorni\_vysetreni\_mne\_ceka.\] for $\Gr$ were verified in Section \[Ceka\_mne\_Psenicka.\].
\[Podivam\_se\_do\_Paramaty?\] Any isomorphism in an operadic category has the local terminal objects as its fibers. Conversely, in a factorizable operadic category in which all are isomorphisms and the blow up axiom is fulfilled, a morphism with terminal fibers is an isomorphism.
Let $\phi:S\to T$ be an isomorphism with the inverse $\psi$. Consider the commutative diagram over $T$: $$\xymatrix@C=2em@R=1.2em{S\ar[r]^{\phi}
\ar[dr]_{\phi} &T\ar[r]^{\psi} \ar[d]^(.4){\id}&\ S. \ar[ld]^{\phi}
\\
&T&
}$$ This diagram induces isomorphisms from the fibers of $\phi$ to the fibers of identity morphism of $T$. Therefore the fibers of $\phi$ are isomorphic to trivial objects, so they are all local terminal.
Conversely, suppose an operadic category $\ttO$ is factorizable with all isomorphisms, and suppose that all fibers of $\phi:A\to T$ are local terminal. By assumption, one can factorize $\phi$ as a $\sigma$ followed by $\xi \in \DO$. The $\sigma$ induces , hence isomorphisms, between the fibers of $\phi$ and $\xi$. So it will be enough to show that any $\xi: R\to S$ in $\DO$ with local terminal fibers is an isomorphism.
Let $F_i$ be, for $i \in |S|$, the $i$th fiber of $\xi$. Since each $F_i$ is local terminal, we have by assumption the unique isomorphism $\xi_i:F_i\to U_{c_i}$ for each $i$, and its inverse $\eta_i : U_{c_i} \to F_i$. By the weak blow up axiom there exists a unique factorization of $\id:S\to S$ as $S\stackrel{a}{\to}
Q\stackrel{b}{\to}
S$ such that $a$ induces the morphisms $\eta_i$ on the fibers. The following diagram $$\xymatrix@C=2em@R=1.2em{R\ar[r]^{\xi}
\ar[dr]_{\xi} &S\ar[r]^{a} \ar[d]^(.3){\id}&Q \ar[ld]^{b}
\\
&S&
}$$ in $\DO$ commutes and by functoriality it induces the identity morphisms between the fibers of $\xi$ and $b$. By the uniqueness part of the weak blow up axiom $Q=R$, $b= \xi$ and $\xi\circ a = \id_R$. Repeating the same argument we find also that $a\circ \xi = \id_S$, hence $\xi$ is an isomorphism.
\[Podival\_jsem\_se\_do\_Paramaty!\] Under the assumptions of the second part of Lemma \[Podivam\_se\_do\_Paramaty?\], each $f \in \ttO$ decomposes as $\psi \circ \omega$, where $\omega$ is an isomorphism and all terminal fibers of $\psi \in
\DO$ are trivial (= chosen local terminal).
Decompose $f$ into $A \stackrel\sigma\to X \stackrel\phi\to B$ with $\sigma$ a and $\phi \in \DO$ using the factorizability in $\ttO$. By the weak blow up axiom, one has the diagram $$\label{v_Koline_s_chripkou}
\xymatrix{X \ar[rr]^{\tilde \sigma} \ar[dr]_\phi && Y \ar[ld]^\psi
\\
&B&
}$$ in which $\psi$ has the same non-terminal fibers as $\phi$ and all its terminal fibers are trivial, and $\tilde \sigma$ induces the identity maps between non-terminal fibers. Clearly, all fibers of $\tilde \sigma$ are local terminal, thus it is an by Lemma \[Podivam\_se\_do\_Paramaty?\]. The desired factorization of $f$ is then given by $\psi$ in (\[v\_Koline\_s\_chripkou\]) and $\omega := \tilde \sigma \circ \sigma$.
\[Kveta\_asi\_spi.\] The [*unique fiber axiom*]{} for an operadic category $\ttO$ requires that, if the fiber of the unique morphism $\Phi:T\to t$ to a local terminal object is $T$, then $t$ is a chosen local terminal object $U_c$. In other words, the only situation when $T \fib T \to t$ is when $t$ is chosen local terminal.
\[Kdy\_konecne\_zas\_budu\_mit\_od\_doktoru\_pokoj?\] If the unique fiber axiom is fulfilled, any inverse to a is again a .
Assume that $\sigma : T \to S$ is a and $\tau : S \to T$ its inverse. There is the induced morphism $\sigma_i : U_c \to
\inv\tau (i)$ for every $i \in |T|$, where $\inv\tau (i)$ is local terminal by Lemma \[Podivam\_se\_do\_Paramaty?\]. By Axiom (iv) of an operadic category, the unique fiber of $\sigma_i$ is $U_c$ again, so we have $U_c \fib U_c \stackrel{\sigma_i}\longrightarrow \inv\tau (i)$. By the unique fiber axiom. $\inv\tau (i)$ is chosen local terminal for each $i \in |T|$, so $\tau$ is a as claimed.
The category $\Gr$ satisfies the unique fiber axiom.
Assume that the directed graph $T$ is given by the left diagram below $$T=
\xymatrix{F \ar[d]_g & \ar@{_{(}->}[l]_u \ar[d]\ \bar n
\\
V \ar@{->>}[r] & \bar 1
},
\hskip 2em
t=\xymatrix{\bar n \ar[d] & \ar@{_{(}->}[l]_\alpha \ar[d]\ \bar n
\\
\bar 1 \ar@{->>}[r] & \bar 1
}$$ and the local terminal object by the right one. A morphism $\Phi:T\to t$ in $\Gr$ is characterized by a monomorphism $\psi : \bar n \to F$ in the diagram $$\label{NOE_v_padesatce}
\xymatrix@R=.3em@C=2.5em{
F \ar[dddd]_{g} && \ar@{_{(}->}[ll]_\psi \ar[dddd] \bar n
\\
&\bar{n} \ar@{^{(}->}[lu]^u \ar[dd] \ar@{_{(}->}[ru]_\alpha &
\\
{}
\\
&\bar{1} &
\\
V \ar@{->>}[ru] \ar[rr] &&\bar 1 \ar@{->>}[lu]
}$$ and its fiber $\inv{\Phi}(1)$ equals, by (\[Jana\_taky.\]), to $$\xymatrix{F \ar[d]_g & \ar@{_{(}->}[l]_\psi \ar[d] \bar n
\\
V \ar@{->>}[r] & \ \bar 1.
}$$ Thus $\inv{\Phi}(1) = T$ if and only if $\psi = u$. On the other hand, the commutativity of the upper triangle in (\[NOE\_v\_padesatce\]) implies that $u = \psi\alpha$. Since $\psi$ is a monomorphism, one sees that $\alpha = \id$, thus $t$ is the chosen local terminal object.
\[twolocal\] Let $F \fib T \stackrel{\phi'}\longrightarrow t'$ and $F \fib T \stackrel{\phi''}\longrightarrow t''$ be morphisms to local terminal objects with the same fiber $F$. If the weak blow up and unique fiber axioms are satisfied then $\phi' = \phi''$.
Consider the commutative triangle $$\xymatrix@C=2em@R=1.2em{T \ar[rr]^{\phi''}
\ar[dr]_{\phi'} && t'' \ar[ld]^{\xi}
\\
&t'&
}$$ in which $\xi$ is the unique map between the local terminal objects. We have the induced morphism of fibers $\phi''_1:F\to t = :
\inv{\xi}(1)$. By Axiom (iv) of an operadic category[^8] the fiber of this morphism is $F$. As $\inv{\xi}(1)$ is local terminal by Lemma \[Podivam\_se\_do\_Paramaty?\], by the unique fiber axiom $\phi''_1 = !:F\to U_c$ is the unique map to a chosen local terminal object. This means that the fiber $\xi^{-1}(1)$ is $U_c$, so $\xi$ is a . By Corollary \[zase\_mam\_nejaky\_tik\], $\xi$ must be an identity.
\[Porad\_nevim\_jestli\_mam\_jit\_na\_ty\_narozeniny.\] An operadic category $\tt O$ is [*rigid*]{} if the only $\sigma$ that makes $$\label{Skrabe_mne_v_krku.}
\xymatrix@C=3.5em{S \ar[d]_{\phi} \ar@{=}[r] & S \ar[d]^{\phi}
\\
T \ar[r]^\sigma_\cong &T
}$$ with $\phi \in \DO$ commutative, is the identity $\id : T = T$.
Notice that the category $\Fin$ is not rigid, but its subcategory $\Surj$ of nonempty finite sets and their surjections is.
\[Podivam\_se\_do\_Paramaty??\] If the weak blow up and the unique fiber axioms are satisfied, the operadic category is rigid,[^9] schematically $$\hbox {\rm {\UFB} \& {\WBU}}
\Longrightarrow \hbox {{\Rig}}.$$
Since the category of finite sets and surjections is obviously rigid, one has $|\sigma|
= \id$ for $\sigma$ in (\[Skrabe\_mne\_v\_krku.\]). For each $i\in |T|$ we have the induced morphism of the fibers $$\phi_i: \phi^{-1}(i) \to \sigma^{-1}(i)$$ whose unique fiber is $\phi^{-1}(i)$ by Axiom (iv) of an operadic category. The fiber $\sigma^{-1}(i)$ is local terminal by Lemma \[Podivam\_se\_do\_Paramaty?\] thus, by the unique fiber axiom, $\sigma^{-1}(i) = U_c$, so $\sigma$ is a . Hence, it must be an identity by Corollary \[zase\_mam\_nejaky\_tik\].
Since the assumptions of Lemma \[Podivam\_se\_do\_Paramaty??\] are satisfied by the operadic category of directed graphs, one has:
The category $\Gr$ is rigid.
Discrete operadic (op)fibrations {#Minulou_sobotu_jsem_odletal_vlekarskou_osnovu.}
================================
In this section we focus on discrete operadic fibrations $p: \ttO \to \ttP$. We show that the operadic category $\ttO$ retains some useful properties of $\ttP$. Since, as we know from [@duodel page 1647], each set-valued $\ttP$-operad determines a discrete operadic fibration $p:
\ttO \to \ttP$, this gives a method to obtain new operadic categories with nice properties from the old ones. In the second part we formulate similar statements for opfibrations and cooperads.
Discrete operadic fibrations {#snad_mne_kotnik_prestane_bolet}
----------------------------
Recall Definition 2.1 of [@duodel]:
\[psano\_v\_Myluzach\] A strict operadic functor $p:\ttO\to \ttP$ is a [*discrete operadic fibration*]{} if
- $p$ induces an epimorphism $\pi_0(\ttO) \twoheadrightarrow
\pi_0(\ttP)$ and
- for any morphism $f : T\to S$ in $\ttP$ and any $t_i, s \in \ttO$, where $i \in |S|$, such that $$p(s) = S \mbox { and } p(t_i) = f^{-1}(i) \mbox { for } i
\in |S|,$$ there exists a unique $\sigma : t\to s$ in $\ttO$ such that $$p(\sigma) = f \mbox { and } t_i = \sigma^{-1}(i)\ \mbox { for } i
\in |S|.$$
\[noha\_stale\_boli\] Let $p:\ttO\to\ttP$ be a discrete operadic fibration and $f:T\stackrel\sim\to
S$ a quasibijection in $\ttP$. Let $s\in \ttO$ be such that $p(s)=S$. Then there exists a unique quasibijection $\sigma$ in $\ttO$ such that $p(\sigma) = f$.
We invoke [@duodel Lemma 2.2] saying that a discrete operadic fibration induces an isomorphism of $\pi_0$’s, plus the fact that strict operadic functors are required to send trivial objects to trivial ones. Therefore $p$ establishes a bijection between the sets of trivial objects of the categories $\ttO$ and $\ttP.$ Hence, we can uniquely complete the data for $s$ by a list of trivial objects in place of the prescribed fibers and construct $\sigma$ as the unique lift of these data.
\[kotnik\] If in $\ttP$ all quasibijections are invertible, the same is true also for in $\ttO.$ In this case we also have that, for any quasibijection $f:T\stackrel\sim\to S$ in $\ttP$ and $t\in\ttO$ such that $p(t)=T$, there exists a unique quasibijection $\sigma:t\stackrel\sim\to s$ such that $p(\sigma)=f.$
Let $\sigma:t\stackrel\sim\to s$ be a quasibijection in $\ttO$. Consider the inverse $g:p(s)\to p(t)$ to the $p(\sigma): p(t) \stackrel\sim\to p(s)$. Notice that $g$ is a by Lemma \[Kdy\_konecne\_zas\_budu\_mit\_od\_doktoru\_pokoj?\]. Using Lemma \[noha\_stale\_boli\], we lift $g$ into a unique $\eta:s'\stackrel\sim\to
t$. The composite $\sigma\circ \eta$ is the lift of the identity $p(s)\to p(s)$ so, by uniqueness, it is the identity as well, in particular, $s =s'$. The composite $\eta\circ\sigma$ is the identity for the same reason.
The second part can be establishes as follows. Let $g:S\stackrel\sim\to T$ be the inverse quasibijection to $f$. We lift it to a $\tilde g : s \stackrel\sim\to t$ in $\ttO$ and define $\sigma:t\stackrel\sim\to s$ to be the inverse of this lift. The uniqueness of the lifting guarantees that $\sigma$ is a lift of $f$.
\[Krtek\_s\_Laurinkou\] Let $\ttP$ be a factorizable operadic category in which all are invertible. Then $\ttO$ is also factorizable.
Let $\xi:t\to s$ be a morphism in $\ttO$ with the fibers $t_i$, $i\in |s|$. Let $f := p(\xi):T\to S$ and let $T\stackrel{h}{\to}
Z \stackrel{g}{\to} S$ be its factorization into a $h$ followed by an order-preserving $g \in {\tt \Delta}\ttP$. Since $h$ is a , it induces $h_i:p(t_i)\stackrel\sim\to g^{-1}(i)$ for each $i\in |S| = |s|$. Using Lemma \[kotnik\] we lift $h_i$’s to $\sigma_i:t_i\stackrel\sim\to \tau_i$. Observe that $p(\tau_i) =
g^{-1}(i).$ We then lift $g$ to a morphism $\eta:z\to s$ in $\ttO.$ Obviously $\eta$ belongs to $\DO$.
Next we lift $h$ to a $\pi: t'\stackrel\sim\to z$ in $\ttO$. Let $\mu$ be the composite $t'\stackrel\pi\to z\stackrel\eta\to s$. We know that $p\big(\mu^{-1}(i)\big)= p(t_i)$ and that the morphism $\pi_i:\mu^{-1}(i)\to
\tau_i$ lifts $h_i:p(t_i)\to g^{-1}(i)$. From the uniqueness of the lifting we get $\mu^{-1}(i) =t_i$. This means that both $\xi$ and $\mu$ are lifts of $f$ with the same fibers. Therefore $\mu=\xi$, thus $\eta \circ \pi$ is the required factorization.
\[Vyhodil\_jsem\_sluchatka.\] If the blow up axiom holds in $\ttP$, it also holds in $\ttO.$
The methods of the proof are very similar to the ones used in the proofs of the previous statements, so we only sketch the main features. We first notice that the input data for the blow up in $\ttO$ are projected by $p$ the the input data in $\ttP$. Using the blow up assumption in $\ttP$, we complete these projected data into a commutative square as in (\[eq:5\]). Invoking the lifting property of operadic fibrations we first lift $f''$ into $\tilde f'' : s'' \to t''$ in $\ttO$, and then $\pi$ into $\tilde \pi : s' \to s''$. Using the uniqueness of lifting we then show that these liftings provide a blow up in $\ttO$.
\[Koupil\_jsem\_si\_silikonove\_podpatenky.\] The unique fiber axiom does not seem to be stable with respect to the discrete fibrations, so it has to be verified separately in each concrete case.
Important examples of discrete operadic fibrations are provided by the [ *operadic Grothendieck construction*]{} introduced in [@duodel page 1647]. Assume that one is given a set-valued $\ttP$-operad $\calO$. One then has the operadic category $\int_\ttP
\calO$ whose objects are $t \in \calO(T)$ for some $T \in \ttP$. A morphism $\sigma : t \to s$ from $t
\in \calO(T)$ to $s \in \calO(S)$ is a couple $(\varepsilon,f)$ consisting of a morphism $f : T \to S$ in $\ttP$ and of some $\textstyle\varepsilon \in \prod_{i \in |S|}
\calO\big(f^{-1}(i)\big)$ such that $$\gamma_f(\varepsilon,s) = t,$$ where $\gamma$ is the structure map of the operad $\calO$. Compositions of morphisms are defined in the obvious manner. The category $\int_\ttP \calO$ thus constructed is an operadic category such that the functor $p :
\int_\ttP \calO \to \ttP$ given by $$\label{bude_205?}
\mbox{$p(t) := T$ for $t \in \calO(T)$ and $p(\varepsilon,f)
:= f$}$$ is a discrete operadic fibration. The trivial objects are given by the operad units . By [@duodel Proposition 2.5], the above construction establishes an equivalence between the category of set-valued $\ttP$-operads and the category of discrete operadic fibrations over $\ttP$.
\[musim\_na\_ten\_odber\] Consider the $\Gr$-operad $C$ in $\Set$ such that $$C(\Gamma) :=
\begin{cases}
1 \hbox { (one point set)}, & \hbox {if $\Gamma$ is connected, and}
\\
\emptyset, & \hbox {otherwise.}
\end{cases}$$ There is a unique way to extend this construction to a $\Gr$-operad. The Grothendieck construction of $C$ produces a discrete operadic fibration $\Grc\to \Gr$. We call $\Grc$ the operadic category of [*connected directed graphs*]{}.
\[Zkusim\_zavolat\_Sehnalovi.\] A construction similar to the one in Example \[musim\_na\_ten\_odber\] produces the operadic category $\Tr$ of trees. We consider the operad $\Pi$ with $$\Pi(\Gamma) :=
\begin{cases}
1, & \hbox {if $\Gamma$ is contractible, and}
\\
\emptyset, & \hbox {otherwise.}
\end{cases}$$ Grothendieck’s construction gives a discrete operadic fibration $\Tr\to \Gr$.
\[Zitra\_s\_Jarkou\_k\_Bartosovi.\] Let us orient edges of a tree $T \in \Tr$ so that they point to the leg which is the smallest in the global order. We say that T is [*rooted*]{} if the outgoing half-edge of each vertex is the smallest in the local order at that vertex. Now define $$R(T) :=
\begin{cases}
1, & \hbox {if $T$ is rooted, and}
\\
\emptyset, & \hbox {otherwise.}
\end{cases}$$ Grothendieck’s construction associated to the operad $R$ gives the operadic category $\RTr$ of [*rooted trees*]{}.
\[Predtim\_s\_Jarkou\_na\_krest\_knihy.\] There is a unique isotopy class of embeddings of $T \in\Tr$ into the plane such that the local orders are compatible with the orientation of the plane. This embedding in turn determines a cyclic order of the legs of $T$. We say that $T$ is [*planar*]{} if this cyclic order coincides with the cyclic order induced by the global order of the legs. The operad $$P(T) :=
\begin{cases}
1, & \hbox {if $T$ is planar, and}
\\
\emptyset, & \hbox {otherwise.}
\end{cases}$$ gives rise to the operadic category $\PTr$ of planar trees. In the similar manner we obtain the operadic category $\PRTr$ of [*planar rooted trees*]{}.
All the above constructions fall into the situation captured by the following lemma whose proof is obvious.
Let $i:\ttC\subset \ttO$ be a full operadic subcategory such that
1. the set of trivial objects of $\ttO$ coincides with the set of trivial objects of $\ttC$, and
2. for any morphism $f$ in $\ttO$ whose codomain and all fibers are in $\ttC$, the domain of $f$ is also in $\ttC$.
Then $i$ is a discrete operadic fibration.
We close this subsection with the following useful
\[onevsO\] Let $\calO$ be a set-valued $\ttP$-operad, $\int_\ttP \calO \to \ttP$ its operadic Grothendieck’s construction and $\Groterm$ the terminal set-valued $\int_\ttP \calO$-operad. Then the categories of $\calO$-algebras and $\Groterm$-algebras are isomorphic, i.e. $$\Alg\calO \cong \Alg\Groterm.$$
The sets of connected components of the categories $\ttP$ and $\int_\ttP \calO$ are canonically isomorphic via the correspondence $$U_c \longleftrightarrow 1_c \in \calO(U_c)$$ of the chosen local terminal objects. We use this isomorphism to identify $\pi_0(\ttP)$ with $\pi_0(\int_\ttP \calO)$. Under this convention, the sets $\pi_0(s(T))$ of connected components of the sources of an object $T \in \ttP$ and the sets $\pi_0(s(t))$ of $t \in \calO(T)$ representing an object of $\int_\ttP
\calO$ are the same, and similarly $\pi_0(T) = \pi_0(t)$. The structure operations of an $\calO$-algebra are by definition
$$\label{Dnes_vecer_s_Jarkou}
\alpha_T : \bigotimes_{c \in \pi_0(s(T))} A_c \ot \calO(T) \longrightarrow
A_{\pi_0(T)}, \ T \in \ttP,$$
which can be interpreted as families $$\alpha_t : \bigotimes_{c \in \pi_0(s(T))} A_c \longrightarrow
A_{\pi_0(T)}, \ t \in \calO(T), \ T \in \ttP,$$ of maps parametrized by $t \in \calO(T)$. Using the above identifications we rewrite the above display as $$\label{do_francouzske_restaurace.}
\alpha_t : \bigotimes_{c \in \pi_0(s(t))} A_c \longrightarrow
A_{\pi_0(t)}, \ t \in \calO(T), \ T \in \ttP$$
which are precisely the structure operations of an $\Groterm$-algebra. It is simple to verify that the correspondence between (\[Dnes\_vecer\_s\_Jarkou\]) and (\[do\_francouzske\_restaurace.\]) extends to an isomorphism of the categories of algebras.
Discrete opfibrations
---------------------
In Subsection \[snad\_mne\_kotnik\_prestane\_bolet\] we recalled how set-valued operads produce discrete operadic fibrations. We are going to present a dual construction for cooperads.
The notion of a cooperad in an operadic category is obtained from that of an operad by reversing the arrows in [@duodel Definition 1.11]. A set-valued [*$\ttP$-cooperad*]{} is thus a collection of sets together with structure maps $$\label{za_chvili_na_ortopedii_s_patou}
\Delta_f: \oC(S) \longrightarrow \oC(T) \times \oC(F_1)\times
\cdots \times \oC(F_s)$$ defined for an arbitrary $f :S \to T$ with fibers $\Rada F1s$. The rôle of counits is played by the unique maps $$\oC(U_c) \to *,\ c \in \pi_0(\ttO),$$ to a terminal one-point set $*$. These operations are required to satisfy axioms dual to those .
A set-valued $\ttP$-cooperad $\oC$ leads to an operadic category $\int^\ttP \hskip -.2em \oC$ via a dual version of the Grothendieck construction recalled in Subsection \[snad\_mne\_kotnik\_prestane\_bolet\].[^10] The objects of $\int^\ttP \hskip -.2em \oC$ are elements $t \in \oC(T)$ for some $T \in \ttP$. A morphism $\sigma : t \to s$ from $t
\in \oC(T)$ to $s \in \oC(S)$ is a morphism $f : T \to S$ in $\ttP$ such that $$\Delta_f(t) = (s,\varepsilon)$$ for some, necessarily unique, $\textstyle\varepsilon \in \prod_{i \in |S|}
\oC\big(f^{-1}(i)\big)$, where $\Delta_f$ is the structure map (\[za\_chvili\_na\_ortopedii\_s\_patou\]).
The category $\int^\ttP \hskip -.2em \oC$ is an operadic category equipped with a functor $p :
\int^\ttP \hskip -.2em \oC \to \ttP$ defined by (\[bude\_205?\]). The trivial objects are all objects of the form $u \in \oC(U_c)$, $c \in\pi_0(\ttP)$. It turns out that the functor $p :
\int^\ttP \hskip -.2em \oC \to \ttP$ is a standard discrete opfibration:
\[zas\_mne\_boli\_zapesti\] A functor $p:\ttO\to \ttP$ is a [*discrete opfibration*]{} if
- $p$ induces an epimorphism $\pi_0(\ttO) \twoheadrightarrow
\pi_0(\ttP)$ and
- for any morphism $f : T\to S$ in $\ttP$ and any $t \in \ttO$ such that $ p(t) = T$ there exists a unique $\sigma : t\to s$ in $\ttO$ such that $p(\sigma) = f$.
Dualizing the steps in the proof of [@duodel Proposition 2.5] one can show that the dual Grothendieck construction is an equivalence between the category of set-valued $\ttP$-cooperads and the category of discrete opfibrations over $\ttP$. As the following statement shows, opfibrations behave nicely with respect to trivial objects.
\[Mam\_tendinopatii.\] Let $p : \ttO \to \ttP$ be a discrete opfibration. If $t \in \ttO$ is local terminal, then $p(t)$ is local terminal, too. Moreover, $t$ is a chosen local terminal if and only if so is $p(t)$.
Notice that the unique map $!_U: p(t) \to U$, where $U$ is some chosen local terminal object in $\ttP$, lifts into a unique map $!_u:
t \to u$ from $t$ to a chosen local terminal object $u$ in $\ttO$. Since $!_u$ is an isomorphism by the terminality of $t$, $!_U = p(!_u)$ is an isomorphism, too. This proves that $p(t)$ is local terminal.
Since operadic functors by definition send chosen local terminal objects to the chosen ones, for the second part of the lemma it suffices to prove that the triviality of $p(t)$ implies the triviality of $t$. To this end, notice that, if $p(t)$ is a chosen local terminal object in $\ttP$, the map $!_U: p(t) \to U$ above is the identity, so its lift $!_u:
t \to u$ is the identity as well.
The next property of opfibrations shall be compared to Remark \[Koupil\_jsem\_si\_silikonove\_podpatenky.\].
The unique fiber axiom in $\ttP$ implies the unique fiber axiom in $\ttO$.
Suppose we have a situation $T \fib T \to t$ in $\ttO$, with $t$ local terminal. By the first part of Lemma \[Mam\_tendinopatii.\], $p(T) \fib p(T) \to p(t)$ in $\ttP$ with $p(t)$ local terminal. By the unique fiber axiom in $\ttP$, $p(t)$ is a chosen local terminal object in $\ttP$, so $t$ is a chosen local object in $\ttO$ by the second part of Lemma \[Mam\_tendinopatii.\].
It turns out that analogs of lemmas \[noha\_stale\_boli\], \[kotnik\] and propositions \[Krtek\_s\_Laurinkou\], \[Vyhodil\_jsem\_sluchatka.\] hold also for discrete opfibrations. As an example, we prove the following variant of Lemma \[noha\_stale\_boli\]:
Let $p:\ttO\to\ttP$ be a discrete opfibration and $f:T\stackrel\sim\to
S$ a in $\ttP$. Let $t\in \ttO$ be such that $p(t)=T$. Then there exists a unique $\sigma$ in $\ttO$ such that $p(\sigma) = f$.
By the lifting property of opfibrations, $f$ lifts to a unique $\sigma$ so we only need to prove that $\sigma$ is a . Since $p$ is an operadic functor, it maps the fibers of $\sigma$ to the fibers of $f$. Since the latter are trivial in $\ttP$, the former must be trivial in $\ttO$ by Lemma \[Mam\_tendinopatii.\]. So $\sigma$ is a .
An analog of Lemma \[kotnik\] for a discrete opfibration $p:\ttO\to\ttP$ reads:
If all quasibijections in $\ttP$ are invertible, so they are also in $\ttO$. Moreover, for each $f:T\stackrel\sim\to S$ in $\ttP$ and $s\in\ttO$ such that $p(s)=S$, there exists a unique quasibijection $\sigma:t\stackrel\sim\to s$ such that $p(\sigma)=f$.
We leave the proof of this lemma as an exercise, as well as the verification that Propositions \[Krtek\_s\_Laurinkou\] and \[Vyhodil\_jsem\_sluchatka.\] hold verbatim for discrete opfibrations as well.
\[Zaletame\_si\_jeste\_do\_konce\_Safari?\] In Example \[musim\_na\_ten\_odber\] we constructed the operadic category $\Grc$ of connected directed graphs. We introduce a set-valued $\Grc$-cooperad $G$ as follows. For $\Gamma = (V,F) \in \Grc$ we put $$G(\Gamma) := {\rm Map}(V,\bbN) = \{g(v) \in \bbN\ | \ v \in V\}.$$ The cooperad structure operations $$\Delta_\Phi : G(\Gamma') \longrightarrow G(\Gamma'') \times
G(\Gamma_1) \times \cdots \times G(\Gamma_s)$$ are, for a map $\Phi : \Gamma' = (V',F') \to \Gamma'' = (V'',F'')$ with fibers $\Gamma_i = (V_i,F_i)$, $i \in V''$, given as $\Delta_\Phi(g') := (g'',g_1,\ldots,g_s)$, where $g_i$ is for $i \in
V''$ the restriction of $g'$ to $V_i \subset V'$ and $$g''(i) :=
\sum_{v \in V_i}g_i(v) + \dim\big(H^1(|\Gamma_i|; {\mathbb Z})\big),
\ i \in V'',$$ where $|\Gamma_i|$ is the obvious geometric realization of $\Gamma_i$.
The Grothendieck construction applied to $G$ produces the operadic category $\ggGrc$ of [*genus graded*]{} connected directed graphs. The morphisms in this category coincide with the morphisms of graphs as introduced in [@getzler-kapranov:CompM98 Section 2], modulo the orders which we used to make $\ggGrc$ an operadic category, cf. also Remark \[zitra\_vycvik\_vlekare\] below.
\[Pani\_Bilkova\_zatim\_neudelala\_rezrvaci.\] We say that a directed graph $\Gamma \in \Gr$ is [*oriented*]{} if
- each internal edge if $\Gamma$ is oriented, meaning that one of the half-edges forming this edge is marked as the input one, and the other as the output, and
- also the legs of $\Gamma$ are marked as either input or output ones.
We will call the above data an [*orientation*]{} and denote the set of all orientations of $\Gamma$ by $\Or(\Gamma)$. It is easy to see that $\Or$ is a cooperad in $\Grc$. The resulting category $\Whe$ consists of [*oriented directed*]{} connected graphs.
\[Travelodge\] Let $C$ be an obvious modification of the operad of Example \[musim\_na\_ten\_odber\] to the category $\Whe$. The Grothendieck’s construction associated to this modified $C$ produces the operadic category $\Dio$ of [*simply connected*]{} oriented directed graphs.
A [*valency*]{} of a vertex $u$ in a graph $\Gamma$ is the number of half-edges adjacent to $u$. For any $v \geq 2$, all operadic categories mentioned above that consist of simply connected graphs, i.e. $\Tr$, $\PTr$, $\RTr$, $\PRTr$ and $\Dio$, posses full operadic subcategories $\Tr_v$, $\PTr_v$, $\RTr_v$, $\PRTr_v$ and $\Dio_v$ of graphs whose all vertices have valency $\geq v$.
\[Zrejme\_brzy\_podlehnu.\] We call a directed simply-connected graph $\Gamma \in \Dio$ a [*$\frac12$graph*]{} if each internal edge $e$ of $\Gamma$ satisfies the following condition: $\bullet$ either $e$ is the unique outgoing edge of its initial vertex, or $\bullet$ $e$ is the unique incoming edge of its terminal vertex. Edges allowed in a $\frac12$graph are portrayed in Figure \[Jitulka\] which we borrowed from [@markl:handbook].
(28,12)(7,15) (14,19)[(0,1)[4]{}]{} (14,23)[(1,1)[3]{}]{} (14,23)[(1,2)[2]{}]{} (14,23)[(0,1)[4.5]{}]{} (14,23)[(-1,2)[2]{}]{} (14,23)[(-1,1)[3]{}]{} (11,21)[(3,2)[3]{}]{} (17,21)[(-3,2)[3]{}]{} (11,16)[(1,1)[3]{}]{} (17,16)[(-1,1)[3]{}]{} (14,16)[(0,1)[3]{}]{} (25,19)[(0,1)[4]{}]{} (25,23)[(1,1)[3]{}]{} (25,23)[(-1,2)[2]{}]{} (25,23)[(0,1)[4.5]{}]{} (25,23)[(1,2)[2]{}]{} (25,23)[(-1,1)[3]{}]{} (25,19)[(1,2)[2]{}]{} (25,19)[(-1,2)[2]{}]{} (22,16)[(1,1)[3]{}]{} (28,16)[(-1,1)[3]{}]{} (25,16)[(0,1)[3]{}]{} (25,19)[(1,1)[3]{}]{} (25,19)[(-1,1)[3]{}]{} (23,15)[(1,2)[2]{}]{} (27,15)[(-1,2)[2]{}]{} (-11,0)[ (23,15)[(1,2)[2]{}]{} (27,15)[(-1,2)[2]{}]{} ]{} (-11,4)[ (23,15)[(1,2)[2]{}]{} (27,15)[(-1,2)[2]{}]{} ]{} (25,14.5)[(0,1)[2]{}]{} (14,14.5)[(0,1)[2]{}]{} (14,23)[(0,0)\[cc\][$\bullet$]{}]{} (14,19)[(0,0)\[cc\][$\bullet$]{}]{} (25,23)[(0,0)\[cc\][$\bullet$]{}]{} (25,19)[(0,0)\[cc\][$\bullet$]{}]{} (14.5,20)[(0,0)\[lc\][$e$]{}]{} (25.5,22)[(0,0)\[lc\][$e$]{}]{}
Let us, for $\Gamma \in \Dio$, define $$\textstyle
\mathtt{\frac12}(\Gamma) :=
\begin{cases}
1, & \hbox {if $\Gamma$ is a $\frac12$graph, and}
\\
\emptyset, & \hbox {otherwise.}
\end{cases}$$ It is easy to verify that the restriction of $\mathtt{\frac12}$ to $\Dio_3 \subset \Dio$ is an operad. Grothendieck’s construction applied to $\mathtt{\frac12}$ produces the operadic category $\hGr$ of $\frac12$graphs.
The constructions above are summarized in the diagram $$\label{Dnes_jsme_byli_na_vylete_u_more.}
\xymatrix@R=1em{&\RTr\ar[dr] &&& \ar[d]\ggGrc&&
\\
\PRTr \ar[ur]\ar[dr] &&\ar[rr] \Tr && \ar[rr]\Grc && \Gr
\\
&\PTr\ar[ur] &&&\ar[u]\Whe&\Dio\ar[l]&\, \Dio_3 \ar@{_{(}->}[l]&\hGr\ar[l]
}$$ in which all arrows are discrete operadic fibrations except for $\ggGrc \to \Grc$ and $\Whe\to \Grc$ which are discrete opfibrations.
Let $\Ord$ be the lluf subcategory of $\Fin$ consisting of order-preserving surjections. It is an operadic category whose operads are the classical constant-free non-$\Sigma$ (non-symmetric) operads [@duodel Example 1.15]. One has the $\Ord$-cooperad $\Su$ with components $$\Su(\bar n) := \coprod_{m \geq n}\Sur(\bar m,\bar
n), \ n \geq 1,$$ where $\Sur(\bar m,\bar n)$ denotes the set of all (not necessarily order-preserving) surjections. Its structure map $\Delta_f
: \Su(\bar n) \to \Su(\bar s) \times \Su(\inv f (1)) \times \cdots
\times \Su(\inv f (s))$ is, for $f : \bar n \to \bar s$, given by $$\Delta_f(\alpha) := \beta \times \alpha_1 \times \cdots \times
\alpha_s,\
\alpha \in \Su(\bar n),$$ where $\beta:= f \circ \alpha$ and $\alpha_i : (f\alpha)^{-1}(i) \to
f^{-1}(i)$ is the restriction of $\alpha$, $i \in \bar s$. Grothendieck’s construction of the cooperad $\Su$ leads to the operadic category $\Per$ related to permutads which we discuss in Subsection \[Michael\_do\_Prahy.\].
-------------------------- ------------------ --------------------------
?
& & &
?
?
$f$ is iso if $e(f) = 0$ ? $f$ is iso if $e(f) = 0$
$\ttP$ is graded $\ttO$ is graded $\ttO$ is graded
-------------------------- ------------------ --------------------------
: How (op)fibrations interact with properties of operadic categories.[]{data-label="analogies"}
\
------------------------------------------------------------------------
The leftmost column of Table \[analogies\] lists properties required in the following sections. Its top four rows record results obtained in this section. The 5th and 6th rows easily follow from the uniqueness of lifts in discrete opfibrations, while the grading in the last row is given by formula (\[Dnes\_plynari\_ale\_ja\_jsem\_v\_Australii.\]) and does not require any additional assumptions on the functor $p : \ttO \to \ttP$.
\[Pujdu\_zitra\_na\_seminar?\] If $p : \ttO \to \ttP$ is a discrete opfibration and $\ttP$ fulfills the properties listed in the leftmost column of Table \[analogies\], then $\ttO$ shares the same properties. If $p : \ttO \to \ttP$ is a discrete operadic fibration, the situation in not so simple. One may however invoke the implication $\hbox {\rm {\UFB} \& {\WBU}}
\Longrightarrow \hbox {{\Rig}}$ of Lemma \[Podivam\_se\_do\_Paramaty??\] and conclude that if one ‘manually’ verifies $\UFB$ and “$f$ is iso if $e(f)=0$,” then $\ttO$ satisfies all the properties in the leftmost column also in the case of opfibrations.
The $+$-construction {#+construction}
====================
Let $\ttO$ be an operadic category. A [*construction*]{} over $\ttO$ is an operadic category $\ttO^+$ together with an equivalence $$\Oper\ttO \cong \Alg{{\sf 1}_{\ttO^+}}$$ of the category of $\ttO$-operads and the category of algebras over the terminal $\ttO^+$-operad ${\sf 1}_{\ttO^+}$.
\[+example\] It is simple to see that the operadic category $\RTr$ of rooted trees introduced in Example \[Zitra\_s\_Jarkou\_k\_Bartosovi.\] is a over $\Fin$. It is equally easy to check that the operadic category $\Deltaalg$ of finite ordinals and their order-preserving maps recalled in Example \[finset+deltaalg\] is over the operadic category $\tt 1$ with one object of cardinality $\bar{1}\in \Fin$ and one arrow. $\tt 1$-operads are monoids [@duodel Example 1.14] while ${\sf 1}_{\Deltaalg}$ is the terminal nonsymmetric (non-$\Sigma$) operad.
We will see in Example \[+econ\] that there might be several different constructions while Theorem \[Gr+\] below shows that there always exists at least one.
\[freeop\] For each operadic category $\ttO$ there exists a $\Bq(\Ob(\ttO))$-operad[^11] $\PO$ such that the category of $\PO$-algebras is isomorphic to the category of $\ttO$-operads.
Recall that a $\Bq(\frC)$-colored operad $\oP$ with the set of colors $\frC$ has components $$\colorop{\oP} (\Rada c1n;c), \ (\Rada c1n;c) \in \Bq(\frC),$$ consisting of operations with the input colors $\Rada c1n$ and the output color $c$. For each $c \in \frC$ one requires the unit $$e_c \in \colorop{\oP} ( c;c).$$
We define a $\Bq(\Ob(\ttO))$-colored operad $\PO$ of the proposition by specifying its generators and relations between them. The generators are of two types.
[*Type 1.*]{} There is a generator $$\Gamma_f \in \colorop{\PO}(S,\Rada F1k;T)$$ for each morphims $f:T
\to S$ in $\ttO$ with fibers $\Rada F1k$.
[*Type 2.*]{} There is a generator $$\Gamma_U \in \colorop\PO(\emptyset;U)$$ for every trivial (= chosen terminal) $U \in \ttO$. The generators are subject to the following relations.
[*Associativity.*]{} Consider the diagram $$\xymatrix{T \ar[rr]^f \ar[dr]_h &&S \ar[dl]^g
\\
&R&
}$$ of morphisms in $\ttO$ and denote by $\Rada G1k$, resp. $\Rada H1k$, the fibers of $g$ (resp. $h$), and by $f_i
: H_i \to G_i$, $1\leq i \leq k$, the induced maps. Let also $\rada{F_i^1}{F_i^{s_i}}$ be the fibers of $f_i$, $1\leq i \leq k$. In shorthand $$\Rada G1k \fib S \stackrel g\to R,\
\Rada H1k \fib T \stackrel h\to R \ \hbox { and }\
\rada{F_i^1}{F_i^{s_i}} \fib H_i \stackrel {f_i}\longrightarrow G_i, \
1 \leq i \leq k.$$ One then has the compositions $$\Gamma_f(\Gamma_g,\rada{e_{F^1_1}}{e_{F^{s_k}_k}}) \in
\colorop\PO(R,\Rada G1k,\rada{F^1_1}{F^{s_k}_k};T)$$ and $$\Gamma_h(e_R,\rada{\Gamma_{f_1}}{\Gamma_{f_k}} ) \in
\colorop\PO(R,\rada{G_1,\rada{F^1_1}{F_1^{s_1}}}{G_k,\rada{F^1_k}{F_k^{s_k}}};T).$$ We require that $$\tag{R1}
\Gamma_f(\Gamma_g,\rada{e_{F^1_1}}{e_{F^{s_k}_k}}) \cdot \sigma
=
\Gamma_h(e_R,\rada{\Gamma_{f_1}}{\Gamma_{f_k}} ),$$ where $\sigma$ is the permutation $$(R,\Rada G1k,\rada{F^1_1}{F^{s_k}_k})
\longmapsto
(R,\rada{G_1,\rada{F^1_1}{F_1^{s_1}}}{G_k,\rada{F^1_k}{F_k^{s_k}}}).$$
[*Unitality 1.*]{} Let $T \in \ttO$ has sources $\Rada U1n$, i.e. $\Rada U1n \fib T \stackrel {\id_T}\longrightarrow T $. We require $$\tag{R2}
\Gamma_{\id_T}(e_T,\rada{\Gamma_{U_1}}{\Gamma_{U_n}}) = e_T.$$
[*Unitality 2.*]{} Consider the unique map $T \fib T \stackrel{!_T}\to U$ to the chosen local trivial object. We ask $$\tag{R3}
\Gamma_{!_T}(\Gamma_U,e_T) = e_T.$$
A direct calculation shows that $\PO$-algebras are indeed $\ttO$-operads. Relation (R1) guarantees axiom (i) of Definition \[Jarca\_u\_mne\_prespala!\], relation (R2) axiom (ii) and relation (R3) axiom (iii).
\[Gr+\] For any operadic category $\ttO$ the domain of the Grothendieck construction $$\int_{\Bq(\Ob(\ttO))} \PO \longrightarrow \Bq(\Ob(\ttO))$$ of the operad $\PO$ constructed in Proposition \[freeop\] is a $+$-construction on $\ttO$.
It follows immediately from Propositions \[onevsO\] and \[freeop\].
\[+econ\] Let again $\tt 1$ be the operadic category with one object and one arrow. The operad $\oP({\tt 1}) $ is the operad of symmetric groups $\Ass$. This is a single-colored operad so its Grothendieck construction $\int_\Fin \Ass$ is an operadic category over $\Bq(\pi_0(1)) = \Fin$. Its objects are pairs $(\bar n,\sigma)$, $\bar n \in \Fin$, $\sigma\in \Sigma_n$, $n \geq 1$, and a morphism $(\bar n,\sigma) \to (\bar k,\tau)$ consists of a morphism $f:\bar n\to \bar k$ in $\Fin$ together with a sequence of permutations $\tau_1,\ldots,\tau_k$ such that $m_f(\tau; \tau_1,\ldots, \tau_k) = \sigma$, where $m_f$ is the operadic composition in $\Ass$ induced by $f$.
There is an inclusion of operadic categories $i:\Deltaalg\to \int_\Fin\Ass$ given by $i(\bar{n}) = (\bar n,1_n),$ with $1_n \in
\Sigma_n$ the identity permutation. It induces an isomorphisms of categories $$\Alg{{\sf 1}_{\Deltaalg}} \cong \Alg{{\sf 1}_{\int_\Fin
\hskip -.1em \Ass}}$$ which are both isomorphic to the category of monoids. Thus $\int_\Fin\Ass$ and $\Deltaalg$ are two non-isomorphic constructions over ${\tt 1}$.
The material below is not necessary for the rest of our paper but we decided to include it for future references and for clarification of the nature of constructions. Let $\SOp$ be the category of classical colored operads. Its objects are couples $(\frakC,\oP)$ formed by a set $\frakC$ and a $\Bq(\frakC)$-operad $\oP$. A morphism $(\frakC',\oP') \to (\frakC'',\oP'')$ is a couple $(\frakc,g)$ of a map $\frakc : \frakC' \to \frakC''$ of sets and of a morphism $g : \oP'
\to \frakc^* (\oP'')$ of $\Bq(\frakC')$-operads, where $$\frakc^* : \Oper {\Bq(\frakC'')} \longrightarrow \Oper {\Bq(\frakC')}$$ is the restriction along the color-change map $\frakc$. It is simple to see that the construction of the operad $\PO$ in Proposition \[freeop\] extends to a functor $$\oP:\OpCat \to \SOp$$
We are going to describe yet another functor $\EuScript{Q}:\OpCat \to \SOp$. The arity functor $Ar(\ttO) = \Ar:\ttO\to \Bq(\pi_0(\ttO))$ recalled in (\[Jarca\]) induces a pair $$\xymatrix@C=1em{\Oper \ttO \ar@/^{1.3em}/[rr]^{Ar(\ttO)_!} & \hskip 1.3em\perp &
\ar@/^{1.5em}/[ll]^{Ar(\ttO)^*} \hskip .5em \Oper {\Bq(\pi_0(\ttO))}
}$$ of adjoint functors between the categories of operads. We then define $$\EuScript{Q}(\ttO) := Ar(\ttO)_!({\sf 1}_{\ttO}) \in {\Bq(\pi_0(\ttO))}.$$ Observe that the operad $\mathcal{Q}(\ttO)$ can be characterised up to isomorphism by the ‘equation’ $$\label{Qeq}
\Alg {\EuScript{Q}(\ttO)}\cong
\Alg{{\sf 1}_{\ttO}}.$$ Indeed, for any $\pi_0(\ttO)$-colored collection $E$, Definition \[Jaruska\_ma\_chripecku\] gives the equation $$\Alg {\EuScript{Q}(\ttO)} = \Oper{{\Bq(\pi_0(\ttO))}}
\big(\EuScript{Q}(\ttO),\End_E^{\, \Bq(\pi_0(\ttO))}\big)$$ whose right hand side equals by the definition of $\EuScript{Q}(\ttO)$ $$\Oper{{\Bq(\pi_0(\ttO))}}
\big( Ar(\ttO)_!({\sf 1}_{\ttO}),\End_E^{\, \Bq(\pi_0(\ttO))}\big).$$ On the other hand $$\Alg{{\sf 1}_{\ttO}} =
\Oper \ttO \big({\sf 1}_{\ttO}, \Ar(\ttO)^*(\End^{\Bq(\frC)}_E)\big).$$ We see that (\[Qeq\]) is precisely the adjunction isomorphism.
Let us verify that the assignment $\ttO \mapsto \EuScript{Q}(\ttO)$ extends to a functor $\OpCat \to \SOp$. Indeed, for any operadic functor $f:\ttO\to\ttP$ the square $$\xymatrix@C = +4em{
\ttO \ar[r]^{f} \ar[d]^{\Ar(\ttO)}& \ttP\ar[d]^{\Ar(\ttP)}
\\
\Bq(\pi_0(\ttO))\ar[r]^{\Bq(\pi_0(f))} & \Bq(\pi_0(\ttP))
}$$ commutes and generates a commutative, up to canonical isomorphism, diagram of adjoint functors $$\xymatrix@C = +4em@R=4em{
\Oper{\ttO} \ar@/^{1.3em}/[r]^{f_!}
\ar@/^{1em}/[d]^{\Ar(\ttO)_!}& \ar@/^{1.3em}/[l]_{f^*}
\Oper{\ttP}\ar@/^{1em}/[d]^{\Ar(\ttP)_!}
\\
\ar@/^{1em}/[u]^{\Ar(\ttO)^*}
\Oper{\Bq(\pi_0(\ttO))}
\ar@/^{1.3em}/[r]^{\Bq(\pi_0(f))_!}
& \Oper{\Bq(\pi_0(\ttP))} \ar@/^{1.3em}/[l]_{\Bq(\pi_0(f))^*}
\ar@/^{1em}/[u]^{\Ar(\ttP)^*}
}$$
Notice also that the right adjoint functor $$\Bq(\pi_0(f))^* : \Oper{\Bq(\pi_0(\ttP))} \longrightarrow
\Oper{\Bq(\pi_0(\ttO))}$$ is just the restriction $\pi_0(f)^*$ along the color-change map $\pi_0(f)$ induced by $f$. Applying $f^*$ to the unit ${\sf 1}_{\ttP}\to \big(\Ar(\ttP)^* \circ
\Ar(\ttP)_!\big)({\sf 1}_{\ttP})$ of the adjunction $\Ar(\ttP)_! \dashv
\Ar(\ttP)^*$ gives the map $${\sf 1}_{\ttO} = f^*({\sf 1}_{\ttP})
\longrightarrow \big(f^* \circ \Ar(\ttP)^*
\circ Ar(\ttP)_!\big)({\sf 1}_{\ttP}) =
\big(\Ar(\ttO)^* \circ \pi_0(f)^* \circ \Ar(\ttP)_!\big)({\sf 1}_{\ttP})$$ whose mate is a morphism $$g:
\EuScript{Q}(\ttO) =\Ar(\ttO)_!({\sf 1}_{\ttO}) \longrightarrow
\pi_0(f)^* \big(\Ar(\ttP)_!({\sf 1}_{\ttP})\big) =
\pi_0(f)^*(\EuScript{Q}(\ttP)).$$ We put $\EuScript{Q}(f) := (\pi_0(f)^*, g)$.
A is a functor $+: \OpCat\to \OpCat$, $\ttO\mapsto \ttO^+$, equipped with a natural isomorphism $$\PO \cong \EuScript{Q}(\ttO^+)$$ of functors $\OpCat \to \SOp$.
Theorem \[Gr+\] shows that the functorial constructions exist. It turns out that there is a natural notion of a morphism between functorial constructions so they form a category. In a follow-up to this work we show that the construction provided by Theorem \[Gr+\] is a terminal object in this category, though it is not the most economical one. For example, $\Deltaalg$ discussed in Example \[+example\] is an instance of a smaller functorial construction which will be described in our future work.
The terminal construction of Theorem \[Gr+\] was inspired by an analogous construction for polynomial monads due to Baez and Dolan, described in [@KJBM]. It associates to a polynomial monad $P$ another polynomial monad $P^+$ such that there is an equivalence of categories $$\label{BD+}
\Alg{P^+} \cong {\tt PolyMon}/P,$$ where ${\tt PolyMon}/P$ is the slice category of polynomial monads and their cartesian morphisms over $P$. Notice that there cannot be many constructions in the polynomial monad setting, as equation (\[BD+\]) determines $P^+$ up to canonical isomorphism. We will show in a future work that there is a comparison functor between polynomial monads and operadic categories under which the Baez-Dolan construction corresponds to the terminal construction of Theorem \[Gr+\].
Markl’s operads {#section-markl}
===============
The aim of this section is to introduce Markl’s operads and their algebras in the context of operadic categories, and formulate assumptions under which these notions agree with the standard ones introduced in [@duodel]. As in Section \[letam\_205\], $\ttO$ will be a factorizable operadic category such that all are invertible, and the blow up and unique fiber axioms are fulfilled, i.e. $$\hbox{\rm \Fac\ \& \SBU\ \& \QBI\ \& \UFB}.$$ We also assume that $\ttO$ is graded, with the property that a morphism $f$ is an isomorphisms if and only if $e(f) =
0$.[^12] In other words, we conclude that $\ttO$ satisfies all the properties listed in the leftmost column of Table \[analogies\] except . Denoting by $\Iso$ the subcategory of $\ttO$ consisting of all isomorphisms we therefore have $$\Iso = \{ f : S \to T;\ e(f) = 0\}
= \{ f : S \to T;\ e(F) = 0 \hbox { for each fiber $F$ of $f$} \}.$$ Another consequence of our assumptions is that $T \in \ttO$ is local terminal if and only if . Operads in this section will live in a cocomplete strict symmetric monoidal category $\ttV$ with a strict monoidal unit $\bfk$ and the symmetry $\tau$.
\[markl\] A [*Markl’s $\ttO$-operad*]{} is a $\ttV$-presheaf $M: \Iso^{\rm op} \to \ttV$ equipped, for each elementary morphism $F\fib T \stackrel\phi\to S$ as in Definition \[plysacci\_postacci\], with the ‘circle product’ $$\label{ten_prelet_jsem_podelal}
\circ_{\phi}: M(S)\otimes M(F)\to M(T).$$ These operations must satisfy the following set of axioms.
- Let $T\stackrel{(\phi,j)}{\longrightarrow} S
\stackrel{(\psi,i)}{\longrightarrow} P$ be elementary morphisms such that $|\psi|(j) = i$ and $\xi: T \to P$ is the composite $\psi \circ \phi$. Then the diagram $$%\label{vymena}
\xymatrix@R = 1em{& { M(P)\otimes M(\xi^{-1}(j))}
\ar[dr]^(.65){\circ_\xi} &
\\
M(P)\otimes M(\psi^{-1}(i))\otimes M(\phi^{-1}(j))
\ar[dr]_{\circ_\psi} \ar[ur]^{\id \ot \circ_{\phi_i}} && M(T)
\\
&{ M(S)\otimes M(\phi^{-1}(j))}\ar[ur]_(.65){\circ_\phi}&
}$$ commutes.
- Let us consider the diagram $$\label{den_pred_Silvestrem_jsem_nachlazeny}
\xymatrix@R = 1em@C=4em{& {P'} \ar[dr]^{(\psi',\, i)} &
\\
T\ar[dr]^{(\phi'',\, l)} \ar[ur]^{(\phi',\, j)} && S
\\
&{P''}\ar[ur]^{(\psi'',\, k)}&
}$$ of elementary morphisms with disjoint fibers as in Corollary \[move\]. Then the diagram $$\xymatrix@R = 2.5em@C=4em{
M(S) \ot M(G') \ot M(F')\ar[r]^(.55){\circ_{\psi'} \ot \id} &
M(P') \ot M(F')\ar[d]^(.5){\circ_{\phi'}}
\\
M(S) \ot M(F'') \ot M(G'')
\ar[u]^{\id \ot (\sigma''^{-1})^* \ot \sigma'^*}
& M(T)
\\
\ar[u]^{\id \ot \tau}
M(S) \ot M(G'') \ot M(F'')\ar[r]^(.55){\circ_{\psi''} \ot \id} &
M(P'') \ot M(F'') \ar[u]_(.5){\circ_{\phi''}}
}$$ in which $
F' \fib T \stackrel{\phi'}\to P',\
F'' \fib T \stackrel{\phi''}\to P',\
G' \fib T \stackrel{\psi'}\to S \hbox { and }
G'' \fib T \stackrel{\psi''}\to S,
$ and the maps $({\sigma''^{-1}})^*$ resp. $\sigma'^*$ are induced by (\[za\_tyden\_poletim\_do\_Prahy\]).
- For every commutative diagram $$\xymatrix@C=4em@R=1.4em{T'\ar[r]_\cong^\omega \ar[d]^{\phi'}
& T'' \ar[d]^{\phi''}
\\
S'\ar[r]^\sigma_\sim & S''
}$$ where $\omega$ is an isomorphisms, $\sigma$ a , and $F' \fib_i T' \stackrel{\phi'}\to S'$, $F'' \fib_j
T'' \stackrel{\phi'}\to S''$, the diagram $$\label{posledni_nedele_v_Sydney}
\xymatrix@C=4em{\ar[d]_{\omega_{(i,j)}^* \ot \sigma^*}
M(F'') \ot M(S'')
\ar[r]^(.65){\circ_{\phi''}}& M(T'') \ar[d]^{\omega^*}_\cong
\\
M(F') \ot
M(S')\ar[r]_(.65){\circ_{\phi'}}& M(T')
}$$ in which $\omega_{(i,j)} : F' \to F''$ is the induced map (\[zitra\_na\_prohlidku\_k\_doktoru\_Reichovi\]) of fibers, commutes.
Markl’s operad $M$ is [*unital*]{} if one is given, for each trivial $U$, a map $\eta_U
:\bfk \to M(U)$ such that the diagram $$\label{Holter_se_blizi.}
\xymatrix{M(U) \ot M(T) \ar[r]^(.6){\circ_!} & M(T)
\\
\ar[u]^{\eta_U \ot \id}
\bfk \ot M(T) \ar@{=}[r]^(.58)\cong & M(T) \ar@{=}[u]
}$$ in which $T$ is such that $e(T) \geq 1$ and $T \fib T \stackrel!\to U$ the unique map, commutes.
Let, as before, $\LT$ be the operadic subcategory of $\ttO$ consisting of its local terminal objects. Denote by $\term_\Term : \LT \to \ttV$ the constant functor, i.e. the functor such that $\term_\Term (u) = \bfk$ for each local terminal $u\in \ttO$. A unital Markl’s operad $M$ is [*extended unital*]{} if the unit maps $\eta_U :\bfk \to M(U)$ extend to a transformation $$\label{za_chvili_pujdu_behat}
\eta : \term_\Term \to \iota^* M$$ from $ \term_\Term$ to the restriction of $M$ along the inclusion $\iota : \LT \hookrightarrow \ttO$.
Transformation (\[za\_chvili\_pujdu\_behat\]) is the same as a family of maps $\eta_u : \bfk \to M(u)$ given for each local terminal $u \in
\ttO$ such that the diagram $$\label{Je_vedro.}
\xymatrix@R=1.2em{M(u) \ar[r]^(.5){!^*} & M(v)
\\
\bfk \ar@{=}[r] \ar[u]^{\eta_u } &
\bfk
\ar[u]_{\eta_v}
}$$ commutes for each (unique) map $! : v \to u$ of local terminal objects.
Transformation (\[za\_chvili\_pujdu\_behat\]) induces, for each $T$ with $e(T) \geq 1$ and $F \fib T \stackrel!\to u$, a map $\vartheta(T,u) : M(F) \to M(T)$ defined by the diagram $$\label{proc_ty_lidi_musej_porad_hlucet}
\xymatrix{M(u) \ot M(F) \ar[r]^(.6){\circ_!} & M(T)
\\
\ar[u]^{\eta_u \ot \id}
\bfk \ot M(F) \ar@{=}[r]^\cong &\ M(F). \ar[u]_{\vartheta(T,u)}
}$$ The extended unitality implies a generalization of Axiom (iii) of Markl’s operads that postulates for each commutative diagram $$\label{moc_se_mi_na_prochazku_nechce}
\xymatrix@C=4em{T'\ar[r]^\omega_\cong \ar[d]^{\phi'}\ar[rd]^{\phi}
& T'' \ar[d]^{\phi''}
\\
S'\ar[r]^\sigma_\cong & S''
}$$ where the horizontal maps are isomorphisms and the vertical maps are elementary, with $F' \fib_i T' \stackrel{\phi'}\to S'$, $F'' \fib_j
T'' \stackrel{\phi''}\to S''$, the commutativity of the diagram $$\label{ve_ctvrtek_letim_do_Prahy}
\xymatrix@C=4em
{M(F) \ot M(S'')&M(F'') \ot M(S'') \ar[l]_{\omega^*_j \ot \id}^\cong
\ar[r]^(.65){\circ_{\phi''}}& M(T'') \ar[d]^{\omega^*}_\cong
\\
\ar[u]^{\vartheta(F,\inv{\sigma}(j)) \ot \id}
M(F') \ot M(S'')\ar[r]^{\id \ot \sigma^*}_\cong &M(F') \ot
M(S')\ar[r]_(.65){\circ_{\phi'}}& M(T')
}$$ in which $F := \inv{\phi}(j)$ and $\omega_j : F \to F''$ is the induced map of fibers. Notice that if $\sigma$ is a , (\[ve\_ctvrtek\_letim\_do\_Prahy\]) implies (\[posledni\_nedele\_v\_Sydney\]).
Markl’s operad $M$ is [*strictly extended unital*]{} if all the maps $\vartheta(T,u)$ in (\[proc\_ty\_lidi\_musej\_porad\_hlucet\]) are identities. It is [*$1$-connected*]{} if (\[za\_chvili\_pujdu\_behat\]) is an isomorphism.
Let us introduce a similar terminology for ‘standard’ $\ttO$-operads. In this framework, $\term_\Term$ will denote the constant $\Term$-operad. An $\ttO$-operad $P$ is [*extended unital*]{} if one is given a morphism $$\label{7_dni_do_odletu_ze_Sydney}
\eta : \term_\Term \to \iota^* P$$ of (unital) $\Term$-operads. As for Markl’s operads, one has an obvious analog of diagram (\[proc\_ty\_lidi\_musej\_porad\_hlucet\]). Also the strict extended unitality and $1$-connectedness for $\ttO$-operads is defined analogously. The main statement of this section reads:
\[Mcat\] There is a natural forgetful functor from the category of strictly extended unital $\ttO$-operads to the category of strictly extended unital Markl’s operads which restricts to an isomorphism of the subcategories of $1$-connected operads.
Constant-free May’s operads recalled in the introduction are operads in the operadic category $\Surj$ of non-empty finite sets and their epimorphisms. Let us analyze the meaning of the above definitions and results in this particular case. With respect to the canonical grading, elementary morphisms in $\Fin$ are precisely order-preserving epimorphisms $$\pi(m,i,n) : [m+n-1] \epi [m],\ m \geq 1,\ n \geq 2,$$ uniquely determined by the property that $$\label{zitra_do_Rataj_slavit_Silvestra}
|\inv{\pi(m,i,n)}(j)|
=
\begin{cases}
\hbox{$1$ if $j \not= i$, and}
\\
\hbox {$n$ if $j=i$.}
\end{cases}$$ Since $[1]$ is the only local terminal object of $\Fin$, the strict extended unitality is the same as the ordinary one and all are . An $\Iso$-presheaf is a collection $\{M(n)\}_{n\geq 1}$ of $\Sigma_n$-modules, while elementary maps (\[zitra\_do\_Rataj\_slavit\_Silvestra\]) induce operations $$\circ_i := \circ_{\pi(m,i,n)} : M(m) \ot M(n) \to M(m+n-1)$$ which satisfy the standard axioms listed e.g. in [@markl:zebrulka Definition 1.1]. Theorem \[Mcat\] in this case states the well-known fact that the category of unital May’s operads with $P(1) = \bfk$ is isomorphic to the category of unital Markl’s operads with $M(1) = \bfk$.
Let $P$ be a strictly extended unital $\ttO$-operad with structure operations $\gamma_f$. If $\omega : T' \to T''$ is an isomorphism, we define $\omega^* : P(T'') \to P(T')$ by the diagram $$\label{dnes_jsem_byl_v_Eastwoodu}
\xymatrix{P(T'') \ot P(\omega) \ar[r]^(.6){\gamma_\omega}
& P(T')
\\
\ar[u]_{\id \ot \eta_\omega}
P(T'') \ot \bfk \ar@{=}[r]^(.58)\cong& P(T'') \ar[u]_{\omega^*}
}$$ in which $P(\omega)$ denotes the product $P(u_1) \ot \cdots \ot P(u_s)$ over the fibers $\Rada u1s$ of $\omega$ and, likewise, $\eta_\omega :=
\eta_{u_1} \ot \cdots \ot \eta_{u_s}$. It is simple to show that this construction is functorial, making $P$ an $\Iso$-presheaf in $\ttV$. In particular, $\omega^*$ is an isomorphism. For an elementary $F \fib_i T
\stackrel\phi\to S$ we define $\circ_\phi : P(S)\ot P(F) \to P(T)$ by the commutativity of the diagram $$\xymatrix{P(S) \ot P(U_1) \!\ot\! \cdots \!\ot\! P(U_{i-1}) \!\ot\! P(F) \!\ot\!
P(U_{i+1}) \!\ot\! \cdots \!\ot\! P(U_{|S|})\ar[r]^(.84){\gamma_\phi} & P(T)
\\
P(S) \ot
\bfk^{\ot (i-1)} \ot P(F) \ot \bfk^{\ot (|T|-i)} \ar[u] \ar@{=}[r]^(.6)\cong
&P(S) \ot P(F) \ar[u]_{\circ_\phi}
}$$ in which the left vertical map is induced by the unit morphism (\[7\_dni\_do\_odletu\_ze\_Sydney\]) of $P$ and the identity automorphism of $P(F)$. We claim that the $\Iso$-presheaf $P$ with operations $\circ_\phi$ defined above is a Markl’s operad.
It is simple to check that these $\circ_\phi$’s satisfy the associativities (i) and (ii) of a Markl’s operad. To prove Axiom (iii), consider diagram (\[moc\_se\_mi\_na\_prochazku\_nechce\]) and invoke Axiom (i) of an operad in an operadic category, see Section \[a0\], once for $\phi =
\sigma\phi'$ and once for $\phi = \phi''\omega$ in place of $h =fg$. We will get two commutative squares sharing the edge $\gamma_\phi$. Putting them side-by-side produces the central hexagon in the diagram in Figure \[Bozi\_oko\]. The remaining arrows of this diagram are constructed using the $\Iso$-presheaf structure of $P$ and the extended units.
$$\xymatrix@R = 1.41em@C=-3em{
&\boxed{P(\phi'') \ot P(S'')} \ar@{=}[rrr] \ar[d]^\cong_{\bigotimes_k \omega_i^* \ot \id}
&&&&\ \ \bfk \ot P(\phi'') \ot P(S'')
\ar[d]^{\eta_\omega \ot \id \ot \id}
&
\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
&\boxed{P(\phi) \ot P(S'')} \ar@{-->}[ddrrr]^{\gamma_\phi}
&&\rule{9em}{0em}&&\ar[llll]
P(\omega) \ot P(\phi'') \ot P(S'') \ar[d]_{\id \ot \gamma_{\phi''}}&
\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
&\bigotimes_k P(\phi_k) \ot P(\sigma)\ar[u]^{\bigotimes_k
\gamma_{\phi'_k} \ot \id}\ar[rd]^{\id \ot \gamma_\sigma} \ot P(S'')
&&&& \boxed{P(\omega) \ot P(T'')}\ar[ld]_{\gamma_\omega}&
\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%
P(\phi')\ot \bfk \ot P(S'') \ar[ur]^(.3){\id \ot \eta_\sigma \ot \id}
&& \boxed{P(\phi') \ot P(S')}
\ar [rr]^{\gamma_{\phi'}} && \boxed{P(T')} && \bfk \ot P(T'')
\ar[lu]_{\eta_\omega \ot \id}
\\
%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\ar@{=}[u]\boxed{P(\phi') \ot P(S'')} \ar[rru]_\cong^{\id \ot \sigma^*} &&&&&&
P(T'')\ar@{=}[u]\ar[llu]_{\omega^*}^\cong
}$$
The boxed terms in Figure \[Bozi\_oko\] form the internal hexagon in Figure \[v\_Ratajich\_na\_Stepana\]. The commutativity of the outer hexagon follows from the commutativity of the internal one. We recognize in it diagram (\[ve\_ctvrtek\_letim\_do\_Prahy\]) with $P$ in place of $A$. Since (\[ve\_ctvrtek\_letim\_do\_Prahy\]) implies (\[posledni\_nedele\_v\_Sydney\]) for $\sigma$ a , Axiom (iii) is verified. To check that the strict extended unit (\[7\_dni\_do\_odletu\_ze\_Sydney\]) is also the one for $P$ considered as a Markl’s operad is simple.
$$\xymatrix@R=1.3em@C=1em{
P(F) \ot P(S'')
\ar[rd]&&
P(F') \ot P(S'')
\ar[ll]^\cong_{\omega_j^* \ot \id} \ar[rr]^(.55){\circ_{\phi''}}\ar[d] && \ar[ld] P(T'') \ar[ddd]_\cong^{\omega^*}
\\
&P(\phi) \ot P(S'')&
P(\phi') \ot P(S'')\ar[l]_\cong \ar[r]
& P(\omega)\ot P (T'') \ar[d]&
\\
&P(\phi') \ot P(S'') \ar[r]^\cong \ar[u]&\ar[r]
P(\phi') \ot P(S'')
& P (T')&
\\
P(F') \ot P(S'') \ar[uuu]^{\vartheta(F,\inv{\sigma}(j)) \ot \id}
\ar[ur] \ar[rr]_\cong^{\id \ot \sigma^*}
&&
P(F') \ot P(S') \ar[rr]^(.55){\circ_{\phi'}} \ar[u] && P(T')\ar@{=}[lu]
}$$
Conversely, let $M$ be a Markl’s operad. We are going to define, for each $f: S \to T$ with fibers $\Rada F1s$, the structure map $$\label{po_navratu_z_Prahy}
\gamma_f:
M(T) \ot M(F) \longrightarrow M(S)$$ where, as several times before, $M(F)$ denotes $M(F_1) \ot \cdots \ot M(F_s)$. If $f$ is an , all its fibers are terminal, so $M(F) \cong \bfk$ by the strict extended unitality and the $1$-connectivity of $M$. In this case we define $\gamma_f$ as the composition $$\label{vecer_Deminka?}
M(T) \ot M(F) \cong M(T)
\stackrel{f^*}\longrightarrow
M(S)$$ that uses the $\Iso$-presheaf structure of $M$.
Assume now that $f \in \DO$ and that all terminal fibers of $f$ are trivial. If $f$ is an it must be the identity by Corollary \[zase\_mam\_nejaky\_tik\]. If it is not the case, at least one fiber of $f$ has grade $\geq 1$ and we decompose $f$, using the blow up axiom, into a chain of elementary morphisms. The operation $\gamma_f$ will then be defined as the composition of $\circ$-operations corresponding to these elementary morphisms. Let us make this procedure more precise.
To understand the situation better, consider two elementary morphisms $\phi$, $\psi$ with [$(i,j)$-disjoint]{} fibers as in Lemma \[l7\] and Figure \[Jarka\_ma\_novy\_pocitac.\], and their composition $\xi = \psi(\phi)$. Notice that by $M(\xi) \cong M(\inv\xi(i)) \ot M(\inv\xi(k))$ by the strict extended unitality and the $1$-connectivity of $M$. In this particular case we define $\gamma_\xi$ by the commutativity of the diagram $$\xymatrix{M(P) \ot M(\inv\xi(i)) \ot M(\inv\xi(k))
\ar[r]^(.65)\cong \ar[d]_{\id \ot (\phi^*_i)^{-1} \ot \id}
& M(P) \ot M(\xi)\ar[dd]^{\gamma_\xi}
\\
\ar[d]_{\id \ot \circ_\psi}
M(P) \ot M(\inv\psi(i)) \ot M(\inv\phi(k))&
\\
M(S) \ot M(\inv\phi(k))\ar[r]^(.6){\circ_\phi}& M(T)
}$$ or, in shorthand, by $\gamma_\xi := \circ_\phi(\id \ot \circ_\psi)$.
Now take $f: S \to T \in \DO$ whose fibers of grade $\geq 1$ are $\Rada F1k$ and the remaining fibers are trivial. Using the blow up axiom we factorize $f$ into a chain $$\label{Z-142}
S = S_1 \stackrel{\phi_1}\longrightarrow
S_2 \stackrel{\phi_2}\longrightarrow \cdots
\stackrel{\phi_k}\longrightarrow S_k = T$$ in which each $\phi_i$ is elementary with the unique nontrivial fiber $F_i$, $1 \leq i \leq k$; we leave the details how to obtain such a factorization to the reader. We then define $$\gamma_f := \circ_{\phi_1} ( \circ_{\phi_2} \ot \id) \cdots
( \circ_{\phi_k} \ot \id^{\ot(k-1)}):
M(T) \ot M(F_k) \ot \cdots \ot M(F_1) \longrightarrow M(S).$$
If $f: S \to T$ is a general morphism in $\ttO$, we use Lemma \[Podival\_jsem\_se\_do\_Paramaty!\] to write it as $f: S \stackrel\omega\to X \stackrel\psi\to T$ with an isomorphism $\omega$ and $\psi \in \DO$ whose all terminal fibers are trivial. Notice that, due to the strict extended unitality and $1$-connectivity, $M(\psi) \cong M(f)$. We then define $\gamma_f$ by the commutativity of the diagram $$\xymatrix{M(\psi) \ot M(T) \ar[r]^\cong \ar[d]^{\gamma_\psi} &
M(f) \ot M(T)\ar[d]_{\gamma_f}
\\
M(X)\ar[r]^{\omega^*}&\ M(S).&
}$$ The extended units are given by the extended units of $M$ in the obvious way.
Our definition of the $\gamma_f$-operations does not depend on the choices – the commutativity of (\[ve\_ctvrtek\_letim\_do\_Prahy\]) that holds for extended unital Markl’s operads guarantees the independence on the factorization $f =
\psi \circ \omega$, while the commutativity of (\[den\_pred\_Silvestrem\_jsem\_nachlazeny\]) the independence on the choice of the decomposition (\[Z-142\]). We leave the tedious but straightforward verification that $M$ with the above structure operations forms an strictly extended unital $\ttO$-operad to the reader.
We close this section by adapting algebras over operads [@duodel Definition 1.20] to the realm of Markl’s operads. Recall that the [*$i$th source*]{} $s_i(T)$ of an object $T \in
\ttO$ is, for $i \in |T|$, defined as the $i$th fiber of the identity automorphism of $T$, i.e. as $\id_T^{-1}(i)$. We denote by $s(T)$ the set of all sources of $T$. For an object $T \in \ttO$ we denote by $\pi_0(T) \in
\pi_0(\ttO)$ the connected component to which $T$ belongs. Similarly, for a subset $X$ of objects of $\ttO$, $$\pi_0(X) := \{\pi_0(T) \ | \ T \in X\} \subset \pi_0(\ttO).$$ Finally, $U_c$ will denote the chosen local terminal object of a component $c \in \pi_0(\ttO)$.
\[Zapomel\_jsem\_si\_pocitac\_v\_Koline\_ja\_hlupak\] An [*algebra*]{} over a $1$-connected Markl’s operad $M$ in a symmetric monoidal category $\ttV$ is a collection $A = \{A_c \ | \ c \in \pi_0(\ttO)\}$ of objects of $\ttV$ together with structure maps $$\label{Ben_Ward_in_Prague}
\big\{
\alpha_T:
M(T) \ot \hskip -.5em
\bigotimes_{c \in \pi_0(s(T))} A_c \longrightarrow
A_{\pi_0(T)}\big\}_{T \in \ttO} \ .$$ These operations are required to satisfy the following axioms.
- Unitality: for each component $c \in \pi_0(\ttO)$ the diagram $$\xymatrix{M(U_c) \ot A_c \ar[r]^(.65){\alpha_{U_c}} & A_c
\\
\bfk \ot A_c\ar[r]^\cong \ar[u]^{\eta_{U_c}} & A_c \ar@{=}[u]
}$$ commutes.
- Equivariance: let $f : S\to T$ be an with fibers $\Rada u1s$. For $1 \leq
i \leq s$ denote $c_i := \pi_0(s_i(S))$ and $d_i :=
\pi_0(s_i(T))$. Then the diagram $$\xymatrix@C=0em{
M(T) \ot A_{c_1} \ot \cdots \ot A_{c_s} \ar[rr]^(.45)\cong
\ar[d]_{f^* \ot \id^{\ot s}} &&
M(T) \ot \bfk \ot A_{c_1} \ot \cdots \ot \bfk \ot A_{c_s}
\ar[d]^{\id \ot \eta_{u_1} \ot \cdots \ot \eta_{u_s}}
\\
M(S) \ot A_{c_1} \ot \cdots \ot A_{c_s}\ar[d]_{\alpha_S} &&
M(T) \ot M(u_1) \ot A_{c_1} \ot \cdots \ot M(u_s) \ot A_{c_s}
\ar[d]^{\id \ot \alpha_{u_1} \ot \cdots \ot \alpha_{u_s}}
\\
A_{\pi_0(S)} \ar@{=}[r]
&A_{\pi_0(T)}& M(T) \ot A_{d_1} \ot \cdots \ot A_{d_s}\ar[l]_(.6){\alpha_T}
}$$ commutes.
- Associativity: for an elementary map $F \fib_i S \stackrel\phi\to T$, the diagram $$\xymatrix{
M(S) \ot A_{c_1} \ot \cdots A_{c_{i-1}} \ot
A_{c_{i}}
\ot \cdots \ot A_{c_{t+s-1}} \ar[r]^(.78){\alpha_S}
& A_{\pi_0(S)}\ar@{=}[ddd]
\\
\ar[u]^{\circ_\phi \ot \id^{\ot t+s-1}}
\ar[d]_{\id \ot \tau \ot \id^{\ot t-i}}
M(T) \ot M(F) \ot A_{c_1} \ot \cdots\ot A_{c_{i-1}} \ot
A_{c_{i}}
\ot \cdots \ot A_{c_{t+s-1}} &
\\
\ar[d]_{\id \ot \id^{\ot i} \ot \alpha_F \ot \id^{\ot t-i}}
M(T)\ot A_{c_1} \ot \cdots \ot A_{c_{i-1}} \ot M(F) \ot
A_{c_{i}}
\ot \cdots \ot A_{c_{t+s-1}}
&
\\
M(T)\ot A_{c_1} \ot \cdots A_{c_{i-1}}\ot
A_{\pi_0(F)}
\ot \cdots \ot A_{c_{t+s-1}}\ar[r]^(.78){\alpha_T}
& A_{\pi_0(T)}
}$$ where $s = |S|$, $t = |T|$, $c_j := \pi_0(s_j)$ for $1\leq j \leq
s+t-1$ and $$\tau : M(F) \ot A_{c_1} \ot \cdots\ot A_{c_{i-1}} \longrightarrow
A_{c_1} \ot \cdots\ot A_{c_{i-1}} \ot M(F)$$ the commutativity constraint in $\ttV$, commutes.
Notice that in the situation of item (ii) of Definition \[Zapomel\_jsem\_si\_pocitac\_v\_Koline\_ja\_hlupak\], $s_i(S) = s(u_i)$, $\pi_0(s_i(T)) = \pi_0(u_i)$ and $\pi_0(S) =
\pi_0(T)$. Likewise in (iii), $$\label{za_necely_tyden_do_Bari}
\pi_0(s_j(T)) =
\begin{cases}
\pi_0(s_{\inv{|\phi|}(j)}(S)) & \hbox { if $j \not= i$, and}
\\
\pi_0(F) & \hbox { otherwise.}
\end{cases}$$
The category of algebras of a strictly extended unital $1$-connected Markl’s operad $M$ is isomorphic to the category of algebras of the corresponding operad $P$.
An exercise in the axioms of operads and their algebras.
Let us close this section by an useful reformulation of Definition \[Zapomel\_jsem\_si\_pocitac\_v\_Koline\_ja\_hlupak\] for algebras in the category of graded vector spaces and related facts.
\[vcera\_s\_Mikesem\_na\_Jazz\_Bluffers\] An [*algebra*]{} over a $1$-connected Markl’s operad $M$ in the category $\Vect$ of graded $\bfk$-vector spaces is a collection $A = \{A_c \ | \ c \in \pi_0(\ttO)\}$ together with structure maps $$\nonumber
M(T) \ot \hskip -.5em
\bigotimes_{c \in \pi_0(s(T))} A_c \ni x \ot a_1 \ot \cdots \ot a_s
\longmapsto x(\Rada a1s) \in A_{\pi_0(T)}$$ given for each $T \in \ttO$. These operations are required to satisfy the following axioms.
- Unitality: for a local terminal $u$, $1 \in \bfk \cong M(u)$ and $a \in
A_{\pi_0(s(u))}$ denote $u a : = 1(a)$. Then $U a = a$ for $U$ a chosen local terminal object.
- Equivariance: for an $f : S\to T$ with fibers $\Rada u1s$ and $x \in M(T)$, $$f^*(x)(\Rada a1s) = x(\rada{u_1a_1}{u_sa_s}).$$
- Associativity: for an elementary map $F \fib_i S \stackrel\phi\to T$, $x\in M(T)$ and $y \in M(F)$, $$\circ_\phi(x,y)(\rada{a_1}{a_{i-1}},a_{i},\ldots,
{a_{t+s-1}})
=
(-1)^\varepsilon \cdot
x(\rada{a_1}{a_{i-1}},y(a_{i},\ldots),\ldots
{a_{t+s-1}}),$$ where $\varepsilon := {|y|(|a_1| + \cdots + |a_{i-1}|)}$, $s = |S|$ and $t = |T|$.
\[coboundary\] Following the terminology of [@getzler-kapranov:CompM98] we call a Markl’s operad $M$ in $\Vect$ such that $M(T)$ is for each $T$ a $1$-dimensional vector space a [*cocycle*]{}. An important cocycle is the operad ${\sf 1}_\ttO$ such that ${\sf
1}_\ttO(T) := \bfk$ for each $T \in \ttO$, with all structure operations the identities. We will call, slightly imprecisely, ${\sf 1}_\ttO$ the [ *terminal*]{} $\ttO$-operad since it is the linearization of the terminal $\ttO$-operad in the Cartesian monoidal category of sets.
Less trivial cocycles can be constructed as follows. We say that a graded vector space $W$ is [*invertible*]{} if $W\ot
W^{-1} \cong \bfk$ for some $W^{-1} \in \Vect$. This clearly means that $W$ is an iterated (de)suspension of the ground field $\bfk$. Suppose we are given a map $\fl : \pi_0(\ttO) \to \Vect$ that assigns to each $c \in \pi_0(\ttO)$ an invertible graded vector space $\fl(c)$. With the notation used in (\[Ben\_Ward\_in\_Prague\]) we introduce the cocycle $\fdl$ by $$\fdl(T) := \fl(\pi_0(T)) \ot \bigotimes_{c \in \pi_0(s(T))} \inv{\fl}(c)$$ with the trivial action of $\Iso$. To define, for $F \fib_i S \stackrel\phi\to T$, the structure operations $$\circ_\phi : \fdl(F) \ot \fdl(T) \to \fdl(S)$$ we need to specify a map $$\fl(\pi_0(F)) \ot \bigotimes_{c \in \pi_0(s(F))} \inv{\fl}(c) \ot
\fl(\pi_0(T)) \ot \bigotimes_{c \in \pi_0(s(T))} \inv{\fl}(c) \longrightarrow
\fl(\pi_0(S)) \ot \bigotimes_{c \in \pi_0(s(S))} \inv{\fl}(c).$$ To do so, we notice that $$\pi_0(s(F)) \sqcup \pi_0(s(T)) = \pi_0(s(S)) \sqcup \{ \pi_0(s_i(T))
\},\
\pi_0(S) = \pi_0(T) \ \hbox { and } \ \pi_0(F) = \pi_0(s_i(T)),$$ cf. (\[za\_necely\_tyden\_do\_Bari\]). Keeping this in mind, the structure operation $\circ_\phi$ is defined as the canonical isomorphism $\fdl(F) \ot \fdl(T) \cong \fdl(S)$.
Cocycles of the above form are called [*coboundaries*]{}. Notice that ${\sf 1}_\ttO = \fdl(T)$ with $\fl$ the constant function such that $\fl(c) := \bfk$ for each $c \in \pi_0(s(T))$.
Markl’s operads in $\Vect$ form a symmetric monoidal category, with the monoidal structure given by the level-wise tensor product and ${\sf 1}_\ttO$ the monoidal unit. As an exercise to Definition \[vcera\_s\_Mikesem\_na\_Jazz\_Bluffers\] we recommend to prove the following very useful
\[za\_chvili\_zavolam\_Jarusce\] The categories of $(M \ot \fdl)$-algebras and of $M$-algebras in $\Vect$ are isomorphic. More precisely, there is a natural one-to-one correspondence between – $M$-algebras with the underlying collection $A = \{A_c \ | \ c \in \pi_0(\ttO)\}$, and – $(M\ot \fdl)$-algebras with the underlying collection $A = \{A_c \ot \inv{\fl}(c) \ | \ c \in \pi_0(\ttO)\}$.
Proposition \[za\_chvili\_zavolam\_Jarusce\] should be compared to Lemma II.5.49 of [@markl-shnider-stasheff:book]. In the classical operad theory, algebras can equivalently be described as morphism to the endomorphism operad. We are going to give a similar description also in our setup. While the classical construction assigns the endomorphism operad $\End_V$ to a vector space $V$, here we start with a collection $$\label{Jarka_hovori}
V = \{V_c \ | \ c \in \pi_0(\ttO)\}$$ of graded vector spaces indexed by the components of $\ttO$. We moreover assume that to each local terminal object $u\in \ttO$ we are given a linear map (denoted $u$ again) $$\label{Asi_jsem_dostal_premii!}
u: V_{\pi_0(s_1(u))} \to V_{\pi_0(u)}$$ such that, for each map $u \to v$ of local terminal objects with fiber $t$, the triangle $$\label{Kdy_zacnu_jezdit_na_kole?}
\xymatrix{
V_{\pi_0(s_1(u))} \ar[rr]^u \ar[rd]^t &&V_{\pi_0(u)}
\\
&V_{\pi_0(t)} \ar[ru]^v &
}$$ commutes. Since $\pi_0(s_1(u)) = \pi_0(s_1(t))$, $\pi_0(u) = \pi_0(v)$ and $\pi_0(s_1(v)) = \pi_0(t)$, the above diagram makes sense. We moreover assume that the maps corresponding to the chosen local terminal objects are the identities. For $T \in \ttO$ we put $$\End_V(T) := \Vect\big(\textstyle
\bigotimes_{c \in \pi_0(s(T))} V_c,V_{\pi_0(T)}\big).$$
We define an action $
\End_V(T) \ni \alpha \mapsto f^*(\alpha) \in \End_V(S)$ of an $f : S\to T$ with fibers $\Rada u1s$ by $$\label{Popozitri_letim_do_Bari.}
f^*(\alpha)(\Rada a1s) := \alpha(\rada{u_1a_1}{u_sa_s}), \
a_1 \ot \cdots \ot a_s \in \textstyle \bigotimes_{c \in \pi_0(s(T))} V_c.$$ This turns $\End_V$ into a functor $\Iso^{\rm op} \to \Vect$. The structure operation $$\circ_\phi: \End_V(F) \ot \End_V(T) \to \End_V(S) .$$ is, for an elementary morphism $F \fib_i S \stackrel\phi\to T$, defined as follows. Assume $$\alpha : \textstyle \bigotimes_{c \in \pi_0(s(F))} V_c \longrightarrow
V_{\pi_0(F)} \in \End_V(F),\
\beta : \textstyle \bigotimes_{c \in \pi_0(s(T))} V_c \longrightarrow
V_{\pi_0(T)}\in
\End_V(T)$$ and notice that $$\pi_0(s(S)) = \pi_0(s(F)) \sqcup (\pi_0(s(T))
\setminus \{ \pi_0(s_i(T))\}) \ \hbox { and } \ \pi_0(F) = \pi_0(s_i(T)).$$ Then $$\circ_\phi(\alpha \ot \beta) : \textstyle \bigotimes_{c \in
\pi_0(s(S))} V_c \to V_{\pi_0(S)} \in \End_V(S)$$ is the map that makes the diagram $$\xymatrix{
\textstyle \bigotimes_{c \in
\pi_0(s(S))} V_c \ar[r]^(.3)\cong
\ar[dd]_{\circ_\phi(\alpha \ot \beta)} & \textstyle \bigotimes_{c \in
\pi_0(s(F))} V_c
\ot
\bigotimes_{c \in
\pi_0(s(S))\setminus \{ \pi_0(s_i(T))\}} V_c \ar[d]^{\alpha \ot \id}
\\
& \hskip 2em \textstyle
V_{\pi_0(s_i(T))}
\ot
\bigotimes_{c \in
\pi_0(s(S))\setminus \{ \pi_0(s_i(T))\}} V_c\ar[d]^\cong
\\
V_{\pi_0(S)} & \textstyle
\bigotimes_{c \in
\pi_0(s(S))} V_c \ar[l]_\beta
}$$ commuting. The result of the above construction is the Markl version of the [*endomorphism operad*]{}. We trust that it will not be confused with the endomorphism operad recalled from [@duodel] in Example \[end\].
Notice that $\End_V$ is extended unital, with the transformations $\eta_u : \bfk \to \End_V(u)$ given by the maps in (\[Asi\_jsem\_dostal\_premii!\]) as $$\eta_u(1) := u :
V_{\pi_0(s_1(u))} \to V_{\pi_0(u)}
\in \End_V(u).$$ It is simple to verify that the commutativity of (\[Je\_vedro.\]) is precisely (\[Kdy\_zacnu\_jezdit\_na\_kole?\]). The induced maps $$\vartheta(T,u) : \End_V(F) \to \End_V(T)$$ in (\[proc\_ty\_lidi\_musej\_porad\_hlucet\]) are given by the composition, $\vartheta(T,u)(\phi) := u \circ \phi$, with the map (\[Asi\_jsem\_dostal\_premii!\]).
The above analysis shows that the morphisms $\vartheta(T,u)$ need not be the identities for a general $\End_V$. Endomorphism operads are therefore examples of unital operads that need not be [*strictly*]{} extended unital.
We have the expected
\[Udelal\_jsem\_si\_ciruvky\_zelanky.\] There is a one-to-one correspondence between $M$-algebras with the underlying collection (\[Jarka\_hovori\]) and morphisms $M \to \End_V$ of Markl’s operads.
Direct verification.
Virtual isomorphisms {#letam_205}
====================
In this section we introduce the category of virtual isomorphism $\VR$ in an operadic category $\ttO$, its extension $\VI$, and its quotient $\QV$ modulo virtual isomorphisms. In the presence of a grading $e$ of $\ttO$ we will further consider the subgroupoid $\VR(e)$ of objects of grade $\geq 1$, the extension $\VIe$ and the related quotient $\QV(e)$. Presheaves on $\QV(e)$ will serve in Section \[Zitra\_bude\_foukat\_a\_ja\_musim\_preletet\_205.\] as the underlying collections for Markl’s operads.
From this moment on, $\ttO$ will be a factorizable operadic category in which all are invertible, and the blow up and unique fiber axioms are fulfilled. Using the abbreviations of Section \[Poletim\_letos\_do\_Sydney?\] we therefore require $$\hbox{\rm \Fac\ \& \SBU\ \& \QBI\ \& \UFB}.$$ We denote by $\LT$ the groupoid of local terminal objects in $\ttO$ and by $\Iso$ the lluf subcategory of its isomorphism.
Let $T\in \ttO$ and let $t\in \ttO$ be a local terminal object in the connected component of $T$. We therefore have a unique morphism $F \fib T\to t$ with the fiber $F$. In this situation we write simply $F\fib T$ and say that there is [*a virtual morphism*]{} from $F$ to $T$.
\[Za\_chvili\_pojedu\_na\_schuzi\_klubu.\] Virtual morphisms in the operadic category $\ttO$ form a groupoid $\VR$.
The virtual identity morphism for $T\in \ttO$ is defined as the fiber $T \fib T\to U_c$ of the map to a trivial object $U_c$. The composition of virtual morphisms is defined as follows.
Let $S\fib T \fib R$ be a chain of virtual morphisms. This means that $S$ is the fiber of the unique morphism $\phi:T\to t$, $t \in \LT$, and $T$ is the fiber of $\psi:R\to r$, i.e. $S \fib T \stackrel\phi\to t$ and $T \fib R \stackrel\psi\to r$. By the blow up exists a unique factorization of $\psi$ as in the diagram $$\label{Musim_Jarce_odvezt_kramy.}
\xymatrix@R=1em@C=1em{
R \ar[rr]^\xi \ar[rd]_\psi&&s \ar[ld]^\delta
\\
&r&
}$$ such that $\xi_1$ is $\phi$. From Axiom (iv) of an operadic category, $\xi^{-1}(1) = \xi_1^{-1}(1) = \phi^{-1}(1) =S,$ that is $S\fib R \stackrel{\xi}{\to} s$. We take the related virtual morphism $S\fib R$ as the composite of $S \fib T$ and $T \fib R$.
To prove the associativity of the composition we observe that two possible compositions of three virtual morphisms give two fiber sequences: $S\fib Q\to q'$ and $S\fib Q \to q''$. By Lemma \[twolocal\] they however coincide. The proof of the unitality is similar.
The invertibility of morphisms in $\VR$ is established as follows. Consider a virtual morphism $S\fib
T$ given by some $T\stackrel{\phi}{\to} t \in \ttO$. The morphism $\phi$ has a unique factorization $T\to U_c\stackrel{\delta}{\to} t$ through a chosen local terminal object; $\delta$ is clearly the inverse to the canonical isomorphism $t\to U_c$. Let $t'\fib
U_c\stackrel{\delta}{\to} t$. The diagram $$\xymatrix@R=1em@C=1em{
T \ar[rr] \ar[rd]&&U_c \ar[ld]^\delta
\\
&t&
}$$ induces a morphism of fibers $S\to t'$ whose fiber is $T$, giving rise to a virtual morphism $T\fib S$. The composition $(T\fib S) \circ (S\fib T)$ is a morphism $T \fib T$, i.e. the identity by the unique fiber axiom, the equality $(S \fib
T) \circ (T\fib S) = \id$ is established similarly. Thus $T\fib S$ is the inverse to $S\fib T$.
\[zitra\_volam\_Machovi\_do\_Jaromere\] For any $R\in \ttO$ one has an isomorphism of groupoids $\VR/R \cong (R/\LT)^{\rm op}$. In particular, if there exists a morphism $S\fib T$ in $\VR$ then it is unique.
Assume that $S \fib R \to t$. On the level of objects, the isomorphism of the lemma sends $S \fib R$ to $t$. To a morphism $S \fib T$ in $\VR$ over $R$ as in the second half of the proof of Lemma \[Za\_chvili\_pojedu\_na\_schuzi\_klubu.\] we associate the (unique) map $\delta: s \to t$ in (\[Musim\_Jarce\_odvezt\_kramy.\]). It is easy to show that this construction gives rise to an isomorphism of categories. Notice that the second part of of the statement also directly follows from Lemma \[twolocal\].
Lemma \[zitra\_volam\_Machovi\_do\_Jaromere\] implies that $\VR$ is equivalent as a category to a discrete set. As the next step towards our construction of $\QV$ we extend $\VR$ into a category $\VI$ which has the same objects as $\ttO$ but whose morphisms $T\to R$ are sequences $$S\stackrel{\phi}{\to} T\fib R$$ where $\phi \in \Iso$ is an isomorphism in $\ttO$. To define the composition, consider a sequence $$S\stackrel{\phi}{\to} T\stackrel h\fib R\stackrel{\psi}{\to} Q\stackrel{f}\fib
P.$$ The virtual morphism $T\fib R$ is related to a fiber sequence $T\fib R\to r$ with a unique $r\in \LT$. The objects $R$ and $Q$ live in the same connected component, so we have a unique $Q\to r$ and the associated virtual morphism $D\stackrel{g}{\fib} Q$. The diagram $$\xymatrix@R=1em@C=1em{
R \ar[rr]^\psi \ar[rd]&&Q \ar[ld]
\\
&r&
}$$ induces the morphism $\xi: = \psi_1 :T\to D$ of fibers. We then define the composition $$(R\stackrel{\psi}{\to} Q\stackrel{f}{\fib}
P) \circ (S\stackrel{\phi}{\to} T\fib R)$$ as the sequence $$\label{v_patek_pak_Psenicka}
S\stackrel{\xi \circ\phi}{\longrightarrow}D\stackrel{f \circ g}{\fib} P.$$ The identity morphism of $S$ is given by $S\stackrel{\id}{\to} S\fib S$. One can easily check that the above structure makes $\VI$ a category.
Let us show that $\VI$ is a groupoid. For a morphism $\Phi: S\stackrel{\phi}{\to} T\stackrel{f}{\fib} R$ in $\VI$ we consider the inverse $R \stackrel g\fib T$ to the virtual morphism $T \stackrel f\fib R$ and the inverse $\psi:T \to S$ of $\phi$. It turns out that the composition $$R\stackrel{\id}{\to} R\stackrel{g}{\fib} T \stackrel{\psi}{\to}
S\fib S$$ in $\VI$ represents the inverse to $\Phi$.
We now consider the quotient $\QV$ of $\VI$ whose objects are classes of objects of $\VI$ with respect to the relation generated by virtual isomorphisms $F \fib T$. That is, two objects are equivalent if there is a virtual isomorphism between them. More precisely, $\QV$ is defined by the pushout $$\label{Musim_napsat_doporuceni_pro_Jovanu.}
\xymatrix@C=3.5em@R=1.2em{\VR\ar[d]_{} \ar[r] & \VI \ar[d]^{}
\\
\pi_0(\VR) \ar[r] &\QV
}$$ in the category of groupoids. Since the left vertical functor is an equivalence and the top horizontal functor a cofibration of groupoids, the right vertical functor is an equivalence of groupoids, too.
It is easy to see that morphisms between objects in $\QV$ are equivalence classes of non-virtual isomorphisms in the following sense. Let $\phi':T'\to S'$ and $\phi'':T''\to S''$ be two isomorphisms in $\ttO$. They are equivalent if there exist a local terminal object $t$ such that $\phi''$ is the induced fiber map in the diagram $$\label{Dnes_jsem_se_koupal_v_Hradistku.}
\xymatrix@R=1em@C=1em{
T' \ar[rr]^{\phi'} \ar[rd]&&\ S'. \ar[ld]
\\
&t&
}$$
\[Budu\_mit\_nova\_sluchatka.\] Assume that each connected component of $\ttO$ contains precisely one terminal object, i.e. all local terminal objects are the trivial (chosen) ones. Then $\QV \cong \Iso$. This is the case of e.g. the category $\Fin$ of finite sets or of the operadic category $\Per$ in Subsection \[Michael\_do\_Prahy.\] governing permutads.
In the operadic category $\Bq(\frC)$ recalled in Example \[puget\] two bouquets are virtually equivalent if they differ only in the last color. The groupoid $\QV$ is in this case the groupoid of strings $(i_1,\ldots,i_k)$, $k \geq 1$, with morphisms arbitrary bijections.
\[Snad\_to\_projde\_i\_formalne.\] Two graphs in the operadic category $\Gr$ of Definition \[Radeji\_bych\_sel\_s\_Jaruskou\_na\_vyhlidku\_sam.\] are virtually equivalent if they differ only in the global orders of their leaves. Morphisms in $\QV$ are isomorphisms of graphs which however need not preserve the global orders.
Assume that $\ttO$ possesses a grading $e : {\rm Objects}(\ttO) \to \bbN$ as in Definition \[grad\]. In this case we denote by $\VR(e)\subset \VR$ the full subgroupoid with objects $T \in \ttO$ such that $e(T) \geq 1$. We construct $\VIe$ out of $\VR(e)$ and $\Iso$ as before, and define its quotient $\QV(e)$ by replacing $\VR$ by $\VR(e)$ in (\[Musim\_napsat\_doporuceni\_pro\_Jovanu.\]).
Markl’s operads and virtual isomorphisms {#Zitra_bude_foukat_a_ja_musim_preletet_205.}
========================================
This section contains preparatory material for Section \[zitra\_letim\_do\_Pragy\]. Its main result, Proposition \[dnes\_schuze\_klubu\], states that strictly extended unital Markl’s operads determine presheaves on the category $\QV(e)$ introduced at the end of Section \[letam\_205\]. The assumptions are again $$\hbox{\rm \Fac\ \& \SBU\ \& \QBI\ \& \UFB},$$ plus the existence of a grading.
\[zitra\_Holter\] Each extended unital Markl’s operad $M$ induces a covariant functor , denoted $M$ again, which is $M$ on objects, and on morphisms is defined by $$M(F \fib T) := \vartheta(T,u),$$ where $\vartheta(T,u)$ is as in (\[proc\_ty\_lidi\_musej\_porad\_hlucet\]). Since $\VR(e)$ is a groupoid, all maps $\vartheta(T,u)$ are invertible.
It follows from the unitality (\[Holter\_se\_blizi.\]) of $M$ that $M(T \fib T) = \id_T$. Let us verify the functoriality $$\label{pozitri_Holter}
M(S \fib R) = M(R \fib T) \circ M(S \fib T).$$ To this end we consider the commutative diagram $$\label{dopocitat_si_zapisnik}
\xymatrix{
&M(T) \ot M(r) \ar[r]^(.6){\circ_!} & M(R)
\\
&M(S) \ot M(r) \ar[u]^{\vartheta(T,r) \ot \id}
\ar[r]^{\id \ot \delta^*} & \ M(S) \ot M(r').\ar[u]^{\circ_!}
\\
M(S)\ar[r]^(.4)\cong &
\ar[u]^{\id \ot \eta_r} \ar[ur]_{\id \ot \eta_{r'}}
M(S) \ot \bfk
}$$ Its upper square is (\[ve\_ctvrtek\_letim\_do\_Prahy\]) applied to the diagram $$\xymatrix@C=3.5em@R=1.5em{
R \ar@{=}[r] \ar[d]_\xi \ar[dr]^\psi & R \ar[d]^\psi
\\
r' \ar[r]^\delta & r
}$$ in place of (\[moc\_se\_mi\_na\_prochazku\_nechce\]), in which the symbols have the same meaning as in (\[Musim\_Jarce\_odvezt\_kramy.\]). The commutativity of the bottom triangle follows from the commutativity of (\[Je\_vedro.\]). It follows from the definition of the maps $\vartheta(T,u)$ that the composition $$\xymatrix@1{M(S) \cong M(S) \ot \bfk \ \ar[r]^(.55){\id \ot \eta_{r'}}
&\ M(S) \ot M(r) \
\ar[r]^(.63){\circ_!} & \ M(R)
}$$ in (\[dopocitat\_si\_zapisnik\]) equals the left hand side of (\[pozitri\_Holter\]), while the composition $$\xymatrix@1{
M(S) \cong M(S) \ot \bfk \
\ar[r]^(.55){\id \ot \eta_{r}} & \ M(S) \ot M(r) \
\ar[rr]^{\vartheta(T,r) \ot \id} &&
\ M(T) \ot M(r) \ \ar[r]^(.63){\circ_!} & \ M(R)
}$$ equals the right hand side of (\[pozitri\_Holter\]).
\[dnes\_schuze\_klubu\] The $\Iso$-presheaf structure of a Markl’s operad $M$ combined with the functor $M: \VR(e) \to \ttV$ of Lemma \[zitra\_Holter\] makes $M$ a $\VIe$-presheaf via the formula $$\label{Pujdu_s_Jarkou_na_demonstraci?}
M\big(S\stackrel{\phi}{\to} T\stackrel{f}{\fib} R\big)
:= \phi^* \circ M(f^{-1}).$$
We are going to use the notation of (\[v\_patek\_pak\_Psenicka\]). Since clearly $M(T\stackrel\id\to T \fib T) = \id_T$, we need only to prove that $$\label{uz_zase_stavebni_stroje_duni}
M\big(S\stackrel{\phi}{\to} T\stackrel{h}{\fib} R\big)
\circ M\big(R\stackrel{\psi}{\to} Q\stackrel{f}{\fib}
P\big) =
M\big(S\stackrel{\xi \circ\phi}{\longrightarrow}D\stackrel{f \circ
g}{\fib}
P\big).$$ Evaluating both sides using definition (\[uz\_zase\_stavebni\_stroje\_duni\]) gives $$\phi^* \circ M(h^{-1}) \circ \psi^* \circ M(f^{-1}) =
(\xi \circ\phi)^* \circ M\big((f \circ g)^{-1}\big)$$ where, by the functoriality established in Lemma \[zitra\_Holter\], $$(\xi \circ\phi)^* \circ M\big((f \circ g)^{-1}\big)
= \phi^* \circ \xi^*
\circ M(g^{-1}) \circ M(f^{-1}).$$ Since all maps involved are isomorphisms, we easily see that (\[uz\_zase\_stavebni\_stroje\_duni\]) is equivalent to $$\label{V_pondeli_plicni.}
M(g) \circ (\inv\xi)^* = (\inv\psi)^* \circ M(h).$$ To prove this equality, consider the diagram $$\label{V_patek_pak_Psenicka.}
\xymatrix{
M(T) \cong M(T) \ot \bfk \ar[r]^(.54){\id \ot \eta_r}
& M(T) \ot M(r)\ar[d]_{\circ_!} &\ar[l]_{\xi^* \ot \id} M(D) \ot
M(r) \ar[d]^{\circ_!}
\\
& M(R)&M(Q)\ar[l]_{\psi^*}
}$$ in which the square is (\[ve\_ctvrtek\_letim\_do\_Prahy\]) associated to $$\xymatrix@C=3.5em@R=1.5em{\ar[dr]^{!}
R\ar[d]_{!} \ar[r]^\psi &Q\ar[d]^{!}
\\
r \ar@{=}[r] &r
}$$ in place of (\[moc\_se\_mi\_na\_prochazku\_nechce\]). It follows from definitions that the composition of the maps $$\xymatrix@1{
M(T) \cong M(T) \ot \bfk \ \ar[r]^(.55){\id \ot \eta_r} &
\ M(T) \ot M(r) \ \ar[rr]^{\inv{(\xi^* \ot \id)}}&&
\ M(D) \ot
M(r) \ \ar[r]^(.65){\circ_!}&\ M(Q)
}$$ in (\[V\_patek\_pak\_Psenicka.\]) equals the left hand side of (\[V\_pondeli\_plicni.\]), while the composition $$\xymatrix{
M(T) \cong M(T) \ot \bfk \ \ar[r]^(.55){\id \ot \eta_r} &
\ M(T) \ot M(r) \ \ar[r]^(.6){\circ_!} &
\ M(R) \ \ar[r]^{\inv{\psi^*}} & \ M(Q)
}$$ equals its right hand side.
\[Za\_14\_dni\_LKDL\] If $M$ is an extended unital operad, then the $\VIe$-presheaf of Proposition \[dnes\_schuze\_klubu\] associated to $M$ functorially descents to a $\QV(e)$-presheaf $\oM$.
Objects of $\QV(e)$ are, by definition, equivalence classes $[T]$ of objects of $\ttO$ modulo the relation $[T'] = [T'']$ if $T'' \fib
T'$. We define $\oM([T])$ as the colimit $$\label{V_sobotu_budu_letat_vycvik_vlekare.}
\oM([T]) : = \colim M(S)$$ over the groupoid of all $S \in \ttO$ virtually isomorphic to $T$. It is clear that the canonical injection $\iota_T : M(T) \hookrightarrow \oM([T])$ is an isomorphism.
Consider a morphism $[\phi] : [T] \to [S]$ in $\QV(e)$ given by an isomorphism $\phi : T \to S$. We define $ \oM([\phi]) :
\oM([S]) \to \oM([T])$ by the diagram $$\xymatrix@C=3em{
\oM([S]) \ar[r]^{\oM([\phi])} & \oM([T])
\\
M(S)\ar[r]^{\phi^*}
\ar@{_{(}->}[u]^{\iota_S}_\cong
& M(T) \ar@{_{(}->}[u]_{\iota_T}^\cong
}$$ in which $\phi^*$ refers to the $\Iso$-presheaf structure of $M$.
We need to show that $\oM([\phi])$ does not depend on the choice of a representative of the map $[\phi]$ under the equivalence that identifies $\phi'$ as in (\[Dnes\_jsem\_se\_koupal\_v\_Hradistku.\]) with the induced map $\phi''$ between the fibers over $t$. To this end, consider the commutative diagram $$\label{Dnes_jde_Jarka_s_bouli_pod_uchem_k_doktorovi.}
\xymatrix{
M(T') \ar@{=}[r] \ar[d]_{\vartheta(T',t)}
& \bfk \ot M(T') \ar[d]_{\eta_t \ot \id} & \ar[d]^{\eta_t \ot \id}
\bfk \ot M(S') \ar@{=}[r]\ar[l]_{\id \ot \phi'^*} &
M(S')
\ar@/_2em/[lll]_{\phi'^*} \ar[d]^{\vartheta(S',t)}
\\
M(T'') &\ar[l]_(.6){\circ_!} M(u) \ot M(T') & M(u) \ot M(S')
\ar[r]^(.63){\circ_!}\ar[l]_{\id \ot \phi'^*}
&
M(S'')\ar@/^2em/[lll]_{\phi''^*}
}$$ in which the extreme left and right squares are instances of (\[proc\_ty\_lidi\_musej\_porad\_hlucet\]). The commutativity of the central square and of the upper part is clear. Finally, the commutativity of the lower part follows from axiom (\[posledni\_nedele\_v\_Sydney\]) of Markl’s operads. An easy diagram chase shows that the commutativity of (\[Dnes\_jde\_Jarka\_s\_bouli\_pod\_uchem\_k\_doktorovi.\]) implies the commutativity of the middle square in $$\xymatrix@C=3em{
\oM([T']) \ar@{^{(}->}[r]^{\iota_{T'}}_\cong \ar@{=}[d]
&M(T') \ar[d]_{\vartheta(T',t)} &
M(S')
\ar[l]_{\phi'^*} \ar[d]^{\vartheta(S',t)}& \oM([S'])
\ar@{_{(}->}[l]_{\iota_{S'}}^\cong \ar@{=}[d]
\\
\oM([T''])\ar@{^{(}->}[r]^{\iota_{T''}}_\cong
&M(T'')
&
M(S'')\ar[l]_{\phi''^*}&\ \oM([S'']) .
\ar@{_{(}->}[l]\ar@{_{(}->}[l]_{\iota_{S''}}^\cong
}$$ The independence of $\oM([\phi])$ on the choice of a representative of $[\phi]$ is now clear.
\[Zitra\_mam\_prednasku\_na\_Macquarie.\] If $M$ is strictly extended unital, the definition (\[V\_sobotu\_budu\_letat\_vycvik\_vlekare.\]) via a colimit can be replaced by $\oM([T]) := M(T)$.
Free Markl’s operads {#zitra_letim_do_Pragy}
====================
This section is devoted to our construction of free strictly extended unital $1$-connected Markl’s operads in $\ttO$ generated by $\QV(e)$-presheaves. In the light of Theorem \[Mcat\] this will also provide free (standard) $\ttO$-operads. As before, $\ttO$ will be a factorizable operadic category in which all are invertible, and the blow up and unique fiber axioms are fulfilled, that is $$\hbox{\rm \Fac\ \& \SBU\ \& \QBI\ \& \UFB}.$$ We moreover assume that $\ttO$ is graded and that a morphism $f$ is an isomorphisms if $e(f) = 0$.
\[Jarka\_dnes\_u\_lekare\] We will call $\QV(e)$-presheaves $E$ in $\ttV$ such that $E([T])=0$ whenever $e(T) = 0$ [*$1$-connected $\ttO$-collections*]{} in $\ttV$ and denote by $\Coll(\ttO)$ or simply $\Coll$ when $\ttO$ is understood the corresponding category. We will often write simply $E[T]$ instead of $E([T])$.
Notice that a $1$-connected $\ttO$-collection can equivalently be defined as an $\QV$-presheaf $E$ such that $E(T) = 0$ if $e(T) = 0$. It follows from definitions that a $1$-connected $\ttO$-collection is the same as an $\Iso$-presheaf $\calE$ in $\ttV$ such that
- $\calE(T) = 0$ if $e(T) = 0$ ($1$-connectivity),
- $\calE(T) = \calE(F)$ whenever $F \fib T
\stackrel!\to u$, and
- $\phi'^* = \phi''^*$, where $\phi'$ is as in (\[Dnes\_jsem\_se\_koupal\_v\_Hradistku.\]) and $\phi''$ is the induced map between the fibers.
It follows from Example \[Budu\_mit\_nova\_sluchatka.\] that the category $\Coll(\Fin)$ is isomorphic to the category of $1$-connected $\Sigma$-modules, i.e. sequences $\{E(n) \in \ttV\}_{n \geq 2}$, with actions of the symmetric groups $\Sigma_n$.
\[Pomuze\_vitamin\_C?\] One has a forgetful functor $M \mapsto \Box M$ from the category of $1$-connected strictly extended unital Markl’s $\ttO$-operads to the category of $1$-connected $\ttO$-collections defined on objects by $$\Box M\big([T]\big) :=
\begin{cases}
M(T) &\hbox {if $e(T) \geq 1$, and}
\\
0 & \hbox {otherwise.}
\end{cases}$$
The functor $\Box$ is the composition of the functor $M \mapsto \oM$ of Proposition \[Za\_14\_dni\_LKDL\][^13] with the functor that replaces the values of the presheaf $\oM$ by $0$ on objects of grade zero.
In the rest of this section we construct a left adjoint $E \mapsto
\Free(E)$ to the forgetful functor of Proposition \[Pomuze\_vitamin\_C?\]. Our strategy will be to construct a Markl’s $\ttO$-operad $\lTw$ with values in the category of groupoids $\Grp$, extend $E$ to a functor $E : \lTw
\to \ttV$ and define $\Free(E)$ as the colimit of this functor. The building blocks of the operad $\lTw$ will be the towers $$\label{t1-1}
\bfT: =
T \stackrel {\tau_1} \longrightarrow T_1 \stackrel {\tau_2} \longrightarrow
T_2\stackrel {\tau_3} \longrightarrow
\cdots \stackrel {\tau_{k-1}} \longrightarrow T_{k-1}
\stackrel {\tau_{k} } \longrightarrow U_c$$ of elementary morphisms as in Definition \[plysacci\_postacci\], with $\tau_k$ the unique morphism to a chosen local terminal object $U_c$. Since $\tau_k$ bears no information, we will sometimes write the tower as $$\label{t1}
\bfT: =
T \stackrel {\tau_1} \longrightarrow T_1 \stackrel {\tau_2} \longrightarrow
T_2\stackrel {\tau_3} \longrightarrow
\cdots \stackrel {\tau_{k-1}} \longrightarrow T_{k-1}.$$ Let $t_1,\ldots,t_{k}$ be the unique nontrivial fibers of $\tau_1,\ldots,\tau_{k}$; notice that $t_k = T_{k-1}$. We will call $t_1,\ldots,t_{k}$ the [*fiber sequence*]{} of the tower $\bfT$.
We will denote by $\Tw(T)$ the set of all towers with the initial term $T$. A [* morphism*]{} $\bfsigma : \bfT' \to \bfT''$ of towers in (\[t1-1\]) is an array $\bfsigma =
(\sigma_1,\sigma_2,\ldots,\sigma_{k})$ of as in Figure \[f1\].
$$\xymatrix{T' \ar[d]_{\tau'_1} \ar[r]^{\sigma_1}_\cong & T'' \ar[d]^{\tau''_1}
\\
T_1' \ar[d]_{\tau'_2} \ar[r]^{\sigma_2}_\cong & T_1'' \ar[d]^{\tau''_2}
\\
\vdots \ar[d]_{\tau'_{k-1}} & \vdots \ar[d]^{\tau''_{k-1}}
\\
T_{k-1}' \ar[r]^{\sigma_{k}}_\cong & T_{k-1}''
}$$
\[zitra\_vedeni\] A [*labelled tower*]{} is a couple $(\ell,\bfT)$ consisting of a tower $\bfT$ as in (\[t1-1\]) together with an (the [*labeling*]{}) $\ell : X \to
T$. We denote by ${\it lTw}(X)$ the set of all labelled towers of this form.
We will equip ${\ltw}(X)$ with the structure of a groupoid with morphisms of two types. Each morphism $\bfsigma : \bfT' \to \bfT''$ of towers in (\[t1-1\]) determines a morphism $(\ell,\bfsigma)
:(\ell,\bfT') \to (\sigma_1\ell,\bfT'')$ of the [*first type*]{}. These morphisms compose in the obvious manner, namely $$(\sigma'_1\ell,\bfsigma'') \circ (\ell, \bfsigma') = (\ell, \bfsigma''
\circ \bfsigma').$$
To define morphisms of the second type, consider two towers of elementary morphisms, $$\bfT':= T \stackrel {\tau_1} \longrightarrow T_1 \stackrel{\tau_2} \longrightarrow
\cdots \stackrel{\tau_{u-1}} \longrightarrow {T_{u-1}}
\stackrel{\tau'_{u}}\longrightarrow T'_{u} \stackrel
{\tau'_{u+1}}\longrightarrow T_{u+1} \stackrel{\tau_{u+2}}
\longrightarrow \cdots \stackrel {\tau_{k-1}} \longrightarrow
T_{k-1}$$ and $$\bfT'':=
T \stackrel {\tau_1} \longrightarrow T_1 \stackrel{\tau_2} \longrightarrow
\cdots
\stackrel{\tau_{u-1}} \longrightarrow {T_{u-1}}
\stackrel{\tau''_{u}}\longrightarrow T''_{u}
\stackrel {\tau''_{u+1}}\longrightarrow T_{u+1}
\stackrel{\tau_{u+2}} \longrightarrow \cdots
\stackrel {\tau_{k-1}} \longrightarrow T_{k-1},$$ as in (\[t1\]). Their associated fiber sequences are clearly of the form $$t_1,\ldots,t_{u-1},t'_u,t'_{u+1},t_{u+2},\ldots,t_k
\ \hbox { resp. } \
t_1,\ldots,t_{u-1},t''_u,t''_{u+1},t_{u+2},\ldots,t_k.$$ Assume that the diagram $$\xymatrix@R = 1em{& {T'_u} \ar[dr]^{\tau'_{u+1}} &
\\
T_{u-1}\ar[dr]^{\tau_u''} \ar[ur]^{\tau_u'} && T_{u+1}
\\
&{T''_u}\ar[ur]^{\tau''_{u+1}}&
}$$ is as in (\[eq:3\]), with $\phi' = \tau'_u$, $\phi'' = \tau''_u$, $\psi' = \tau'_{u+1}$ and $\psi'' = \tau''_{u+1}$. The above situation, by definition, determines an invertible morphism $(\ell,\bfT') \to (\ell,\bfT'')$ of the [*second type*]{}.
The above morphism are subject to the obvious commutativity relations. The resulting groupoid will be denoted $\lTw(X)$.
\[vcera\_jsem\_bezel\_36\_minut\] Since morphisms of both types preserve the height of towers, the groupoid $\lTw(X)$ is graded, $$\lTw(X) = \textstyle\coprod_{h \geq 1}\lTw^h(X).$$ It is clear that $\lTw^1(X)$ is the category $X/\Iso$ of in $\ttO$ under $X$.
In $\lTw^2(X)$, only morphisms of the first type exist. Therefore, labelled towers $(\ell',\bfT')$ and $(\ell'',\bfT'')$ are connected by a morphism if and only if one has a commuting diagram $$\label{vecer_ustrice}
\xymatrix{&X \ar[ld]_{\ell'}^\cong \ar[rd]^{\ell''}_\cong &
\\
T' \ar[d]_{\tau'} \ar[rr]^{\sigma_1}_\cong &&T'' \ar[d]^{\tau''}
\\
T'_1 \ar[rr]^{\sigma_2}_\cong &&T''_1
}$$ with $\sigma_1$ and $\sigma_2$.
For an $\omega : X' \iso X''$ one has the induced map $\omega^*: \ltw(X'') \to \ltw(X')$ that sends the labelled tower $(\ell'',\bfT'') \in \ltw(X'')$ into $(\ell''\omega,\bfT'') \in
\ltw(X')$ which clearly extends to a functor (denoted by the same symbol) $\omega^*: \lTw(X'') \to \lTw(X')$. This makes the collection of categories $\lTw(X)$ a $\Grp$-presheaf on $\Iso$. Our next move will be to construct, for each $G \fib W \stackrel\phi\to X$, a functor $$\label{mohl_bych_si_dat_i_12_ustric}
\circ_\phi : \lTw(X) \times \lTw(G) \to \lTw(W).$$
As the first step in this construction we will prove that each labelled tower $(\ell, \bfT)$ can be functorially replaced by one in which $\ell$ is a quasibijection. To this end we prove a couple of auxiliary lemmas.
The factorization $\xi =\phi \circ \sigma$, $\phi \in \DO$, $\sigma \in \QO$, of an isomorphism $\xi :A \to B$ guaranteed by the factorization axiom is unique, and both $\phi$ and $\sigma$ are isomorphisms, too.
Consider two factorizations, $\phi' \circ \sigma'$ and $\phi''
\circ \sigma''$, of $\xi$. Since $\sigma'$ and $\sigma''$ are , they are invertible, so one may define $u$ by the commutativity of the diagram $$\xymatrix@R=1em{&X' \ar[rd]^{\phi'}_\cong &
\\
A\ar[ur]^{\sigma'}_\sim \ar[rd]^{\sigma''}_\sim&&\ B.
\\
&X'' \ar[ru]^{\phi''}_\cong \ar[uu]_u &
}$$ By the left triangle, $u$ is a while it belongs to $\DO$ by the right triangle. The uniqueness follows from Corollary \[zase\_mam\_nejaky\_tik\]. The invertibility of $\phi'$ and $\phi''$ is clear.
\[zitra\_na\_ustrice\] Each corner $$\xymatrix@R=1.2em@C=2.5em{\tilde T \ar[r]^\omega_\cong & T \ar[d]^\phi
\\
&S
}$$ in which $\omega$ is an from $\DO$ and $\phi$ is elementary, can be functorially completed to the square $$\label{uz_se_to_krati}
\xymatrix@R=1.2em@C=2.5em{\tilde T \ar[d]_{\tilde \phi} \ar[r]^\omega_\cong & T \ar[d]^\phi
\\
\tilde S\ar[r]_{\tilde \omega}^\cong&S
}$$ with $\tilde \omega$ an from $\DO$ and $\tilde \phi$ elementary.
The map $\tilde \phi := \phi \circ \omega$ belongs to $\DO$. It has precisely one fiber $G$ such that $e(G) \geq 1$, its other fibers $\Rada u1{s-1}$ are terminal. The blow up axiom produces a unique diagram (\[uz\_se\_to\_krati\]) such that the maps of fibers induced by $\tilde \omega$ are $$!: G \to U,\ ! : u_1 \to u_1, \ldots, ! : u_{s-1} \to u_{s-1},$$ with $U$ the chosen local terminal object of $\ttO$. It is clear that the maps thus constructed have the properties stated in the lemma.
\[22\_hodin\_cesty\] Each $(\ell, \bfT)\in \lTw(X)$ can be functorially replaced within its isomorphism class by some $(\tilde\ell, \tilde\bfT)$ in which $\tilde\ell$ is a .
Let $\bfT$ be as in (\[t1-1\]) and $\ell : X \to T$ be an . We decompose $\ell$ as $\sigma_1 \circ \tilde\ell$, with $\tilde\ell$ a and $\sigma_1$ an in $\DO$. Lemma \[zitra\_na\_ustrice\] gives a canonical square $$\xymatrix@R=1.2em@C=2.5em{\tilde T\ar[d]_{\tilde \tau_1}
\ar[r]^{\sigma_1}_\cong & T \ar[d]^{\tau_1}
\\
\tilde T_1\ar[r]^{\sigma_2}_\cong&T_1
}$$ in which $\tilde \tau_1$ is elementary and $\sigma_2$ an in $\DO$. Repeating this process produces a tower $\tilde \bfT$ labelled by the $\tilde\ell : X \to \tilde T$.
Proposition \[22\_hodin\_cesty\] implies that the graded category $\lTw(X)$ contains a reflexive graded subcategory $\tlTw(X)$ whose objects are towers $(\ell,\bfT)$ labelled by a . Morphisms of the first type in $\tlTw(X)$ are those $(\ell,\bfsigma)
:(\ell,\bfT') \to (\sigma_1\ell,\bfT'')$ in which $\sigma_1$ is a . Morphisms of the second type are the same as those in $\lTw(X)$.
\[pisu\_druhy\_den\_v\_Srni\] Labelled towers $(\ell',\bfT'), (\ell'',\bfT'') \in\tlTw^2(X)$ are isomorphic if and only if one has the commuting diagram (\[vecer\_ustrice\]) in which the maps in the upper triangle are .
Let $G \fib W \stackrel\phi\to X$ be elementary. Assume we are given a labelled tower $(\ell',\bfF) \in
\ltw(G)$, where $$\label{17b}
\bfF: =
F \stackrel {\varphi_1} \longrightarrow F_1 \stackrel {\varphi_2} \longrightarrow
F_2\stackrel {\varphi_2} \longrightarrow
\cdots \stackrel {\varphi_{l-1}} \longrightarrow F_{l-1}$$ is a tower with the associated fibers $f_1,\ldots,f_l$, with the labeling $\ell' : G \to F$. Assume we are also given a labeled tower $(\tilde \ell,\bfT) \in \tlTw(X)$, with $\tilde\ell$ a . The blow up axiom gives a unique diagram $$\xymatrix@C=3.5em@R=1.2em{W \ar[d]_{\phi} \ar[r]^{\ell''}_\cong & S \ar[d]^{\rho}
\\
X \ar[r]^{\tilde \ell}_\sim &T
}$$ in which $F \fib S \stackrel\rho\to T$ is elementary and $\ell''$ an inducing the map $\ell' : G \to F$ between the unique nontrivial fibers of $\phi$ and $\rho$, respectively. In this situation we have the composite tower $$\label{Taronaga_Zoo}
\bfT \circ_\rho \bfF:=
S \stackrel {\rho_1} \longrightarrow S_1 \stackrel {\rho_2} \longrightarrow
S_2 \stackrel {\rho_3} \longrightarrow \cdots \stackrel
{\rho_{l-1}} \longrightarrow S_{l-1} \stackrel {\rho_l} \longrightarrow T
\stackrel {\tau_1}\longrightarrow T_1 \stackrel {\tau_2} \longrightarrow
T_2\stackrel {\tau_2} \longrightarrow
\cdots \stackrel {\tau_{k-1}} \longrightarrow T_{k-1}$$ whose initial part is (\[eq:2\]), and the composite labeled tower $$\label{Safari_ani_ne_za_tyden.}
(\ell,\bfT) \circ_\phi (\ell',\bfF) := (\ell'',\bfT \circ_\rho \bfF)
\in \ltw(W).$$ The above construction clearly extends to a functor $$\circ_\phi : \tlTw(X) \times \lTw(G) \to \lTw(W)$$ which, precomposed with the equivalence $\lTw(X) \to \tlTw(X)$ in the first variable, gives (\[mohl\_bych\_si\_dat\_i\_12\_ustric\]).
Let $E \in \Coll$ be a $1$-connected collection. For a tower (\[t1\]) we define $$E(\bfT) := E[t_1] \otimes \cdots \otimes E[t_k]
\in \ttV.$$ We will show how the rule $E(\ell,\bfT) := E(\bfT)$ extends to a functor $E
: \lTw(X) \to \ttV$. Consider a morphism $(\ell,\bfsigma)
:(\ell,\bfT') \to (\sigma_1\ell,\bfT'')$ of the first type, with $\bfsigma : \bfT' \to \bfT''$ a map of towers as in Figure \[f1\]. For each $0 \leq s \leq k$ one has the commutative diagram $$\xymatrix@C=4em{T_s' \ar[d]_{(\tau_s',i)}\ar[dr]^{\tau_s}
\ar[r]^{\sigma_{s+1}}_\cong
& T_s'' \ar[d]^{(\tau_s'',j)}
\\
T_{s-1}' \ar[r]^{\sigma_s}_\cong &T_{s-1}''
}$$ in which $\tau_s := \sigma_s \circ \tau'_s = \tau''_s \circ
\sigma_{s+1}$.[^14] Lemma \[l3\] provides us with $$t'_s \fib \tau^{-1}_s(j)
\stackrel{(\tau'_s)_j}\longrightarrow \sigma_s^{-1}(j)
\ \hbox { and } \
(\sigma_{s+1})_j: \tau_s^{-1}(j) \to t''_s$$ so we can define $\sigma_s^\star : E[t''_s] \to E[t'_s]$ as the composition $$\sigma_s^\star : E[t''_s] \stackrel{(\sigma_{s+1})_j^*}\longrightarrow E[
\tau_s^{-1}(j)] = E[t'_s]\footnote{Notice that the equality uses the
fact that $E$ is constat along virtual isomorphisms.}$$ which in turn induces a map $$\bfsigma^\star : E(\bfT'')
= E[t'_1] \otimes \cdots \otimes E[t'_k] \longrightarrow
E(\bfT')
= E[t''_1] \otimes \cdots \otimes E[t''_k]$$ by $\bfsigma^\star := {\sigma}_1^\star \ot \cdots \ot
{\sigma}_k^\star$. Define finally $E(\ell,\bfsigma) : E(\ell,\bfT'') \to E(\sigma_1\ell,\bfT')$ as $E(\ell,\bfsigma):= \bfsigma^\star$.
Let us discuss morphisms of the second type. Corollary \[move\] gives identities $$t'_u = t''_{u+1} \ \mbox { and } \
t''_u = t'_{u+1}.$$ We define the $E$-image of this map as the identification of $$e_1 \ot \cdots\ot e''_{u+1} \ot e'_{u+1} \ot \cdots \ot e_k
\in E[t_1] \ot \cdots \ot E[t'_u] \ot E[t'_{u+1}] \ot
\cdots \ot E[t_k]$$ in $E(\bfT')$ with $$e_1 \ot \cdots\ot e_{u+1}'\ot e''_{u+1} \ot \cdots \ot e_k
\in E[t_1] \ot \cdots \ot E[t''_u] \ot E[t''_{u+1}] \ot
\cdots \ot E[t_k]$$ in $E(\bfT'')$ given by the symmetry constraint in $\ttV$.
\[popozitri\_Janacek\] The diagram of functors $$\xymatrix{& \ttV &
\\
\lTw(X) \times \lTw(G) \ar[ur]^{E \ot E} \ar[rr]^(.6){\circ_\phi} &&
\lTw(W) \ar[ul]_E
}$$ commutes for an arbitrary elementary morphism $G \fib W \stackrel\phi\to X$.
Assume that $(\ell,\bfT) \in \lTw(X)$ and $(\ell',\bfF) \in \lTw(G)$, with $\bfT$ as in (\[t1\]) and $\bfF$ as in (\[17b\]). Recall that then $(\ell,\bfT) \circ_\phi
(\ell',\bfF) \in \lTw(W)$ is given by formula (\[Safari\_ani\_ne\_za\_tyden.\]). The crucial fact is that the fiber sequence of $\bfT \circ_\rho \bfF$ is $$f_1,\ldots,f_l,t_1,\ldots,t_{k},$$ where $f_1,\ldots,f_l$ resp. $t_1,\ldots,t_k$ is the fiber sequence of $\bfF$ resp. of $\bfT$. The canonical isomorphism $$E(\ell,\bfT) \ot E(\ell',\bfF)
\cong E\big((\ell,\bfT) \circ_\phi (\ell',\bfF)\big)$$ then follows immediately from the definition of the functor $E$ as given above.
\[dnes\_bylo\_40\] Let $E \in \Coll$ be a $1$-connected collection in $\ttV$. Then the formula $$\label{Konci_Neuron?}
\Free(E)(X) :=
\begin{cases}\displaystyle
\colim_{(\ell,\bfT) \in\lTw(X)} E(\ell,\bfT)&\hbox{if $e(X) \geq 1$,
and}
\\
\bfk&\hbox{if $e(X) = 0$,}
\end{cases}$$ defines a left adjoint $E \mapsto \Free(E)$ to the forgetful functor $\Box$ of Proposition \[Pomuze\_vitamin\_C?\]. Therefore $\Free(E)$ is the free $1$-connected strictly extended unital Markl’s operad generated by $E$.[^15]
Assume that $X\in \ttO$ is such that $e(X) \geq 1$. It is clear that the vector space $\Free(E)(X)$ is graded by the height $k$ of the underlying towers so that it decomposes as $$\label{co_udela_ta_noha_na_Safari}
\Free(E)(X) \cong\textstyle \bigoplus_{k \geq 1} \Free^k(E)(X).$$ Elements of $\Free^k(E)(X)$ are equivalence classes $[\ell,e]$ consisting of a labeling $\ell :X \iso T$ and of an element $e \in E(\bfT)$ associated with a labeled tower $(\ell,\bfT)$ as in Definition \[zitra\_vedeni\]. For an isomorphism $\omega : Y \iso X$ one puts $\omega^*[\ell,e] :=
[\ell\omega,e] \in \Free^k(E)(Y)$. This turns $\Free^k(E)$ into an $\Iso$-presheaf of vector spaces. Defining formally $\Free^0(E)$ to be the trivial presheaf $\bfk$, one thus has a decomposition $$\Free(E) \cong\textstyle \bigoplus_{k \geq 0} \Free^k(E)$$ of $\Iso$-presheaves of vector spaces.
This in particular shows that $\Free(E)$ is an $\Iso$-presheaf as required in the definition of Markl’s operad. The structure operation $$\label{Zavolam_ty_Sarce?}
\circ_\phi: \Free^k(E)(X) \otimes \Free^l(E)(G) \longrightarrow
\Free^{k+l}(E)(W),\
G \fib W \stackrel\phi\to X,$$ is, for $k,l \geq 1$, is defined as the colimit of the natural isomorphisms between the functors $$E \ot E: \lTw(X) \times \lTw(G) \to \ttV \hbox { and }
E : \lTw(W) \to \ttV$$ described in Lemma \[popozitri\_Janacek\]. One must also define the structure operation in (\[Zavolam\_ty\_Sarce?\]) for $k=0$, i.e. to specify a map $$\label{zavolal_jsem_ji}
\circ_\phi: \Free^l(E)(G) \cong \Free^0(E)(X) \otimes \Free^l(E)(G)
\longrightarrow \Free^{l}(E)(W).$$
Notice first that the grade of $X$ must be zero thus, by our assumptions on $\ttO$, $X$ is a local terminal object. Consider an element $[\ell',e] \in \Free^l(E)(G)$ with $\ell' : G \iso F$ and $e \in
E(\bfF)$ with $\bfF$ as in (\[17b\]). Using the blow up axiom we embed $\phi : W \to X$ into the diagram $$\xymatrix@C=1em{
W \ar[rr]^{\ell''}_\cong\ar[rd]_\phi &&S\ar[ld]^\rho
\\
&X&
}$$ in which $\ell''$ induces the map $\ell' : G \iso F$ between the fibers. Let ${\bfS}$ be the tower as in (\[eq:2\]) with $X$ in place of $T$. Then $(\ell'',\bfS) \in \lTw(W)$. Since by construction the associated fiber sequence of $\bfS$ is the same as the associated fiber sequence of $\bfF$, one has $E(\bfF) = E(\bfS)$, thus it makes sense to define $\circ_\phi$ in (\[zavolal\_jsem\_ji\]) by $\circ_\phi ([\ell',e]) := [\ell'',e]$.
Notice that one cannot have $l = 0$ in (\[Zavolam\_ty\_Sarce?\]), since the fiber of an elementary map has always positive grade. We leave to the reader to verify that the above constructions make $\Free(E)$ a Markl’s operad.
Let us describe $\Free^1(E)$ explicitly. As noticed in Example \[vcera\_jsem\_bezel\_36\_minut\], $\lTw^1(X)$ is the category $X/\Iso$ of in $\ttO$ under $X$. Elements of $\Free^1(E)(X)$ are equivalence classes $[\omega,e]$ of pairs $\omega : X \iso T$, $e \in E[T]$, modulo the identification $[\sigma\omega' ,e''] = [\omega' ,\sigma^*e'']$ for each diagram $$\xymatrix@C=1em{
&X \ar[ld]_{\omega'}^\cong\ar[rd]^{\omega''}_\cong &
\\
T' \ar[rr]^\sigma_\cong && T''
}$$ of isomorphisms in $\ttO$. Since $\lTw^1(X)$ is connected, with a distinguished object $\id : X \to X$, the map $i : E[X] \to \Free^1(E)(X)$ given by $i(e) := [\id,e]$ for $e \in E[X]$, is an isomorphism of vector spaces. These isomorphisms assemble into an isomorphism $E \cong \Free^1(E)$ of collections. Let us finally denote by $\iota : E \hookrightarrow \Box \Free(E)$ the composition $$\iota:
E \iso \Free^1(E) \hookrightarrow \Box \Free(E).$$
To establish the freeness of Theorem \[dnes\_bylo\_40\] means to prove that, for each $1$-connected strictly extend unital Markl operad $M$ and a map of collections $y : E \to \Box M$, there exists precisely one map $\hat y : \Free(E) \to M$ of strictly extend unital Markl’s operads making the diagram $$\xymatrix@C=3em{E\ \ar@{^{(}->}[r]^(.4){\iota}\ar[rd]^y &
\Box \Free(E) \ar[d]^{\Box
\hat y}
\\
&\Box M
}$$ commutative.
Assume that such a map $\hat y : \Free(E) \to M$ exists and prove that it is unique. To this end consider an arbitrary element $[\ell, e] \in
\Free(E)(X)$ given by a pair $\ell : X \iso T$, $e \in E(\bfT)$ for a labelled tower $(\ell, \bfT) \in \lTw(X)$ as in Definition \[zitra\_vedeni\]. For $$e = e_1 \ot \cdots \ot e_k \in E[t_1] \ot \cdots \ot E[t_k]$$ it immediately follows from the definition of the operad structure of $\Free(E)(X)$ that $$[\ell, e] = \ell^*\big(e_1 \circ_{\tau_1}(e_2 \circ_{\tau_2}
\cdots(e_{k-1}
\circ_{\tau_{k-1}} e_{k}) \cdots )\big)$$ where we used the notation $$x \circ_{\tau_i} y := (-1)^{|x||y|} \circ_{\tau_i}(y \otimes x)$$ for $x \in E[T_i],\ y \in E[t_i]$ and $1 \leq i \leq k-1$. We moreover considered $\Rada e1k$ as elements of $\Free^1(E)$ via the isomorphism $i : E \iso \Free^1(E)$. Since $\hat y$ is a morphism of operads, one has $$\label{dneska_musim_zvladnout_jeste_toho_Slovaka}
\hat y([\ell, e]) = \ell^*\big(y(e_1) \circ_{\tau_1}(y(e_2) \circ_{\tau_2}
\cdots(y(e_{k-1})
\circ_{\tau_{k-1}} y(e_{k})) \cdots )\big).$$
On the other hand, one may verify that (\[dneska\_musim\_zvladnout\_jeste\_toho\_Slovaka\]) indeed defines a morphism of operads with the required property. This finishes the proof.
Quadratic Markl’s operads and duality {#pairing}
=====================================
The goal of this section is to introduce quadratic Markl’s operads in operadic categories and define their Koszul duals. The basic monoidal category $\ttV$ will be the category $\Vect$ of graded vector spaces over a field $\bfk$ of characteristic $0$. All operads will be tacitly assumed to be strictly extended unital. The basic operadic category $\ttO$ shall fulfill the same additional axioms as in Section \[zitra\_letim\_do\_Pragy\], plus the rigidity of Definition \[Porad\_nevim\_jestli\_mam\_jit\_na\_ty\_narozeniny.\]. In brief, $$\hbox{\rm \Fac\ \& \SBU\ \& \QBI\ \& \UFB\ \& \Rig},$$ plus the assumption that $\ttO$ is graded and that a morphism $f$ is an isomorphisms if $e(f) = 0$.
An [*ideal*]{} $\calI$ in a Markl’s operad $M$ is a sub-$\Iso$-presheaf of $M$ which is simultaneously an ideal with respect to the circle products (\[ten\_prelet\_jsem\_podelal\]), i.e.$$\circ_\phi(a \ot b) \in \calI(T)
\hbox { if $a\in \calI(S)$ or $b \in \calI(F)$.}$$ For a sub-$\Iso$-presheaf $R$ we denote by $(R)$ the component-wise intersection of all ideals containing $R$. We call $(R)$ the ideal [*generated*]{} by $R$.
\[zitra\_seminar\] A [*quadratic data*]{} consists of a $1$-connected collection $E \in \Coll$ and an sub-$\Iso$-presheaf $R$ of $\Free^2(E)$. A $1$-connected Markl’s operad $A$ is [*quadratic*]{} if it is of the form $$A = \Free(E)/(R).$$ It is [*binary*]{} if the generating collection $E$ is such that $E[T] \not= 0$ implies that $e(T)= 1$.
Many examples of binary operads will be given in the following sections. Let us proceed to our generalization of the operadic Koszul duality of [@ginzburg-kapranov:DMJ94] for operads in general operadic categories.
We start by noticing that the piece $\lTw^k(X)$ of height $k$ of the groupoid $\lTw(X)$ constructed in Section \[zitra\_letim\_do\_Pragy\] decomposes into the coproduct $$\lTw^k(X) = \coprod_{c \in \pi_0(\lTw^k(X))} \lTw^k_c(X)$$ over the set $ \pi_0(\lTw^k(X))$ of connected components of $\lTw^k(X)$, so it does also the $k$th piece of the $X$-component of the free operad, $$\label{Do_konce_Safari_zbyvaji_tri_dny.}
\Free^k(E)(X) = \bigoplus_{c \in \pi_0(\lTw^k(X))} \Free^k_c(E)(X).$$ Chose a labelled tower $(\ell^c,\bfT^c)$ in each connected component $c$ of $\lTw^k(X)$ and assume the notation $$\bfT^c: =
T^c \stackrel {\tau^c_1} \longrightarrow T^c_1 \stackrel
{\tau^c_2} \longrightarrow
T^c_2\stackrel {\tau^c_2} \longrightarrow
\cdots \stackrel {\tau^c_{k-1}} \longrightarrow T^c_{k-1},$$ with the associated fiber sequence $t^c_1,\ldots,t^c_{k}$. Since there are no automorphisms of $(\ell^c,\bfT^c)$ in $\lTw^k(X)$ by the rigidity of $\ttO$, $$\Free^k_c(E)(X) \cong E[t^c_1] \otimes \cdots \otimes E[t^c_k]$$ so we have an isomorphism of graded vector spaces
$$\label{dnes_bude_vedro}
\Free^k(E)(X) \cong
\bigoplus_{c \in \pi_0(\lTw^k(X))} E[t^c_1] \otimes \cdots \otimes E[t^c_k],$$
c.f. the similar presentation [@markl-shnider-stasheff:book formula (II.1.51)] for ‘ordinary’ free operads. In the light of Proposition \[22\_hodin\_cesty\], one may assume that the tower $(\ell^c,\bfT^c)$ in (\[Do\_konce\_Safari\_zbyvaji\_tri\_dny.\]) belongs to $\tlTw^k(X)$, therefore (\[dnes\_bude\_vedro\]) can be reduced into the direct sum $$\label{dnes_bude_vedro_b}
\Free^k(E)(X) \cong
\bigoplus_{c \in \pi_0(\tlTw^k(X))} E[t^c_1] \otimes \cdots \otimes E[t^c_k]$$
over isomorphism classes of objects of $\tlTw^k(X)$.
Let $\susp E^*$ be the suspension of the componentwise linear dual of the collection $E$. With the above preliminaries, it is easy to define a pairing $$\label{jdu_si_lepit_142}
\langle - | - \rangle :
\Free^2(\susp E^*)(X) \ot \Free^2(E)(X) \longrightarrow
\bfk,\ (\alpha,x) \longmapsto \alpha(x)$$ as follows. If $c' \not= c''$ we declare the subspaces $\Free_{c'}^2(\susp E^*)(X)$ and $\Free_{c''}^2(E)(X)$ of $\Free^2(\susp E^*)(X)$ resp. $\Free^2(E)(X)$ to be orthogonal. If $c := c' = c''$, we invoke that, by (\[dnes\_bude\_vedro\]), $$\Free_c^2(\susp E^*)(X) \cong \susp E^*[t^c_1] \ot \susp E^*[t^c_2]\
\hbox { and }\
\Free_c^2(E)(X) \cong E[t^c_1] \ot E[t^c_2].$$ The pairing between $\Free_c^2(\susp E^*)(X)$ and $\Free_c^2(E)(X)$ is defined as the canonical evaluation map $$\susp E^*[t^c_1] \ot \susp E^*[t^c_2] \ot E[t^c_1] \ot E[t^c_2]
\longrightarrow \bfk.$$ We leave as an exercise to show that this definition does not depend on the choices of the representatives $(\ell^c,\bfT^c)$.
Let $A$ be a quadratic Markl’s operad as in Definition \[zitra\_seminar\]. Its [*Koszul dual*]{} $A^!$ is the quadratic Markl operad defined as $$A^! = \Free(\susp E^*)/(R^\perp),$$ where $R^\perp$ denotes the component-wise annihilator of $R$ in $\Free^2(\susp E^*)$ under the pairing (\[jdu\_si\_lepit\_142\]).
\[Uz\_je\_o\_te\_medaili\_zapis.\] A quadratic Markl’s operad $A$ is [*self-dual*]{} if the associated categories of $A$- and $A^!$-algebras in $\Vect$ are isomorphic.
All assumptions of this section are met by the operadic category $\Surj$ of finite non-empty sets and their surjections. The operads for this category are the classical constant-free operads for which Koszul duality is the classical heritage [@ginzburg-kapranov:DMJ94]. A similar example is the operadic category $\Ord$ of non-empty ordered finite sets and their order-preserving surjections. Our theory in this case recovers Koszul duality for non-$\Sigma$ operads.
Modular and odd modular operads {#zitra_budu_pit_na_zal}
===============================
In this section we analyze binary quadratic operads in the operadic category $\ggGrc$ of connected genus-graded directed graphs introduced in Example \[Zaletame\_si\_jeste\_do\_konce\_Safari?\]. Recall that $\ggGrc$ was constructed from the basic category $\Gr$ of ordered graphs via the iterated Grothendieck’s construction combining a discrete operadic fibration and an opfibration, see diagram (\[Dnes\_jsme\_byli\_na\_vylete\_u\_more.\]). Since, as we established in Sections \[Ceka\_mne\_Psenicka.\] and \[Poletim\_letos\_do\_Sydney?\], the category $\Gr$ is graded and satisfies , and , $\ggGrc$ has the same properties by the results collected in Table \[analogies\]. Properties $\UFB$ and “$f$ is iso if $e(f)=0$” for $\ggGrc$ can easily be checked directly thus, by Remark \[Pujdu\_zitra\_na\_seminar?\], $\ggGrc$ has all the properties required for Koszul duality, namely $$\hbox{\rm \Fac\ \& \SBU\ \& \QBI\ \& \UFB\ \& \Rig},$$ plus the grading with the property that a morphism $f$ is an isomorphisms if $e(f) = 0$.
We will prove that the terminal operad $\term_\ggGrc$ in the category is binary quadratic and describe its Koszul dual $\oddGr :=\term_\ggGrc^!$. We then show that algebras for $\term_\ggGrc$ are modular operads of [@getzler-kapranov:CompM98] while algebras for $\term_\ggGrc^!$ are their suitably twisted versions. We start by analyzing graphs in $\ggGrc$ with small number of internal edges.
\[Snad\_se\_ta\_Achylovka\_trochu\_lepsi.\] Local terminal objects of $\ggGrc$ are genus-graded corollas $c(\sigma)^g$ for a permutation $\sigma =(\Rada\sigma1n) \in \Sigma_n$ and a genus $g \in \bbN$ depicted in Figure \[Dasa\_Vokata\]. The chosen local terminal objects are the genus-graded corollas $c_n^g :=c(\id_n)^g$ with $\id_n \in \Sigma_n$ the identity permutation.
(0,-2.6672062)(11,1.672062) (5.275,-0.8477939) (5.475,-1.2477939)[$g$]{} (5.275,1.5)[$1$]{} (5.275,1.1522061)(5.275,-0.8477939) (6.275,0.95220613)(5.275,-0.8477939) (4.275,0.95220613)(5.275,-0.8477939) (8.177993,-5.332993)[(5.275,-0.8477939)[1.2]{}[0.0]{}[220]{}]{} (6.875,0.3522061)(5.275,-0.8477939) (7.275,-0.44779387)(5.275,-0.8477939) (6.3,1.1)[$2$]{} (7,0.22061)[$3$]{} (7.4,-0.54779387)[$4$]{} (4.075,1.1)[$n$]{} (5.375,0.7522061)[$\sigma_1$]{} (6.275,0.422061)[$\sigma_2$]{} (6.875,-1)[$\sigma_4$]{} (6.75,-0.247793884)[$\sigma_3$]{} (4.475,0.7522061)[$\sigma_n$]{}
Any directed connected genus-graded graph with one internal edge and one vertex looks as $\xi(\Rada \lambda1k | \lambda_{k+1},\lambda_{k+2})^g$ in Figure \[dnes\_na\_Rusalku\] (left) with half-edges labelled by a permutation $\{\Rada \lambda1{k+2}\}$ of $\{\rada1{k+2}\}$. Its automorphism group equals $\Sigma_2$ that interchanges the half-edges forming the loop. Each two graphs of this kind are isomorphic.
(0,-1.7275)(15.055,1.9275) (5.3,-4.332993)[(5.275,-0.8477939)[1.2]{}[40]{}[95]{}]{} (2.5,-.55)[$g$]{} (11.5779,-4.332993)[(5.275,-0.8477939)[1]{}[40]{}[95]{}]{} (15.1,-4.332993)[(5.275,-0.8477939)[1]{}[205]{}[-95]{}]{} (9,-.55)[$g_u$]{} (11.6,-.55)[$g_v$]{} (4.275,0.0125)[1.0]{}[230]{}[123]{} (2.675,0.0125)(1.075,1.6125)(1.075,1.6125) (2.675,0.0125)(1.075,-1.3875)(1.075,-1.3875) (5.075,0.0125)(5.475,0.0125) (2.675,0.0125) (2.675,0.0125)(0.475,1.0125) (3.075,-0.1875)[$\relax$]{} (3.7,0.85)(2.675,0.0125) (2.675,0.0125)(3.675,-0.7875) (0.875,1.8125)[$1$]{} (0.25,1.0125)[$2$]{} (0.875,-1.5875)[$k$]{} (1.675,1.6125)[$\lambda_1$]{} (1,1.1)[$\lambda_2$]{} (1.275,-0.7875)[$\lambda_k$]{} (4.175,1.39)[$\lambda_{k+1}$]{} (4.175,-0.5875)[$\lambda_{k+2}$]{} (9.075,0.0125)(7.475,1.6125)(7.475,1.6125) (9.075,0.0125)(7.475,-1.3875)(7.475,-1.3875) (9.075,0.0125)(6.875,1.0125) (7.275,1.8125)[$l_1$]{} (6.6,1.0125)[$l_2$]{} (7.275,-1.5875)[$l_k$]{} (8,1.6125)[$\lambda^u_1$]{} (7.375,1.125)[$\lambda^u_2$]{} (7.675,-0.7875)[$\lambda^u_k$]{} (9.075,0.0125) (13.475,-1.5875)(11.875,0.0125)(11.875,0.0125) (13.475,1.4125)(11.875,0.0125)(11.875,0.0125) (14.075,-0.9875)(11.875,0.0125) (13.075,1.6125)[$\lambda^v_l$]{} (13.675,-0.3875)[$\lambda^v_2$]{} (13.475,-1.1875)[$\lambda^v_1$]{} (11.875,0.0125) (13.975,-1.7875)[$l_{k+1}$]{} (14.575,-0.9875)[$l_{k+2}$]{} (13.875,1.6125)[$l_{k+l}$]{} (10.475,0.2125)(10.475,-0.1875) (9.075,0.0125)(11.875,0.0125)(11.875,0.0125) (9.875,0.4125)[$\lambda^u_{k+1}$]{} (11.075,0.4125)[$\lambda^v_{l+1}$]{} (9.075,0.4125)[$u$]{} (11.875,0.4125)[$v$]{}
Figure \[dnes\_na\_Rusalku\] (right) depicts a general graph $\nu(\Rada {\lambda^u}1k | \lambda^u_{k+1},\lambda^v_{l+1}|\Rada{\lambda^v}1l)^{g_u|g_v}$ with one internal edge and two vertices labelled by $u,v \in \{1,2\}$ with genera $g_u,g_v \in \bbN$. Its global order is determined by a $(k,l)$-shuffle $$\{l_1 < \cdots < l_k, \ l_{k + 1} < \cdots < l_{k + l}\} = \{\rada 1{k+l}\}.$$ Its half-edges adjacent to $u$ are labelled by a permutation $\lambda^u$ of $\{\rada 1{k+1}\}$, the half-edges adjacent to $v$ by a permutation $\lambda^v$ of $\{\rada {1}{l+1}\}$. Two such graphs with the same global orders and the same genera are always isomorphic. There are no nontrivial automorphisms except for the case $k=l=0$ and $g_u = g_v$ when the graph is an interval with no legs. Then one has the automorphism flipping it around its middle.
\[chteji\_po\_mne\_abych\_velel\_grantu\] A general graph with two internal edges and one vertex is depicted in Figure \[pozitri\_do\_Srni\]. Its local order at its single vertex is determined by a permutation $\lambda$ of $\{\rada 1 {k\+4}\}$. Its automorphism group equals the semidirect product $\Sigma_2 \semidirect ( \Sigma_2 \times
\Sigma_2) $. We leave the similar detailed analysis of the remaining graphs with two internal edges as an exercise.
(2.5,-2.1219277)(8.795,2.1219277) (6.275,0.0030721934)(4.675,1.6030722)(4.475,1.8030722) (3.5,0)[ (5.3,-4.332993)[(5.275,-0.8477939)[1.2]{}[40]{}[95]{}]{} (2.6,-.65)[$g$]{} ]{} (6.275,0.0030721934) (6.275,0.0030721934)(4.075,1.2030722) (6.075,-0.5969278)[$\relax$]{} (4.275,-1.9969279)[$k$]{} (4.275,1.9969279)[$1$]{} (4.875,1.8030722)[$\lambda_1$]{} (4.475,1.30722)[$\lambda_2$]{} (4.675,-1.05)[$\lambda_k$]{} (6.275,1.6030722)[$\lambda_{k+1}$]{} (7.475,0.60307217)[$\lambda_{k+2}$]{} (-0.025787057,0.21568896)[(7.475,1.0030721)[0.8]{}[280]{}[170]{}]{} (6.275,0.0030721934)(7.675,0.20307219) (8.6,-0.3969278)[$\lambda_{k+3}$]{} (7.475,-1.3969278)[$\lambda_{k+4}$]{} (3.875,1.4030722)[$2$]{} (6.675,1.2030722)(6.275,0.0030721934) (8.106839,6.5177426)[(7.475,-0.9969278)[0.8]{}[280]{}[170]{}]{} (7.675,-0.1969278)(6.275,0.0030721934) (6.675,-1.1969278)(6.275,0.0030721934) (8, 1.3930722)(8.275,1.5969278) (8,-1.3969278)(8.275,-1.5969278) (6.275,0.0030721934)(4.475,-1.7969278)
Our next task will be to describe free operads in $\ggGrc$ using formula (\[dnes\_bude\_vedro\_b\]). As the first step towards this goal we describe isomorphism classes of labelled towers $(\ell,\bfT) \in \tlTw^2(X)$ for the directed graph $$X := \xi(\rada 1k
|{k\+1},{k\+2}|{k\+3},{k\+4})^g$$ i.e. for the graph in Figure \[pozitri\_do\_Srni\] with $\lambda$ the identity. As observed in Example \[pisu\_druhy\_den\_v\_Srni\], it suffices to consider diagrams (\[vecer\_ustrice\]) in which all maps in the upper triangle are . Since the graphs $X, T'$ and $T''$ in that triangle have one vertex only, all $\ell', \ell''$ and $\sigma_1$ belong to ${\tt \Delta}\Gr$, therefore they are the identities by Corollary \[zase\_mam\_nejaky\_tik\]. Isomorphism classes in $\tlTw^2(X)$ are thus represented by maps $$\label{dnes_na_Breznik}
\tau: \xi(\rada 1k
|{k\+1},{k\+2}|{k\+3},{k\+4} )^g
\longrightarrow \xi(\Rada \nu1k | \nu_{k+1},\nu_{k+2})^{g+1}$$ modulo the equivalence that identifies $\tau'$ with $\tau''$ if and only if there exists an isomorphism $\sigma$ such that $\tau'' = \sigma \circ
\tau'$. Notice that a map as in (\[dnes\_na\_Breznik\]) is automatically elementary, and that all elementary maps from $X$ decreasing the grade by $1$ are of this form. Now define the ‘canonical’ maps $$p_i:
\xi(\rada 1k |{k\+1},{k\+2}|{k\+3},{k\+4})^{g} \longrightarrow
\xi(\rada 1k |{k\+1},{k\+2})^{g+1}, \ i = 1,2,$$ by postulating that $p_1$ (resp. $p_2$) contracts the loop $\{{k\+1},{k\+2}\}$ (resp. $\{{k\+3},{k\+4}\})$ leaving the other loop unaffected. In other words, the injection $\psi_1$ (resp. $\psi_2$) of half-edges defining $p_1$ (resp. $p_2$) is the order-preserving injection $$(\rada 1{k\+2}) \hookrightarrow (\rada 1{k\+4})$$ that misses the subset $\{k\+1,k\+2\}$ (resp. $\{k\+3,k\+4\})$.
We claim that for each $\tau$ in (\[dnes\_na\_Breznik\]) there exist a unique $i \in \{1,2\}$ and a unique isomorphism $\sigma$ making the diagram $$\label{lyze_mi_nejedou}
\xymatrix@R=3em@C=-6em{
&\xi(\rada 1k |{k\+1},{k\+2}|{k\+3},{k\+4})^g\ar[ld]_{p_i}\ar[dr]^\tau&
\\
\xi(\rada 1k |{k\+1},{k\+2})^{g+1}\ar[rr]^\sigma_\cong
&& \xi(\Rada \nu1k | \nu_{k\+1},\nu_{k\+2})^{g+1}
}$$ commutative. Since, by definition, morphisms in $\ggGrc$ preserve global orders, one has for the injections $\psi_\tau$ resp. $\psi_\sigma$ of half-edges defining $\tau$ resp. $\sigma$, $$\psi_\tau(\nu_j) = \psi_\sigma(\nu_j) = j \ \hbox { for } \ 1\leq j \leq k.$$ Since $\psi_\tau$ must further preserve the involutions on the sets of half-edges, there are only two possibilities:
[*Case 1: $\psi_\tau\{\nu_{k+1},\nu_{k+2}\} = \{k\+3,k\+4\}$.*]{} In this case we take $i=1$ in (\[lyze\_mi\_nejedou\]) and define $$\psi_\sigma(\nu_{k+1}) := \psi_\tau(\nu_{k+1}) -2,\
\psi_\sigma(\nu_{k+2}) := \psi_\tau(\nu_{k+2}) -2.$$ It is clear that with this choice the diagram in (\[lyze\_mi\_nejedou\]) is commutative and that it is the only such a choice.
[*Case 2: $\psi_\tau\{\nu_{k+1},\nu_{k+2}\} = \{k\+1,k\+2\}$.*]{} In this case we take $i=2$ and define $$\psi_\sigma(\nu_{k+1}) := \psi_\tau(\nu_{k+1}),\
\psi_\sigma(\nu_{k+2}) := \psi_\tau(\nu_{k+2}).$$ Intuitively, in Case 1 the map $\tau$ contracts the loop $\{k\+1,k\+2\}$, in Case 2 the loop $\{k\+3,k\+4\}$. The isomorphism $\sigma$ is in both cases uniquely determined by the behavior of $\tau$ on the non-contracted edge.
The above calculation shows that there are precisely two isomorphism classes of objects of $\tlTw^2(X)$, namely those of $p_1$ and $p_2$. Notice that $$p^{-1}_1(1) = \xi(\rada 1k,k\+3,k\+4 |{k\+1},{k\+2})^g
\ \hbox { and } \
p^{-1}_2(1) = \xi(\rada 1k,{k\+1},{k\+2} |{k\+3},{k\+4})^g.$$ Let $E\in \Coll$ be a $1$-connected $\ggGrc$-collection as in Definition \[Jarka\_dnes\_u\_lekare\]. Formula (\[dnes\_bude\_vedro\]) gives $$\begin{aligned}
%\label{po_navratu_ze_Srni}
\Free^2(E)(X) \cong\ &
E[\xi(\rada 1k,{k\+3},{k\+4} |{k\+1},{k\+2})^g]
\ot E[\xi(\rada 1k |{k\+1},{k\+2})^{g+1}]
\\
\nonumber
\oplus\ & E[\xi(\rada 1k,{k\+1},{k\+2} |{k\+3},{k\+4})^g]
\ot E[\xi(\rada 1k |{k\+1},{k\+2})^{g+1}].\end{aligned}$$ Analogous expressions for $X = \xi(\Rada \lambda1k |
\lambda_{k+1},\lambda_{k+2}|\lambda_{k+3},\lambda_{k+4} )^g$ can be obtained from the above ones by substituting $j \mapsto \lambda_j$ for $1 \leq j
\leq k+4$. The result is $$\begin{aligned}
\label{po_navratu_ze_Srni_b}
\Free^2(E)(X) \cong\ &
E[\xi(\Rada \lambda1k,\lambda_{k+3},\lambda_{k+4} |\lambda_{k+1},\lambda_{k+2})^g]
\ot E[\xi(\Rada \lambda1k |\lambda_{k+1},\lambda_{k+2})^{g+1}]
\\
\nonumber
\oplus\ & E[\xi(\Rada \lambda1k,\lambda_{k+1},\lambda_{k+2} |\lambda_{k+3},\lambda_{k+4})^g]
\ot E[\xi(\Rada \lambda1k |\lambda_{k+1},\lambda_{k+2})^{g+1}].\end{aligned}$$
\[toho\_snehu\_je\_moc\] The right hand side of (\[po\_navratu\_ze\_Srni\_b\]) depends only on the virtual isomorphism classes in $\QV(e)$ of the graphs involved. By the observations made in Example \[Snad\_to\_projde\_i\_formalne.\], these classes do not depend on the global orders. In this particular case means that they do not depend on the indices $\Rada \lambda1k$; we can therefore simplify the exposition by removing them from notation and drawings. We also replace $\Rada
\lambda{k+1}{k+4}$ by less clumsy symbols $a,b,c$ and $d$. With this convention, we write the two representatives of isomorphism classes in $\tlTw^2(X)$ as: $$\begin{aligned}
%\label{dnes_volne_dopoledne}
\xi(*,c,d |a,b)^g& \fib \xi(*|a,b|c,d)^g
\stackrel {p_1} \longrightarrow
\xi(*|c,d)^{g+1},\ \hbox { and}
\\
\xi(*,a,b |c,d)^{g}& \fib \xi(*|a,b|c,d)^{g}
\stackrel {p_2} \longrightarrow
\xi(*|a,b)^{g+1},\end{aligned}$$ where $*$ stands for unspecified labels The right hand side of (\[po\_navratu\_ze\_Srni\_b\]) now takes the form $$\begin{aligned}
\label{Bude_Ovar_dalsich_pet_let prezidentem?}
\big\{
E[\xi(*,c,d |a,b)^g] \ot E[\xi(* |c,d)^{g+1}]\big\}
\oplus\big\{ E[\xi(*,a,b |c,d)^g]\ot E[\xi(* |a,b)^{g+1}] \big\}\end{aligned}$$ with the first summand corresponding to the class of $p_1$ and the second to the class of $p_2$.
We also noticed that the maps $p_1$ and $p_2$ are determined by specifying which of the two loops of $\xi(*|a,b|c,d)^g$ they contract. The map $p_1$ and its unique nontrivial fiber is thus encoded by the picture in Figure \[fig:streda\_v\_Srni\].
(0,-1.4)(14.850119,1.4) (2.4,.3)[$g$]{} (8,.3)[$g$]{} (12.4,.3)[$g\+1$]{} (0.4305688,0.822039)[(1.0,0.0)[0.8]{}[108.04867]{}[355.90524]{}]{} (0.4,0.0)(0.0,0.0) (2.4,0.0) (6.7916665,0.0)(1.9,1.4) (0.9583333,1.0)[$a$]{} (0.89166665,-0.6)[$b$]{} (7.1583333,0.8)[$a$]{} (6.2916665,-0.6)[$b$]{} (2.825,.95)[$c$]{} (2.825,-1.0666667)[$d$]{} (13.625,1.0)[$c$]{} (9.425,1.0)[$c$]{} (9.425,-0.6666667)[$d$]{} (13.625,-0.6666667)[$d$]{} (3.425,0.8)(2.425,0.0) (2.425,0.0)(3.425,-0.8) (11.225,0.2)[$p_1$]{} (4,-0.2)[$\fib$]{} (11.891666,0.06666667)(10.825,0.06666667) (2.425,0.0)(1.225,0.8) (2.425,0.0)(1.225,-0.8) (24.849737,-0.082693085)[(12.425036,0.008896015)]{} (21.6421,-11.452748)[(13.849989,-0.002628368)[0.8]{}[108.04867]{}[355.90524]{}]{} (14.449969,-0.00748074)(14.8499565,-0.010715654) (12.425036,0.008896015)(13.618526,-0.80078256) (12.425036,0.008896015)(13.631467,0.7991651) (8.025,0.0) (2.8417542,5.4254575)[(6.6,0.0)[0.8]{}[108.04867]{}[355.90524]{}]{} (6.0,0.0)(5.6,0.0) (8.025,0.0)(6.825,0.8) (8.025,0.0)(6.825,-0.8) (16.049881,-0.047109026)[(8.025036,0.008896015)]{} (14.765935,-7.815632)[(9.449989,-0.002628368)[0.8]{}[108.04867]{}[355.90524]{}]{} (10.04997,-0.00748074)(10.449957,-0.010715654) (8.025036,0.008896015)(9.218527,-0.80078256) (8.025036,0.008896015)(9.231466,0.7991651)
The pictorial expression of $p_2$ is similar.
We will use similar pictures as a language for free operads in $\ggGrc$. Thus Figure \[dnes\_jsem\_objel\_kanal\] is a pictorial version of (\[Bude\_Ovar\_dalsich\_pet\_let prezidentem?\]). It features [*souls*]{} of the relevant graphs, i.e. objects obtained by amputating their legs. The $E$’s inside the dashed circles indicate the decoration of the fiber represented by the subgraph inside the circle, while the $E$’s outside the circles the decoration of the images. Thus the left object in Figure \[dnes\_jsem\_objel\_kanal\] represents the left summand of (\[Bude\_Ovar\_dalsich\_pet\_let prezidentem?\]) and the right object the right one. This description shall be compared to the description of free ‘classical’ operads in terms of trees with decorated vertices, cf. [@markl-shnider-stasheff:book Section II.1.9]. Here we have graphs instead of trees and ‘nests’ of subgraphs directed by inclusion in place of vertices.
(0,-1.4019446)(12.320392,1.4019446) (3.2,.3)[$g$]{} (9.2,.3)[$g$]{} (5.133333,-0.0019446583)(5.5333333,-0.0019446583) (3.1333332,-0.0019446583) (0.0016315024,0.072724074)[(1.7083331,-0.0019446583)[0.8]{}[55.279263]{}[317.44205]{}]{} (1.1083331,-0.0019446583)(0.70833313,-0.0019446583) (1.8999997,-0.0019446583)(1.9,1.4) (6.0,-0.13527799)[$\oplus$]{} (3.1333332,-0.0019446583)(1.9333332,0.79805535) (3.1333332,-0.0019446583)(1.9333332,-0.8019447) (9.025159,-0.39291084)[(4.5204096,-0.01659256)[0.8]{}[55.279263]{}[317.44205]{}]{} (3.0963898,0.03624849)(4.265899,-0.807699) (3.1333332,-0.0019446583)(4.333333,0.79805535) (18.374098,0.030811006)[(9.187082,-0.0043584616)]{} (21.216465,-0.40032578)[(10.612064,0.002911593)[0.8]{}[55.279263]{}[317.44205]{}]{} (11.212056,0.0059726685)(11.612051,0.008013386) (20.840694,0.048701778)[(10.4204,0.0019337495)(1.9,1.4)]{} (9.187082,-0.0043584616)(10.391149,-0.7982259) (9.187082,-0.0043584616)(10.382985,0.8017533) (0.021902805,0.5873463)[(7.799949,0.003212673)[0.8]{}[55.279263]{}[317.44205]{}]{} (9.22422,-0.042362634)(8.050421,0.7956073) (9.187082,-0.0043584616)(7.9911795,-0.8104702) (6.833333,-0.0019446583)(7.133333,-0.0019446583) (0.33333313,0.19805534)[$E$]{} (0,0.99805534)[$E$]{} (11.6,0.19805534)[$E$]{} (12,0.99805534)[$E$]{} (2.1333332,0.79805535)[$a$]{} (7.5333333,0.99805534)[$a$]{} (1.5333332,-0.6019447)[$b$]{} (7.733333,-0.6019447)[$b$]{} (9.533333,0.59805536)[$c$]{} (4.733333,0.99805534)[$c$]{} (4.333333,-0.6019447)[$d$]{} (10.533333,-0.6019447)[$d$]{}
\[Spregner\] Using the same reasoning as in Examples \[chteji\_po\_mne\_abych\_velel\_grantu\] and \[toho\_snehu\_je\_moc\], we can draw similar pictures describing $\Free^2(E)(X)$ for $X$ a graph with two internal edges and two vertices with genera $g_1$ and $g_2$. Their souls are treated in Figure \[Zeman\_nebo\_Drahos?\].
(0,-1.9412143)(13.26,1.9412143) (4.696944,1.3197962)[(3.0,-1.6587857)[2.6]{}[130.21674]{}[230.1]{}]{} (1.0,-0.058785707) (5.0,-0.058785707) (0.2,-2.1254523)[0.26666668]{}[131.51372]{}[131.67642]{} (-0.299674,-1.0204285)[(2.9666667,-1.4254524)[1.7666667]{}[15.0]{}[121.2717]{}]{} (8.168101,-4.099964)[(5.0333333,-0.1587857)[0.6333333]{}[89.0]{}[260.0]{}]{} (1.5806634,-1.0495273)[(1.0333333,-0.1587857)[0.6333333]{}[21.63039]{}[204.26073]{}]{}
(3.0333333,-1.692119)[3.1]{}[39.0]{}[143.0]{} (5.7254524,8.741215)[(7.233333,1.5078809)[0.26666668]{}[131.51372]{}[131.67642]{}]{} (4.489549,-1.4847707)[(3.0,1.5412143)[2.6]{}[132.3]{}[230.1]{}]{} (25.716059,4.203225)[(12.861101,2.0828211)[0.26666668]{}[131.51372]{}[131.67642]{}]{} (11.644145,-8.692223)[(10.266661,1.6078811)[2.6]{}[130.21674]{}[230.1]{}]{} (24.532843,-0.102453604)[(12.266668,0.007890621)]{} (16.533028,-0.06393689)[(8.266668,0.007871618)]{} (19.667742,5.8717227)[(10.299994,1.374548)[1.7666667]{}[15.0]{}[121.2717]{}]{} (2.8853579,6.368748)[(8.233335,0.10787146)[0.6333333]{}[89.0]{}[260.0]{}]{} (4.327428,9.44399)[(12.233335,0.107890464)[0.6333333]{}[21.63039]{}[204.26073]{}]{} (20.481997,3.1837165)[(10.233327,1.6412143)[3.1]{}[39.0]{}[143.0]{}]{} (12.133192,8.628233)[(10.266676,-1.5921189)[2.6]{}[132.3]{}[230.1]{}]{} (4.4,-1.2587857)[$E$]{} (5.2,-0.4587857)[$E$]{} (12.4,-0)[$E$]{} (13.0,0.5412143)[$E$]{} (3.0,1.1412143)(3.0,0.7412143) (10.2,-0.8587857)(10.2,-1.2587857) (10.2,1.1412143)(10.2,0.7412143) (3.0,-0.8587857)(3.0,-1.2587857) (2.0,0.35)[$a$]{} (2.0,-0.6587857)[$c$]{} (4.2,0)[$b$]{} (3.8,-0.6587857)[$d$]{} (9.0,0.9412143)[$a$]{} (11.2,0.9412143)[$b$]{} (9.2,-0.6587857)[$c$]{} (11.0,-0.6587857)[$d$]{} (6.4,-0.2587857)[$\oplus$]{}
(0,-1.4000019)(12.4,1.4000019) (-.2,0)[ (4.8,-0.0038784079)(5.2,-0.0038784079) (2.8,-0.0038784079) (6.266667,-0.13721174)[$\oplus$]{} (8.359396,-0.3702423)[(4.1870766,-0.01852631)[0.8]{}[55.279263]{}[317.44205]{}]{} (2.7630565,0.03431474)(3.9325662,-0.8096328) (2.8,-0.0038784079)(4.0,0.7961216) (18.17405,0.039097566)[(9.087067,0.0)(1.7,1.0)]{} (0.0,0.19612159)[$E$]{} (3.0,-1.0038784)[$E$]{} (0.6,-0.0038784079) (0.6,-0.0038784079)(2.8,-0.0038784079) (1.8,0.19612159)(1.8,-0.2038784) (12.0,-0.0038784079)(12.4,-0.0038784079) (10.0,-0.0038784079) (22.736555,-0.9432886)[(11.387076,-0.01852631)[0.8]{}[55.279263]{}[317.44205]{}]{} (9.963057,0.03431474)(11.132566,-0.8096328) (10.0,-0.0038784079)(11.2,0.7961216) (8.2,-0.6038784)[$E$]{} (10.4,0.9961216)[$E$]{} (7.8,-0.0038784079) (7.8,-0.0038784079)(10.0,-0.0038784079) (9.0,0.19612159)(9.0,-0.2038784) (7.7740974,0.016724302)[(3.8870666,0.0)(1.5,1.4)]{} (1.2,0.19612159)[$a$]{} (8.4,0.19612159)[$a$]{} (2.2,0.19612159)[$b$]{} (9.4,0.19612159)[$b$]{} (4.0,0.9961216)[$u$]{} (4.2,-0.6038784)[$v$]{} (11.4,0.9961216)[$u$]{} (11.2,-0.6038784)[$v$]{} ]{}
The picture in the upper half represents an analog of (\[Bude\_Ovar\_dalsich\_pet\_let prezidentem?\]): $$\begin{aligned}
\Free^2(E)(X) \cong& \
\big\{
E[\nu(*,c|a,b|d,*)^{g_1|g_2}] \ot E[\xi(*|c,d)^{g_1+g_2}]\big\}
\\
\oplus & \ \big\{ E[\nu(*,a|c,d|b,*)^{g_1|g_2}] \ot
E[\xi(* |a,b)^{g_1+g_2}]\big\}\end{aligned}$$ in which the notation $\nu(*,c|a,b|d,*)$ resp. $\nu(*,a|c,d|b,*)$ refers to the right graph in Figure \[dnes\_na\_Rusalku\]. The lower half of Figure \[Zeman\_nebo\_Drahos?\] symbolizes $$\begin{aligned}
\Free^2(E)(X) \cong& \
\big\{
E[\xi(*,b|u,v)^{g_1}] \ot E[\nu(*|a,b|*)^{g_1|g_2+1}]\big\}
\\
\oplus& \ \big\{ E[\nu(*|a,b|u,v,*)^{g_1|g_2}]
\ot E[\xi(* |u,v)^{g_1+g_2}]\big\}.\end{aligned}$$
The last relevant case is when $X$ is a directed graph with two internal edges and three vertices with genera $g_1,g_2$ and $g_3$. The situation is portrayed in Figure \[Mozna\_bude\_Ovara\_volit\_i\_Jana.\].
(0,-.9412143)(13.052372,1) (0.8,-0.058785707) (1.4,-0.6587857)[$E$]{} (0.3,0.74)[$E$]{} (3.2,-0.058785707) (0.2,-2.1254523)[0.26666668]{}[131.51372]{}[131.67642]{}
(5.7254524,8.741215)[(7.233333,1.5078809)[0.26666668]{}[131.51372]{}[131.67642]{}]{} (25.716059,4.203225)[(12.861101,2.0828211)[0.26666668]{}[131.51372]{}[131.67642]{}]{} (5.6,-0.058785707) (0.8,-0.058785707)(5.6,-0.058785707) (7.2,-0.058785707) (12.4,0.7412143)[$E$]{} (11.0,-0.6587857)[$E$]{} (9.6,-0.058785707) (12.0,-0.058785707) (7.2,-0.058785707)(12.0,-0.058785707) (6.266667,-0.2587857)[$\oplus$]{} (2.0,-0.058785707)(1.48,1) (10.8,-0.058785707)(1.48,1) (4.4,0.1412143)(4.4,-0.2587857) (8.4,0.1412143)(8.4,-0.2587857) (10.8,0.1412143)(10.8,-0.2587857) (2.0,0.1412143)(2.0,-0.2587857) (7.8,0.1412143)[$a$]{} (1.4,0.1412143)[$a$]{} (2.4,0.1412143)[$b$]{} (8.8,0.1412143)[$b$]{} (3.8,0.1412143)[$c$]{} (10.2,0.1412143)[$c$]{} (5.0,0.1412143)[$d$]{} (11.4,0.1412143)[$d$]{}
The resulting formula is $$\begin{aligned}
\Free^2(E)(X) \cong& \
\big\{
E[\nu(*|a,b|c,*)^{g_1|g_2}] \ot E[\nu(* |c,d|*)^{g_1+g_2|g_3}]\big\}
\\
\oplus &\
\big\{ E[\nu(*,b|c,d|*)^{g_2|g_3}] \ot E[\nu(* |a,b|*)^{g_1|g_2+g_3}]\big\}.\end{aligned}$$
Observations in Examples \[toho\_snehu\_je\_moc\] and \[Spregner\] easily generalize to descriptions of isomorphism classes of labeled towers in $\lTw(\Gamma)$ for an arbitrary graph $\Gamma \in \ggGrc$. Since we will be primarily interested in free operads generated by $1$-connected collections. i.e. collections that are trivial on graphs with more than one internal edge, we will consider only towers whose associated fiber sequence consists of graphs with one internal edge. Let $$\Gamma \stackrel {\tau_1} \longrightarrow \Gamma_1 \stackrel {\tau_2} \longrightarrow
\Gamma_2\stackrel {\tau_2} \longrightarrow
\cdots \stackrel {\tau_{k-1}} \longrightarrow \Gamma_{k-1}$$ be such a tower. By the definition of graph morphisms, one has the associated sequence $$\label{ta_zruda_bude_prezident}
\edg(\Gamma) \supset \edg(\Gamma_1) \supset \cdots \supset \edg(\Gamma_{k-1})$$ of inclusions of the sets of internal edges. Since the cardinalities of the sets in (\[ta\_zruda\_bude\_prezident\]) decrease by one, there is an obvious one-to-one correspondence sequences (\[ta\_zruda\_bude\_prezident\]) and linear orders on $\edg(\Gamma)$ such that $x > y$ if $y\in \edg(\Gamma_i)$ while $x \not\in \edg(\Gamma_i)$ for some $i$, $1 \leq i \leq k-1$. We formulate:
\[uz\_melo\_prvni\_cislo\_davno\_vyjit\] The isomorphism classes of labeled towers in $\ltw(\Gamma)$ whose associated fiber sequence consists of graphs with one internal edge are in one-to-one correspondence with linear orders of $\edg(\Gamma)$ modulo the relation $\bowtie$ that interchanges two adjacent edges that do not share a common vertex in $\Gamma$.
One has two isomorphism classes of towers for the graph in Figure \[pozitri\_do\_Srni\]. Let, in the notation of Figure \[dnes\_jsem\_objel\_kanal\], $x$ be the edge $\{a,b\}$ and $y$ the edge $\{c,d\}$. Then the left picture in that figure corresponds to the order $x > y$ ($x$ is contracted first), the right one to $y > x$.
Using the same arguments as in Examples \[toho\_snehu\_je\_moc\] and \[Spregner\] we show that each tower can be replaced withing its isomorphism class by the one whose all morphisms are pure contractions, in the sense of Definition \[Ron\_v\_Praze\], of internal edges. Such towers are determined by the order in which the edges are contracted. The relation $\bowtie$ reflects morphisms of towers of the second type introduced in Section \[zitra\_letim\_do\_Pragy\].
\[v\_nedeli\_si\_snad\_zaletam\_s\_NOE\] The terminal $\ggGrc$-operad $\termGr$ having $\termGr(\Gamma) := \bfk$ for each $\Gamma \in \ggGrc$ and constant structure operations is quadratic binary.
Let us define a collection $E \in \Coll$ by $$\label{vcera}
E[\Gamma] :=
\begin{cases}
\bfk & \hbox {if $\Gamma$ has exactly one internal edge, and}
\\
0 & \hbox {otherwise}
\end{cases}$$ with the constant $\QV(e)$-presheaf structure. As we already noticed, the quadratic part $\Free^2(E)$ of the free operad may be nontrivial only for graphs with precisely two internal edges, i.e. those analyzed in Examples \[toho\_snehu\_je\_moc\] and \[Spregner\]. For $X$ as in Figure \[pozitri\_do\_Srni\], formula (\[Bude\_Ovar\_dalsich\_pet\_let prezidentem?\]) describes $\Free^2(E)(X)$ as the two-dimensional space $\bfk \oplus \bfk$ with the basis $$\label{zni_mi_hlavou_Joe_Karafiat}
b_1^1 : =[ 1\ot 1] \oplus [0 \ot 0] \ \hbox { and } \
b_2^1 : = [0\ot 0] \oplus [1 \ot 1].$$ For the situations portrayed in Figures \[Zeman\_nebo\_Drahos?\] and \[Mozna\_bude\_Ovara\_volit\_i\_Jana.\] we get similar spaces, with bases $(b^t_1,b^t_2)$, $2 \leq t \leq 4$. We define $R$ to be the subspace of $\Free^2(E)$ spanned by $$\label{Krammer}
r_1:= b_1^1 - b_2^1, \ r_2:= b_1^2 - b_2^2, \ r_3:= b_1^3 - b_2^3
\hbox { and } \
r_4:= b_1^4 - b_1^4$$ so that – $r_1$ belongs to the direct sum in Figure \[dnes\_jsem\_objel\_kanal\], – $r_2$ belongs to the direct sum in the upper part of Figure \[Zeman\_nebo\_Drahos?\], – $r_3$ belongs to the direct sum in the lower part of Figure \[Zeman\_nebo\_Drahos?\] and – $r_4$ belongs to the direct sum in Figure \[Mozna\_bude\_Ovara\_volit\_i\_Jana.\]. We are going to prove that $$\label{dnes_jsem_predsedal_vedecke_rade}
\termGr \cong \Free(E)/(R).$$
By Proposition \[uz\_melo\_prvni\_cislo\_davno\_vyjit\] combined with formula (\[dnes\_bude\_vedro\]), the vector space $\Free(E)(\Gamma)$ is spanned by the set of total orders of $\edg(\Gamma)$ modulo the relation $\bowtie$ that interchanges arbitrary two edges $x,y \in \edg(\Gamma)$ that [*do not*]{} share a common vertex in $\Gamma$.
All possible relative configurations of edges $x,y$ that [*do share*]{} a common vertex are in Figures \[dnes\_jsem\_objel\_kanal\]–\[Mozna\_bude\_Ovara\_volit\_i\_Jana.\]. Relations in (\[Krammer\]) guarantee that two orders that differ by the interchange $x \leftrightarrow y$ agree in the quotient (\[dnes\_jsem\_predsedal\_vedecke\_rade\]). We conclude that all orders of $\edg(\Gamma)$ are mutually equivalent modulo $(R)$, so $\Free(E)/(R)(\Gamma) \cong \bfk$ as required.
\[pozitri\_turnaj\_v\_Patku\] Algebras over the terminal $\ggGrc$-operad $\termGr$ are modular operads.
The key ingredients of the proof are presentation (\[dnes\_jsem\_predsedal\_vedecke\_rade\]) together with Proposition \[Udelal\_jsem\_si\_ciruvky\_zelanky.\] that describes $\termGr$-algebras as morphisms into the endomorphism operad. We start by determining what the underlying collection (\[Jarka\_hovori\]) of the endomorphism operad is in this case.
We noticed in Example \[Snad\_se\_ta\_Achylovka\_trochu\_lepsi.\] that the local terminal objects of $\ggGrc$ are the $n$-corollas $c(\sigma)^g$ with the vertex of genus $g$ and the local order given by a permutation $\sigma \in \Sigma_n$. The chosen local terminal objects are the $n$-corollas $c_n^g:=c(\id_n)^g$, $n,g \in \bbN$. Therefore the set $\pi_0(\ggGrc)$ is indexed by couples $(n;g)$ of natural numbers consisting of an ‘arity’ $n$ and ‘genus’ $g$. The underlying collection of the endomorphism operad is thus a family $$\oM = \big\{\oM(n;g) \in \Vect\ \vrt \ (n;g) \in \bbA \times \bbA \big\}.$$ Actions (\[Asi\_jsem\_dostal\_premii!\]) of the groupoid of local terminal objects in this particular case give rise to actions of the symmetric group $\Sigma_n$ on each $\oM(n;g)$. We recognize $\oM$ as the skeletal version of a modular module recalled in Appendix \[a1\]. Proposition \[Udelal\_jsem\_si\_ciruvky\_zelanky.\] now identifies $\termGr$-algebras with morphisms $$\label{nevydrzim_dele_nez_tri_dny}
a: \Free(E)/(R) \to \End_\oM,$$ where $E$ is as in (\[vcera\]) and $R$ is spanned by relations (\[Krammer\]).
By Proposition \[Za\_14\_dni\_LKDL\], the extended unital operad $\End_\oM$ determines a $\QV(e)$-presheaf $\oEnd_\oM$. Although $\End_\oM$ is not strictly extended unital, morphism (\[nevydrzim\_dele\_nez\_tri\_dny\]) is still uniquely determined by a map $\tilde a : E \to \oEnd_\oM$ of $\QV(e)$-presheaves given by a family $$\label{slepit_vrtulku_nebude_snadne}
\tilde a_{[\Gamma]} : E[\Gamma] \to \oEnd_\oM([\Gamma]),\ [\Gamma] \in \QV(e).$$
By definition, the generating collection $E$ is supported on graphs with one internal edge portrayed in Figure \[dnes\_na\_Rusalku\], whose souls are shown in Figure \[hluboka\_deprese\_z\_voleb\].
(0,-1.0434394)(7.1108775,1.0434394) (.15,0.2)[$g_1$]{} (2.3,0.2)[$g_2$]{} (4.7,0.2)[$g$]{} (2.3108778,-0.11656067) (0.110877685,-0.11656067) (0.110877685,-0.11656067)(2.3108778,-0.11656067) (1.3108777,0.08343933)(1.3108777,-0.31656066) (0.71087766,0.08343933)[$a$]{} (1.7108777,0.08343933)[$b$]{} (6.710878,-0.11656067)(7.1108775,-0.11656067) (4.710878,-0.11656067) (12.166121,-0.7473357)[(6.0979543,-0.13120857)[0.8]{}[55.279263]{}[317.44205]{}]{} (4.6739345,-0.07836752)(5.843444,-0.92231506) (4.710878,-0.11656067)(5.9108777,0.6834393) (5.9108777,0.8834393)[$u$]{} (5.9108777,-0.71656066)[$v$]{} (3.3108778,-0.11656067)[or]{}
The operations $\tilde a_{[\Gamma]}$ may therefore be nontrivial only for graphs of this form.
Let us analyze the operation induced by the virtual isomorphism class of the left graph $\Gamma := \xi(\Rada \lambda1k | \lambda_{k+1},\lambda_{k+2})$ in Figure \[dnes\_na\_Rusalku\]. One clearly has $\pi_0(s_1(\Gamma)) = (k\+2;g)$ and $\pi_0(\Gamma) = (k;g\+1)$, therefore $\tilde a_{[\Gamma]}$ is by (\[V\_sobotu\_budu\_letat\_vycvik\_vlekare.\]) a map $$\tilde a_{[\Gamma]} : E[\Gamma]
= \bfk \longrightarrow
\colim_{\sigma \in\Sigma_k}
\Vect\big(\oM(k\+2;g),\oM(k;g\+1)_\sigma\big),$$ where $\sigma = (\Rada \sigma1k)$ and $\oM(k;g\+1)_\sigma$ is the copy of $\oM(k;g\+1)$ corresponding to the graph $\xi(\Rada \sigma 1k | \lambda_{k+1},\lambda_{k+2})$, virtually isomorphic to $\xi(\Rada \lambda1k | \lambda_{k+1},\lambda_{k+2})$. The map $\tilde a_{[\Gamma]}$ is determined by $$\tilde a_{[\Gamma]}(1) : \oM(k\+2;g)\to \colim_{\sigma\in \Sigma_k} \oM(k;g\+1)_\sigma$$ which is the same as a collection of morphisms $$\xxi^\sigma_{uv} :\oM(k\+2;g)
\longrightarrow \oM(k;g\+1),\ u : = \lambda_{k+1},\ v :=
\lambda_{k+2},\
\sigma \in \Sigma_k,$$ satisfying $$\xxi^{\sigma \delta}_{uv}(x) = \sigma \xxi^{\delta}_{uv}(x),\
x \in \oM(k\+2;g), \ \sigma, \delta \in \Sigma_k.$$
The operation $\xxi_{uv} := \xxi^{\id_k}_{uv}$ is the skeletal version of the contraction (\[Galway\]). The identity $\xxi_{uv} = \xxi_{vu}$ follows from the $\Sigma_2$-symmetry of the graph $\Gamma$. In exactly the same manner, the right graph in Figure \[dnes\_na\_Rusalku\] gives rise to the operations in (\[v\_Galway1\]).
The map $\tilde a$ determines a morphism (\[nevydrzim\_dele\_nez\_tri\_dny\]) if and only if it sends the generators (\[Krammer\]) of $R$ to $0$. The vanishing $\tilde
a(r_i) = 0$ for $\leq i \leq 4$ corresponds to the remaining axiom of modular operads:
[ll]{} , & ,\
&
------------------------------------------------------------------------
This finishes the proof.
\[vcera\_jsem\_podlehl\] The Koszul dual of the operad $\termGr$ is the operad $\oddGr$ whose algebras are odd modular operads.
The Koszul dual $\oddGr := \termGr^!$ is, by definition, generated by the collection $$\susp E^* :=
\begin{cases}
\susp\, \bfk & \hbox {if $\Gamma$ has exactly one internal edge, and}
\\
0 & \hbox {otherwise} .
\end{cases}$$ We get the similar type of generators $d^i_1,d^i_2$, $1 \leq i \leq
4$, for $\Free^2(\susp E^*)$ as in the proof of Theorem \[v\_nedeli\_si\_snad\_zaletam\_s\_NOE\] except that now they will be in degree $2$. The pairing (\[jdu\_si\_lepit\_142\]) is in this particular case given by $$\langle \ b^i_k\ |\ d^j_l\ \rangle =
\begin{cases}
1 & \hbox {if $i=j$, $k=l$, and}
\\
0 & \hbox {otherwise}.
\end{cases}$$ therefore the annihilator $R^\perp$ of the relations (\[Krammer\]) is spanned by $$o_1:= d_1^1 + d_2^1, \ o_2:= d_1^2 + d_2^2, \ o_3:= d_1^3 + d_2^3
\hbox { and } \
o_4:= d_1^4 + d_2^4.$$ Repeating the arguments in the proof of Theorem \[v\_nedeli\_si\_snad\_zaletam\_s\_NOE\] we identify algebras over $\Free(\susp E^*)/(R^\perp)$ with odd modular operads whose definition is recalled in Appendix \[a1\].
\[zitra\_vycvik\_vlekare\] As observed in Example \[Zaletame\_si\_jeste\_do\_konce\_Safari?\], the category $\ggGrc$ is similar to the category of graphs of [@getzler-kapranov:CompM98 §2.15]. The difference is the presence of the local orders of graphs in $\ggGrc$ manifested e.g. by the fact that, while the category in [@getzler-kapranov:CompM98 §2.15] has only one local terminal object for each arity $n$ and genus $g$, the local terminal objects in $\ggGrc$ are indexed by and by a permutation $\sigma \in \Sigma_n$, cf. Example \[Snad\_se\_ta\_Achylovka\_trochu\_lepsi.\]. Up to this subtle but important difference, operads for the operadic category $\ggGrc$ are hyperoperads in the sense of [@getzler-kapranov:CompM98 §4.1].
This relation enables one compare the operad $\oddGr$ of Theorem \[vcera\_jsem\_podlehl\] to a similar object considered in [@getzler-kapranov:CompM98]. Recall that a [*determinant*]{} $\det(S)$ of a finite set $S$ is the top-dimensional piece of the exterior (Grassmann) algebra generated by the elements of $S$ placed in degree $+1$. In particular, $\det(S)$ is an one-dimensional vector space concentrated in degree $k$, with $k$ the cardinality of $S$. Mimicking the arguments in the second half of the proof of Theorem \[v\_nedeli\_si\_snad\_zaletam\_s\_NOE\] one can establish that $\oddGr(\Gamma) \cong \det(\edg(\Gamma))$, the determinant of the set of internal edges of $\Gamma$. This relates $\oddGr$ directly to the [*dualizing cocycle*]{} of [@getzler-kapranov:CompM98 §4.8], cf. also Example II.5.52 of [@markl-shnider-stasheff:book].
Other operad-like structures {#Asi_pojedu_vecer.}
============================
In this section we analyze other operad-like structures whose pasting schemes are obtained from the basic operadic category $\Gr$ of graphs by means of the iterated Grothendieck construction. For all these categories the properties $\UFB$ and “$f$ is iso if $e(f)=0$” can be easily checked ‘manually.’ By the reasoning of the beginning of Section \[zitra\_budu\_pit\_na\_zal\] they thus fulfill all the properties required for Koszul duality
${{\hbox{$(\!($}}}\def$
Cyclic operads
--------------
Cyclic operads introduced in [@getzler-kapranov:CPLNGT95] are, roughly speaking, modular operads without the genus grading and contractions (\[Galway\]). Explicitly, a cyclic operad is a functor $\oC : \Fin \to
\Vect$ along with operations $$\label{v_Galway_jsem_byl}
{\sideset{_{a}}{_{b}}{\mathop{\circ}}}:\oC\big(S_1 \sqcup \stt a \big)
\otimes \oC\big(S_2\sqcup \stt b\big)
\longrightarrow \oC ( S_1\sqcup S_2)$$ defined for arbitrary disjoint finite sets $S_1$, $S_2$ and symbols $a,b$. These operations shall satisfy axioms (\[piji\_caj\_z\_Jarcina\_hrnku\]), (\[eq:24\]) and (\[Dnes\_s\_Jaruskou\_k\_Pakouskum\]) of modular operads (without the genus grading). Let $\CGr$ be the full subcategory of $\Gr$ consisting of graphs of genus zero whose geometric realizations are contractible, i.e. which are trees. The local terminal objects of $\Tr$ are corollas $c(\sigma)$, $\sigma \in
\Sigma_n$, as in Figure \[Dasa\_Vokata\] but without the genus labeling the vertex. The chosen local terminal objects are corollas $c_n := c(\id_n)$, $n \in \bbN$.
\[Zitra\_letim\_do\_Bari.\] The terminal $\CGr$-operad $\termCGr$ is binary quadratic. Its algebras are cyclic operads. Its Koszul dual $\termCGr^!$ is the operad $\oddCGr$ governing anticyclic operads.
Anticyclic operads introduced in [@getzler-kapranov:CPLNGT95 §2.11] are ‘odd’ versions of cyclic operads, see also [@markl-shnider-stasheff:book Definition II.5.20]. Due to the absence of the operadic units in our setup, the category of anticyclic operads is however isomorphic to the category of ordinary cyclic operads, via the isomorphism given by the suspension of the underlying collection.
The proof is a simplified version of calculations in Section \[zitra\_budu\_pit\_na\_zal\]. The soul of the only graph in $\CGr$ with one internal edge is the left one in Figure \[hluboka\_deprese\_z\_voleb\] (without the genera, of course), the corresponding operation is (\[v\_Galway\_jsem\_byl\]). The souls of the only graphs in $\CGr$ with two internal edges are portrayed in Figure \[Mozna\_bude\_Ovara\_volit\_i\_Jana.\]. Let $E$ be the restriction of the collection (\[vcera\]) to the virtual isomorphism classes of trees in $\CGr$. If $R$ denotes the subspace of $\Free^2(E)$ spanned by $r_2$ in (\[Krammer\]), then $\termCGr
\cong \Free(E)/(R)$. The arguments are the same as in the proof of Theorem \[v\_nedeli\_si\_snad\_zaletam\_s\_NOE\]. With the material of Section \[zitra\_budu\_pit\_na\_zal\] at hand, the identification of $\termCGr$-algebras with cyclic operads is immediate.
Algebras over $\oddCGr = \termCGr^!$ can be analyzed in the same way as $\oddGr$-algebras in the proof of Theorem \[vcera\_jsem\_podlehl\]. $\oddCGr$-algebras posses degree $+1$ operations $$\label{v_Galway_jsem_byl_jednou}
{\sideset{_{a}}{_{b}}{\mathop{\bullet}}}:\oC\big(S_1 \sqcup \stt a \big)
\otimes \oC\big(S_2\sqcup \stt b\big)
\longrightarrow \oC ( S_1\sqcup S_2)$$ satisfying non-genus graded variants of (\[neni\]), (\[kdy\_zacnu\_byt\_rozumny\]) and (\[neni1\]). The level-wise suspension $\susp \oC$ with operations $${\sideset{_{a}}{_{b}}{\mathop{\circ}}}:\susp \oC\big(S_1 \sqcup \stt a \big)
\!\otimes\! \susp \oC\big(S_2\sqcup \stt b\big) \longrightarrow
\susp \oC ( S_1\sqcup S_2)$$ defined as the composition $$\susp \oC\big(S_1 \sqcup \stt a \big)
\!\otimes\! \susp \oC\big(S_2\sqcup \stt b\big) \stackrel{\susp \ot
\susp}\longrightarrow \oC\big(S_1 \sqcup \stt a \big)
\!\otimes\! \oC\big(S_2\sqcup \stt b\big) \stackrel{{\sideset{_{a}}{_{b}}{\mathop{\bullet}}}}\longrightarrow
\oC ( S_1\sqcup S_2) \stackrel{\susp}\to \susp \oC ( S_1\sqcup S_2)$$ can easily be shown to be an anticyclic operad [@markl-shnider-stasheff:book Definition II.5.20].
As in Remark \[zitra\_vycvik\_vlekare\], one may observe that $\oddCGr(T)$ equals the determinant of the set of internal edges of the tree $T$. Our description of anticyclic operads as $\oddCGr$-algebras is therefore parallel to their definition as ${\mathbb T}_-$-algebras given in [@getzler-kapranov:CPLNGT95 page 178].
Ordinary operads {#Cinani_jsou_hovada.}
----------------
Let us consider a variant $\RTre$ of the operadic category $\Tre$ consisting of trees that are rooted in the sense explained in Example \[Zitra\_s\_Jarkou\_k\_Bartosovi.\]. By definition, the output half-edge of each vertex is the minimal element in the local order; we will denote this minimal element in the context of rooted trees by $0$. We use the same convention also for the smallest leg in the global order, i.e. for the root. Since $\RTre$ was obtained from the basic operadic category $\Gr$ by Grothendieck’s construction, it is again an operadic category sharing all nice properties of $\Gr$.
\[markl1\] The terminal $\RTre$-operad $\termRTre$ is binary quadratic. Its algebras are nonunital Markl’s operads recalled in Definition \[b3\] of Appendix \[a1\]. The category of algebras over its Koszul dual $\oddRTre := \termRTre^!$ is isomorphic to the category of Markl’s operads, via the isomorphism given by the suspension of the underlying collection.
(0,-0.32000703)(8.167188,0.32000703) (0.083594486,-0.121504106) (3.283593,-0.11848185) (1.7738773,0.08000693)(1.7741597,-0.3199929) (0.08359377,-0.11999298)(3.2835937,-0.11999298) (0.08359377,-0.11999298)(1.8,-0.11999298) (2.4835937,0.08000702)[$i$]{} (0.8835938,0.08000702)[$0$]{} (5.2835946,-0.121504106) (8.083593,-0.11848185) (5.2835937,-0.11999298)(8.083593,-0.11999298) (6.6835938,0.08000702)[$i$]{}
The soul of graphs in $\RTre$ with one internal edge is the oriented interval consisting of two oriented half-edges portrayed in Figure \[je\_patek\_a\_melu\_z\_posledniho\] (left). Since the label of the out-going half-edge is always the minimal one in the local order, we omit it from pictures and draw the internal edges as arrows acquiring the label of the in-going half-edge, see Figure \[je\_patek\_a\_melu\_z\_posledniho\] (right).
(0,-1.7993219)(12.822262,1.7993219) (0.36226174,-0.6006781)(1.5622617,0.5993219) (1.5622624,0.59781075) (2.5622618,-0.6006781)(1.5622617,0.5993219) (3.7622616,-0.6006781)(4.9622617,0.5993219) (4.9622626,0.59781075) (5.9622617,-0.6006781)(4.9622617,0.5993219) (0.26521635,-0.66406024)[(0.96226174,0.0)(1.2,0.6)]{} (8.761535,-4.147821)[(5.362262,0.0)(1.2,0.6)]{} (0.36226246,-0.60218924) (5.9622626,-0.60218924) (3.7622623,-0.60218924) (2.5622625,-0.60218924) (2.9622617,0.1993219)[$\oplus$]{} (9.762261,-1.4006782)(9.762261,0.0) (9.762262,1.3978108) (9.762262,-0.0021892267) (9.762262,-1.4021893) (9.762261,0.0)(9.762261,1.3993219) (11.962262,-1.4006782)(11.962262,0.0) (11.962262,1.3978108) (11.962262,-0.0021892267) (11.962262,-1.4021893) (11.962262,0.0)(11.962262,1.3993219) (9.762261,-0.7006781)(0.6,1.1) (11.962262,0.6993219)(0.6,1.1) (10.762261,0.0)[$\oplus$]{} (0.56226176,0.0)[$a$]{} (2.1622617,0.0)[$b$]{} (3.9622617,0.0)[$c$]{} (5.5622616,0.0)[$d$]{} (9.362262,0.5993219)[$e$]{} (9.362262,-0.8006781)[$f$]{} (11.562262,0.5993219)[$g$]{} (11.562262,-0.8006781)[$h$]{} (0.96226174,-0.4006781)[$E$]{} (5.2262,-0.6)[$E$]{} (5.762262,0.7993219)[$E$]{} (0.36226174,0.7993219)[$E$]{} (12.2262,-.9)[$E$]{} (10.3,-1.6006781)[$E$]{} (12.1,.3219)[$E$]{} (9.85,-1.0006781)[$E$]{}
Let $E$ be an obvious modification of the constant collection (\[vcera\]) to the category $\RTre$. Figure \[za\_chvili\_sraz\_s\_mymi\_studenty\] features souls of rooted trees with two internal edges. It shows that $\Free^2(E)$ has two families of bases, $(b^1_1,b^1_2)$ corresponding to the direct sum in the left part of Figure \[za\_chvili\_sraz\_s\_mymi\_studenty\], and $(b^2_1,b^2_2)$ corresponding to the right direct sum. Let $R$ be the subspace of $\Free^2(E)$ spanned by the relations $$r_1 := b^1_1 - b^1_2 \ \hbox { and } \ r_2 := b^2_1 - b^2_2.$$ The isomorphism $\termRTre \cong \Free(E)/(R)$ can be established as in the proof of Theorem \[v\_nedeli\_si\_snad\_zaletam\_s\_NOE\].
(0,-1.34791)(3.84,1.6534791) (2.0,0.14652084) (0.6,-1.0534792)(2.0,0.14652084) (1.2,-1.2534791)(2.0,0.14652084) (3.2,-1.0534792)(2.0,0.14652084) (2.0,0.14652084)(2.0,1.465209) (0.4,-.8)[$\sigma_1$]{} (.8,-1.2)[$\sigma_2$]{} (2.7,-1.2)[$\sigma_n$]{} (.3,-1.4)[$1$]{} (1.0,-1.791)[$2$]{} (3.2,-1.4792)[$n$]{} (2.2,0.6)[$0$]{} (2,1.6)[$0$]{} (-3.177993,1)[(5.275,-0.8477939)[1]{}[-110]{}[-60]{}]{}
To identify $\termRTre$-algebras with Markl’s operads we proceed as in the proof of Theorem \[pozitri\_turnaj\_v\_Patku\]. We start by realizing that the local terminal objects are rooted corollas $c^\uparrow(\sigma)$, $\sigma \in
\Sigma_n$, shown in Figure \[Ward\_zitra\_odleti.\] while the chosen local terminal objects are $c^\uparrow_n := c^\uparrow(\id_n)$. The set $\pi_0(\RTre)$ of connected components is therefore identified with the natural numbers $\bbN$. Analyzing action (\[Asi\_jsem\_dostal\_premii!\]) of local terminal objects we conclude that the underlying collections for $\termRTre$-algebras are sequences $\calS(n)$, $n \in \bbN$, of $\Sigma_n$-modules.
As in the proof of Theorem \[pozitri\_turnaj\_v\_Patku\] we establish that the value of the generating collection $E$ on graphs whose soul is the right arrow in Figure \[je\_patek\_a\_melu\_z\_posledniho\] produces partial compositions (\[zitra\_s\_Mikesem\_na\_Jazz\]), that relation $r_1$ expresses the parallel associativity, i.e. the first and the last cases of (\[zitra\_s\_Mikesem\_do\_Salmovske\]), and $r_2$ the sequential associativity, i.e. the middle case of (\[zitra\_s\_Mikesem\_do\_Salmovske\]). We are sure that at this stage the reader will easily describe the annihilator $R^\perp$ of the space $R$ of relations and identify algebras the Koszul dual $$\oddRTre := \termRTre^! = \Free(\susp E^*)/(R^\perp)$$ as structures with degree $+1$ operations $$\label{Byl_jsem_se_142_v_Benesove.}
\bullet_i : \calS(m) \ot \calS(n) \to \calS(m +n -1)$$ satisfying (\[Krtecek\_na\_mne\_kouka.\]) and the associativities (\[zitra\_s\_Mikesem\_do\_Salmovske\]) with the minus sign. It can be verified directly that the level-wise suspension of such a structure is an ordinary Markl’s operad. However, a more conceptual approach based on coboundaries introduced in Example \[coboundary\] is available.
As in the cases of modular and cyclic operads we notice that, for a rooted tree $T \in \RTre$, $\oddRTre(T) \cong \det(\edg(T))$, the determinant of the set of internal edges of $T$. On the other hand, the correspondence that assigns to each vertex of $T$ its out-going edge is an isomorphism $$\label{Dnes_mam_vedecke_narozeniny.}
\edg(T) \cong \{\hbox {vertices of $T$}\} \setminus \{\hbox {the root}\}$$ which implies that $\det(\edg(T))$ is isomorphic to $\fdl(T)$, where $\fdl$ is the coboundary with $\fl: \pi_0(\RTre) \to
\Vect$ the constant function with value the desuspension $\hbox {$\downarrow\!\bfk$}$ of the ground field. Therefore $$\oddRTre = \termRTre \ot \fdl$$ and the identification of $\oddRTre$-algebras with Markl’s operads via the suspension of the underlying collection follows from Proposition \[za\_chvili\_zavolam\_Jarusce\].
Similar statements can be proved also for the operadic categories $\PTr$ and $\PRTr$ of planar resp. planar rooted trees introduced in Example \[Predtim\_s\_Jarkou\_na\_krest\_knihy.\]. The corresponding terminal operads ${\sf 1}_\PTr$ resp. ${\sf 1}_\PRTr$ will again be self-dual binary quadratic, with algebras non-$\Sigma$ cyclic operads [@markl-shnider-stasheff:book page 257] resp. non-$\Sigma$ Markl’s operads [@markl-shnider-stasheff:book Definition II.1.14]. We leave the details to the reader.
[-.5em **Pre-permutads**]{} introduced in [@loday-ronco] form a link between non-$\Sigma$ operads and permutads. They are structures satisfying all axioms of Markl’s operads as recalled in Definition \[b3\] except the parallel associativity, i.e. the first and the last cases of (\[zitra\_s\_Mikesem\_do\_Salmovske\]). Pre-permutads are algebras over a certain operad in the category $\RTre$ of rooted trees which is very far from being Koszul self-dual:
\[Vcera\_s\_Mikesem\_v\_Matu.\] Pre-permutads are algebras over a binary quadratic $\RTr$-operad $\pperm$. The category of algebras over its Koszul dual $\pperm^!$ is isomorphic to the category of structures satisfying all axioms of Markl’s operads, except the associativity (\[zitra\_s\_Mikesem\_do\_Salmovske\]) which is replaced by $$(f \circ_j g)\circ_i h =
\begin{cases}
0,& \mbox{for } 1\leq i< j,
\\
f \circ_j(g \circ_{i-j+1} h),& \mbox{for }
j\leq i~< b+j, \mbox{ and}
\\
0,& \mbox{for }
j+b\leq i\leq a+b-1.
\end{cases}$$
The generating collection $E$ is the same as the one for ordinary Markl’s operads used in the proof of Theorem \[markl1\]. Referring to the notation used in that proof, we define $R$ to be the subspace of $\Free^2(E)$ spanned by $r_2$ belonging to the second direct sum of Figure \[je\_patek\_a\_melu\_z\_posledniho\]. It is clear that $\pperm \cong \Free(E)/(R)$. Let $d^1_1,d^1_2$ resp. $d^2_1,d^2_2$ be the bases of $\Free^2(\susp E^*)$ dual to $b^1_1,b^1_2$ resp. $b^2_1,b^2_2$. Then the annihilator $R^\perp$ is clearly spanned by $$o := d^2_1 + d^2_2,\ d^1_1 \hbox { and } d^1_2.$$ As before we identify algebras over $\pperm^! =
\Free(\susp^*E)/{R^\perp}$ with structures equipped with degree $+1$ operations (\[Byl\_jsem\_se\_142\_v\_Benesove.\]) satisfying $$(f \bullet_j g)\bullet_i h =
\begin{cases}
0,& \mbox{for } 1\leq i< j,
\\
- f \bullet_j(g \bullet_{i-j+1} h),& \mbox{for }
j\leq i < b+j, \mbox{ and}
\\
0,& \mbox{for }
j+b\leq i\leq a+b-1,
\end{cases}$$ whose first case corresponds to $d^1_1$, the middle to $o$ and the last one to $d^1_2$. The level-wise suspension of this object is the structure described in the theorem.
PROP-like structures and permutads {#Ta_moje_lenost_je_strasna.}
==================================
In this section we treat some important variants of PROPs governed by operadic categories that are sundry modifications of the category $\Whe$ of connected directed oriented graphs introduced in Example \[Pani\_Bilkova\_zatim\_neudelala\_rezrvaci.\]. The orientation divides the set of half-edges adjacent to each vertex of the graphs involved into two subsets – inputs and outputs of that vertex. The local terminal objects in these categories will thus be the directed corollas $c{\sigma \choose \lambda}$, $\sigma \in \Sigma_k$, $\lambda \in
\Sigma_l$, as in Figure \[Mikes\], the chosen local terminal objects the directed corollas $c^k_l
:= c{\id_k \choose \id_l}$, $k,l \in \bbN$. The underlying collections of the corresponding algebras will be families $$\label{Dnes_jsem_byl_s_NOE_v_Pribyslavi.}
D(m,n),\ m,n \in \bbN,$$ of $\Sigma_m \times \Sigma_n$-modules. We will see that the orientation of the underlying graphs implies that the corresponding terminal operads are self-dual.
(0,-1.88)(3.64,1.88) (1.8,-0.08) (0.4,-1.28)(1.8,-0.08) (1.0,-1.48)(1.8,-0.08) (3.0,-1.28)(1.8,-0.08) (0.5,.7)[$\sigma_1$]{} (1.05,1.1)[$\sigma_2$]{} (2.7,1.1)[$\sigma_k$]{} (0.2,1.32)[$1$]{} (1.2,1.6)[$2$]{} (3.3,1.3)[$k$]{} (1.8,-0.08)(0.4,1.32) (1.8,-0.08)(3.2,1.12) (1.8,-0.08)(1.2,1.52) (0.5,-.7)[$\lambda_1$]{} (.8,-1.3)[$\lambda_2$]{} (2.5,-1.18)[$\lambda_l$]{} (0.2,-1.28)[$1$]{} (1.1,-1.6)[$2$]{} (3.2,-1.38)[$l$]{} (-3.4,.9)[(5.275,-0.8477939)[1]{}[-110]{}[-60]{}]{} (-3.4,.7)[(5.275,-0.8477939)[1]{}[55]{}[105]{}]{}
Wheeled properads
-----------------
These structures were introduced in [@mms] as an extension of Vallette’s properads [@vallette:TAMS07] that allowed ‘back-in-time’ edges in order to capture traces and therefore also master equations of mathematical physics. Surprisingly, this extended theory is better behaved than the theory of properads in that the structure operations are iterated compositions of elementary ones meaning, in terms of pasting schemes, of those given by contraction of a single edge.
The guiding operadic category for wheeled properads is the category $\Whe$ of oriented connected directed graphs. Since $\Whe$ was in Example \[Pani\_Bilkova\_zatim\_neudelala\_rezrvaci.\] constructed from the basic operadic category $\Gr$ by iterating Grothendieck’s construction and since it clearly satisfies the conditions $\UFB$ and “$f$ is iso if $e(f)=0$,” we conclude as in the previous sections that our theory of Koszul duality applies to it.
The terminal $\Whe$-operad $\Wheterm$ is binary quadratic. Its algebras are wheeled properads introduced in [@mms Definition 2.2.1]. The operad $\Wheterm$ is self-dual in the sense of Definition \[Uz\_je\_o\_te\_medaili\_zapis.\].
It goes along the same lines as the proofs of similar statements in the previous sections, so we will be telegraphic. As before, for a wheeled graph $\Gamma$, ${\mathfrak {K}}_\Whe (\Gamma):=\Wheterm^!(\Gamma) \cong \det(\edg(\Gamma))$, the determinant of the set of internal edges of $\Gamma$. On the other hand, the correspondence that assigns to each vertex $v$ of $\Gamma$ the set $\out(v)$ of its out-going edges defines an isomorphism $$\label{Dnes_mam_vedecke_narozeniny_bis.}
\textstyle
\edg(\Gamma) \cong \bigcup_{v \in \Vert(\Gamma)} \out(v) \setminus \out(\Gamma)$$ which implies that $\det(\edg(\Gamma))$ is isomorphic to $\fdl(\Gamma)$, where $\fdl$ is the coboundary with $\fl: \pi_0(\RTre) \to
\Vect$ the function defined by $$\fl(c^k_l) : = \downarrow^k \bfk,$$ the desuspension of the ground field iterated $k$ times. Therefore $${\mathfrak {K}}_\Whe \cong \Wheterm \ot \fdl$$ which, by Proposition \[za\_chvili\_zavolam\_Jarusce\], implies the self-duality of $ \Wheterm$.
It follows from the description of the local terminal objects in $\Whe$ that the underlying structure of an $\Wheterm$-algebra is a collection of bimodules as in (\[Dnes\_jsem\_byl\_s\_NOE\_v\_Pribyslavi.\]). The structure operations are given by wheeled graphs with one internal edge whose souls are depicted in Figure \[Za\_necely\_tyden\_zavody\_v\_Boleslavi.\].
(0,-1.0554386)(2.5108774,1.0554386) (0.11087738,-0.9445613) (0.11087738,-0.9445613)(0.11087738,0.8554387) (0.11087738,0.8554387) (2.0108774,0.05543869)(0.5,1.0) (1.5108774,0.05543869) (1.5108774,0.2554387)(1.7108774,0.8554387)
We recognize them as the operations
$$\begin{aligned}
\label{musim_to_dokoukat}
\circ^i_j :\ & D(m,n) \ot D(k,l)
\longrightarrow D(m\! +\! k\! -\!1,n\!+ \!l \!-\!1),\
1 \leq i \leq n, \ 1 \leq j \leq k, \hbox { and}
\\
\label{Budu_mit_silu_to_dokoukat?}
\xi^i_j :\ & D(m,n) \longrightarrow D(m\!-\!1,n\!-\!1),\
1\leq i \leq m, \ 1 \leq j \leq n\end{aligned}$$
in formulas (16) and (17) of [@mms].
As in the previous cases, the axioms that these operations satisfy are determined by graphs with two internal edges whose souls are depicted in Figure \[Ten\_film\_mne\_oddelal.\]. The graphs with three vertices induce the parallel and sequential associativity of the $\xi$-operations, similar to that for Markl’s operads (\[zitra\_s\_Mikesem\_do\_Salmovske\]). They were explicitly given in the dioperadic context as axioms (a) and (b) in [@gan page 111].
The circles in Figure \[Ten\_film\_mne\_oddelal.\] represent the rules of the type $
\circ_1 \xi_2 = \circ_2 \xi_1,
$ where $\xi_1$ resp. $\xi_2$ is the operation corresponding to the shrinking of the edge labelled $1$ resp. $2$, and similarly for $\circ_1$ and $\circ_2$. The lollipops in Figure \[Ten\_film\_mne\_oddelal.\] force the interchange rule $\xi_2 \circ_1 = \circ_1 \xi_2$, and the eyes the rule $\circ_1 \circ_2 = \circ_1 \circ_2$. To expand these remaining axioms into explicit forms similar to that on [@gan page 111] would not be very helpful, we thus leave it as an exercise for a determined reader.
(0,-2.9354386)(10.510878,2.9354386) (1.7108774,0.8245613) (0.9108774,2.2245612) (7.7108774,1.4245613)(0.8,1.0) (0.11087738,0.8245613)(0.9108774,2.2245612) (1.7108774,0.8245613)(0.9108774,2.2245612) (2.9108775,0.2245613)(2.9108775,1.6245613) (2.9108775,1.6245613)(2.9108775,2.8245614) (2.9108775,2.8245614) (2.9108775,1.6245613) (0.11087738,0.8245613) (4.1108775,2.2245612) (4.910877,0.8245613) (5.7108774,2.2245612)(4.910877,0.8245613) (4.1108775,2.2245612)(4.910877,0.8245613) (5.7108774,2.2245612) (2.9108775,0.2245613) (1.3858774,-1.7754387)(0.5,1.0) (1.9108773,-1.7754387) (1.8858774,-1.5754387)(1.6858773,-0.9754387) (1.8858774,-1.7754387)(1.8858774,-0.17543869) (8.810878,-1.3754387)(0.5,1.0) (8.310878,-1.3754387) (8.310878,-1.1754386)(8.510878,-0.5754387) (7.785877,-1.3754387)(0.5,1.0) (8.285877,-1.1754386)(8.085877,-0.5754387) (7.7108774,0.42456132) (7.7108774,2.4245613) (7.7108774,2.4245613)(7.3108773,2.2245612) (7.7108774,2.4245613)(8.110877,2.2245612) (9.710877,1.4245613)(0.8,1.0) (9.710877,2.4245613) (9.710877,2.4245613)(9.310878,2.2245612) (9.710877,0.42456132) (9.710877,0.42456132)(10.110877,0.6245613) (3.5858774,-1.1754386)(0.5,1.0) (4.1108775,-1.1754386) (4.0858774,-0.9754387)(3.8858774,-0.3754387) (4.0858774,-2.7754388)(4.0858774,-1.1754386) (6.7108774,1.4245613)[$1$]{} (8.310878,1.4245613)[$2$]{} (3.3108773,-1.1754386)[$1$]{} (1.1108774,-1.7754387)[$1$]{} (9.110877,1.4245613)[$1$]{} (7.5108776,-1.3754387)[$1$]{} (2.1108773,-0.9754387)[$2$]{} (4.3108773,-2.1754386)[$2$]{} (9.110877,-1.3754387)[$2$]{} (10.310878,1.4245613)[$2$]{} (2.9108775,-2.7754388)[lollipops]{} (8.310878,-2.7754388)[eyes]{} (8.710877,0.2245613)[circles]{}
Dioperads
---------
They were introduced in [@gan] as tools governing structures like Lie or infinitesimal bialgebras (called mock bialgebras in [@markl:dl]). A short definition is that a dioperad is a wheeled properad without the $\xi^i_j$-operations (\[Budu\_mit\_silu\_to\_dokoukat?\]). The underlying operadic category is the category $\Dio$ of directed simply connected oriented graphs introduced in Example \[Travelodge\]. As before one may check that $\Dio$ meets all requirements of our theory. One has the expected:
Dioperads are algebras over the terminal $\Dio$-operad ${\sf 1}_\Dio$, which is binary quadratic and self-dual.
The proof is a simplified version of the wheeled case. The self-duality of ${\sf 1}_\Dio$ is established in precisely the same way as the self-duality of the terminal $\Whe$-operad $\Wheterm$; the existence of the relevant coboundary is given by isomorphism (\[Dnes\_mam\_vedecke\_narozeniny\_bis.\]) which clearly holds in $\Dio$ as well. The soul of graphs in $\Dio$ with one internal edge is the oriented interval, with the corresponding operation as in (\[musim\_to\_dokoukat\]). The souls of graphs in $\Dio$ with two internal edges are the three upper left graphs in Figure \[Ten\_film\_mne\_oddelal.\]. The resulting axioms are the parallel and sequential associativities which are the same as for $\xi^i_j$-operations of wheeled properads, see [@gan §1.1].
\#1[[\_[\#1]{}.2mm]{}]{}
$\frac12$-PROPs
---------------
These structures were introduced, following a suggestion of M. Kontsevich, in [@mv] as a link between dioperads and PROPs. A $\frac12$-PROP is a collection of bimodules (\[Dnes\_jsem\_byl\_s\_NOE\_v\_Pribyslavi.\]) which is [ *stable*]{} in that is fulfills $$D(m,n) = 0\ \hbox { if } \ m+n < 3,$$ together with partial vertical compositions $$\begin{aligned}
\circ_i:\ & D(m_1,n_1) \ot D(1,l)
\to D(m_1,n_1+l-1),\ 1 \leq i \leq n_1, \hbox { and}
\\
\jcirc j :\ & D(k,1) \ot D(m_2,n_2)
\to D(m_2 + k -1,n_2),\ 1 \leq j \leq m_2,\end{aligned}$$ that satisfy the axioms of vertical compositions in PROPs. The corresponding operadic category $\hGr$ is introduced in Example \[Zrejme\_brzy\_podlehnu.\]. We have the expected statement whose proof is left to the reader.
$\frac12$PROPs are algebras over the terminal $\hGr$-operad ${\sf 1}_\hGr$. This operad is binary quadratic and self-dual.
Operadic categories considered so far in this section were based on graphs. Let us give one example where it is not so.
Permutads {#Michael_do_Prahy.}
---------
They are structures introduced by Loday and Ronco in [@loday11:_permut] to handle the combinatorial structure of objects like the permutahedra. We will describe an operadic category $\Per$ such that permutads are algebras over the terminal operad for this category.
Let $\underline n$ for $n \geq 1$ denote the finite ordered set $(\rada 1n)$. Objects of $\Per$ are epimorphisms $\alpha: \underline n \epi \underline k$, $n \geq 1$, and the morphisms are diagrams $$\label{nikdo_mi_nepopral_k_narozeninam}
\xymatrix@C=2em{\underline n \ar@{=}[r]
\ar@{->>}[d]_{\alpha'} & \underline n \ar@{->>}[d]^{\alpha''}
\\
\underline k' \ar[r]^\gamma & \underline k''
}$$ in which $\gamma$ is order-preserving (and necessarily an epimorphism).
The cardinality functor is defined by $|\alpha:
\underline n \epi \underline k| := k$. The $i$-th fiber of the morphism in (\[nikdo\_mi\_nepopral\_k\_narozeninam\]) is the epimorphism $\gamma(\alpha')^{-1}(i) \epi \gamma^{-1}(i)$, $i \in \underline
k$. The only local terminal objects are $\underline n \to \underline
1$, $n \geq 1$, which are also the chosen ones. The category $\Per$ is graded by $e(\underline n \epi \underline k) :=
k-1$. All , and isomorphisms in general, are the identities.
\[Treti\_den\_je\_kriticky.\] Algebras over the terminal $\Per$-operad $\Pterm$ are the permutads of [@loday11:_permut]. The operad $\Pterm$ is binary quadratic. It is self-dual in the sense that the category of algebras over $\Pterm^!$ is isomorphic to the category permutads via the functor induced by the suspension of the underlying collection.
Let us give a quadratic presentation of the terminal operad $\Pterm$. As noticed in Example \[Budu\_mit\_nova\_sluchatka.\], the category $\QV(e)$ of virtual isomorphisms related to $\Per$ is isomorphic to the category $\Iso$ of isomorphisms in $\Per$. Since all isomorphisms in $\Per$ are the identities, we infer from this that in fact $\QV(e) \cong \Per_{\rm disc}$, the discrete category with the same objects as $\Per$. Therefore a $\QV(e)$-presheaf is just a rule that assigns to each $\alpha \in \Per$ a vector space $E(\alpha) \in \Vect$. Let us define a $1$-connected $\Per$-collection, in the sense of Definition \[Jarka\_dnes\_u\_lekare\], by $$\label{po_GGP}
E(\alpha) :=
\begin{cases}
\bfk & \hbox { if $|\alpha| = 2$, and}
\\
0 & \hbox {otherwise}
\end{cases}$$ and describe the free operad $\Free(E)$ generated by $E$.
The first step is to understand the labelled towers in $\ltw(\alpha)$. As all in $\Per$ are the identities, the labeling is the identity map so these towers are of the form $${\mathbf \alpha}: =
\alpha \stackrel {\tau_1} \longrightarrow \alpha_1 \stackrel {\tau_2} \longrightarrow
\alpha_2\stackrel {\tau_3} \longrightarrow
\cdots \stackrel {\tau_{s-1}} \longrightarrow \alpha_{s-1}.$$ Since the generating collection $E$ is such that $E(\alpha)
\not=0$ only if $|\alpha| =2$, we may consider only towers in which each $\tau_i$, $1 \leq i \leq s-1$, decreases the cardinality by one. For $\alpha : \underline n \epi \underline k$, such a tower is a diagram $$\label{Michael_chce_do_Prahy.}
\xymatrix{
\underline n \ar@{=}[r] \ar@{->>}[d]_\alpha
& \underline n \ar@{=}[r] \ar@{->>}[d]_{\alpha_1}&
\underline n \ar@{=}[r]\ar@{->>}[d]_{\alpha_2} & \cdots \ar@{=}[r]
&\ar@{->>}[d]_{\alpha_{k-1}}
\underline n
\\
\underline k \ar@{->>}[r]^{\nu_1}& \underline {k\!-\!1} \ar@{->>}[r]^{\nu_2}
& \underline {k\!-\!2}
\ar@{->>}[r]^{\nu_3}
& \cdots \ar@{->>}[r]^{\nu_{k-2}}& \underline 2
}$$ with $\Rada \nu1{k-2}$ order-preserving epimorphism. Notice that all vertical maps are determined by $\alpha$ and $\Rada \nu1{k-2}$. It will be convenient to represent $\underline k$ by a linear graph with $k$ vertices:
(0,-0.3167067)(6.8168273,0.3167067) (4.5,-0.2082933)(5.108413,-0.2082933) (0.10841339,0.1917067)[$1$]{} (1.5084134,0.1917067)[$2$]{} (2.9084134,0.1917067)[$3$]{} (1.5084134,-0.2082933)(0.10841339,-0.2082933) (1.5084134,-0.2082933) (6.7084136,-0.2082933)(5.3084135,-0.2082933) (0.10841339,-0.2082933) (4.3084135,-0.2082933)(2.9084134,-0.2082933) (6.7084136,-0.2082933) (2.9084134,-0.2082933)(1.5084134,-0.2082933) (2.9084134,-0.2082933) (6.7084136,0.1917067)[$k$]{}
and denote by $\edg(\underline k)$ or $\edg(\alpha)$ the set of $k\!-\!1$ edges of this graph. In this graphical presentation, each $\Rada \nu1{k-2}$ contracts one of the edges of our linear graph, thus $\Rada \nu1{k-2}$ and therefore also the tower (\[Michael\_chce\_do\_Prahy.\]) is determined by the linear order of $\edg(\underline k)$ in which the edges are contracted. We readily get the following analog of Proposition \[uz\_melo\_prvni\_cislo\_davno\_vyjit\]:
\[Uz\_melo\_prvni\_cislo\_davno\_vyjit.\] The isomorphism classes of labeled towers (\[Michael\_chce\_do\_Prahy.\]) are in one-to-one correspondence with the linear orders of $\edg(\underline k)$ modulo the relation $\bowtie$ that interchanges two edges adjacent in this linear order that do not share a common vertex.
Let us continue the proof of Proposition \[Treti\_den\_je\_kriticky.\]. By Proposition \[Uz\_melo\_prvni\_cislo\_davno\_vyjit.\], $\Free(E)(\alpha)$ equals the span of the set of linear orders on $\edg(\underline k)$ modulo the equivalence $\bowtie$. Let us inspect in detail its component $\Free^2(E)(\alpha)$. It might be nonzero only for $\alpha : \underline n \to \underline k
\in \Per$ with $k=3$, for which (\[Michael\_chce\_do\_Prahy.\]) takes the form $$\xymatrix@C=3em{
\underline n \ar@{=}[r] \ar@{->>}[d]_\alpha
& \underline n \ar@{->>}[d]_{\alpha_1}
\\
\underline 3 \ar@{->>}[r]^{\nu}& \underline 2
}$$ and the relation $\bowtie$ is vacuous.
There are two possibilities for the map $\nu$ and therefore also for $\alpha_1$. The map $\nu$ may either equal $\nu_{\{1,2\}}: \underline 3 \to \underline 2$ defined by $$\nu_{\{1,2\}}(1) = \nu_{\{1,2\}}(2) :=1,\ \nu_{\{1,2\}}(3) := 2$$ which corresponds to the linear order
(-4.5,-0.3167067)(6.8168273,0.3167067) (0.8,0.14)[$1$]{} (2.2,0.14)[$2$]{} (1.5084134,-0.2082933)(0.10841339,-0.2082933) (1.5084134,-0.2082933) (0.10841339,-0.2082933) (2.9084134,-0.2082933)(1.5084134,-0.2082933) (2.9084134,-0.2082933)
of $\edg(3)$, or $\nu_{\{2,3\}}: \underline 3 \to \underline 2$ defined by $$\nu_{\{2,3\}}(1) :=1,\ \nu_{\{2,3\}}(2) = \nu_{\{2,3\}}(3) := 2,$$ corresponding to the order
(-4.5,-0.3167067)(6.8168273,0.3167067) (0.8,0.14)[$2$]{} (2.2,0.14)[$1$]{} (1.5084134,-0.2082933)(0.10841339,-0.2082933) (1.5084134,-0.2082933) (0.10841339,-0.2082933) (2.9084134,-0.2082933)(1.5084134,-0.2082933) (2.9084134,-0.2082933) (3.2,-.3)[.]{}
The fiber sequence associated to $\nu_{\{1,2\}}$ is $
\alpha|_{\alpha^{-1}\{1,2\}}, \nu_{\{1,2\}}\alpha
$, the one associated to $\nu_{\{2,3\}}$ is $
\alpha|_{\alpha^{-1}\{2,3\}}, \nu_{\{2,3\}}\alpha,
$ therefore $$\Free^2(E)(\alpha) \cong \{ E(\alpha|_{\alpha^{-1}\{1,2\}}) \ot
E(\nu_{\{1,2\}}\alpha)\} \oplus \{ E(\alpha|_{\alpha^{-1}\{2,3\}}) \ot
E(\nu_{\{2,3\}}\alpha)\}.$$ Since $ E(\alpha|_{\alpha^{-1}\{1,2\}}) =
E(\alpha|_{\alpha^{-1}\{2,3\}}) = E(\nu_{\{1,2\}}\alpha) =
E(\nu_{\{2,3\}}\alpha) =\bfk$ by definition, $\Free^2(E)(\alpha)$ admits a basis formed by $$b_1 : =[ 1\ot 1] \oplus [0 \ot 0] \ \hbox { and } \
b_2 : = [0\ot 0] \oplus [1 \ot 1].$$ Let $R$ be the subspace of $\Free^2(E)$ spanned by $b_2-b_1$. Quotienting by the ideal $(R)$ generated by $R$ extends the relation $\bowtie$ of Proposition \[Uz\_melo\_prvni\_cislo\_davno\_vyjit.\] by allowing edges that do share a common vertex, thus $\Free(E)/(R)(\alpha) \cong \bfk$ for any $\alpha$, in other words, $$\Pterm \cong \Free(E)/(R).$$
Now we describe $\Pterm$-algebras. Since $\pi_0(\Per) = \{1,2,\ldots\}$, their underlying collections are sequences of vector spaces $P(n)$, $n\geq 1$. As we saw several times before, the structure operations of $\Pterm$-algebras are parameterized by the generating collection $E$, therefore, by (\[po\_GGP\]), by epimorphisms $r : \underline n \epi \underline 2 \in \Per$. If $n_i := |r^{-1}(i)|$, $i
=1,2$, the operation corresponding to $r$ is of the form $$\label{Nevim_jestli_do_ty_Ciny_opravdu_pojedu.}
\circ_r : P(n_1) \ot P(n_2) \to P(n_1+n_2)$$ by (\[Ben\_Ward\_in\_Prague\]). It is easy to verify that the vanishing of the induced map $\Free(E) \to \End_P$ on the generator $b_2-b_1$ of the ideal of relations $(R)$ is equivalent to the associativity $$\label{Oslava_premie_bude_za_14_dni.}
\circ_t(\circ_s \ot \id) = \circ_u(\id \ot \circ_v)$$ with $s := \alpha|_{\alpha^{-1}\{1,2\}},\ t:= \nu_{\{1,2\}}\alpha,\
u := \alpha|_{\alpha^{-1}\{2,3\}}$ and $v:= \nu_{\{2,3\}}\alpha$. We recognize it as the associativity of [@loday11:_permut Lemma 2.2] featuring in the biased definition of permutads.
It can be easily seen that ${\mathfrak{K}}_\Per(\alpha) := \Pterm^!(\alpha) \cong
\det(\edg(\alpha))$. As in §\[Cinani\_jsou\_hovada.\] we identify ${\mathfrak{K}}_\Per$-algebras as structures with degree $+1$ operations $$\bullet_r : P(n_1) \ot P(n_2) \to P(n_1+n_2)$$ with $r$ as in (\[Nevim\_jestli\_do\_ty\_Ciny\_opravdu\_pojedu.\]) satisfying an odd version $$\bullet_t(\bullet_s \ot \id) + \bullet_u(\id \ot \bullet_v) =0$$ of (\[Oslava\_premie\_bude\_za\_14\_dni.\]). It is elementary to show that the structure induced on the component-wise suspension of the underlying collection is that of a permutad.
In [@markl:perm] we proved the following theorem:
The terminal $\ttP$-operad $\Pterm$ is Koszul.
Its meaning is that the canonical map $\Omega(\Pterm^!) \to \Pterm$ from a suitably defined bar construction of $\Pterm^!$ to $\Pterm$ is a component-wise homology equivalence. In other words, the dg-$\Per$ operad $\Omega(\Pterm^!)$ is the minimal model of $\Pterm$ therefore, according to the philosophy of [@markl:zebrulka Section 4], $\Omega(\Pterm^!)$-algebras are [*strongly homotopy*]{} permutads. An explicit description of these objects is given in [@markl:perm] as well.
Recollections {#a1}
=============
In this part of the appendix we recall various operad-like structures referred to in this work. All definitions given here are standard today, see e.g. [@getzler-kapranov:CompM98; @markl-shnider-stasheff:book], so the purpose is merely to fix the notation and terminology.
Recall that a [*modular module*]{} is a functor $\Fin \times \bbN \to
\Vect$, with $\bbN$ interpreted as a discrete category with objects called [*genera*]{} in this context.
\[modular\] A [*modular operad*]{} is a modular module $$\oM = \big\{\oM(S;g) \in \Chain\ \vrt \ (S;g) \in \Fin \times \bbA \big\}$$ together with degree $0$ morphisms (compositions) $$\label{v_Galway1}
{\sideset{_{a}}{_{b}}{\mathop{\circ}}}:\oM\big(S_1 \sqcup \stt a;g_1\big)
\otimes \oM\big(S_2\sqcup \stt b;g_2\big)
\to \oM ( S_1\sqcup S_2;g_1\+g_2)$$ defined for arbitrary disjoint finite sets $S_1$, $S_2$, symbols $a,b$, and arbitrary genera $g_1,g_2 \in \bbA$. There are, moreover, degree $0$ contractions $$\label{Galway}
\xxi_{uv} = \xxi_{vu} : \oM\bl S \sqcup \stt {u,v} ;g\br
\to \oM(S ;g\+1)$$ given for any finite set $S$, genus $g \in \bbA$, and symbols $u,v$. These data are required to satisfy the following axioms.
1. For arbitrary isomorphisms $\rho : S_1 \sqcup \stt a \to T_1$ and $\sigma : S_2 \sqcup \stt b \to T_2$ of finite sets and genera $g_1$, $g_2 \in \bbA$, one has the equality $$\label{piji_caj_z_Jarcina_hrnku}
\oM\bl\rho|_{S_1}\sqcup\sigma|_{S_2}\br
{\sideset{_{a}}{_{b}}{\mathop{\circ}}} =
{\sideset{_{\rho(a)}}{_{\sigma(b)}}{\mathop{\circ}}} \ \big(\oM(\rho)\ot\oM(\sigma)\big)$$ of maps $$\oM\bl S_1 \sqcup \stt a ;g_1\br \otimes \oM \bl S_2 \sqcup \stt b
;g_2\br
\to
\oM\bl T_1\sqcup T_2
\setminus \{\rho(a),\sigma(b)\};g_1+g_2\br.$$
2. For an isomorphism $\rho : S \sqcup \stt {u,v} \to T
$ of finite sets and a genus $g \in \bbA$, one has the equality $$\label{eq:27}
\oM\bl\rho|_S\br \ \xxi_{uv} =
\xxi_{\rho(u)\rho(v)}\oM(\rho)$$ of maps $\oM \bl S \sqcup \stt {u,v};g \br \to \oM\bl T \setminus
\{\rho(u),\rho(v)\} ;g+s\br$.
3. For $S_1$, $S_2$, $a$, $b$ and $g_1$, $g_2$ as in (\[v\_Galway1\]), one has the equality $$\label{eq:24}
{\sideset{_{a}}{_{b}}{\mathop{\circ}}} = {\sideset{_{b}}{_{a}}{\mathop{\circ}}} \tau$$ of maps $\oM(S_1 \sqcup \stt a;g_1)\otimes \oM(S_2 \sqcup \stt b;g_2)
\to \oM\bl S_1\sqcup S_2;g_1+g_2\br$.[^16]
4. For mutually disjoint sets $S_1,S_2,S_3$, symbols $a, b,c, d$ and genera $g_1,g_2,g_3 \in \bbA$, one has the equality $$\label{Dnes_s_Jaruskou_k_Pakouskum}
{\sideset{_{a}}{_{b}}{\mathop{\circ}}} (\id \ot {\sideset{_{c}}{_{d}}{\mathop{\circ}}}) = {\sideset{_{c}}{_{d}}{\mathop{\circ}}} ({\sideset{_{a}}{_{b}}{\mathop{\circ}}} \ot
\id)$$ of maps from $\oM \bl S_1 \sqcup \stt a;g_1 \br \ot \oM\bl S_2 \sqcup \stt {b,c};g_2\br
\ot \oM\bl S_3 \sqcup \stt d;g_3\br$ to the space $\oM\bl S_1 \sqcup S_2 \sqcup S_3;
g_1\!+\!
g_2\! +\! g_3\br$.
5. For a finite set $S$, symbols $a,b,c,d$ and a genus $g \in \bbA$ one has the equality $$\label{eq:33}
\xxi_{ab} \ \xxi_{cd} = \xxi_{cd} \ \xxi_{ab}$$ of maps $\oM\bl S \sqcup \{a,b,c,d\};g\br
\to \oM (S;g+2s)$.
6. For finite sets $S_1, S_2$, symbols $a,b,c,d$ and genera $g_1,g_2
\in \bbA$, one has the equality $$\label{eq:38}
\xxi_{ab} \ {\sideset{_{c}}{_{d}}{\mathop{\circ}}} = \xxi_{cd} \ {\sideset{_{a}}{_{b}}{\mathop{\circ}}}$$ of maps $\oM\bl S_1
\sqcup \stt {a,c};g_1\br \ot \oM \bl S_2 \sqcup \stt {b,d};g_2\br
\to \oM( S_1 \sqcup S_2;g_1+g_2+s)$.
7. For finite sets $S_1, S_2$, symbols $a,b,u,v$, and genera $g_1,g_2 \in \bbA$, one has the equality $$\label{eq:39}
{\sideset{_{a}}{_{b}}{\mathop{\circ}}} \ (\xxi_{uv}\ot\id) = \xxi_{uv} \ {\sideset{_{a}}{_{b}}{\mathop{\circ}}}$$ of maps $\oM\bl
S_1 \sqcup \stt{a,u,v};g_1\br \ot \oM \bl S_2 \sqcup
\stt b;g_2\br \to \oM (S_1 \sqcup S_2 ;g_1+g_2+s)$.
\[odd\_modular\] An [*odd modular* ]{} is a modular module $$\oM = \big\{\oM(S;g) \in \Vect\ \vrt \ (S;g) \in \Fin \times \bbA \big\}$$ together with degree $+1$ morphisms (${\sideset{_{a}}{_{b}}{\mathop{\bullet}}}$-operations) $$\label{v_GGalway}
{\sideset{_{a}}{_{b}}{\mathop{\bullet}}}:\oM\big(S_1 \sqcup \stt a;g_1\big)
\otimes \oM\big(S_2\sqcup \stt b;g_2\big)
\to \oM ( S_1\sqcup S_2;g_1+g_2)$$ defined for arbitrary disjoint finite sets $S_1$, $S_2$, symbols $a,b$, and arbitrary $g_1,g_2 \in \bbA$. There are, moreover, degree $1$ morphisms (the contractions) $$\twxxi_{uv} = \twxxi_{vu} : \oM\bl S \sqcup \stt {u,v} ;g\br
\to \oM(S ;g+1)$$ given for any finite set $S$, $g \in \bbA$, and symbols $u,v$.[^17] These data are required to satisfy the following axioms.
1. For arbitrary isomorphisms $\rho : S_1 \sqcup \stt a \to T_1$ and $\sigma : S_2 \sqcup \stt b \to T_2$ of finite sets and $g_1$, $g_2 \in \bbA$, one has the equality $$\label{neni}
\oM\bl\rho|_{S_1}\sqcup\sigma|_{S_2}\br
{\sideset{_{a}}{_{b}}{\mathop{\bullet}}} =
{\sideset{_{\rho(a)}}{_{\sigma(b)}}{\mathop{\bullet}}} \ \big(\oM(\rho)\ot\oM(\sigma)\big)$$ of maps $$\oM\bl S_1 \sqcup \stt a ;g_1\br \otimes \oM \bl S_2 \sqcup \stt b
;g_2\br
\to
\oM\bl T_1\sqcup T_2
\setminus \{\rho(a),\sigma(b)\};g_1+g_2\br.$$
2. For an isomorphism $\rho : S \sqcup \stt {u,v} \to T
$ of finite sets and $g \in \bbA$, one has the equality $$\oM\bl\rho|_S\br \ \twxxi_{uv} =
\twxxi_{\rho(u)\rho(v)}\oM(\rho)$$ of maps $\oM \bl S \sqcup \stt {u,v};g \br \to \oM\bl T \setminus
\{\rho(u),\rho(v)\} ;g+1\br$.
3. For $S_1$, $S_2$, $a$, $b$ and $g_1$, $g_2$ as in (\[v\_GGalway\]), one has the equality $$\label{kdy_zacnu_byt_rozumny}
{\sideset{_{a}}{_{b}}{\mathop{\bullet}}} = {\sideset{_{b}}{_{a}}{\mathop{\bullet}}} \tau$$ of maps $\oM(S_1 \sqcup \stt a;g_1)\otimes \oM(S_2 \sqcup \stt b;g_2)
\to \oM\bl S_1\sqcup S_2;g_1+g_2\br$.
4. For mutually disjoint sets $S_1,S_2,S_3$, symbols $a, b,c, d$ and $g_1,g_2,g_3 \in \bbA$, one has the equality $$\label{neni1}
{\sideset{_{a}}{_{b}}{\mathop{\bullet}}} (\id \ot {\sideset{_{c}}{_{d}}{\mathop{\bullet}}}) = -{\sideset{_{c}}{_{d}}{\mathop{\bullet}}} ({\sideset{_{a}}{_{b}}{\mathop{\bullet}}} \ot
\id)$$ of maps from $\oM \bl S_1 \sqcup \stt a;g_1 \br \ot \oM\bl S_2 \sqcup \stt {b,c};g_2\br
\ot \oM\bl S_3 \sqcup \stt d;g_3\br$ to the space $\oM\bl S_1 \sqcup S_2 \sqcup S_3;
g_1\!+\!
g_2\! +\! g_3\br$.
5. For a finite set $S$, symbols $a,b,c,d$ and $g \in \bbA$ one has the equality $$\twxxi_{ab} \ \twxxi_{cd} =- \twxxi_{cd} \ \twxxi_{ab}$$ of maps $\oM\bl S \sqcup \{a,b,c,d\};g\br
\to \oM (S;g+2)$.
6. For finite sets $S_1, S_2$, symbols $a,b,c,d$ and $g_1,g_2
\in \bbA$, one has the equality $$\twxxi_{ab} \ {\sideset{_{c}}{_{d}}{\mathop{\bullet}}} = -\twxxi_{cd} \ {\sideset{_{a}}{_{b}}{\mathop{\bullet}}}$$ of maps $\oM\bl S_1
\sqcup \stt {a,c};g_1\br \ot \oM \bl S_2 \sqcup \stt {b,d};g_2\br
\to \oM( S_1 \sqcup S_2;g_1+g_2+1)$.
7. For finite sets $S_1, S_2$, symbols $a,b,u,v$, and $g_1,g_2 \in \bbA$, one has the equality $${\sideset{_{a}}{_{b}}{\mathop{\bullet}}} \ (\twxxi_{uv}\ot\id) = -\twxxi_{uv} \ {\sideset{_{a}}{_{b}}{\mathop{\bullet}}}$$ of maps $\oM\bl
S_1 \sqcup \stt{a,u,v};g_1\br \ot \oM \bl S_2 \sqcup
\stt b;g_2\br \to \oM (S_1 \sqcup S_2 ;g_1+g_2+1)$.
Odd modular operads appeared in [@getzler-kapranov:CompM98 Section 4] as [*modular ${\mathfrak
{K}}$-operads*]{} for the dualizing cocycle ${\mathfrak {K}}$. The terminology we use was suggested by Ralph Kaufmann. A discussion of odd modular operads and similar structures can be found e.g. in [@markl:odd].
\[b3\] A [*Markl’s operad*]{} is a collection $\calS =\{\calS(n)\}_{n\geq 0}$ of right $\bfk[\Sigma_n]$-modules, together with $\bfk$-linear maps ($\circ_i$-compositions) $$\label{zitra_s_Mikesem_na_Jazz}
\circ_i : \calS(m) \ot \calS(n) \to \calS(m +n -1),$$ for $1 \leq i \leq m$ and $n \geq 0$. These data fulfill the following axioms.
1. For each $1 \leq j \leq a$, $b,c \geq 0$, $f\in
\calS(a)$, $g \in \calS(b)$ and $h \in \calS(c)$, $$\label{zitra_s_Mikesem_do_Salmovske}
(f \circ_j g)\circ_i h =
\begin{cases}
(f \circ_i h) \circ_{j+c-1}g,& \mbox{for } 1\leq i< j,
\\
f \circ_j(g \circ_{i-j+1} h),& \mbox{for }
j\leq i~< b+j, \mbox{ and}
\\
(f \circ_{i-b+1}h) \circ_j g,& \mbox{for }
j+b\leq i\leq a+b-1,
\end{cases}$$
2. For each $1 \leq i \leq m$, $n \geq 0$, $\tau \in \Sigma_m$ and $\sigma \in \Sigma_n$, let $\tau \circ_i
\sigma \in \Sigma_{m+n -1}$ be given by inserting the permutation $\sigma$ at the $i$th place in $\tau$. Let $f \in \calS(m)$ and $g \in
\calS(n)$. Then $$\label{Krtecek_na_mne_kouka.}
(f\tau)\circ_i(g\sigma) = (f\circ_{\tau(i)} g)(\tau \circ_i
\sigma).$$
Blow up axiom, , page
Canonical grading, page
Constant-free operadic category, Definition \[constant-free\]
Contraction, Definition \[Ron\_v\_Praze\]
Discrete operadic fibration Definition \[psano\_v\_Myluzach\]
Disjoint fibers, Definition \[d3\]
Elementary morphism, Definition \[plysacci\_postacci\]
Factorizable operadic category, , Definition \[dnes\_prednaska\_na\_Macquarie\]
Grading of an operadic category, Definition \[grad\]
Harmonic pair, Definition \[har\]
Local isomorphism, Definition \[pisu\_jednou\_rukou\]
Local reordering morphism, Definition \[pisu\_jednou\_rukou\]
Order preserving morphism, Definition \[pisu\_jednou\_rukou\]
Ordered graph, Definition \[Radeji\_bych\_sel\_s\_Jaruskou\_na\_vyhlidku\_sam.\]
Preodered graph, Definition \[pre\]
Pure contraction, Definition \[Ron\_v\_Praze\]
Rigid operadic category, , Definition \[Porad\_nevim\_jestli\_mam\_jit\_na\_ty\_narozeniny.\]
Strongly factorizable operadic category, , Definition \[zase\_jsem\_podlehl\]
Unique fiber axiom, , Definition \[Kveta\_asi\_spi.\]
Weak blow up axiom, , page
[10]{}
C. Barwick. . , 22(4):1893-1959, 2018.
J.C. Baez and J. Dolan. . , 135(2):145–206, 1998.
M.A. Batanin. , 136(1):39-103, 1998.
M.A. Batanin. . , 211(2):684–725, 2007.
M.A. Batanin. . , 217(1):334–385, 2008.
M.A. Batanin. , v.4 , n.2 , 237-265, [ 2010]{}.
M.A. Batanin and M. Markl. . , 285:1630–1687, 2015.
M.A. Batanin and C. Berger. , Vol. 32, No. 6, 148-253, 2017.
M.A. Batanin, J. Kock and M. Weber. , vol. 33, 6-7, p.148-192.
C. Berger and R.M. Kaufmann. . Preprint [arXiv:1710.09438]{}, October 2017.
C. Berger and I. Moerdijk. Resolution of coloured operads and rectification of homotopy algebras. In [*Categories in algebra, geometry and mathematical physics*]{}, volume 431 of [*Contemp. Math.*]{}, pages 31–58. Amer. Math. Soc., Providence, RI, 2007.
J.M. Boardman and R.M. Vogt. . Springer-Verlag, 1973.
P. Boavida de Brito and M. Weiss. Spaces of smooth embeddings and configuration categories. , 11(1): 65–143, 2018.
D. Borisov and Y. Manin. . Preprint [math.CT/0609748]{}, September 2006.
Z. Fiedorowicz. , 1992.
N. Gambino and J. Kock. . , Soc. 154(1): 153–192, 2013.
W.L. Gan. . , 10(1):109–124, 2003.
R. Garner, J. Kock and M. Weber. Operadic categories and décalage, Preprint [arXiv:1812.01750 \[math.CT\]]{}, December 2018.
E. Getzler. . In [*[Algebra, arithmetic, and geometry: in honor of [Y]{}u. [I]{}. [M]{}anin. [V]{}ol. [I]{}]{}*]{}, volume 269 of [*[Progr. Math.]{}*]{}, pages 675–698. Birkhäuser Boston Inc., Boston, MA, January 2009. Preprint [math.CT/0701767]{}.
E. Getzler and M.M. Kapranov. . In S.-T. Yau, editor, [*[Geometry, Topology and Physics for [Raoul]{} [Bott]{}]{}*]{}, volume 4 of [*[Conf. Proc. Lect. Notes. Geom. Topol.]{}*]{}, pages 167–201. International Press, 1995.
E. Getzler and M.M. Kapranov. . , 110(1):65–126, 1998.
V. Ginzburg and M.M. Kapranov. . , 76(1):203–272, 1994.
R.M. Kaufmann and B.C. Ward. . , 387:x+161, 2017.
R.M. Kaufmann, B.C. Ward, and J.J. Zuniga. . , 56:103504, 2015.
J. Kock, A. Joyal, M.A. Batanin and J. Mascari, , , 224(6):2690-2737, 2010.
S. Lack. , , 2(1): 1–29, 2018.
T.G. Lavers. . 25:1257–1284, 1997.
J. Leray. PhD thesis, University of Angers. J.-L. Loday. . Slides of a talk given at HOGT Lille on 23th of March 2012.
J.-L. Loday and M.O. Ronco. . , 333(2):81–86, 2001.
J.-L. Loday and M.O. Ronco. . , 120(2):340–365, 2013.
S. [Mac Lane]{}. . , 49(4):28–46, 1963.
M. Markl. . , 46(4):307–323, 1996.
M. Markl. . , 24(4):1471–1500, 1996.
M. Markl. . In [*[Handbook of algebra. [V]{}ol. 5]{}*]{}, pages 87–140. Elsevier/North-Holland, Amsterdam, 2008.
M. Markl. . , 27(2):567–1580, 2017.
M. Markl. .
M. Markl, S.A. Merkulov, and S. Shadrin. . , 213:496–535, 2009.
M. Markl, S. Shnider, and J.D. Stasheff. , volume 96 of [ *[Mathematical Surveys and Monographs]{}*]{}. American Mathematical Society, Providence, RI, 2002.
M. Markl and A.A. Voronov. . In [*[Algebra, arithmetic, and geometry: in honor of [Y]{}u. [I]{}. [M]{}anin. [V]{}ol. [II]{}]{}*]{}, volume 270 of [*[Progr. Math.]{}*]{}, pages 249–281. Birkhäuser Boston, Inc., Boston, MA, 2009.
J.P. May. . Springer-Verlag, Berlin, 1972. Lectures Notes in Mathematics, Vol. 271.
B. Vallette. . , 359(10):4865–4943, 2007.
M. Weber. . , 50, 1713–1792.
M. Weber. . , 30, 1659–1712.
J. Yoshida. . Preprint
[^1]: As everywhere in this work, we do not consider the constants $P(0)$.
[^2]: In [@duodel] such morphisms were called [*trivial*]{}.
[^3]: Denoted in [@duodel] by $\sFSet$.
[^4]: We refer to the axioms of operadic categories recalled in Appendix \[a0\].
[^5]: Also called [*half-edges*]{}.
[^6]: Thus $\psi$ maps injectively flags to flags and bijectively legs to legs.
[^7]: We denote by the same symbol both the map of flags and the obvious induced map of edges.
[^8]: See Section \[a0\].
[^9]: Recall that all operadic categories in this work are assumed to be constant-free.
[^10]: It will always be clear whether we apply Grothendieck’s construction to an operad or cooperad.
[^11]: That is a classical colored operad with objects of $\ttO$ as its set of colors.
[^12]: Recall that by Lemma \[Podivam\_se\_do\_Paramaty?\], $f$ is an isomorphisms if and only if all its fibers are terminal.
[^13]: Cf. Remark \[Zitra\_mam\_prednasku\_na\_Macquarie.\].
[^14]: We put $T'_0 := T'$, $T''_0 := T''$.
[^15]: Adjoining the ground field in (\[Konci\_Neuron?\]) should be compared to adjoining the unit to the free nonunital operad in formula (II.1.58) of [@markl-shnider-stasheff:book].
[^16]: Recall that $\tau$ is the commutativity constraint.
[^17]: We are using the notation for structure operations of odd modular operads introduced in [@KWZ].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Lax Logical Framework, , was introduced, by a team including the last two authors, to provide a conceptual framework for integrating different proof development tools, thus allowing for [*external evidence*]{} and for [*postponing, delegating*]{}, or [*factoring-out*]{} side conditions. In particular, allows for [*reducing*]{} the number of times a [*proof-irrelevant*]{} check is performed. In this paper we give a shallow, actually [*definitional*]{}, implementation of in [`Coq`]{}, we use [`Coq`]{} both as host framework and oracle for . This illuminates the principles underpinning the mechanism of [*Lock-types*]{} and also suggests how to possibly extend [`Coq`]{} with the features of . The derived proof editor is then put to use for developing case-studies on an emerging paradigm, both at logical and implementation level, which we call [*fast and loose reasoning*]{} following Danielsson et [*alii*]{} [@FL]. This paradigm trades off efficiency for correctness and amounts to postponing, or running in parallel, tedious or computationally demanding checks, until we are really sure that the intended goal can be achieved. Typical examples are branch-prediction in CPUs and optimistic concurrency control.'
author:
- Fabio Alessi Alberto Ciaffaglione Pietro Di Gianantonio Furio Honsell Marina Lenisa
bibliography:
- 'biblio.bib'
nocite: '[@*]'
title: 'A Definitional Implementation of the Lax Logical Framework in [`Coq`]{}, for Supporting Fast and Loose Reasoning'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this article we investigate the relation between consequences of Dirichlet boundary conditions (momenta noncommutativity and parameters of the effective theory) and background fields of fermionic T-dual theory. We impose Dirichlet boundary conditions on the endpoints of the open string propagating in background of type IIB superstring theory with constant background fields. We showed that on the solution of the boundary conditions the momenta become noncommutative, while the coordinates commute. Fermionic T-duality is also introduced and its relation to noncommutativity is considered. We use compact notation so that type IIB superstring formally gets the form of the bosonic one with Grassman variables. Then momenta noncommutativity parameters are fermionic T-dual fields. The effective theory, the initial theory on the solution of boundary conditions, is bilinear in the effective coordinates, odd under world-sheet parity transformation. The effective metric is equal to the initial one and terms with the effective Kalb-Ramond field vanish.'
author:
- |
B. Nikolić [^1] and B. Sazdović [^2]\
[*Institute of Physics*]{}\
[*University of Belgrade*]{}\
[*P.O.Box 57, 11001 Belgrade, Serbia*]{}
title: ' Dirichlet boundary conditions in type IIB superstring theory and fermionic T-duality [^3]'
---
\#1\#2 \#1[.3cm[**\#1**]{}]{}
§ Ø c
: 11.10.Nx, 04.20.Fy, 11.10.Ef, 11.25.-w
Introduction
============
In string theory T-duality is an important tool to show the equivalence of different geometries and topologies and in determining some of the genuinely stringy implications on structure of the low energy quantum field theory limit [@rabin]. The T-duality rules that relate different curved backgrounds with an Abelian isometry were first constructed by Buscher [@buscher]. These rules are known as the Buscher T-duality rules.
Up to few years ago only T-duality along bosonic coordinates has been considered. Analyzing the gluon scattering amplitudes in $N=4$ super Yang-Mills theory, a new kind of T-dual symmetry, fermionic T-duality, was discovered [@ferdual; @MW1]. Mathematically, fermionic T-duality is realized within the same procedure as bosonic one, except that dualization is performed along fermionic directions, $\theta^\alpha$ and $\bar\theta^\alpha$, so it can be considered as a generalization of Buscher T-duality to theories with fermionic variables. The fermionic T-duality maps superstring in certain supersymmetric background to superstring in another supersymmetric background.
On the other hand, in the open string theory besides equations of motion there appear boundary conditions also. In this article we will derive the boundary conditions in pure canonical manner and treat them as canonical constraints as in Refs.[@kanonski; @BNBS]. Checking consistency of the constraints, it turns out that there are infinite number of consistency conditions. Using Taylor expansion we will rewrite them in compact $\sigma$ dependent form. For the case of Dirichlet boundary conditions all constraints are of the second class and we solve them. In this way we will obtain initial coordinates and momenta in terms of the effective variables.
In the previous articles [@bnbsjhep; @bnbsnpb] we considered type IIB superstring theory in pure spinor formulation and applied Busher T-duality along all bosonic directions $x^\mu$. We obtained dual theory and its background fields which represented $N=2$ supermultiplet. This supermultiplet consists of two $N=1$ supermultiplets [@bnbsjhep]. The $N=1$ supermultiplet, which is odd under world-sheet parity transformation $\Omega:\sigma\to-\sigma$, in fact contains noncommutativity parameters corresponding to Neumann boundary conditions. Effective theory is initial theory on the solution of boundary conditions with $\Omega$ even coordinates. The other $N=1$ supermultiplet, which is $\Omega$ even, contains background fields of the effective theory.
In Ref.[@bnbsjhep] we found that, for the specific boundary conditions some of the bosonic T-dual background fields were equal to the noncommutativity parameters and the other ones were equal to the effective background fields. Motivated by this fact, in the paper [@bnbsfd] we found such boundary conditions which produced noncommutativity parameters equal to the fermionic T-dual background fields. In that article we used light-cone canonical analysis in order to find suitable variables (currents) which enable us to perform calculations.
In the present paper we investigate the relation between consequences of Dirichlet boundary conditions (noncommutativity parameters and effective background fields) and background fields of the fermionic T-dual theory. The paper is organized in the following way. First, we introduce the action of the pure spinor formulation for type IIB superstring theory keeping quadratic and neglecting ghost terms. We assume that all background fields of type IIB theory: metric tensor $G_{\mu\nu}$, antisymmetric field $B_{\mu\nu}$, two gravitinos $\Psi^\alpha_\mu$ and $\bar\Psi^\alpha_\mu$ and bispinor $F^{\alpha\beta}$ are constant. Also we assume that bispinor $F^{\alpha\beta}$ is invertible and that generalized metric $G_{AB}$ introduced in Section 4 is nonsingular. Then we perform fermionic T-duality and find the explicit expressions for T-dual background fields.
From this point we change approach of Ref.[@bnbsfd]. Instead light-cone canonical analysis we rewrite the action in the compact form in terms of the generalized coordinates $x^A=(x^\mu,\theta^\alpha,\bar\theta^\alpha)$, metric tensor $G_{AB}$ and Kalb-Ramond field $B_{AB}$. It is unexpected that the components of the generalized metric $G_{AB}$ and Kalb-Ramond field $B_{AB}$ are equal to the fermionic T-dual background fields of the paper [@bnbsfd]. Because the action has the same form as in bosonic case we define the currents analogous with those of Ref.[@BNBS] but now in the extended space. It turns out that components of these currents are equal to the currents obtained by light-cone canonical analysis.
Choosing Dirichlet boundary conditions, $\dot
x^A|_0^\pi=0$, and treating them as canonical constraints, we find the initial coordinates and momenta in terms of the effective ones. Note that in this case the effective variables are $\Omega$ odd parts of the initial variables. It turns out that coordinates are commutative, while the momenta are noncommutative. The reason for noncommutativity is the presence of the effective coordinates in the expressions for initial momenta. The noncommutativity parameter is proportional to the generalized Kalb-Ramond field $B_{AB}$.
We find the form of the initial theory on the solution of the boundary conditions, which we will call effective theory. It turns out that the effective generalized coordinates are $\Omega$ odd, while the Lagrangian, which is bilinear in generalized coordinates, is $\Omega$ even. The effective metric is equal to initial one $G_{AB}$, while the terms with effective Kalb-Ramond field vanish.
It is known that noncommutativity parameters and background fields obtained on the solution of Neumann boundary conditions correspond to T-dual background fields [@bnbsjhep; @bnbsnpb; @bnbsfd], if T-duality is performed along bosonic coordinates. In this paper we will present similar result just considering Dirichlet boundary conditions instead Neumann ones and fermionic T-duality instead bosonic one.
At the end we give some concluding remarks.
Type IIB superstring and fermionic T-duality
============================================
In this section we will introduce the action of type IIB superstring theory in pure spinor formulation and perform fermionic T-duality.
Action
------
The action of type IIB superstring theory in pure spinor formulation up to the quadratic terms [@berko; @susyNC; @BNBSPLB] neglecting ghost terms as in Ref.[@susyNC] is of the form $$\begin{aligned}
\label{eq:SB}
&{}&S(\partial_\pm x, \partial_\pm \theta, \partial_\pm \bar\theta)=\kappa \int_\Sigma d^2\xi \partial_{+}x^\mu
\Pi_{+\mu\nu}\partial_- x^\nu \\&+&\int_\Sigma d^2 \xi \left[
-\pi_\alpha
\partial_-(\theta^\alpha+\Psi^\alpha_\mu
x^\mu)+\partial_+(\bar\theta^{\alpha}+\bar \Psi^{\alpha}_\mu
x^\mu)\bar\pi_{\alpha}+\frac{1}{2\kappa}\pi_\alpha F^{\alpha
\beta}\bar \pi_{\beta}\right ]\, ,\nonumber\end{aligned}$$ where $G_{\mu\nu}$ is metric tensor, $B_{\mu\nu}$ antisymmetric Kalb-Ramond field, $\Psi^\alpha_\mu$ and $\bar\Psi^\alpha_\mu$ are gravitino fields and bispinor $F^{\alpha\beta}$ is RR field strength. In oreder to simplify calculation we suppose that all background fields are constant. The world sheet $\Sigma$ is parametrized by $\xi^m=(\xi^0=\tau\, ,\xi^1=\sigma)$ and $\partial_\pm=\partial_\tau\pm\partial_\sigma$. Superspace is spanned by bosonic coordinates $x^\mu$ ($\mu=0,1,2,\dots,9$) and fermionic ones, $\theta^\alpha$ and $\bar\theta^{\alpha}$ $(\alpha=1,2,\dots,16)$. The variables $\pi_\alpha$ and $\bar
\pi_{\alpha}$ are canonically conjugated momenta to the coordinates $\theta^\alpha$ and $\bar\theta^\alpha$, respectively. All spinors are Majorana-Weyl ones and $\Pi_{\pm \mu\nu}=B_{\mu\nu}\pm\frac{1}{2}G_{\mu\nu}$.
On the equations of motion for fermionic momenta $\bar\pi_\alpha$ and $\pi_\alpha$ we have respectively $$\label{eq:impulsi}
\pi_\alpha=-2\kappa \partial_+(\bar\theta^\beta+\bar\Psi^\beta_\mu x^\mu)(F^{-1})_{\beta\alpha}\, ,\quad \bar\pi_\alpha=2\kappa (F^{-1})_{\alpha\beta}\partial_-(\theta^\beta+\Psi^\beta_\mu x^\mu)\, ,$$ and the action gets the form $$\begin{aligned}
\label{eq:lcdejstvo}
&{}&S=\kappa \int_\Sigma d^2\xi \partial_+ x^\mu \left[\Pi_{+\mu\nu}+2\bar\Psi^\alpha_\mu(F^{-1})_{\alpha\beta}\Psi^\beta_\nu\right]\partial_-x^\nu \\ &{}& +2\kappa \int_\Sigma d^2\xi \left[ \partial_+\bar\theta^\alpha (F^{-1})_{\alpha\beta}\partial_-\theta^\beta+\partial_+\bar\theta^\alpha (F^{-1})_{\alpha\beta}\Psi^\beta_\nu\partial_-x^\nu+\partial_+x^\mu \bar\Psi^\alpha_\mu (F^{-1})_{\alpha\beta}\partial_-\theta^\beta\right]\, .\nonumber\end{aligned}$$ Here we assume that RR field strength $F^{\alpha\beta}$ is invertible.
Fermionic T-duality
-------------------
Because the action has a global shift symmetry in $\theta^\alpha$ and $\bar\theta^\alpha$ directions, we introduce gauge fields, $(v^\alpha_+, v^\alpha_-)$ and $(\bar v^\alpha_+, \bar v^\alpha_-)$, to get a local symmetry $$\partial_\pm\theta^\alpha \to D_\pm \theta^\alpha\equiv\partial_\pm\theta^\alpha+v_\pm^\alpha\, , \quad \partial_\pm\bar\theta^\alpha \to D_\pm
\bar\theta^\alpha\equiv\partial_\pm\bar\theta^\alpha+\bar v_\pm^\alpha\, .$$ Also we introduce the Lagrange multipliers $\vartheta_\alpha$ and $\bar\vartheta_\alpha$ which will impose the field strengths of gauge fields $v_\pm^\alpha$ and $\bar
v_\pm^\alpha$ to vanish $$S_{gauge}(\vartheta, \bar\vartheta, v_\pm,\bar v_\pm)=\frac{1}{2}\kappa \int_\Sigma d^2\xi \bar \vartheta_\alpha (\partial_+
v_-^\alpha-\partial_- v^\alpha_+)+\frac{1}{2}\kappa \int_\Sigma d^2\xi (\partial_+
\bar v_-^\alpha-\partial_- \bar v^\alpha_+)\vartheta_\alpha\, .$$ So, we do not introduce new degrees of freedom and the full action is of the form $$\label{eq:auxdejstvo}
{}^\star S(x, \theta,\bar\theta,\vartheta,\bar \vartheta,v_\pm,\bar v_\pm)=S(\partial_\pm x, D_- \theta, D_+ \bar\theta)+S_{gauge}(\vartheta,\bar \vartheta, v_\pm, \bar v_\pm)\, .$$
Now we fix $\theta^\alpha$ and $\bar\theta^\alpha$ to zero and obtain the action quadratic in the fields $v_\pm$ and $\bar v_\pm$ $$\begin{aligned}
\label{eq:JED}
&{}&{}^\star S=\kappa \int_\Sigma d^2\xi \partial_+ x^\mu \left[\Pi_{+\mu\nu}+2\bar\Psi^\alpha_\mu(F^{-1})_{\alpha\beta}\Psi^\beta_\nu\right]\partial_-x^\nu \\ &{}& +2\kappa \int_\Sigma \left[ \bar v_+^\alpha (F^{-1})_{\alpha\beta}v_-^\beta+\bar v_+^\alpha (F^{-1})_{\alpha\beta}\Psi^\beta_\nu\partial_-x^\nu+\partial_+x^\mu \bar\Psi^\alpha_\mu (F^{-1})_{\alpha\beta}v_-^\beta\right]\nonumber \\ &{}& +\frac{\kappa}{2} \int_\Sigma d^2\xi \left[\bar \vartheta_\alpha (\partial_+
v_-^\alpha-\partial_- v^\alpha_+)+ (\partial_+
\bar v_-^\alpha-\partial_- \bar v^\alpha_+)\vartheta_\alpha\right]\, .\nonumber\end{aligned}$$
On the equations of motion for multipliers $\vartheta_\alpha$ and $\bar \vartheta_\alpha$ we obtain $\partial_+
v^\alpha_--\partial_- v^\alpha_+=0$ and $\partial_+
\bar v^\alpha_--\partial_- \bar v^\alpha_+=0$ which gives $$\label{eq:vteta}
\bar v_\pm^\alpha=\partial_\pm \bar\theta^\alpha\, ,\quad v_\pm^\alpha=\partial_\pm \theta^\alpha\, .$$ Substituting these expression in (\[eq:JED\]) we obtain the initial action (\[eq:lcdejstvo\]).
On the equations of motion for $v_\pm^\alpha$ and $\bar v_\pm^\alpha$ we obtain, respectively $$\label{eq:jed1}
\partial_- \bar \vartheta_\alpha=0\, ,\quad \bar v_+^\alpha=\frac{1}{4}\partial_+ \bar \vartheta_\beta F^{\beta\alpha}-\partial_+ x^\mu \bar\Psi^\alpha_\mu\, ,$$ $$\label{eq:jed2}
\partial_+ \vartheta_\alpha=0\, ,\quad v_-^\alpha=-\frac{1}{4}F^{\alpha\beta}\partial_-\vartheta_\beta-\Psi^\alpha_\mu \partial_- x^\mu\, .$$ Substituting these expression in the action ${}^\star S$ we obtain the dual action $$\begin{aligned}
\label{eq:dualnodej}
&{}& {}^\star S(\partial_\pm x, \partial_- \vartheta, \partial_+
\bar \vartheta)=\kappa\int_\Sigma d^2\xi \partial_+ x^\mu \Pi_{+\mu\nu}\partial_- x^\nu\, ,\\ &{}&+\frac{\kappa}{8}\int_\Sigma d^2\xi\left[\partial_+\bar \vartheta_\alpha F^{\alpha\beta}\partial_-\vartheta_\beta -4\partial_+x^\mu\bar\Psi^{\alpha}_{\mu} \partial_-\vartheta_\alpha+4\partial_+\bar \vartheta_\alpha\Psi^{\alpha}_{\mu}\partial_- x^\mu\right]\, .\nonumber\end{aligned}$$ Demanding that dual action has the same form as initial one (\[eq:lcdejstvo\]) $$\begin{aligned}
\label{eq:lcdejstvodual}
&{}&{}^\star S=\kappa \int_\Sigma d^2\xi \partial_+ x^\mu \left[{}^\star \Pi_{+\mu\nu}+2{}^\star\bar\Psi^\alpha_\mu({}^\star F^{-1})_{\alpha\beta}{}^\star\Psi^\beta_\nu\right]\partial_-x^\nu \\ &{}& +2\kappa \int_\Sigma d^2\xi \left[ \partial_+\bar\vartheta^\alpha ({}^\star F^{-1})_{\alpha\beta}\partial_-\vartheta^\beta+\partial_+\bar\vartheta^\alpha ({}^\star F^{-1})_{\alpha\beta}\Psi^\beta_\nu\partial_-x^\nu+\partial_+x^\mu {}^\star\bar\Psi^\alpha_\mu ({}^\star F^{-1})_{\alpha\beta}\partial_-\vartheta^\beta\right]\, ,\nonumber\end{aligned}$$ and comparing it with (\[eq:lcdejstvo\]), we read the dual background fields (with stars) $$\label{eq:GBdual}
{}^\star B_{\mu\nu}=B_{\mu\nu}+\left[ (\bar\Psi F^{-1}\Psi)_{\mu\nu}-(\bar\Psi F^{-1}\Psi)_{\nu\mu}\right] \, , {}^\star G_{\mu\nu}=G_{\mu\nu}+2\left[ (\bar\Psi F^{-1}\Psi)_{\mu\nu}+(\bar\Psi F^{-1}\Psi)_{\nu\mu}\right]\, ,$$ $$\label{eq:Psidual}
{}^\star\Psi_{\alpha \mu}=4(F^{-1}\Psi)_{\alpha\mu}\, ,\quad {}^\star\bar\Psi_{\mu\alpha}=-4(\bar\Psi F^{-1})_{\mu\alpha}\, ,$$ $$\label{eq:Fdual}
{}^\star F_{\alpha\beta}=16(F^{-1})_{\alpha\beta}\, .$$ They are well defind because we assume that bispinor $F^{\alpha\beta}$ is invertible. Let us note that two successive dualizations give the initial background fields.
Canonical analysis of type IIB theory in compact notation
=========================================================
We are going to find such noncommutativity parameters corresponding to some boundary conditions which can be related with fermionc T-dual fields (\[eq:GBdual\])-(\[eq:Fdual\]).
If we introduce the supercoordinates $x^A=(x^\mu,\theta^\alpha,\bar\theta^\alpha)$ and supermatrices $$\Pi_{\pm AB}=B_{AB}\pm\frac{1}{2}G_{AB}\, ,$$ as $$\Pi_{+AB}=\left(
\begin{array}{ccc}
{}^\star\Pi_{+\mu\nu} & -\frac{1}{2}{}^\star\bar\Psi_{\mu\beta} & 0\\
0 & 0 & 0\\
\frac{1}{2}{}^\star\Psi_{\alpha\nu} & \frac{1}{8}{}^\star F_{\alpha\beta} & 0
\end{array}
\right)\, ,\quad \Pi_{-AB}=\left(
\begin{array}{ccc}
{}^\star\Pi_{-\mu\nu} & 0 & \frac{1}{2}({}^\star\Psi^T)_{\mu\beta}\\
-\frac{1}{2}({}^\star\bar\Psi^T)_{\alpha\nu} & 0 & \frac{1}{8}({}^\star F^T)_{\alpha\beta}\\
0 & 0 & 0
\end{array}
\right)\, ,$$ then the action (\[eq:lcdejstvo\]) can be rewritten in the form $$\label{eq:Dejstvo}
S=\kappa\int_\Sigma d^2\xi \partial_+ x^A \Pi_{+AB} \partial_- x^B=-\kappa\int_\Sigma d^2\xi \partial_- x^A \Pi_{-AB} \partial_+ x^B\, .$$ From the expression for $\Pi_{\pm AB}$ we read the supersymmetric generalization of the metric, $G_{AB}$, and antisymmetric Kalb-Ramond field, $B_{AB}$, $$\label{eq:GAB}
G_{AB}=\Pi_{+AB}-\Pi_{-AB}=\left(
\begin{array}{ccc}
{}^\star G_{\mu\nu} & -\frac{1}{2}{}^\star \bar\Psi_{\mu\beta} & -\frac{1}{2}({}^\star \Psi^T)_{\mu\beta}\\
\frac{1}{2}({}^\star \bar\Psi^T)_{\alpha\nu} & 0 & -\frac{1}{8}({}^\star F^T)_{\alpha\beta}\\
\frac{1}{2}{}^\star\Psi_{\alpha\nu} & \frac{1}{8}{}^\star F_{\alpha\beta} & 0
\end{array}
\right)\, ,$$ $$\label{eq:BAB}
B_{AB}=\frac{1}{2}(\Pi_{+AB}+\Pi_{-AB})=\left(
\begin{array}{ccc}
{}^\star B_{\mu\nu} & -\frac{1}{4}{}^\star \bar\Psi_{\mu\beta} & \frac{1}{4}({}^\star \Psi^T)_{\mu\beta}\\
-\frac{1}{4}({}^\star \bar\Psi^T)_{\alpha\nu} & 0 & \frac{1}{16}({}^\star F^T)_{\alpha\beta}\\
\frac{1}{4}{}^\star\Psi_{\alpha\nu} & \frac{1}{16}{}^\star F_{\alpha\beta} & 0
\end{array}
\right)\, .$$ Note that $G_{BA}=(-)^{A+B+AB}G_{AB}$ and $B_{BA}=-(-)^{A+B+AB}B_{AB}$. Consequently, we have $x^A G_{AB}y^B=y^A G_{AB}x^B$ and $x^A B_{AB} x^B=0$.
The momenta canonically conjugated to the coordinates $x^A$ are $$\label{eq:impulsA}
\pi_A=\frac{\partial_L \mathcal L}{\partial \dot x^A}=\kappa(G_{AB}\dot x^B-2B_{AB}x'^B)\, .$$ The basic Poisson algebra is of the form $$\left\lbrace x^A(\sigma),\pi_B(\bar\sigma) \right\rbrace=(-)^A \delta^A{}_B \delta(\sigma-\bar\sigma)\, ,$$ where $\delta^A{}_B$ is unity operator in the superspace $$\delta^A{}_B=\left(
\begin{array}{ccc}
\delta^\mu{}_\nu & 0 & 0 \\
0 & \delta^\alpha{}_\gamma & 0\\
0 & 0 & \delta^\beta{}_\delta
\end{array}
\right)\, .$$ Let us now introduce the currents $$\label{eq:gdstruja}
J_{\pm A}=\pi_A+2\kappa\Pi_{\pm AB}x'^B\, ,\quad J^A_\pm\equiv (G^{-1})^{AB}J_{\pm B}\, ,$$ which satisfy Abelian Kac-Moody algebra $$\label{eq:algebraJ}
\left\lbrace J_{\pm A}(\sigma), J_{\pm B}(\bar\sigma)\right\rbrace=\pm 2\kappa G^{st}_{AB}\delta'\, ,\quad \left\lbrace J_{\pm A}(\sigma), J_{\mp B}(\bar\sigma)\right\rbrace =0\, .$$ Here we introduced supertransposition $$X_{AB}=\left(\begin{array}{cc}
A & B\\
C & D
\end{array}
\right)\, ,\quad X^{st}_{AB}=(-)^{A+AB}X_{BA}=\left(\begin{array}{cc}
A^T & C^T\\
-B^T & D^T
\end{array}
\right)\, ,$$ where $T$ is related to the ordinary transposition. Note that $G_{AB}^{st}=(-)^B G_{AB}$.
Using the expression for canonical momenta the currents get the form $$\label{eq:Astruja}
J_{\pm A}=\kappa G_{AB}\partial_{\pm} x^B\, ,$$ from which we can express $\tau$ and $\sigma$ derivative of $x^A$ in terms of the currents $$\label{eq:izvodi}
\dot x^A=\frac{(G^{-1})^{AB}}{2\kappa}(J_{+B}+J_{-B})\, ,\quad x'^A=\frac{(G^{-1})^{AB}}{2\kappa}(J_{+B}-J_{-B})\, .$$ From the definition of the canonical Hamiltonian $$\mathcal H_c=\dot x^A\pi_A-\mathcal L\, ,$$ using the expression (\[eq:izvodi\]), (\[eq:impulsA\]) and (\[eq:Dejstvo\]), we obtain canonical Hamiltonian in terms of the currents $$\label{eq:noviHam}
H_c=\int d\sigma \mathcal H_c\, ,\quad \mathcal H_c=T_--T_+\, ,\quad T_{\pm}=\mp\frac{1}{4\kappa}J^A_\pm G_{AB} J^B_\pm\, .$$ Using the definition of the current with upper index (\[eq:gdstruja\]) we rewrite the energy-momtum tensor components in the form $$\label{eq:tei}
T_{\pm}=\mp\frac{1}{4\kappa}J_{\pm A}[(G_{st})^{-1}]^{AB}J_{\pm B}\, ,$$ Let us stress that $[(G_{st})^{-1}]^{AB}=(-)^{A+B}[(G^{-1})_{st}]^{AB}$.
The energy-momentum tensor components are in Sugawara form and satisfy two independent Virasoro algebras $$\left\lbrace T_\pm(\sigma),T_\pm(\bar\sigma)\right\rbrace=-\left[T_{\pm}(\sigma)+T_\pm(\bar\sigma)\right]\delta'\, ,\quad \left\lbrace T_\pm(\sigma),T_\mp(\bar\sigma)\right\rbrace=0\, .$$ For further analysis it is useful the following relation $$\label{eq:HJ}
\left\lbrace H_c, J_{\pm A}\right\rbrace=\mp J'_{\pm A}\, .$$
Boundary conditions and fermionic T-duality
===========================================
We are looking for such boundary conditions that corresponding noncommutativity parameters are just the background fields of the fermionic T-dual theory (\[eq:GBdual\])-(\[eq:Fdual\]).
Varying the Hamiltonian (\[eq:noviHam\]) we obtain $$\label{eq:korisno2}
\delta H_c=\delta H_c^{(R)}-\tilde\gamma_A^{(0)}\delta x^A|_0^\pi\, ,$$ where $\delta H_c^{(R)}$ has a form $\delta H_c^{(R)}=C_A \delta x^A+D^A \delta \pi_A$. It contains variations $\delta x^A$ and $\delta\pi_A$ but does not contain corresponding variations of the $\sigma$ derivatives $\delta x'^A$ and $\delta \pi'_A$. The boundary term was obtained, after partial integration, from parts of the form $E_A\delta x'^A$ and it yields $$\tilde\gamma_A^{(0)}=\Pi_{+AB} J^B_-+\Pi_{-AB} J^B_+\, .$$ Because the Hamiltonian is time translation generator it must have well defined functional derivatives with respect to the coordinates and momenta. Consequently, boundary term must vanish $$\label{eq:GU}
\tilde\gamma_A^{(0)}\delta x^A|_0^\pi=0\, .$$ We choose Dirichlet boundary conditions $$\label{eq:gu1}
\gamma_A^{(0)}|_0^\pi=0\, ,\quad \gamma_A^{(0)}=2\kappa G_{AB}\dot
x^B=J_{+A}+J_{-A}\, .$$ The fact that velocity $\dot x^A$ is zero at string endpoints means that the endpoints do not move. Consequently, string endpoints are fixed, $\delta x^A|_0^\pi=0$, and they solve boundary conditions (\[eq:GU\]). Applying Dirac consistency procedure we obtain infinite set of the constraints $$\gamma^{(n)}_A|_0^\pi=0\, ,\quad \gamma_A^{(n)}=\left\lbrace H_c, \gamma_A^{(n-1)} \right\rbrace\, .(n=1,2,3,\dots)$$ With the help of the relation (\[eq:HJ\]) using Taylor expansion $$\Gamma_A(\sigma)=\sum_{n=0}^\infty \frac{\sigma^n}{n!}\gamma_A^{(n)}|_0\, ,$$ we rewrite these infinite sets of consistency conditions at $\sigma=0$ in compact, $\sigma$ dependent form $$\Gamma_A(\sigma)=J_{+A}(-\sigma)+J_{-A}(\sigma)\, .$$ In the similar way we can write the consistency conditions at $\sigma=\pi$. If we impose $2\pi$ periodicity of the canonical variables, the solution of the constraints at $\sigma=0$ also solve the constraints at $\sigma=\pi$.
Using the algebra of the currents (\[eq:algebraJ\]) we obtain the algebra of the constraints $$\left\lbrace \Gamma_A(\sigma),\Gamma_B(\bar\sigma) \right\rbrace=-4\kappa G^{st}_{AB}\delta'\, .$$ Because we assume that metric $G_{AB}$ is nonsingular, the constraints are of the second class.
Solution of the boundary conditions and noncommutativity
========================================================
Appearance of the second class constraints in the theory means that we have either to introduce Dirac brackets or to solve the constraints. The result will not depend on the choice, but for the practical reasons, we will solve the constraints.
Solving the constraint equations $$\Gamma_A(\sigma)=0\, ,$$ we obtain initial variables in terms of the effective ones $$\label{eq:x}
x^A(\sigma)=\tilde q^A(\sigma)\, ,\quad \pi_A=\tilde p_A-2\kappa B_{AB}\tilde q'^B\, .$$ Here we introduced new variables, symmetric and antisymmetric under world-sheet parity transformation $\Omega:\sigma\to -\sigma$. For bosonic variables we use standard notation [@BNBS] $$\begin{aligned}
\label{eq:bv1}
q^A(\sigma)&=&P_s x^A(\sigma)\, ,\quad \tilde
q^A(\sigma)=P_a x^A(\sigma)\, ,\nonumber\\ p_A(\sigma)&=&P_s \pi_A(\sigma)\, ,\quad \tilde
p_A(\sigma)=P_a \pi_A(\sigma)\, ,\end{aligned}$$ where $P_s$ and $P_a$ are projectors on the $\Omega$ even and odd parts, respectively, $$P_s=\frac{1}{2}(1+\Omega)\, ,\quad P_s=\frac{1}{2}(1-\Omega)\, .$$ Note that all effective independent variables are $\Omega$ odd. Also the initial momenta (not the coordinates) are linear combination of the effective momenta and effective coordinates.
From basic graded Poisson bracket $$\{x^A(\sigma), \pi_B(\bar\sigma)\}=(-)^A\delta^A{}_B
\delta(\sigma-\bar\sigma)\, ,$$ we obtain the corresponding one in $\Omega$ antisymmetric subspace $$\label{eq:pz0}
\{\tilde q^A(\sigma)\, ,\tilde p_B(\bar\sigma)\}=2(-)^A\delta^A{}_B
\delta_a(\sigma\, ,\bar\sigma)\, ,$$ where $$\delta_a(\sigma,\bar\sigma)=\frac{1}{2}\left[
\delta(\sigma-\bar\sigma)-\delta(\sigma+\bar\sigma)\right]\, ,$$ is antisymmetric delta function. Therefore, the momenta $\tilde p_A$ are canonically conjugated to the coordinates $\tilde q^A$.
We conclude that all supercoordinates are commutative, while the Poisson brackets of momenta are nonzero $$\left\lbrace \pi_A(\sigma), \pi_B(\bar\sigma)\right\rbrace = 4\kappa B^{st}_{AB}\partial_\sigma \delta(\sigma+\bar\sigma)\, .$$ In fact variables $\tilde p$ are $\Omega$ odd parts of the momenta density. Consequently, if we define the momenta as $$\tilde P_A(\sigma)=\int_0^\sigma d\sigma_1 \tilde p_A(\sigma_1)\, ,$$ the noncommutativity relations get the form $$\left\lbrace \tilde P_A(\sigma), \tilde P_B(\bar\sigma)\right\rbrace = 4\kappa B^{st}_{AB}\theta(\sigma+\bar\sigma)\, ,$$ where $$\label{eq:fdelt}
\theta(x)=\left\{\begin{array}{ll}
0 & \textrm{if $x=0$}\\
1/2 & \textrm{if $0<x<2\pi$}\, .\\
1 & \textrm{if $x=2\pi$} \end{array}\right .$$ Therefore, the background fields of the fermionic T-dual theory (\[eq:GBdual\])-(\[eq:Fdual\]) (except ${}^\star G_{\mu\nu}$) are noncommutativity parameters for boundary conditions (\[eq:gu1\]).
Effective theory
================
Effective theory is the initial theory on the solution of boundary conditions. In this section we will find the expressions for the Hamiltonian and Lagrangian of the effective theory.
Because the Hamiltonian is bilinear expression in currents, first we have to find the currents $J_{\pm A}$ on the solution of boundary conditions. Substituting the solution (\[eq:x\]) in the expressions for the currents $J_{\pm A}$ (\[eq:gdstruja\]) and energy-tensor components (\[eq:tei\]), we obtain $$J^{eff}_{\pm A}\equiv J_{\pm A}|_{\Gamma=0}=\tilde p_A\pm \kappa G_{AB}\tilde q'^B\, ,$$ $$T^{eff}_{\pm}\equiv T_{\pm}|_{\Gamma=0}=\mp\frac{1}{4\kappa}J^{eff}_{\pm A}[(G_{st})^{-1}]^{AB}J^{eff}_{\pm B}\, .$$ Comparing these relations with (\[eq:gdstruja\]) and (\[eq:tei\]), we find the transition rule from the initial to the effective theory $$\label{eq:trule}
x^A\to\tilde q^A\, ,\quad \pi_A\to \tilde p_A\, ,\quad G_{AB} \to G_{AB}\, ,\quad B_{AB}\to 0\, .$$ Let us stress that this is only prescription which enables us to obtain effective from initial theory. Actually, $B_{AB}$ does not vanish but the term in the action with effective antisymmetric background field, $\int d^2\xi \partial_+ \tilde q^A B^{eff}_{AB}\partial_- \tilde q^B$, disappears.
Using the expression for effective Hamiltonian $$\mathcal H_c^{eff}=T^{eff}_--T^{eff}_+=\frac{1}{2\kappa}\tilde p_A [(G_{st})^{-1}]^{AB} \tilde p_B+\frac{\kappa}{2}\tilde q'^A G_{AB}\tilde q'^B\, .$$ we can find effective Lagrangian $$\mathcal L^{eff}=\dot {\tilde q}^A \tilde p_A-\mathcal H_c^{eff}\, .$$ On the equation of motion for momenta $\tilde p_A$ we have $$\dot {\tilde q}^A=\frac{1}{\kappa}(G^{-1})^{AB}\tilde p_B \, ,$$ which gives $$\tilde p_A=\kappa G_{AB}\dot {\tilde q}^B \, .$$ Substituting this relation in the expression for effective Lagrangian we finally obtain $$\begin{aligned}
&{}&\mathcal L^{eff}=\frac{\kappa}{2}\partial_+ \tilde q^A G_{AB} \partial_- \tilde q^B=\\ &=&\frac{\kappa}{2}\left(\partial \tilde q^\mu {}^\star G_{\mu\nu} \partial_- \tilde q^\nu+\frac{1}{4}\partial_+ \bar\theta_a^\alpha {}^\star F_{\alpha\beta}\partial_- \theta_a^\beta+\partial_+ \bar\theta_a^\alpha {}^\star \Psi_{\alpha\mu}\partial_- \tilde q^\mu-\partial_+ \tilde q^\mu {}^\star \bar\Psi_{\mu\alpha}\partial_-\theta^\alpha_a\right)\, ,\nonumber\end{aligned}$$ where index $a$ in $\theta^\alpha_a$ and $\bar\theta^\alpha_a$ in according with the definition (\[eq:bv1\]) means antisymmetrization under world-sheet parity transformation $\Omega$.
The form of the effective Lagrangian confirms transition rule (\[eq:trule\]). The effective theory is $\Omega$ even. The result formally has the same form as in pure bosonic case [@BNBS] up to the change of indices $A\to\mu$ and take care of order of anticommuting variables. In component notation it means $$\begin{aligned}
&{}& x^\mu\to \tilde q^\mu\, ,\quad \theta^\alpha\to \theta^\alpha_a\, ,\quad \bar\theta^\alpha\to \bar\theta^\alpha_a\, ,\nonumber\\ &{}& G_{\mu\nu}\to G_{\mu\nu} \, ,\quad \Psi^\alpha_\mu\to \Psi^\alpha_\mu\, ,\quad \bar\Psi^\alpha_\mu\to \bar\Psi^\alpha_\mu\, ,\quad F^{\alpha\beta}\to F^{\alpha\beta}\, ,\end{aligned}$$ while $B_{\mu\nu}\to 0$.
Concluding remarks
==================
In the present article we found some unexpected relation between quite different approaches. On the one side, we investigated influence of the Dirichlet boundary conditions on the endpoints of the open string moving in type IIB superstring background and obtained the noncommutativity properties and effective theory. On the other side, we constructed the fermionic T-dual theory and found explicit expressions for T-dual background fields. Finally, we established the relation of the noncommutativity parameters to fermionic T dual fields. We used the pure spinor formulation of the theory keeping all terms up to the quadratic ones and neglecting ghost terms in the action.
Then we performed fermionic T-duality in the way described in Refs.[@ferdual; @bnbsfd]. Comparing initial and T-dual theory, we found the expressions for fermionic T-dual background fields.
On the other side we performed canonical analysis of the theory. To simplify calculations we introduced the extended space-time coordinates $x^A=(x^\mu,\theta^\alpha,\bar\theta^\alpha)$, generalized metric $G_{AB}$ and Kalb-Ramond field $B_{AB}$, and rewrote the action in the form of bosonic string theory. We introduced the currents $J_{\pm A}$ in the extended space-time following the analogy with bosonic string theory. It turned out that they are equal to the currents obtained by light-cone analysis from Ref.[@bnbsfd]. Varying the canonical Hamiltonian and demanding that it has well defined functional derivatives with respect to the coordinates and momenta, we obtained boundary term. We chose Dirichlet boundary conditions, $\dot x^A|_0^\pi\sim (J_{+A}+J_{-A})|_0^\pi=0$ and checked their consistency. We found infinite number of constraints at both string endpoints and introduced one $\sigma$ dependent constraint at each endpoint. The algebra of the $\sigma$ dependent boundary conditions closes on the generalized metric $G_{AB}$. Because it is nonsingular, all constraints are of the second class. Solving the constraints we found that Poisson brackets of the momenta are nonzero, while the Poisson brackets of the coordinates are zero. The noncommutativity parameter is generalized Kalb-Ramond field $B_{AB}$. Taking into account that componets of the the $B_{AB}$ are fermionic T-dual fields (except dual metric ${}^\star G_{\mu\nu}$), we establish relation between fermionic T-duality and noncommutativity.
Plugging the solution of the constraints in the initial theory we obtained effective theory. The effective Lagrangian, bilinear in the effective coordinates, is $\Omega$ even. So, term with generalized Kalb-Ramond field (or all $\Omega$ odd terms) are absent from the effective Lagrangian. Only the terms with generalized metric tensor $G_{AB}$ (\[eq:GAB\]) ($\Omega$ even terms) survive. Note that components of generalized metric of the initial theory are just fermionic T-dual background fields. Therefore, the effective background fields (except $B_{\mu\nu}$) are equal to the initial background fields and there are no bilinear corrections with $\Omega$ odd fields as in the previous cases with Neumann boundary conditions [@bnbsnpb; @BNBSPLB]. Only Kalb-Ramond field $B_{\mu\nu}$ is projected out from the effective theory. Let us stress some differences of the present case obtained on the solution of the Dirichlet boundary conditions and all previous cases (bosonic and type IIB superstring) obtained on the solution of Neumann boundary conditions. In all previous cases effective variables, $q^A$ and $p_A$, are $\Omega$ even, while in the present case they are $\Omega$ odd, $\tilde q^A$ and $\tilde p_A$. Also, here we obtained the momenta noncommutativity instead of standard coordinate noncommutativity.
We found how boundary conditions determine type of T-duality and noncommutative variables. In great majority of papers Neumann boundary conditions are connected with bosonic T-duality and coordinate noncommuattivity. In the present paper Dirichlet boundary conditions produce fermionic T-duality and momenta noncommutativity. It gives some interesting consequences. For example, the dual RR field strength ${}^\star F_{\alpha\beta}$ is inverse of the initial one $F^{\alpha\beta}$. In bosonic T-duality similar relation holds for the corresponding metric tensors ${}^\star G^{\mu\nu}$ and $G_{\mu\nu}$. Therefore, the noncommutativity parameter of fermionic momenta $\Theta_{\alpha\beta}$, which is proportional to ${}^\star F_{\alpha\beta}$, is inverse of background field $F^{\alpha\beta}$ in fermionic part of the dual action $\partial_+\bar\vartheta_\alpha F^{\alpha\beta}\partial_-\vartheta_\beta$. On the other hand, momenta noncommutativity means that product of uncertainties in two momenta related by nontrivial noncommutativity relation can not be arbitrary small. So, this gives condition in the infrared region instead of the coordinate noncommutativity which gives conditions in ultraviolet region. Consequently, it could affect the long distance physics.
[99]{}
E. Alvarez, L. Alvarez-Gaume, Y. Lozano, An Introduction to T-Duality in String Theory, arxiv: hep-th/9410237; A. Giveon, M. Porrati and E. Rabinovici, “Target space duality in string theory”, Phys. Rept. 244 (1994), 77-202, arXiv:hep-th/9401139; I. Bandos and B. Julia, [*JHEP*]{} [**08**]{} (2003) 032; D. Luest, [*JHEP*]{} [**12**]{} (2010) 084. T. H. Buscher, [*Phys. Lett.*]{} [**B194**]{} (1987) 59; T. H. Buscher, [*Phys. Lett.*]{} [**B201**]{} (1988) 466. N. Berkovits and J. Maldacena, [*JHEP*]{} [**09**]{} (2008) 062; I. Bakhmatov and D. S. Berman, [*Nucl. Phys.*]{} B[**832**]{} (2010) 89-108; K. Sfetsos, K. Siampos and D. C. Thompson, QMUL-PH-10-08, arXiv:1007.5142; C. Hao, B. Chen and X. Song, [*JHEP*]{} [**12**]{} (2009) 051. N. Beisert, R. Ricci, A. A. Tseytlin and M. Wolf, [*Phys.Rev.*]{} [**D78**]{} (2008) 126004; R. Ricci, A. A. Tseytlin and M. Wolf, [*JHEP*]{} [**12**]{} (2007) 082. F. Ardalan, H. Arfaei, M. M. Sheikh-Jabbari, [*Nucl. Phys.*]{} B[**576**]{}, 578 (2000); C. S. Chu, P. M. Ho, [*Nucl. Phys.*]{} B[**568**]{}, 447 (2000); T. Lee, [*Phys. Rev.*]{} [**D62**]{} (2000) 024022. B. Sazdović, Eur. Phys. J. **C44** (2005) 599; B. Nikolić and B. Sazdović, [*Phys. Rev.*]{} [**D 74**]{} (2006) 045024; B. Nikolić and B. Sazdović, [*Phys. Rev.*]{} [**D 75**]{} (2007) 085011; B. Nikolić and B. Sazdović, [*Adv. Theor. Math. Phys.*]{} [**14**]{} (2010) 1. B. Nikolić and B. Sazdović, [*JHEP*]{} [**08**]{} (2010) 037. B. Nikolić and B. Sazdović, [*Nucl. Phys.*]{} [**B 836**]{} (2010) 100-126. B. Nikolić and B. Sazdović, [*Phys.Rev.*]{} [**D84**]{} (2011) 065012.
N. Berkovits, hep-th/0209059; P. A. Grassi, G. Policastro and P. van Nieuwenhuizen, [*JHEP*]{} [**10**]{} (2002) 054; P. A. Grassi, G. Policastro and P. van Nieuwenhuizen, [*JHEP*]{} [**11**]{} (2002) 004; P. A. Grassi, G. Policastro and P. van Nieuwenhuizen, [*Adv. Theor. Math. Phys.*]{} [**7**]{} (2003) 499; P. A. Grassi, G. Policastro and P. van Nieuwenhuizen, [*Phys. Lett.*]{} [**B553**]{} (2003) 96. J. de Boer, P. A. Grassi and P. van Nieuwenhuizen, [*Phys. Lett.*]{} [**B574**]{} (2003) 98. B. Nikolić and B. Sazdović, [*Phys. Lett.*]{} [**B666**]{} (2008) 400.
[^1]: e-mail address: [email protected]
[^2]: e-mail address: [email protected]
[^3]: This work was supported in part by the Serbian Ministry of Education and Science, under contract No. 171031.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It was previously shown that any two-colour colouring of $K(C_n)$ must contain a monochromatic planar $K_4$ subgraph for $n \ge N^*$, where $6 \le N^* \le N$ and $N$ is Graham’s number. The bound was later [@aemp] improved to $11 \le N^* \le N$. In this article, it is improved to $13 \le N^* \le N$.'
author:
- Jerome Barkley
title: Improved lower bound on an Euclidean Ramsey problem
---
makecaption\#1\#2[ tempboxa[[[**]{}\#1: \#2]{}]{} tempboxa > to ]{}
Introduction
============
Consider an $n$-dimensional hypercube $C_n$. Consider the complete graph $K(C_n)$ connecting the vertices of the $n$-cube $C_n$, and consider a two-colour colouring of $K(C_n)$. Let $N^*$ be the smallest integer, such that any two-colour colouring of $K(C_{N^*})$ must contain a monochromatic planar $K_4$ (4 vertices, complete) subgraph.
It was shown [@rtfnps] that $6 \le N^* \le N$, where $N$ is Graham’s number, a very large number. It was later shown [@aemp] that $11 \le N^*$, by constructing a colouring of $K(C_{10})$ which doesn’t contain a monochromatic planar $K_4$ subgraph. According to Exoo, at least two other people have constructed unpublished colourings of $K(C_{11})$, showing that $12 \le N^*$.
*Fig. \[cubes\]a* shows $K(C_3)$, *Fig. \[cubes\]b* shows $K(C_3)$ with one planar $K_4$ subgraph in red, and *Fig. \[cubes\]c* shows a colouring with no monochromatic planar $K_4$ subgraphs, showing that $4 \le N^*$. Colourings in higher dimensions are hard to illustrate with images.
\[cubes\] 
Construction of a colouring of $K(C_{10})$
==========================================
In [@aemp], a colouring of $K(C_{10})$ is constructed by colouring all edges the same colour, cycling through the list of edges in some order, and flipping the colour of each edge, whenever flipping the colour of the edge would not create new monochromatic planar $K_4$ subgraphs. The algorithm terminates when there are no monochromatic planar $K_4$ subgraphs left. Such an algorithm took over 30 hours on a 1.4 GHz CPU.
In the article at hand, also, is found a colouring of $K(C_{10})$ using an almost as simple, nondeterministic, algorithm as follows. The probability function is arbitrarily chosen, perhaps almost any would do, and a better one probably exists.
Colour all edges the same colour. While there are monochromatic planar $K_4$ subgraphs remaining, randomly pick an edge $e$. Let $n_B$ be the number of monochromatic planar $K_4$ subgraphs containing $e$ and $n_G$ be the number of planar $K_4$ subgraphs containing $e$ which would become monochromatic if the edge colour was flipped. Let $P = \min(\frac{n_B}{10+100n_G}, 1)$. Flip the edge colour with probability $P$.
The implementation of this algorithm took 2 minutes and 39 seconds to colour $K(C_{0})$ up to $K(C_{10})$, on a single 2.4 GHz core of a Core2 Q6600 processor. It is not known how the processor or memory bandwidth compares to the processor and memory bandwidth used for the previous construction. ($K(C_{2})$ is the first non-trivial graph which actually contains (and is) a planar $K_4$ subgraph.)
Construction of a colouring of $K(C_{11})$
==========================================
In [@aemp] was estimated that an attempt at constructing a colouring of $K(C_{11})$ would take over 1000 hours, and over half a gigabyte of memory, which was too much memory to make the attempt.
Here, the same non-deterministic algorithm was used to construct a colouring of $K(C_{11})$ as was used here for $K(C_{10})$. The implementation of this algorithm took 48 minutes and 6 seconds to colour $K(C_{11})$, on the same processor as before, using about 17MB of memory. Since the algorithm picks edges randomly, and does not need to contain an ordered list of all edges, the algorithm needs less than the previously estimated half a gigabyte of memory.
Since the algorithm terminated, $12 \le N^*$. The colouring of $K(C_{11})$ is available at http://www.nbi.dk/\~barkley/graham/ . When running the same algorithm to colour $K(C_{12})$, it appeared that it would take a very long time.
Estimates of difficulty of colouring $K(C_n)$
=============================================
The $K(C_n)$ graph has $\frac{2^n(2^n-1)}2 = 2^{2n-1}-2^{n-1} \equiv n_E$ edges. When specifying a colouring, there are thus $n_E$ bits of freedom in specifying the colouring.
In [@aemp], the $K(C_n)$ graph is shown to have $2^{n-3}(3^n-2^{n+1}+1) \equiv n_K$ planar $K_4$ subgraphs. There are $2^6 = 64$ ways of colouring a $K_4$ graph, $62$ of which are not monochromatic. That is, $\frac{31}{32}$ of all possible colourings of a $K_4$ graph are monochromatic. The constraint that one particular $K_4$ subgraph be monochromatic thus constrains $-\log_2 \frac{31}{32} \approx 0.0458$ of the bits of freedom. Under the very naïve assumption that each constraint, that each planar $K_4$ subgraph not be monochromatic, is independent of each other constraint, $-n_K\log_2 \frac{31}{32}$ bits of freedom would be constrained. The fraction of constrained bits is then $-\frac{n_K}{n_E}\log_2 \frac{31}{32} \equiv n_F$. (So, for example, if the assumption was correct, and $n_E$ was 10 and $n_K$ was 30%, then 7 bits would be required to specify a particular of $2^7$ solutions.)
$n$ $n_E$ $n_K$ $n_F$
----- ----------- ------------ ----------
2 6 1 0.763%
3 28 12 1.963%
4 120 100 3.817%
5 496 720 6.649%
6 2016 4816 10.942%
7 8128 30912 17.420%
8 32640 193600 27.168%
9 130816 1194240 41.815%
10 523776 7296256 63.805%
11 2096128 44301312 96.805%
12 8386560 267904000 146.317%
13 33550336 1615810560 220.594%
14 134209536 9728413696 332.016%
Under the assumption, the problem is thus overspecified for $n \ge 12$, making a solution for $n \ge 12$ seem unlikely.
The fraction is trivially correct for $n = 2$. Out of the $2^{28} = 268435456$ colourings for $K(C_3)$, $182596118$ of them are without monochromatic planar $K_4$ subgraphs. The correct fraction for $n = 3$ is thus actually $-\frac1{28}\log_2 \frac{182596118}{2^{28}} \approx 1.985\%$, which means there are slightly fewer solutions for $n = 3$ than if the assumption were correct. It is currently impractical to test all possible colourings of $K(C_n)$ and count the solutions, for $n \ge 4$.
Symmetries
----------
A note on notation — $C_n$ is used here to refer to the $n$-element cyclic group. (As opposed to the $C_n$ in $K(C_n)$, where $C_n$ is used to refer to the $n$-dimensional cube.)
It is possible to require that the colouring be symmetric, such that the colouring does not change under some subgroup of the automorphism group of the problem. The automorphism group of the $n$-cube is the signed permutation group $C_2 \wr_n S_n$ (notation[^1]) and the colour flipping symmetry $C_2$. (This risks turning a solvable problem into an unsolvable one.)
Since using colour flipping symmetry or the signedness of the signed permutation group didn’t seem to help, only subgroups of $S_n$, the (unsigned) permutation group will be considered here.
Under such a symmetry, many $K_4$ subgraphs will be congruent to each other, and the constraints imposed by them are equivalent. This means means less constraints.
If two or more edges on the same $K_4$ subgraph are congruent to each other under the symmetry, or can somehow be shown to have the same colour, the constraint is reduced from a constraint on 6 edges, to a constraint on 5 or less edges (which means harder constraints). If a constraint can be reduced in that way to a constraint on 1 edge, the edge must be both colours, which is not possible, and a solution with the given symmetry is not possible.
If there are a constraint on only edge $a$ and edge $b$ (that is, a constraint that edge $a$ and edge $b$ have opposite colour), and another constraint on only edge $b$ and edge $c$, then since both edge $a$ and edge $c$ have the opposite colour as edge $b$, edge $a$ must then have the same colour as edge $c$, so they may be considered equivalent. (And thus the two constraints become equivalent.)
If there is a constraint on a set $A$ of edges and another constraint on a set $B$ of edges, and if $A \subseteq B$, then the constraint on $A$ implies the constraint on $B$, and the constraint on $B$ is redundant.
A fraction $\frac{2^\nu-2}{2^\nu}$ of colourings (respecting the symmetry) will satisfy a constraint on $\nu$ edges. Under the naïve assumption that after applying the symmetry, the remaining constraints are (still) independent, each constraint on $\nu$ edges will constrain $-\log_2\frac{2^\nu-2}{2^\nu}$ of the bits of freedom.
### Some arbitrarily chosen symmetries of $K(C_{9})$
One possible symmetry of a colouring of $K(C_{9})$ is the group $S_9$. There is a colouring of $K(C_{9})$ with $S_9$ symmetry with no monochromatic planar $K_4$ subgraphs, therefore there is a colouring (the same one) with any possible (unsigned) permutation symmetry.
### Some arbitrarily chosen symmetries of $K(C_{10})$
There are no solutions for $K(C_{10})$ with $S_{10}$ (or $A_{10}$) symmetry. There is a solution with $S_5$ symmetry, where $S_5$ is represented as a primitive permutation group on 10 coordinates.
Generators for $S_5$ found with GAP[^2] are\
$(1\ 5\ 7)(2\ 9\ 4)(3\ 8\ 10)$ and $(1\ 8)(2\ 5\ 6\ 3)(4\ 9\ 7\ 10)$. This group has order $5! = 120$.
### Some arbitrarily chosen symmetries of $K(C_{11})$
Three possible symmetries of a colouring of $K(C_{11})$ are the group $M_{11}$ represented as a permutation of 11 coordinates [@m11], a Sylow 3-subgroup (GAP[^3])of the permutation group $S_{11}$ (which is equivalent to $C_3 \wr_3 C_3$) and the projective special linear group $L_2(11)$ (GAP[^4]). For $M_{11}$, an almost-solution which violates just one constraint (which corresponds to multiple planar $K_4$ subgraphs) exists.
Group Generators Order
----------------- --------------------------------------------------------- -----------------
$M_{11}$ $(2\ 10)(4\ 11)(5\ 7)(8\ 9)$ $(1\ 4\ 3\ 8)(2\ 5\ 6\ 9)$ $11!/7! = 7920$
$Syl_3(S_{11})$ $(3\ 9\ 7)(6\ 11\ 10)$ $(3\ 6\ 4)(5\ 9\ 11)(7\ 10\ 8)$ $3^4 = 81$
$L_2(11)$ $(1\ 5)(2\ 4)(3\ 10)(7\ 11)$
$(3\ 11\ 5)(4\ 7\ 9)(6\ 8\ 10)$ $660$
### Some arbitrarily chosen symmetries of $K(C_{12})$
Five possible symmetries of a colouring of $K(C_{12})$ are the group $M_{11}$ represented as a permutation of 12 coordinates [@m11], a Sylow 3-subgroup of the permutation group $S_{12}$ (GAP[^5]) = $(C_3 \wr_3 C_3) \times C_3$, $(D_4)^3$, $AGL_1(5) \times L_3(2)$ (GAP[^6]) and $S_3 \times S_9$. No solution exists for $M_{12}$.
Group Generators Order
-------------------------- ---------------------------------------------- ------------------------
$M_{11}$ $(1\ 6)(2\ 9)(5\ 7)(8\ 10)$
$(1\ 6\ 7\ 4)(2\ 8)(3\ 9)(5\ 11\ 12\ 10)$ $11!/7! = 7920$
$Syl_3(S_{12})$ $(1\ 7\ 11)(3\ 4\ 10)$
$(1\ 10\ 12)(2\ 7\ 3)(4\ 8\ 11)$ $(5\ 6\ 9)$ $3^5 = 243$
$(D_4)^3$ $(1\ 2)$ $(1\ 3)(2\ 4)$ $(5\ 6)$
$(5\ 7)(6\ 8)$ $(9\ 10)$ $(9\ 11)(10\ 12)$ $8^3 = 512$
$AGL_1(5) \times L_3(2)$ $(2\ 3\ 4\ 5)$ $(1\ 2\ 3\ 5\ 4)$
$(6\ 9)(11\ 12)$ $(6\ 8\ 7)(9\ 12\ 10)$ $20 \times 168 = 3360$
### Some arbitrarily chosen symmetries of $K(C_{13})$
One possible symmetry is $L_3(3)$, found with GAP[^7]. Generators are\
$(1\ 10\ 4)(6\ 9\ 7)(8\ 12\ 13)$ and $(1\ 3\ 2)(4\ 9\ 5)(7\ 8\ 12)(10\ 13\ 11)$. This group has order $2^4 \cdot 3^3 \cdot 13 = 5616$. A colouring exists which violates 142 constraints.
### Some arbitrarily chosen symmetries of $K(C_{14})$
One possible symmetry is $S_5 \times S_9$. This group has order $5! \cdot 9! = 43545600$. A colouring exists which violates 83 constraints.
### Revised estimates of difficulty of colouring $K(C_{n})$
The number of constraints with each number of edges is shown, along with the fraction of constrained bits, for the identity (same as before), and for the groups. A solution has been found for groups marked with green.
Group $n$ $n_E$ , 2, 3, 4, 5, 6 $n_F$
------------------ ----- ----------- ------------------------------ ---------------
9 130816 0, 0, 0, 0, 1194240 41.815%
9 111 6, 0, 106, 0, 141 29.620%
$S_{10}$ 10 142 , 6, 0, 106, 0, 211 [$\infty$%]{}
10 5432 12, 64, 3090, 420, 62015 64.681%
10 523776 0, 0, 0, 0, 7296256 63.805%
11 2096128 0, 0, 0, 0, 44301312 96.805%
11 4034 4, 66, 3168, 1340, 66857 94.912%
$M_{11}$ 11 562 12, 38, 901, 516, 4881 84.156%
11 36944 0, 16, 168, 23744, 616272 82.496%
$M_{12}$ 12 429 , 11, 33, 655, 172, 2607 [$\infty$%]{}
12 8386560 0, 0, 0, 0, 267904000 146.317%
$M_{11}$ 12 1969 17, 97, 3104, 1801, 34550 122.165%
12 14138 10, 100, 1950, 15068, 324759 118.159%
12 41588 16, 74, 1985, 35762, 984959 117.519%
12 55440 0, 24, 768, 49912, 1312120 117.073%
$(D_4)^3$ 12 84070 127, 2671, 98585, 0, 1449256 103.020%
12 10168 52, 388, 12350, 4858, 143437 94.556%
$S_3 \times S_9$ 12 2234 24, 80, 2894, 0, 11442 50.976%
$I$ 13 33550336 0, 0, 0, 0, 1615810560 220.594%
$L_3(3)$ 13 9174 34, 196, 11857, 5340, 301940 182.327%
$I$ 14 134209536 0, 0, 0, 0, 9728413696 332.016%
$S_5 \times S_9$ 14 6256 72, 292, 10748, 0, 54524 76.105%
A colouring of $K(C_{11})$ with no monochromatic planar $K_4$ subgraphs can be found in under an hour with no symmetry ($I$), or under a minute[^8] with $Syl_3(S_{11})$ symmetry, or under five seconds with $L_2(11)$ symmetry.
After applying the $S_3 \times S_9$ symmetry, the problem of finding a colouring of $K(C_{12})$ with no monochromatic planar $K_4$ subgraphs appears easier and smaller than the the problem for $K(C_{11})$ with the three mentioned symmetries. However, no solution for $K(C_{12})$ was found for this particular symmetry. The solutions found for $K(C_{12})$ in the next section have $C_3 \wr_4 C_4$ and $C_3 \wr_4 A_4$ symmetry, which according to the previous table were unlikely to have solutions.
Construction of a colouring of $K(C_{12})$
==========================================
A relative probability of flipping edges, $P = \min\left(\frac{n_B}{10+100\max(5n_G-n_B, 0)}, 1\right)$, is used. This arbitrarily chosen probability seems to function better than the previous arbitrarily chosen one. Also, a blacklist of 3 recently flipped edges is used, to avoid flipping back and forth. A random entry in the blacklist is overwritten each flip.
After applying a symmetry, some edges have more constraints than other edges. Let the value $x$ of a constraint $C$ be the maximum of the numbers of constraints affecting one of edges affected by $C$. For a given cutoff $\kappa$, ignore all constraints with $x \leq \kappa$.
The following algorithm is used. Start with all edges the same colour. Let $\kappa$ be the maximum $x$. While all un-ignored constraints are satisfied, reduce $\kappa$. Flip an edge. If flipping 2000000 times without reducing $\kappa$, run 2000000 more flips with $\kappa = 0$ and repeat with $\kappa$ back to the maximum $x$.
The idea is that some constraints are “harder” to satisfy than other constraints, and that if trying to first satisfy the “harder” constraints without the “easier” constraints getting in the way, the “easier” constraints will then be easy to satisfy.
The symmetry used here is $C_3 \wr_4 C_4$. This group has order $3^4 \cdot 4 = 324$.
The implementation of this algorithm took 2 hours, 31 minutes and 39 seconds, or 81658217 edge colour flips, to colour $K(C_{12})$ with no monochromatic planar $K_4$ subgraphs, on the same processor as before. (Other attempts were running at the same time on other processors, which may have reduced memory bandwidth and increased time for this attempt.)
Since the algorithm terminated, $13 \le N^*$. The colouring of $K(C_{12})$ is available at http://www.nbi.dk/\~barkley/graham/ . The existence of a solution with $C_3 \wr_4 C_4$ symmetry implies the existence of, and is, a solution with $I$ symmetry, since $I \leq C_3 \wr_4 C_4$.
The closest to solutions that four later (simultaneous) attempts with the same symmetry, algorithm and implementation got after about 14 hours were almost-solutions that violated 9, 1, 10 and 4 constraints. This suggests that luck was a major factor in the 2 hours, 31 minutes and 39 seconds time of the first attempt with that particular symmetry, algorithm and implementation.
With four simultaneous attempts at colouring with $C_3 \wr_4 A_4$ symmetry, one attempt found a solution after 40 hours, 20 minutes and 14 seconds, after 1961430488 edge colour flips, another found a solution after 37 hours, 4 minutes and 12 seconds, after 1969734275 edge colour flips, the third was down to two violated constraints after four days, and the last was down to one violated constraint after four days.
The existence of a solution with $C_3 \wr_4 A_4$ symmetry implies the existence of, and is, a solution with $I$ symmetry and a solution with $Syl_3(S_{12})$ symmetry, since $I \leq Syl_3(S_{12}) \leq C_3 \wr_4 A_4$.
Conclusion
==========
It is possible to bi-colour $K(C_{12})$ with no monochromatic planar $K_4$ subgraphs, therefore $13 \le N^*$.
An argument has been given why it should not be possible to bi-colour $K(C_{n})$ for $n \gtrapprox 12$ with no monochromatic planar $K_4$ subgraphs. An argument has also been given why it should be possible to bi-colour $K(C_{12})$ with no monochromatic planar $K_4$ subgraphs, with certain symmetries. However, no colouring was found for the symmetries that seem easiest — only for a symmetry which was still estimated to be impossible. It therefore seems that no estimates were accurate.
It is not surprising that no estimates seem accurate. The estimates are based on the naïve assumtion that the constraints are independent, when the constraints are actually far from being independent. A small group of constraints together may be more or less constraining than the the small group constraints would be, had they been independent, causing an error in the estimate. Especially since the problem is highly symmetric, there seems to be no reason why the errors in an estimate would tend to cancel out — the estimates could easily be wildly off.
Maybe a better estimate is possible — the current upper bound for $N^*$ seems a bit big.
[99]{} R. L. Gragam and B. L. Rothschild, Ramsey’s Theorem for $n$-Parameter Sets, Trans. Amer. Math. Soc., 159 (1971), 257-292 Exoo, Geoffrey (2003). “A Euclidean Ramsey Problem”. Discrete Computational Geometry 29: 223–227. http://brauer.maths.qmul.ac.uk/Atlas/v3/spor/M11/
[^1]: $A \wr_n B$ is the wreath product — the direct product $A^n$ combined with $B$ represented as a permutation of the $n$ elements of $A^n$.
[^2]: [PrimitiveGroup(10, 2);]{}
[^3]: [SylowSubgroup(Group((1,2,3,4,5,6,7,8,9,10,11), (1,2)), 3);]{} or\
[SylowSubgroup(SymmetricGroup(IsPermGroup, 11), 3);]{} or\
[WreathProduct(CyclicGroup(IsPermGroup, 3), CyclicGroup(IsPermGroup, 3));]{}
[^4]: [PrimitiveGroup(11, 5);]{}
[^5]: [SylowSubgroup(Group((1,2,3,4,5,6,7,8,9,10,11,12), (1,2)), 3);]{} or\
[SylowSubgroup(SymmetricGroup(IsPermGroup, 12), 3);]{} or\
[DirectProduct(WreathProduct(CyclicGroup(IsPermGroup, 3), CyclicGroup(IsPermGroup, 3)), CyclicGroup(IsPermGroup, 3));]{}
[^6]: [DirectProduct(PrimitiveGroup(5, 3), PrimitiveGroup(7, 5));]{}
[^7]: [PrimitiveGroup(13, 7);]{}
[^8]: Using $P = \min\left(\frac{n_B}{10+100\max(5n_G-n_B, 0)}, 1\right)$ and writing randomly into a blacklist of 3 recently flipped edges, to avoid flipping back and forth.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Dihadron correlations are reported for peripheral and central [$d$+Au]{} collisions at ${\sqrt{s_{_{NN}}}}=200$ GeV from STAR. The ZYAM background-subtracted correlated yields are larger in central than peripheral collisions. The difference is mainly caused by centrality biases to jet-like correlations. Fourier coefficients of the raw dihadron correlations are also reported. It is found that the first harmonic coefficient is approximately inversely proportional to event multiplicity, whereas the second harmonic coefficient is approximately independent of event multiplicity.'
address: 'Physics Department, Purdue University, West Lafayette, Indiana 47907, USA'
author:
- 'Fuqiang Wang (for the STAR Collaboration)'
title: 'Dihadron correlations in [$d$+Au]{} collisions from STAR'
---
[$d$+Au]{}, dihadron correlations, ridge
Introduction
============
Normally, [$d$+Au]{} (and [$pp$]{} and [$pA$]{}) collisions are used as reference for heavy-ion collisions. For example, [$d$+Au]{} data were essential in establishing jet-quenching at RHIC–that the observed high-${p_T}$ suppression [@Adcox:2001jp; @STARsuppression130GeV] is not due to initial-state differences in parton distributions inside proton and nucleus but final-state parton-parton interactions and partonic energy loss [@Adler:2003ii; @STARdAu03]. This, in part, led to the paradigm of strongly interacting quark-gluon plasma [@GyulassyMcLerran].
Surprisingly, a long-range pseudorapidity (${\Delta\eta}$) dihadron correlation at small azimuthal difference (${\Delta\phi}$) was observed, above a uniform background, at high ${p_T}$ in high-multiplicity [$pp$]{} collisions at the LHC [@CMSppRidge]; it was later observed in [$p$+Pb]{} collisions at essentially all ${p_T}$ and multiplicity (except very low multiplicity) [@CMSpPbRidge; @ALICEpPbRidge; @ATLASpPbRidge]. This long-range ${\Delta\eta}$ correlation is called the “ridge,” in analogy to the similar structure observed in heavy-ion collisions; see below. This motivated further studies of those small-system collisions, beyond just for their use as reference for heavy-ion collisions. In fact, prior to the ridge observation in small systems at the LHC, dihadron correlations were extensively studied in [$d$+Au]{} collisions at RHIC [@STARridge09; @Horner; @Nattrass]. No ridge correlations were observable in [$d$+Au]{}. The [$d$+Au]{} dihadron correlations were similar to those in [$pp$]{} collisions, although slight modifications were seen. A difference was observed by using the cumulant variable between [$pp$]{} and [$d$+Au]{} collisions [@STARflow05], qualitatively consistent with a slight difference in dihadron jet-like correlations.
Similar ridge correlations had been observed before in heavy-ion collisions after a subtraction of elliptic anisotropy background [@PRL95; @STARridge09; @PHOBOSridge10]. The heavy-ion ridge correlations were attributed primarily to triangular anisotropy [@AlverV3]. The similarity of the ridge in [$pp$]{} and [$p$+Pb]{} collisions, where only a uniform background is subtracted, suggests that elliptic anisotropy may be responsible for the ridge in these small systems. In fact, hydrodynamic calculations with event-by-event geometry fluctuations can qualitatively and semi-quantitatively describe the observed ridge in [$pp$]{} and [$p$+Pb]{} collisions [@Bozek:2010pb; @Bozek:2011if; @Bozek:2012gr]. Whether the experimentally measured azimuthal anisotropies are of hydrodynamic flow origin remains a quantitative open question.
Hydrodynamic flow is not the only explanation for the [$pp$]{} and [$p$+Pb]{} ridge correlations. They can also be described by the Color Glass Condensate, where two-gluon density is relatively enhanced at small ${\Delta\phi}$ over a wide range of ${\Delta\eta}$ [@Dumitru:2010iy; @Dusling:2012iga; @Dusling:2012wy; @Dusling:2013oia].
Recently, a back-to-back double ridge, resembling an elliptic/quadrupolar shape, was observed by subtracting the per-trigger normalized dihadron correlated yield in peripheral [$p$+Pb]{} collisions from that in central collisions [@ALICEpPbRidge; @ATLASpPbRidge]. If the correlated jet fragments–dominating the away-side dihadron correlations at large $|{\Delta\eta}|$–are equal in number between peripheral and central collisions, then the observed double ridge would be an indication of new physics. Jet correlations are due to hard scattering and are not expected to differ, in leading order, over [$p$+Pb]{} collision centrality, except that the centrality definition, usually by measured multiplicity in the final state, can bias the selection of events with varying magnitudes of jet correlations. For example, events selected by higher multiplicity could contain jets, both near and away side–because they contribute to the overall multiplicity measurement–originating from partons with larger energy or softer fragmentation. Such biases were estimated to be 10-20% [@ALICEpPbRidge] of the observed ridge yield above a uniform pedestal, indicating that the double ridge may indeed be due to new physics other than jets.
PHENIX analyzed [$d$+Au]{} data using the same technique of “central $-$ peripheral” dihadron correlations in the limited ${\Delta\eta}$ acceptance of $|{\Delta\eta}|<0.7$ with the central arm detector [@PHENIXdAuRidge]. They observed a double ridge in the “central $-$ peripheral” dihadron correlations. It is unclear how much centrality biases there are on jet correlations within the PHENIX acceptance.
STAR, with its large acceptance, is suitable to investigate centrality biases to dihadron correlations. STAR has extensively studied [$d$+Au]{} collisions [@STARdAu03; @STARridge09; @Horner]. The recent development of LHC and PHENIX data called for a more detailed study of the STAR data. This contribution reports the status of such a study.
Data Sample
===========
The data presented here were taken during the [$d$+Au]{} run in 2003 [@Levente09]. The coincidence of the signals from the Zero Degree Calorimeters (ZDC) and the Beam-Beam Counters (BBC) selects minimum-bias (MB) events of [$d$+Au]{} collisions, corresponding to ($95\pm3$)% of the total hadronic cross-section[@STARdAu03]. Events used in this analysis are required to have a primary vertex position $|z_{\rm vtx}|<30$ cm from the TPC center. A total of approximately 10 million events were used. TPC(FTPC) tracks are required to have at least 25(5) out of maximum possible 45(10) hits and a distance of closest approach to the primary vertex within 3 cm.
Three quantities were used to define [$d$+Au]{} centrality: charged particle multiplicity within $|\eta|<1$ measured by the TPC, charged particle multiplicity within $-3.8<\eta<-1.8$ measured by the FTPC in the Au-beam direction (FTPC-Au) [@Levente09], and neutral energy measured in the ZDC of the Au-beam direction (ZDC-Au). The correlations between between each pair of the three observables are shown in Fig. \[fig:centrality\]. Positive correlations are observed but the correlations are quite broad. The same percentile centralities defined by different centrality measures correspond to significantly different event samples of [$d$+Au]{} collisions.
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
Data Analysis
=============
Two sets of dihadron correlations are analyzed: TPC-TPC correlations where the trigger and associated particles are both from the TPC within $|\eta|<1$, and TPC-FTPC correlations where the trigger particle is from the TPC and the associated particle is from either the FTPC-Au within $-3.8<\eta<-2.8$ or the FPTC-d within $2.8<\eta<3.8$. The ${p_T}$ ranges of trigger and associated particles reported here are both $1<{p_T}<3$ [GeV/$c$]{}. (For TPC-TPC correlations the trigger and associated particles are, as such, the same set of particles in each event but each pair is counted only once.) The associated particle tracking efficiencies, $\epsilon_{\rm assoc}=85\%$ for TPC tracks and $70\%$ for FTPC tracks (both with relative systematic uncertainty of $\pm5\%$), are corrected. The correlated yields are normalized by the number of trigger particles, $N_{\rm trig}$.
The two-particle acceptance correction is obtained from the mixed-events technique. The mixed events are required to be within 5 cm in $z_{\rm vtx}$, with the same multiplicity (for the TPC and FTPC-Au centrality measures) or within 10 attenuated ADC counts[^1] (for the ZDC-Au centrality measure). The mixed-events acceptance is normalized to 100% at ${\Delta\eta}|_{100\%}=0$ for TPC-TPC correlations and at ${\Delta\eta}|_{100\%}=\pm3.3$ for TPC-FTPC correlations.
Figure \[fig:2d\] shows the two-particle acceptance corrected dihadron correlations in $({\Delta\eta}=\eta_{\rm assoc}-\eta_{\rm trig},{\Delta\phi}=\phi_{\rm assoc}-\phi_{\rm trig})$ between associated and trigger particle pseudo-rapidities and azimuthal angles, respectively, in peripheral and central [$d$+Au]{} collisions. Namely, $$\frac{1}{N_{\rm trig}}\frac{d^2N}{d{\Delta\eta}d{\Delta\phi}}=\frac{1}{N_{\rm trig}}\frac{S({\Delta\eta},{\Delta\phi})/\epsilon_{\rm assoc}}{B({\Delta\eta},{\Delta\phi})/B({\Delta\eta}|_{100\%},{\Delta\phi})}\,,$$ where $S=\frac{1}{N_{\rm trig}}\frac{d^2N^{\rm same}}{d{\Delta\eta}d{\Delta\phi}}$ is the raw pair density from same event and $B=\frac{1}{N_{\rm trig}}\frac{d^2N^{\rm mix}}{d{\Delta\eta}d{\Delta\phi}}$ is the counterpart from mixed event. In the rest of this article, correlation functions projected onto ${\Delta\phi}$ and ${\Delta\eta}$ are studied. Two approaches are taken to analyze the correlation functions. One is to analyze the correlated yields after subtracting a uniform combinatorial background. The background is normalized by the Zero-Yield-At-Minimum (ZYAM) assumption [@ZYAM] in each ${\Delta\eta}$ bin. The other is to decompose the correlation functions into Fourier series and study the Fourier coefficients. No background subtraction is required; the interpretation of the Fourier coefficients, however, requires a physical model and possibly background consideration.
{width="40.00000%"} {width="40.00000%"}
Systematic uncertainties are assessed by varying the ZYAM normalization ${\Delta\phi}$ range from the default of 0.4 to 0.2 and 0.6 radian. In addition a 5% systematic uncertainty from the tracking efficiency is applied on the correlated yield. For the Fourier coefficients, the systematic uncertainties are expected to be small compared to statistical uncertainties, but a thorough study of systematic uncertainties has not been done yet.
Results on correlated yields
============================
Figure \[fig:dphi\] shows the TPC-TPC ${\Delta\phi}$ correlations in three ranges of ${\Delta\eta}$. Both peripheral and central collisions are shown; centrality is determined by the FTPC-Au. It is observed that the correlated yields are larger in central than peripheral [$d$+Au]{} collisions. Difference between central and peripheral data will be shown in Fig. \[fig:dphiDiff\].
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
In order to investigate the source of the differences between central and peripheral collisions, ${\Delta\eta}$ correlations for near side ($|{\Delta\phi}|<0.8$) and away side ($|{\Delta\phi}-\pi|<1$) are shown in Fig. \[fig:deta\]. The near-side correlations exhibit a Gaussian peak and the away-side correlations are approximately uniform. Gaussian+constant fits to the near-side correlations indicate a difference of 20% in the Gaussian area between central and peripheral collisions. The difference between central and peripheral collisions, shown in the right panel of Fig. \[fig:deta\], exhibits a near-side Gaussian peak and an approximate uniform away-side. These resemble the jet-correlation features, suggesting that the “central $-$ peripheral” difference is mainly due to a difference in jet-like correlations. Given that little difference was observed between minimum-bias [$pp$]{} and minimum-bias [$d$+Au]{} data and significant differences appear between central and peripheral [$d$+Au]{} data after event selections by centrality, we conclude that the difference is most likely caused by biases in the centrality determination–although FTPC-Au is used for centrality which is 3 units away from the correlation measurement, away-side jet-correlations can still contribute to the overall multiplicity in FTPC-Au.
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
Results on Fourier coefficients
===============================
Figure \[fig:VnVsDeta\] shows the second harmonic Fourier coefficients of the ${\Delta\phi}$ distributions of the associated particle yields, $V_n={\langle\cos n{\Delta\phi}\rangle}$ $(n=2)$ where the average is taken over all histogram bins of the ${\Delta\phi}$ distribution, as a function of ${\Delta\eta}$ for both peripheral and central collisions. The third harmonic Fourier coefficient ($n=3$) is consistent with zero. The centrality is determined by TPC in the left panel and FTPC-Au in the right panel; thus in the left panel the data points above $|{\Delta\eta}|>2$ (from TPC-FTPC correlations) are more relevant because they are less biased by the centrality measure, and in the right panel those at $|{\Delta\eta}|<2$ (from TPC-TPC correlations) are more relevant. The Fourier coefficients decrease with increasing $|{\Delta\eta}|$. This is consistent with a jet-like contribution to be primarily responsible for the measured $V_2$.
{width="32.00000%"} {width="32.00000%"}
Figure \[fig:VnVsMult\] shows the Fourier coefficients, $V_1$ and $V_2$. ($V_3$ is consistent with zero.) Three ranges of ${\Delta\eta}$ are shown, from left to right, for TPC-FTPC-Au, TPC-TPC with negative ${\Delta\eta}$ (the TPC-TPC positive ${\Delta\eta}$ results are similar), and TPC-FTPC-d correlations. Results with all three centrality determinations are shown, plotted at the corresponding measured mid-rapidity charged particle multiplicity density $dN/d\eta$. The $V_1$ is observed to approximately vary as $(dN/d\eta)^{-1}$, while the $V_2$ is approximately independent of $dN/d\eta$. $V_2$ is finite at all measured ${\Delta\eta}$; it is larger at mid-rapidity than forward/backward rapidities; $V_2$ from TPC-FTPC-d correlation may be even larger than that from TPC-FTPC-Au correlation.
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
In fact, the Fourier coefficients of the “central $-$ peripheral” correlations are no different from those of the peripheral and central collisions. See Fig. \[fig:dphiDiff\] which shows the difference between the raw correlation functions in central and peripheral collisions corresponding to those in Fig. \[fig:dphi\]. This can be expressed in math as: “central $-$ peripheral” $=N_{\rm cent}(1+2V_1^{\rm cent}\cos{\Delta\phi}+2V_2^{\rm cent}\cos2{\Delta\phi})-N_{\rm peri}(1+2V_1^{\rm peri}\cos{\Delta\phi}+2V_2^{\rm peri}\cos2{\Delta\phi})\approx(N_{\rm cent}-N_{\rm peri})(1+2V_2\cos2{\Delta\phi})$ where $N_{\rm cent}V_1^{\rm cent}\approx N_{\rm peri}V_1^{\rm peri}$ ($N_{\rm cent}$ and $N_{\rm peri}$ are the numbers of associated particles in central and peripheral collisions, respectively) and $V_2\approx V_2^{\rm cent}\approx V_2^{\rm peri}$. However, the underlying physics mechanisms for the large Fourier coefficients of the “central $-$ peripheral” correlations are not entirely clear. Whether there are additional sources, except the aforementioned difference in jet-like correlations due to centrality biases, remains an open question. One of the future studies is to better quantify the centrality biases to jet-like correlations, and then investigate any additional physics mechanisms for the “central $-$ peripheral” difference.
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
Summary
=======
Dihadron ${\Delta\phi}$ and ${\Delta\eta}$ correlations are reported for peripheral and central [$d$+Au]{} collisions at ${\sqrt{s_{_{NN}}}}=200$ GeV from STAR. The ZYAM background-subtracted correlated yields are larger in central than peripheral collisions. The “central $-$ peripheral” differences resemble jet-like correlations, Gaussian peaked on the near side and approximately uniform on the away side. The difference is mainly caused by difference in jet-like correlations due to centrality biases. Fourier coefficients of the raw dihadron correlations are also reported. The first harmonic coefficient is found to be approximately inversely proportional to event multiplicity. The second harmonic coefficient is found to decrease with ${\Delta\eta}$, but finite at forward/backward rapidity of $|{\Delta\eta}|\approx3$; it is approximately independent of the event multiplicity.
The large acceptance of STAR allows detailed investigation of dihadron correlations and their centrality biases. The [$d$+Au]{} data seem to be mainly consistent with jet phenomenology. The next step is to quantify “central $-$ peripheral” differences caused by centrality biases, and hopefully isolate possible additional contributions unrelated to jets.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author thanks Dr. Carlos Salgado and the other organizers of IS-2013 for invitation and the stimulating conference. This work was supported by U.S. Department of Energy under Grant No. DE-FG02-88ER40412.
[25]{} natexlab\#1[\#1]{}\[1\][`#1`]{} \[2\][\#2]{} \[1\][\#1]{} \[1\][[](http://dx.doi.org/#1)]{} \[1\][[](pmid:#1)]{} \[2\][\#2]{} , et al. (), () . , et al. (), () . , et al. (), () . , et al. (), () . , , () . , et al. (), () . , et al. (), () . , et al. (), () . , et al. (), () . , et al. (), () . , et al. (), () . , et al. (), () . , et al. (), () . , et al. (), () . , et al. (), () . , , () . . , () . , () . , , () . , , , , , et al., () . , , () . , , () . , , () . , et al. (), () . , et al. (), () . , , , , , et al., () .
[^1]: The ZDC-Au attenuated ADC dynamic range is 0-255 counts.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Nikos Prantzos
date: 'accepted: October 7, 2005'
title: 'The energetics, evolution, and stellar depletion of $^6$Li in the early Galaxy'
---
Introduction
============
The idea that the light and fragile elements Li, Be and B are produced by the interaction of the energetic nuclei of galactic cosmic rays (CRs) with the nuclei of the interstellar medium (ISM) was introduced 35 years ago (Reeves et al. 1970, Meneguzzi et al. 1971, hereafter MAR). In those early works it was shown that, by taking into account the relevant spallation cross-sections and with plausible assumptions about the CR properties (injected and propagated spectra, intensities etc.; see Sect. 2) one can reproduce the abundances of those light elements observed in meteorites and in CRs, i.e after $\sim$10$^{10}$ yr of galactic evolution (see, e.g. Reeves 1994, for a review on LiBeB). However, the earliest evolution of the light element abundances was not considered in those works.
One of the major cosmological developments of the 1980s was the discovery of the Li plateau in low metallicity halo stars (Spite and Spite 1982). The unique behavior of that element, i.e. the constancy of the Li/H ratio with metallicity, strongly suggests a primordial origin. Its major isotope, $^7$Li, is indeed found to be produced in standard primordial nucleosynthesis calculations, at a level close to the one observed in halo stars (e.g. Lambert 2004 and references therein).
The other isotope of Li, , is produced in extreme low levels in Big Bang nucleosynthesis (/H$<$10$^{-14}$, e.g. Serpico et al. 2004). Its only known source at present is non-thermal nucleosynthesis through the interaction of CRs with the ISM. Its abundance is expected to rise continuously during galactic evolution, similar to the one displayed by Be, another light element solely synthesized by CRs. Taking into account the observed abundance of Be in stars of metallicity \[Fe/H\]$\sim$–3 and the respective production cross-sections, one expects that the /H ratio at such low metallicities would be considerably less than 10$^{-12}$.
The recent reports of the detection of in halo stars (Asplund et al. 2004, 2005) give a new twist to the LiBeB saga. The reported $^6$Li/H value at \[Fe/H\]$\sim$–2.7 is much larger than expected if standard galactic CRs are the only source of . This problem was already noticed by Ramaty et al. (2000) after preliminary reports of detection in very low metallicity halo stars. Equally suprising is the report (Asplund et al. 2004, 2005) of a plateau, at the level of $^6$Li/H$\sim$10$^{-11}$ and in the metallicity range –2.7 $<$ \[Fe/H\] $<$ –0.6; such a plateau is reminiscent of the Spite Li plateau and suggests a pre-galactic origin for . It should be stressed, at this point, that the reality of the $^6$Li plateau or even its absolute level, is not well established at present. However, a recent preliminary analysis of [*SUBARU*]{} data (Inoue et al. 2005) corroborates the findings of Asplund et al. (2005) for $^6$Li in halo stars, although the derived abundance could be twice as small.
Those intriguing, albeit not yet fully established, observational results have already prompted a few ideas using a pre-galactic origin of that light isotope:
1\) Primordial, non-standard production during Big Bang Nucleosynthesis (Jedamzik 2004): the decay/annihilation of some massive particle (e.g. neutralino) releases energetic nucleons/photons that produce $^3$He or $^3$H by spallation/photodisintegration of $^4$He, while subsequent fusion reactions between $^4$He and $^3$He or $^3$H create $^6$Li. This scenario may have undesirable consequences on the abundances of other primordial nuclei (e.g. the D/$^3$He ratio), as criticized by Ellis et al (2005), and will not be discussed here.
2\) Pre-galactic by fusion reactions of $^4$He nuclei, accelerated by the energy released during cosmic structure formation (Suzuki and Inoue 2002); in that case, CR energetics are decoupled from the energetics of supernovae (SN), the latter bein at the origin of the failure of the conventional scenario (see Sect. 3.1).
3\) Finally, Rollinde et al (2004) postulated an early pre-galactic burst of CRs with appropriately tuned intensity in order to justify the reported level of early , but without considering potential sources or the corresponding energetics.
All scenarii of production involving energetic particles are constrained by energy requirements: what is the source accelerating the energetic particles (EP) and is the provided energy sufficient to justify the reported abundance of at very low metallicities? This is the main subject of this work. In Sec. 2 we evaluate the energy requirements for production. In Sec. 3 we compare these requirements with the energy potential of various candidate sources of pre-galactic CR and show that, in general, they are hard to meet. Finally, in Sec. 4 we assume a pre-galactic at the reported level and study the evolution of the light element abundances with a detailed model of galactic chemical evolution, coupled to CR propagation and nucleosynthesis. In agreement with previous works, we find that the existence of a plateau is hard to justify, unless a fine-tuned metallicity-dependent depletion mechanism of in stellar envelopes is invoked. All those result together cast some doubt, either on the reality of the reported detection or on our understanding of production and evolution.
Energy requirements for production
===================================
The energy requirements for the production of light nuclei (Li, Be, B) through spallation of CNO nuclei have been thoroughly studied in Ramaty et al. (1997). Here a slightly different formulation of the problem is presented, based on the formalism of CR propagation developed in MAR.
After acceleration cosmic ray nuclei obtain an [*injection spectrum*]{} $Q(E)$ as a function of energy $E$. While propagating through the ISM, they suffer various losses (ionization, nuclear reactions, and escape from the Galaxy, in the framework of the [*leaky box*]{} model). The [*propagated spectrum*]{} $N(E)$ is assumed to reach [*equilibrium*]{} ($\vartheta N/\vartheta t$=0) rapidly:
$$\frac{\vartheta N(E)}{\vartheta t} \ = \ Q(E) \ - \frac{\vartheta}{\vartheta E} [ \ b(E) \ N(E) \ ] \ -
\frac{1}{\tau(E)} N(E) \ = \ 0$$
where $b(E)$ represents ionization losses, and $\tau(E)$ is the effective timescale for losses through nuclear reactions and escape from the Galaxy: $$\frac{1}{\tau} \ = \ \frac{1}{\tau_{NUC}} \ + \frac{1}{\tau_{ESC}}
\ \sim \frac{1}{\tau_{ESC}}.$$
The functions $b(E)$ and $\tau(E)$ are determined from basic physics and from the observed properties of the ISM (density, composition, ionization stage) and of the CR (abundance ratios of primary to secondary and of unstable to stable nuclei). For [*primary*]{} nuclei, like H, He, C, N, O (the abundances of which are little affected by their propagation through the interstellar medium), the solution of Eq. 1 is: $$N(E) \ = \ \frac{1}{b(E)} \ \int_E^{\infty} Q(E') \ {\rm exp}
\left[-\frac{R(E')-R(E)}{\Lambda} \right] \ dE' ,$$ where $$R(E)=\int_0^E \frac{\rho \ v(E')}{b(E')} \ dE'$$ is the [*ionization range*]{}, with $\rho$ the ISM mass density and $v(E)$ the particle velocity, while $\Lambda = \rho v \tau_{ESC}$ is the [*escape length*]{} from the Galaxy.
The resulting equilibrium spectrum can also be expressed in terms of the omnidirectional particle [*flux*]{} $\Phi(E) = N(E) v(E)$. That quantity is then folded with the relevant spallation cross-sections $\sigma(E)$ in order to calculate the yield of the LiBeB nuclei: $$\frac{\vartheta y_k}{\vartheta t} \ = \sum_j y_j^{ISM} \ \sum_i \int_T^{\infty} \Phi_i^{GCR}(E)
\ \sigma_{ij}^k(E) \ P_{ij}^k(E_C) \ dE.$$ In this expression, $y_k$ is the abundance (by number) of the light nucleus $k$ ($k$=1,...,5 for $^6$Li,$^7$Li, $^9$Be, $^{10}$B,$^{11}$B). The indices $i$ and $j$ run over the range 1,...,5 for H, $^4$He, $^{12}$C, $^{14}$N, and $^{16}$O. The cross sections $\sigma_{ij}^k(E)$ represent the probability of producing nucleus $k$ through the interaction of nuclei $i$ and $j$, and they have a threshold $T$. The quantities $ P_{ij}^k(E_C)$ represent the fraction of light nuclei $k$ that are produced at energy $E_C$ and are incorporated in the ISM at time $t$. They are given by $$P_{ij}^k(E_C) \ = \ {\rm exp}\left[-\frac{R_k(E_C)}{\Lambda}\right]$$ where $R_k(E)$ is the ionization range of nucleus $k$. The energy $E_C$ is close to zero when a fast proton or alpha hits a CNO nucleus of the ISM (i.e. the resulting light nucleus is created at rest and $P\sim$1), and $E_C=E$ when fast CNO nuclei are spallated by ISM protons and alphas (i.e. the resulting light nuclei inherit the same energy per nucleon). In the case of the fusion reaction $\alpha + \alpha \longrightarrow$ $^{6,7}$Li ($i=j=$2) the resulting Li nuclei are created with a velocity about half the one of the fast $\alpha$ particles, and $E_C=E/4$ (see Eq. (6) in MAR).
The total power (energy per unit time) in accelerated particles is $$\dot{W} \ = \frac{\vartheta W}{\vartheta t} \
= \ \sum_i A_i \ \int^{\infty}_0 \ E \ Q_i \ dE,$$ where multiplication by the mass number $A_i$ accounts for the fact that energy $E$ is always expressed in units of energy/nucleon. Obviously, by dividing Eq. (7) by Eq. (5) one obtains the energy of accelerated particles of a given composition that is required to produce one nucleus of species $k$. The result essentially depends on two factors: the form of the injection spectrum $Q(E)$ and the composition of that spectrum. The composition of the ISM is also involved, but it is always taken as equal to solar today, while its evolution is constrained well by observations of low-metallicity stars.
The CR equilibrium spectrum is known very poorly at low energies, precisely those that are important for Li production (in view of the relevant production cross sections, see Fig. 1, upper panel). The reason is the poorly understood modulation effects of the solar wind. Instead of using a demodulated spectrum (e.g. Ip and Axford 1985), in most studies of Li production, a theoretical injection spectrum is adopted and propagated in the Galaxy, in order to recover the equilibrium spectrum through Eq. (3). The form of the injection spectrum is motivated by theories of collisionless shock acceleration (e.g. Ellison and Ramaty 1985). Two popular spectra adopted in most studies in the field (Prantzos et al. 1993, Fields et al. 1994, Ramaty et al. 1997, 2000) are
$$Q(E) \ \propto \ \frac{E+E_p}{[E(E+2E_p)]^{1.5}}$$
where $E_p$=938 MeV is the proton rest mass-energy, and
$$Q(E) \ \propto \ \frac{p^{-s}}{\beta} {\rm exp}(-E/E_0)$$
where $\beta=v/c$ is the velocity expressed as a fraction of the light velocity, $p$ the particle momentum per nucleon, the factor $s$ is usually 2$<s<$3 (in the case of strong shocks), and $E_0$ is a cut-off energy. In view of the form of the $\alpha+\alpha$ cross sections, one might think that a much steeper spectrum than those two may favor the energetics of $^6$Li production. However, ionization losses increase as $E^{-1}$ and are so important in the energy range of few tens of MeV/nucleon that steeper spectra lead to much larger energy demands. Even if the energy spectra of energetic particles in the early Galaxy or in the pre-galactic era are poorly known (see e.g. Gabici and Blasi 2003, Inoue et al. 2004), we feel that those adopted here represent the various possibilities reasonably well, at least as far as energetics is concerned (see below).
The adopted spectra appear in Fig. 1 (lower panel). The energetics of $^6$Li production also depends on the adopted CR source composition. Today, that composition is very close to solar, once effects of propagation and various biases are taken into account (e.g. Wiedenbeck et al. 2001). Intuitively, it appears that the CR source composition should always follow the one of the intestellar medium. However, the observed linearity of Be vs Fe in Galactic stars (e.g. Primas et al. 1999) strongly suggests that the CR composition varied little during the Galactic history; this was first suggested by Duncan et al. (1992) and convincingly demonstrated by Ramaty et al. (1997) on the basis of energetics arguments. Note that the composition of CRs in the early Galaxy affects the $^6$Li energetics relatively little, since a large fraction of that isotope is produced by $\alpha+\alpha$ fusion reactions, especially at low metallicities (Steigman and Walker 1992).
The discussion of this section is summarized in Fig. 2, where the energy required to produce a $^6$Li nucleus by EP hitting the ISM is displayed as a function of the evolving composition of the ISM, represented by \[Fe/H\]. It is assumed that the abundances of N and C in the ISM follow exactly the one of Fe, while the abundance of O evolves differently (O/Fe) is 3 times solar for \[Fe/H\]$<$-1 and declines smoothly to its solar value for higher \[Fe/H\]); this assumption is based on the observed evolution of those elements. The momentum spectrum of Eq. (9) is more efficient than the energy spectrum of Eq. (8). Note that increasing the slope $s$ of the momentum spectrum to higher values than the adopted value of $s$=3 does not improve the energetics, because of the resulting high ionization losses[^1]. Thus, the two lowest curves in Fig. 2 are close to the maximum efficiency for $^6$Li production. Pre-galactic CRs were obviously devoid of CNO nuclei, so the energy requirement for $^6$Li production in the case of pure $\alpha+\alpha$ is $\varepsilon_6\sim$16 erg/nucleus. We shall adopt it as a canonical value in the following.\[[*Note:*]{} Increasing the value of the escape length $\Lambda_{ESC}$ from 10 gr cm$^{-2}$ adopted here to 100-200 gr cm$^{-2}$ (i.e. assuming a more efficient confinement of the CRs, which could be the case in the early Galaxy) slightly improves the energetics of $^6$Li production, by less than $\sim$40%; however, this small difference does not affect the conclusions of Sec. 3\].
The tentative plateau value in the data of Asplund et al. (2005) is at ($^6$Li/H)$_{PL} \sim$10$^{-11}$ (where $PL$ stands for “plateau”). Taking the discussion on energetics into account, one sees that $\varepsilon_6 \sim$16 ergs in EP are required in order to produce 1 nucleus of $^6$Li for every 10$^{11}$ nuclei of the ISM; this energetic requirement can be expressed as $$w_6 \ = \ \varepsilon_6 \ \left(\frac{\rm ^6Li}{\rm H}\right)_{PL}
\ \frac{1}{m_P} \ \sim 10^{14} \ {\rm erg \ gr^{-1}}$$ where $m_P$ is the proton mass. This means that $\sim 10^{14}$ ergs of EP are required to pollute each gr of the ISM to the level of ($^6$Li/H)$_{PL} \sim$10$^{-11}$ through $\alpha+\alpha$ reactions. We note that similar values are found in Reeves (2005), through a different evaluation. In the next section we explore the energetic potential of various CR sources and compare it to the requirement for early $^6$Li production of Eq. (10). One should note, however, that if the true pregalactic value is higher than 10$^{-11}$ (see Sect. 4.2), the value of $w_6$ should be revised upwards accordingly.
Candidate sources for particle acceleration in the early Galaxy
===============================================================
Typical core collapse supernovae
--------------------------------
Core collapse supernovae release a “canonical” energy of $E_{SN}\sim$1.5 10$^{51}$ erg, of which $\sim$10-20% may accelerate CR particles, i.e. the efficiency of turning SN shock energy into EP is $f_{SN}\sim$0.1-0.2. The efficiency value results from the total power of Galactic CR, estimated to 2 10$^{41}$ erg s$^{-1}$ (e.g. Longair 1992), and from the statistics of supernovae in the Milky Way (about 3-4 supernovae per century are expected on average, on the basis of observed SN frequencies in Milky Way type galaxies, e.g. Mannucci et al. 2005).
Taking Eq. (10) into account, one sees that a mass of the ISM $M_{ISM} = \ f_{SN} \ E_{SN}/w_6 \sim$ 10$^3$ can be polluted by a single SN to the level of ($^6$Li/H)$_{PL} \sim$10$^{-11}$. However, a “canonical” core-collapse SN, i.e. resulting from the explosion of a not too massive star, also ejects $M_{Fe}\sim$0.07 of Fe, (e.g. Woosley and Weaver 1995) polluting the ISM to the level of
$$%\begin{equation}
{\rm X_{Fe}} \ = \ \frac{M_{Fe}}{M_{ISM}} \ \sim$$ $$5 \ 10^{-5} \frac{M_{Fe}}{0.07 \ M_{\odot}}
\left(\frac{f_{SN}}{0.1}\right)^{-1} \frac{w_6}{10^{14}
{\rm erg/g}} \left(\frac{E_{SN}}{1.5 \ 10^{51} {\rm erg}}\right)^{-1}$$ where ${\rm X_{Fe}}$ is the mass fraction of Fe. This corresponds to a metallicity \[Fe/H\]$\sim$-1.4 (adopting a solar mass fraction of ${\rm X_{Fe,\odot}}$=1.25 10$^{-3}$ for Fe, following Lodders 2003). Thus, “canonical” SN may indeed pollute the ISM to the level of $^6$Li/H$\sim$10$^{-11}$, but only for metallicities as high as \[Fe/H\]$\sim$-1.4. At metallicities \[Fe/H\]$\sim$-2.7 (the lowest metallicity point in the data of Asplund et al. 2005), a simple scaling shows that the expected level of pollution is only $^6$Li/H$\sim$5 10$^{-13}$, i.e. a factor of $\sim$20 below the observations. Note that this a real upper limit to the level of $^6$Li enrichment that can be obtained by a “canonical ”supernova, because it is assumed that the EP accelerated by the supernova produce $^6$Li only inside the matter that is enriched in Fe. In actual reality, those EP diffuse much further and produce $^6$Li in a much larger region, but to a proportionally lower level (since the total number of $^6$Li atoms is determined by the energy of the explosion and is a constant: $\sim M_{ISM} (^6Li/H)_{PL}/ m_P \sim$2 10$^{49}$ $^6$Li atoms).
The situation may be even worse if the level of primordial $^7$Li is as high as suggested by the WMAP data. Indeed, the baryonic density of the universe derived by observations of the cosmic microwave background corresponds to $^7$Li/H$\sim$4 10$^{-10}$, according to standard calculations of primordial nucleosynthesis (e.g. Serpico et al. 2004). That value is 2-3 times higher than the observed plateau of Li/H in the low metallicity stars of the Milky Way, and the discrepancy may be attributed to our presently poor understanding of Li depletion in stellar envelopes (e.g. Lambert 2004 and references therein). If that explanation is correct, an even greater depletion of the more fragile $^6$Li is expected; the true value of $^6$Li/H should then be at least 2-3 times higher than the one measured by Asplund et al. (2005), making the problem of its production by supernovae even worse.
Energetic (and/or low Fe yield) supernovae
------------------------------------------
The arguments of the previous section may not hold if the energetics of SN is decoupled from the Fe yield: if ($E_{SN}/M_{Fe}) >>$ (1.5 10$^{51}$ erg/0.07 ), Eq. (11) shows that a high $^6$Li/H could be obtained for a small \[Fe/H\]. Such a decoupling might be justified on both observational and theoretical grounds; however, as is usually the case, observations leave less room for optimism than does the theory.
Observations of extragalactic core collapse SN suggest a clear, but not exactly linear, correlation between $M_{Fe}$ and $E_{SN}$, as shown in the review by Hamuy( 2003). In Fig. 15 of that review, it can be seen that SN98bw, a Type Ic supernova, may have ejected 0.5 of Fe (i.e. 7 times the canonical value adopted here) for an energy of 2-5 10$^{52}$ erg, which is 15-30 times the canonical energy of core collapse supernovae[^2]. If such events dominated in the early Galaxy, the discrepancy with the energetics of production would be reduced to $\sim$4-9 (the original discrepancy of a factor of 20 is reduced by a factor of 15-30 because of the increased energy, and again augmented by a factor of 7 due to the increased Fe yield). In their analysis, Nomoto et al. (2000) find that the progenitor mass of SN1998bw had $\sim$40 . One might think that such energetic supernovae with large $E_{SN}$/$M_{Fe}$ values would be common among massive stars of the so-called “hypernova” branch (e.g. Nomoto et al. 2005). However, although the energy of such explosions is indeed larger than the canonical value, the $E_{SN}$/$M_{Fe}$ value is not always large. Thus, in the case of SN1999as, again a type Ic supernova, the energy is estimated as 3 10$^{52}$ erg (see e.g. Fig. 1 in Maeda and Nomoto 2003), but the Fe yield is also quite large, around 4 , so that the resulting $E_{SN}$/$M_{Fe}$ ratio is even lower than in the canonical case. Thus observations of energetic supernova in the nearby universe do not support $E_{SN}$/$M_{Fe}$ ratios large enough to explain the $^6$Li/H plateau.
From the theoretical point of view, the relation between $E_{SN}$ and $M_{Fe}$ is impossible to derive at present, due to our very poor understanding of the explosion mechanism of core collapse supernovae (see e.g. Janka et al. 2003 for a review). In current 1D models of nucleosynthesis in supernovae, some assumptions have to be made about the way the shock wave induces explosive nucleosynthesis in the Si layers (see e.g. Limongi and Chieffi 2003); comparison with observations (e.g. SN1987A in the LMC, a 20 star that ejected 0.07 of Fe) may then help to fix some of the parameters of the model, e.g. the so-called ’mass-cut’, the fiducial surface separating expanding material from material falling back to the compact object. Standard 1D calculations of that type usually produce a few 10$^{-2}$-10$^{-1}$ of Fe for a few 10$^{50}$-10$^{51}$ erg of kinetic energy, at least for stars in the mass range 12-35 (e.g. Thielemann et al. 1996, Woosley and Weaver 1996, Limongi and Chieffi 2003).
For larger stellar masses, in the range 30-100 , simulations suggest that collapse to a black hole should be the general outcome, either directly or after the formation of a weak shock (see Heger et al. 2003 for a review of the situation for 1D simultions with no rotation). Even if the shock is weak, the almost total absence of metals in the ejecta of such explosions makes them suitable candidates for the production of early $^6$Li, free of the problem of the associated production of Fe. Further support for that possibility comes from the very short lifetimes of such objects (less than a few million yrs) and their potentially large number in a zero-metallicity stellar generation: indeed, some theoretical arguments suggest that the stellar IMF at zero metallicity was considerably skewed towards massive stars, in the 100 range (Nakamura and Umemura 2002). If such objects are at the origin of the observed early $^6$Li/H, one can easily evaluate their population in the Milky Way halo (with a rather large uncertainty, though).
As discussed in Sec. 3.1, the mass of the ISM that a normal SN may “pollute” to the level of the observed early $^6$Li/H is $M_{ISM}\sim$10$^3$ . We assume here that each of these explosions leaves a black hole with mass $M_B \sim$50-100 . The stellar mass of the Galactic halo is $M_H\sim$2 10$^9$ (Bullock and Johnston 2005 and references therein), and this may well be (within an order of magnitude) the mass of the gas from which the halo was formed (see Fig. 1, right top panel, in Prantzos 2003). The number of black holes is then $N_B = M_H/M_{ISM}\sim$10$^6$, and their total mass $N_B \times M_B \sim$10$^8$ , i.e. $\sim$10% of the stellar halo mass. One immediately sees that only a “top-heavy” IMF can produce such a large fraction of massive stars, since in a normal IMF (i.e. Salpeter or Kroupa et al. 1993), only $\sim$1% of the mass is in 50-100 stars. Unfortunately, the required total mass in black holes is far below the microlensing detection limits of the MACHO or EROS2 experiments: EROS2 places an upper limit of $\sim$20% on the fraction of the Milky Way’s [*dark halo*]{} which could be in the form of such objects (Tisserand, PhD Thesis 2004 and private communication). For a 10$^{12}$ dark halo, this corresponds to $\sim$2 10$^{11}$ . Thus, the idea that early $^6$Li is produced by mildly energetic stellar explosions of a first generation of (non-rotating) massive stars that lead to black holes and produce too few metals does not violate current constraints from microlensing experiments.
At this point, it should be noted that the two most metal-poor stars of the Galaxy (as far as their Fe content is concerned) are at present HE 1326-2326 (a subgiant or main sequence star) and HE 0107-5240 (a red giant). They both exhibit low Fe abundances, \[Fe/H/H\]$\sim$-5.2, but quite large C/Fe and N/Fe ratios (around 10$^4$ times solar) and high ratios of Na/Fe, Mg/Fe and, rather surprisingly, Sr/Fe (around 10-100 times solar), while other abundance ratios X/Fe are compatible with solar values (Frebel et al. 2005). These peculiar abundance patterns may indeed be explained (at least qualitatively) by “faint” SN, assuming that some mixing of the inner Fe peak nuclei with the exterior layers has occurred and that most of the Fe peak elements fall back onto the black hole. This mixing scheme, dictated on purely observational grounds, might correspond to asymmetric explosions of rotating stars (e.g. Nomoto et al. 2005). One might think that such explosions also fulfill the requirements of “large energy and/or low Fe yield” explosions for early $^6$Li production. However, no Li has been found in HE 1327-2326 and only an upper limit of $\sim$0.5 times the Spite plateau value is given in Frebel et al. (2005); in all probability, $^6$Li should also be absent from its atmosphere. Of course, one might argue that $^6$Li was present during the star’s formation but subsequently depleted in its envelope. In any case, there is no observational proof at present that the supernovae responsible for the abundance pattern of the most metal poor Galactic stars were at the origin of the early $^6$Li observed.
Shocks from cosmic structure formation
--------------------------------------
The currently popular scenario of hierarchical structure formation in the Universe suggests that large scale objects, such as galaxies and clusters, are formed from the merging of smaller subsystems, which are moving and virialised in the gravitational potential wells of dark matter haloes. During those mergers, shocks should develop in the gaseous component of the merging subsystems. The kinetic energy of those shocks from structure formation should be, on average, $$E_{SF} \ = \ \frac{1}{2} \ M \ v_{VIR}^2$$ where $M$ is the mass of the baryonic gaseous component and the virial velocity is $$v_{VIR} \ \sim \ \left(\frac{G M_{DH}}{R}\right)^{1/2} \ \sim \ 400
\ \left(\frac{M_{DH}}{10^{13} M_{\odot}}\right)^{1/4} \ {\rm km/s},$$ $G$ being the gravitational constant, $M_{DH}$ the dark halo mass, and $R$ the virial radius. The numerical values in Eq. (12 ) result from cosmological simulations of large-scale structure formation (Christophe Pichon, private communication). In the case of the Milky Way, the present-day dark halo mass is evaluated to $M_{DH}
\sim$10$^{12}$ (Battaglia et al. 2005), leading to a virial velocity $v_{VIR, MW} \sim$220 km/s today.
The energy of those shocks [*per unit mass of gas*]{} is $w_{SF}$=1/2 $v_{VIR}^2$, and for the case of the Milky Way one gets $w_{SF,MW} \sim$2 10$^{14}$ erg gr$^{-1}$. In order to satisfy the energetics requirement for $^6$Li production of Eq. (10), the kinetic energy of shocks from the formation of the Milky Way should be converted to energetic particles with an efficiency of 50%, which is an extremely high value. Suzuki and Inoue (2002), who first suggested the idea that shocks from cosmological structure formation may be at the origin of early $^6$Li production, assumed a higher value for $M_{DH}$ (3 10$^{12}$ ) and found a reasonable acceleration efficiency of 15%. However, this estimate is overly optimistic, because it is based on the assumption that [*the dark halo is fully formed before the first stars appear polluted by* ]{} $^6$Li. Taking the age of the low metallicity halo stars of the Milky Way into account (larger than 11 Gyr, corresponding to a redshift $z>$3), such a hypothesis is extremely improbable. Indeed, a simple application of the Press-Shechter formalism (e.g. in the Appendix of van den Bosch 2002) shows that a dark halo of 10$^{12}$ is only assembled at redshift $\sim$1.1-1.3. In the earliest phases of our Galaxy, only much smaller dark haloes (less than 10$^{10}$ ) may have existed; the corresponding virial velocity being lower than 200 km/s, the resulting kinetic energy was too low to fulfill the energy requirement for $^6$Li production.
One may, in fact, obtain an upper limit to that energy by noting that the mass of the baryonic halo today is $M_H\sim$2 10$^9$ . Assuming a baryonic/dark matter ratio of 0.1, this corresponds to a dark matter halo of 2 10$^{10}$ [^3] and to a corresponding virial velocity of $\sim$100 km/s (from Eq. 13). This leads to a total kinetic energy per unit mass of baryons of $\sim$0.5 10$^{14}$ erg gr$^{-1}$ and, adopting a standard conversion efficiency to energetic particles ($f\sim$0.1), one sees that no more than 5% of the observationally required $w_6$ value can be provided by that mechanism. Thus, shocks from early structure formation cannot be at the origin of early $^6$Li.
Note that, in the framework of the hierarchical structure formation paradigm, mergers occurred during a large fraction of the Galaxy’s history (e.g. Helmi et al. 2003). If the corresponding shocks accelerated EP, the proposed scenario cannot be characterized as “pre-galactic” and it would lead to a continuous rise of the $^6$Li/H ratio, not to the observed “plateau”. A metallicity-dependent mechanism of $^6$Li depletion inside stars should then also be introduced to account for the “plateau” (see also Sec. 4.2).
Accretion onto the Galactic black hole
--------------------------------------
Accretion onto black holes may have provided another energy source for particle acceleration in the early galactic (or pre-galactic) era. Both theory and observations suggest that a large fraction of the black hole mass-energy may be extracted in that case, either in the form of a jet or a wind; this fraction may be as large as $\eta\sim$0.1.
The largest black hole in the vicinity of the Milky Way is the one laying in the Galactic center (GC). Its mass is estimated to $M_{BH}\sim$3 10$^6$ (Melia and Falcke 2001). The energy that could be released by accretion onto it is at most $E_{BH}$= $\eta M_{BH} c^2$, where $c$ is the light velocity. Assuming a high value for the energy extraction efficiency ($\eta$=0.1) and a typical value for the efficiency of conversion of that energy to energetic particles ($f$=0.1) one finds that $$E_{BH} \ \sim \ 5 \ 10^{58} \ \frac{\eta}{0.1} \ \frac{f}{0.1} \
\frac{M_{BH}}{3 \ 10^6
{\rm M}_{\odot}} \ {\rm erg}.$$
Assuming that the black hole was already in place before the formation of the Milky Way’s halo, one finds that a mass of $E_{BH}/w_6 \sim$2 10$^{11}$ could have been polluted to the level of the observed early $^6$Li/H. This is $\sim$3 times larger than the total stellar mass of the Milky Way (around 6-7 10$^{10}$ , including the bulge), comparable to the total stellar mass of the Local Group, which is dominated by the Milky Way and Andromeda, and 100 times larger than the stellar mass of the Galactic halo $M_H\sim$2 10$^9$ . Thus, the Galactic center black hole apparently fulfills the energy requirement of Eq. (14) for early $^6$Li production; however, this occurs only under extreme assumptions about 1) the time of its formation and 2) the efficiency of extracting its mass-energy.
The first of those points is, perhaps, not that crucial: observations of luminous quasars at high redshift (e.g. in the Sloan Digital Sky Survey, Fan et al. 2004) imply that massive black holes were already in place in the first billion years of the cosmic evolution. Similar conclusions are drawn in a recent study of the accretion history of supermassive black holes based on observations of the current population and accretion rates of those objects (Hopkins et al. 2005b).
The second assumption is perhaps more difficult to justify, in view of the well-known (and poorly undestood) inefficiency of the Galactic black hole for converting accreted matter into radiation. Indeed, the bolometric luminosity of Sgr A$^*$ is $\sim$10$^5$ times lower than what is expected from the well-determined Bondi accretion rate of $\sim$10$^{-5}$ /yr (e.g. Feng, Quataert and Narayan 2003 and references therein). If the same low efficiency characterised the GC black hole in its earliest life (i.e. if $\eta$=10$^{-5}$ in Eq. 14), and if we assume that the kinetic energy of the matter escaping the black hole (in the form of a jet or wind) is equivalent to the energy radiated by the black hole, then the corresponding energetic particles could enrich just the mass of a large globular cluster ($\sim$10$^6$ ) with $^6$Li. Of course, one may assume that, for some reason, the efficiency of the GC black hole was much higher in its early youth, in which case the energetics of early $^6$Li production could be satisfied. However, even in that case, it is difficult to imagine that cosmic rays accelerated from a single object could produce a uniform abundance of over the whole Galactic volume (a radial gradient is rather expected).
The conclusion of this section is that, although promising in principle, the idea that accretion on the GC black hole is at the origin of early $^6$Li encounters some serious difficulties.
Accretion onto supermassive black holes in the Universe
-------------------------------------------------------
On a larger scale than the one of the Milky Way, accretion onto supermassive black holes may provide an important source of energy for particle acceleration in the Universe. Such objects are now routinely found in the centers of galaxies (see Ferrarese and Ford 2005 for a recent review). One may obtain a useful upper limit to their capacity to produce significant amounts of by considering the following:
The present day cosmic density of those objects (in units of the critical density) is evaluated to $\Omega_{SMBH}$=4 10$^{-6}$, while the corresponding total baryon density is $\Omega_{BAR}$=4.5 10$^{-2}$ (Fukugita and Peebles 2004). As in the previous subsection, assuming that a fraction $\eta\sim$0.1 of the black hole rest mass can be usefully extracted as mechanical/radiative energy, and that a fraction $f\sim$0.1 of it can be used in accelerating particles, one sees that the corresponding energy in such particles per unit baryon mass is: $$% \begin{equation}
w_{SMBH} \ = \frac{\eta \ f \ \Omega_{SMBH} \ c^2}{\Omega_{BAR}} \ =$$ $$10^{15} \ \frac{\eta}{0.1} \ \frac{f}{0.1} \
\frac{\Omega_{SMBH}/\Omega_{BAR}}{10^{-4}} \ {\rm erg \ g^{-1}}$$ i.e. $w_{SMBH} \sim$10 $w_6$. It then appears that supermassive black holes should be able to provide the energy needed to produce early ($^6$Li/H)$_{PL}$, since the available energy per unit baryon mass is about ten times larger than required. However, one should again consider the meaning of those numbers. They imply that those supermassive black holes were fully formed and released the accreted energy [*before the formation of the first stars of the Galactic halo*]{}, i.e. at redshifts $z>3$.
The recent study of Hopkins et al. (2005b), based on observations of the nearby supermassive black hole population, indeed suggests that the bulk of the supermassive black hole mass was accreted early on in a radiatively efficient accretion phase. However, it is not clear exactly when that happened. For instance, studies of the evolution of the quasar population suggest that the maximum in their number density in any luminosity interval was at redshift $z<$2 (see e.g. Fig. 4 in Hopkins et al. 2005a). If this evolution also characterizes the formation history and accretion rate of supermassive black holes, then those sources cannot be at the origin of observed early .
Evolution of
=============
In this section we [*assume*]{} that $^6$Li was already present in the earliest moments of the formation of the Milky Way, at an abundance level at least as high as suggested by the observations of Asplund et al. (2005). We study its subsequent evolution with a detailed model of galactic chemical evolution, including its production by fusion of energetic alpha particles and spallation of CNO nuclei (see Sec. 4.2) in a self-consistent way; our aim is not to fully reassess the whole subject of LiBeB production (which is still poorly understood, despite the large amount of theoretical work devoted to it; see Prantzos 2004 for a short review) but rather to study the implications of a pregalactic $^6$Li component.
Pre-galactic $^6$Li: cosmic or just local ?
-------------------------------------------
At this point, a (quite useful) distinction should be made between the terms primordial, cosmic pre-galactic, and local pre-galactic.
The term [*primordial*]{} implies production of $^6$Li in the early Universe, either during the period of Big Bang nucleosynthesis or shortly after, i.e. through the decay of an unstable (super-)particle e.g. Jedamzik (2004). The resulting abundance of $^6$Li is then characteristic of the [*total baryonic content*]{} of the Universe, i.e. of the baryonic fraction $\Omega_{BAR}$.
The term [*cosmic pre-galactic*]{} implies production of $^6$Li prior to star or galaxy formation [*everywhere*]{} in the Universe; i.e. the resulting abundance of $^6$Li is again characteristic of the [*total baryonic content*]{} of the Universe, so both the intergalactic medium (IGM) and the star-forming galaxies have the same $^6$Li/H ratio. This is assumed in e.g. the scenario of Rollinde et al. (2005).
The term [*local pre-galactic*]{} implies that $^6$Li has only (or mostly) polluted the baryonic gas that [participates in galaxy formation]{} and not at all the baryons of the intergalactic medium (or very little). This picture corresponds to the realistic case where accelerating sources and the energetic particles producing $^6$Li are mostly confined inside the high density gas that decouples from the Hubble flow and forms galaxies. Indeed, except for the case of supermassive black holes, all the energy sources explored in Sec. 3 (including shocks from structure formation) belong to this class. Moreover, the magnetic field is much more intense within the dense gas surrounding such sources than in the rarefied IGM, and it certainly produces some local confinement, albeit to a degree that is very hard to evaluate at present (in view of our poor understanding of the origin and evolution of galactic magnetic fields).
The distinction between [*cosmic*]{} and [*local*]{} pre-galactic is important for reasons related to the energetics of $^6$Li production. The local scenario requires less energy than the cosmic one, since in the former case only a fraction of the baryonic matter (the one participating in early galaxy formation) is involved. For instance, one can assume that only the $\sim$2 10$^9$ of the halo mass was polluted with ($^6$Li/H)$_{PL}$ early on; the $\sim$5 10$^{10}$ of the Milky Way disk were accreted much later from the intergalactic medium and were enriched to much lower levels of $^6$Li (produced from the small fraction of energetic particles that escaped confinement in the halo region).
Evolution of and Be in the Milky Way
--------------------------------------
The evolution of the light isotopes $^{6,7}$Li, $^9$Be, $^{10,11}$B is followed with a detailed model of the Milky Way’s chemical evolution. The model, presented in Goswami and Prantzos (2000), satisfactorily reproduces all the major observational constraints in the solar neighborhod, and in particular, the metallicity distributions of halo and local disk stars. The only difference with that model is that the recent, metallicity dependent, massive star yields of Chieffi an Limongi (2003) are adopted here. They differ from those of Woosley and Weaver (1995), which were adopted in GP2000, in several respects, and in particular in the absence of neutrino-induced nucleosynthesis, so no primary $^{11}$B and $^7$Li are present in the CL2003 yields. However, the yields of C, N, O, and Fe, most important for following the overall metallicity and the evolution of ligth isotopes by spallation, display only small differences between the two sets (see Goswami and Prantzos 2003 for a comparison).
![Evolution of total Li, $^6$Li, and Be in the Milky Way. Observations for Li are: at low metallicities from Melendez and Ramirez (2004, [*small dots*]{}) and from Bonifacio et al (2002, [*large dot*]{}); and at high metallicities from Chen et al. (2001, [*small dots*]{}) and Boesgaard and King (1993, [*large dots*]{}). At high metallicities, only values of Li/H$>$–9.8 are plotted, to avoid confusion with $^6$Li values. Observations for Be ([*open symbols*]{}) are from Garcia-Lopez et al. (1998) and Primas et al. (1999). Observations for $^6$Li ([*with error bars*]{}) are from Asplund et al. (2005) and references therein. The Li abundance corresponding to the baryonic density of the Universe derived by WMAP is indicated as [*dashed horizontal line*]{}. The curves correspond to a simple chemical evolution model with metallicity dependent stellar yields, GCR composition [*assumed primary*]{} (to reproduce the Be observations) and GCR spectra given by Eq. (9) with $s$=3. The contribution of the GCR component of $^6$Li is indicated by a [*dashed curve*]{}. The primordial $^7$Li is assumed to be either low (at the level of the plateau of Charbonnel and Primas 2005, slightly lower than derived by Melendez and Ramirez 2004) or high (at the level of WMAP); only production by GCR is considered for Li7, i.e. no stellar source. Similarly, the pre-galactic $^6$Li is assumed to be either “low” (at the level observed by Asplund et al. 2005), or “high” (by 0.$\sim$0.4 dex, for consistency with assumed high Li value, which implies depletion of in the stellar envelopes by at least that value) ](Prantzos_Obs.ps){width="40.00000%"}
The production of the light isotopes is followed as in Prantzos et al. (1993), i.e. using the formalism presented in Sec. 2 and, in particular, the momentum injection spectrum of Eq. (9). A major difference with Prantzos et al. (1993) is the adopted normalization precedure for the light element abundances: instead of normalizing the model Be abundance 4.5 Gyr ago to its solar value (and scaling all other abundances of light elements accordingly) we assume here that a fraction $f_{SN}$=0.1 of the kinetic energy $E_{KIN}$=1.5 10$^{51}$ erg of each SN is used to accelerate CR particles, i.e. $$\dot{W} \ = \ f_{SN} \ E_{KIN} \ R_{SN}$$ where $\dot{W}$ is given by Eq. (7) and $R_{SN}$ is the SN rate (number of SN per unit of time) given by the model. As shown by Ramaty et al. (1997), this procedure is the only one that guarantees consistency between the energetics of CRs and SN. Another difference with Prantzos et al. (1993) is that the light isotopes are [*assumed*]{} to be produced as primaries by CR, i.e. that the CR composition does not vary with time. This is, indeed, the only way to reproduce the observed linearity between Be and Fe (see Fig. 3), since secondary production fails energetically at low metallicities. Even the total SN energy turned into CR is not sufficient to produce the required number of Be atoms, as convincingly argued by Ramaty et al. (1997). It should be noted though that the reason for such $\sim$constant CR composition has not been satisfactorily explained up to now (see Prantzos 2004 for a short critical assessment).
The results of our calculation concerning the evolution of Be, , and total Li, appear in Fig. 3. We adopt a pre-galactic Li value that is either “low”’, i.e. at the level of the low-metallicity Li plateau reported by Charbonnel and Primas (2005), or “high”, i.e. at the level suggested by WMAP data plus Standard Big Bang Nucleosynthesis. In the latter case, currently observed Li in halo stars has been depleted by about 0.4 dex. Similarly, and for consistency, a “low” and a “high” value are adopted for pregalactic $^6$Li, respectively /H=10$^{-11}$ (at the level of the “plateau” reported by Asplund et al. 2005) and 0.4 dex higher (assuming its depletion has been equal to the one of total Li). Note that the latter value corrsponds to the minimal possible amount of depletion, since this isotope is more fragile than $^7$Li and should be more depleted (see Table 6 in Asplund et al. 2005, based on calculations by Richard et al. 2005). It can be seen that:
- The evolution of Be is satisfactorily reproduced; both its solar value and the slope of Be vs Fe are reproduced with the energy normalization of Eq. (15). We stress, however, that the slope is “naturally” produced only under the assumption of a time invariant CR composition (as suggested first by Duncan et al. 1992), which has no sound theoretical justification at present.
- The CR component of is sufficient to produce the solar value of that isotope (as already found analytically in MAR) but fails to reproduce the lowest metallicity value reported by Asplund et al. (2005) by a factor of $\sim$10 (compared to the factor of 20, which was analytically derived in Sec. 3.1).
- Assuming that the plateau is real and extends to metallicities as high as \[Fe/H\]=–0.6, one sees that the CR component of alone (dotted curve in Fig. 3) crosses that plateau value slightly earlier (around \[Fe/H\]=–1.8). A depletion mechanism [*depending on metallicity*]{} should be introduced to justify a plateau in the range –1.8$<$\[Fe/H\]$<$ –0.6 (see also Rollinde et al. 2005). When the assumed pre-galactic component is also taken into account (either “low” or “high”), the abundance curve leaves the plateau value even earlier, around \[Fe/H\]=–2.4.
![[*Upper panel:*]{} Evolution of Li and , observations vs theoretical estimates. Data points are as in Fig. 3. [*Dashed*]{} curves corresponding to observations are: for , at the plateau level suggested by Asplund et al. (2004, 2005, and references therein), for metallicities lower than \[Fe/H)=–0.6; and for Li, the plateau value of Charbonnel and Primas (2005) for halo stars, plus the upper envelope of the data points for higher metallicities (of course, there is no unique way to draw that envelope, especially because of the lack of data in the region around \[Fe/H\]=–1). [*Solid*]{} curves corresponding to the “true” run of the abundances (as expected from theoretical considerations) are the upper curves of Fig. 3 for both Li and . Note, however, that no late stellar production is considered for Li; consequently, the corresponding curve underestimates the “true” Li abundance in disk stars. [*Lower panel:*]{} Corresponding depletion that Li and have suffered in stellar envelopes. Depletion of at the lowest metallicities is assumed to be equal to depletion of Li (i.e. the difference of $\sim$0.35 dex between values found by WMAP data analysis and by Charbonnel and Primas 2005), which is the minimum possible value for depletion in view of the greater fragility of . ](Prantzos_Obs2.ps){width="40.00000%"}
A mechanism that depletes in the stellar envelope with a metallicity dependent efficiency is, perhaps, not all that hard to find. Indeed, it is well known that convection sets in more easily in stellar envelopes with higher opacity, that is, higher metallicity. More metallic stars have deeper convective envelopes, which could bring to higher temperatures and destroy it more efficiently. However, it is much harder to imagine that such a mechanism preserves the putative plateau value [*exactly*]{}. Also, note that depletion of total Li is mandatory when one assumes a high primordial Li abundance (as suggested by WMAP, see Fig. 3) in order to bring that value to the observed plateau level of Li; here the depletion mechanism is metallicity independent for halo stars, but it also has to maintain a plateau value, without significant dispersion (see Lambert 2004 and Charbonnel and Primas 2005).
Assuming that both Li and have been equally depleted from their pre-galactic values in the atmospheres of the lowest metallicity stars and that the observed plateau values are those given by Charbonnel and Primas (2005) for Li and by Asplund et al. (2005) for , one can quantitatively trace the evolution of the stellar depletion of those elements, which is required to keep their abundances within the observed levels. This is done in Fig. 4. For total Li, the “observed” abundance level is situated at the plateau value for halo stars plus the upper envelope of the data for disk stars (the sparsness of data around \[Fe/H\]=–1 making a precise definition of the transition metallicity difficult). As can be seen in the lower panel of Fig. 4, the depletion of has to be increasing with metallicity in order to preserve the observed plateau, from the original depletion value of $\sim$0.4 dex to almost its double at \[Fe/H\]=–1.2, and to about three times a higher value at \[Fe/H\]=–0.6. On the other hand, depletion of Li has to remain essentially constant in halo stars, but necessarily to decrease with metallicity at some point. This is even truer if it is confirmed that the Li abundances in halo stars do not form a real plateau but increase slightly with metallicity (e.g. Ryan et al. 1999, Boesgaard et al. 2005, Asplund et al. 2005, Charbonnel and Primas 2005).
Stated in a different way, Fig. 4 suggests that standard CRs play an important early role in the case of , but not in the case of the much more abundant $^7$Li; the primordial component dominates the abundance of the latter in halo stars. In order to cancel the effect of the CR contribution and to keep at the level of the observed “plateau”, stellar depletion has to be progressively greater in the case of . In the case of Li, a metallicity independent depletion (or slowly decreasing with metallicity) has to be invoked, to bring agreement the primordial (WMAP) value and observed plateau values into agreement. Whether a realistic stellar environment can indeed produce such a differential (and fine-tuned to preserve the plateau values) depletion, remains to be discovered.
Summary
=======
In this work, we reassess the problem of evolution in the Milky Way, motivated by the recently reported existence of a “plateau”-like behavior of its isotopic abundance in old and metal poor halo stars (Asplund et al. 2005).
At first, we calculate the energy requirements for production through fusion and spallation reactions, during the propagation of energetic particles in the ISM. We show that, even under most favorable conditions (i.e. steep CR spectra), it takes at least 10$^{14}$ erg gr$^{-1}$ to justify the reported value of (/H)$_{PL}\sim$10$^{-11}$ (10$^{14}$ erg in accelerated particles for each gr of the ISM).
We proceed then by examining the energy performances of various candidate acceleration sources that may have operated in the early Galaxy: normal core collapse SN, atypical SN (energetic and/or having a low Fe yield), shocks from cosmic structure formation (an interesting suggestion of Suzuki and Inoue 2002), and the supermassive black hole lying in the Galactic Center. We find that:
- normal SN producing $\sim$0.07 of Fe could satisfy the energetics, but they should also simultaneously enrich the ISM to a level of \[Fe/H\]$\sim$–1.4; they fail to produce the earliest reported value (at \[Fe/H\]$\sim$–2.7) by a factor of $\sim$10.
- energetic SN with low Fe yield could certainly satisfy all the constraints; we note, however, that most energetic SN observed today (sometimes called “hypernovae”) have rather high Fe yields and, despite their high energy, they fail for the same reason as normal SN.
- shocks from structure formation are not powerful enough, since in the early times of Galaxy formation, the masses of the assembling dark haloes were still quite small and the corresponding virial velocities insufficient; in the most optimistic case (towards the end of halo formation), derived energies are lower than required by a large factor (on the order of 20, i.e. they fail by a factor larger than normal SN).
- the Galactic black hole satisfies the energy requirements, assuming that a) it is formed to a large extent even before the metal poor stars of the halo (perhaps not an unreasonable assumption) and b) its efficiency in converting accretion energy into accelerated particles was then much higher than its present-day notorious inefficiency in turning accretion energy into radiation.
- supermassive black holes in galaxies could pollute the total baryonic content of the Universe with ($^6$Li/H)$_{PL}$, provided that they they were mostly formed before the formation of the very low-metallicity halo stars. This requirement is not compatible with recent ideas about the quasar luminosity evolution (Hopkins et al. 2005a), suggesting that the peak in the number density of those objects occurred at redshift $z<$2.
We note that, in the case of early , a useful distinction should be made between [*cosmic*]{} and [*local*]{} pre-galactic values. The former implies pollution of the total baryonic content of the Universe, where the latter only concerns the baryons assembled in galaxies; obviously, the energy requirements are easier to fulfill in the latter case than in the former.
Finally, we study the evolution of the light elements with a full-scale galactic chemical evolution model, which satisfies all the major observational constraints. We assume momentum CR spectra, time-independent CR composition, a pre-galactic Li value either “low” (at the level reported by Charbonnel and Primas 2005) or “high” (WMAP value) and a pre-galactic value again either “low” (at the level reported by Asplund et al. 2005) or “high” (by $\sim$0.4 dex, minimal amount of depletion to be consistent with the assumption of “high” Li).
The model reproduces the observed linearity between Be and Fe abundances “by construction”. The associated production of by CR is found to “break” the reported plateau as early as \[Fe/H\]=–2.4. A fine-tuned and metallicity-dependent mechanism of depletion in stellar envelopes would then be required in order to preserve the plateau value. We quantitatively evaluate the amount of required stellar depletion, and show that in the case of , it should increase with metallicity, while for Li it should roughly be metallicity-independent (or even decreasing with metallicity).
In summary, the present study suggests that 1) the energy requirements for large early production are very constraining and hard to fulfill by the currently suggested sources, and that 2) contrary to the case of Li, the reported presence of a plateau in halo stars is “threatened” by the production of by ordinary CRs (the same that produce the observed Be) and requires some “fine-tuning”. In view of these implications, an unambiguous determination of the presence of in halo stars (absolute abundance values and reality of the plateau) is urgently required. In that respect, we note that Inoue et al. (2005) recently report tentative detection of in halo stars, at a level roughly compatible with the one reported by Asplund et al. (2005).
Asplund, M., et al., 2004, in “ESO-ARCETRI workshop on: Chemical Abundances and Mixing in Stars in the Milky Way”, Eds. L. Pasquini and S. Randich, in press
Asplund, M., Lambert, D., Nissen, P., et al., 2005, astro-ph/0510636
Battaglia, G. Helmi, A., Morrison, H., et al. (2005), astro-ph/0506102
Boesgaard, A. M., King, J., 1993, 106, 2309
Boesgaard, A. M., Novicki, M., Stephens, A., 2005, in IAU Symp. 228 “From Li to U: element tracers of early cosmic evolution”, Eds. V. Hill et al., in press
Bonifacio, P., Pasquini, L., Spite, F., 2002, 390, 91
Bullock, J., Thornston, K., 2005, astro-ph/0506467
Charbonnel, C., Primas, F., 2005, in press, astro-ph/0505247
Chen, Y. Q., Nissen, P. E., Benoni, T., Zhao, G., 2001, 371, 943
Duncan, D., Lambert, D., Lemke, M., 1992, 584, 595
Ellis, J., Olive, K., Vangioni, E., 2005, PhL B 619, 30
Ellison, D., Ramaty, R., 1985, 298, 400
Fan, X., Hennawi, J. F., Richards, G. T, 2004, AJ 128, 515
Feng, Y., Quataert, E., Narayan, R., 2003, 598, 301
Ferrarese, L., Ford, H., 2005, SpScRev 116, 523
Fields, B., Olive, K., Schramm, D., 1994, 435, 185
Frebel, A., Aoki, W., Christlieb, N., et al. 2005, Nature 434, 871
Fukugita, M., Peebles, P., J. E., 2004, 616, 643
Gabici, S., Blasi, P., 2003, , 583, 695
Garcia Lopez, R., Lambert, D., Edvardsson, B., 1998, 500, 241
Goswami, A., Prantzos N., 2000, 359, 191
Hamuy, M., 2003, 582, 905
Heger, A., Fryer, C. L., Woosley, S. E., et al., 2003, 591, 298
Helmi, A., White, S., Springel, V., 2003, MNRAS 339, 834
Hopkins, P., Hernquist, L., Cox, T., 2005a, astro-ph/0508299
Hopkins, P., Narayan, R., Hernquist, L., 2005b, astro-ph/0510369
Inoue, S., Nagashima, M., Suzuki, T., Aoki, W., 2004, J. Kor. Astr. Soc. 447, 2004
Inoue S., et al., 2005, in IAU Symp. 228 “From Li to U: element tracers of early cosmic evolution”, Eds. V. Hill et al., in press
Ip, W., Axford, W., 1985, 149, 7
Iwamoto, K., Mazzali, P. A., Nomoto, K., et al., 1998, Nature 395, 672
Janka, H.-Th., Buras, R., Kifonidis, K., et al., 2003, in From Twilight to Highlight: The Physics of Supernovae, Proceedings of the ESO/MPA/MPE Workshop, p. 39
Jedamzik, K., 2004, PhysRev D 70, 063524
Kroupa, P., Tout, C., Gilmore, G., 1993, MNRAS 262, 545
Lambert, D., 2004, astro-ph/0410418
Limongi, M., Chieffi, A., 2003, 592, 404
Longair, M., 1992, High Energy Astrophysics, Cambridge University Press
Lodders, K., 2003, 591, 1220
Maeda, K., Nomoto, K., 2003, 598, 1163
Mannucci, F., della Valle, M., Panagia, N., et al. 2005, 433, 807
Melendez, J., Ramirez, I., 2004, 615, L33
Melia, F., Falcke H., 2001, ARAA 39, 309
Meneguzzi, M., Audouze, J., Reeves H., 1971, 15, 337
Mercer, D., Austin, S., Brown, J., et al., 2001, PhysRev C, 63f5805M
Nakamura, F., Umemura M., 2002, 569, 549
Nomoto, K., Maeda, K., Nakamura, T., et al., 2000, in Gamma-Ray Bursts, Eds. R. M. Kippen et al., p. 622
Nomoto, K., 2005, in IAU Symp. 228 “From Li to U: element tracers of early cosmic evolution”, Eds. V. Hill et al., in press
Prantzos, N., Cassé, M., Vangioni-Flam, E., 1993, 403, 630
Prantzos, N., 2003, 404, 211
Prantzos, N., 2004, in “ESO-ARCETRI workshop on: Chemical Abundances and Mixing in Stars in the Milky Way”, Eds. L. Pasquini and S. Randich, in press (astro-ph/0411569)
Primas, F., Duncan, D., Peterson, R., Thorburn, J., 1999, 343, 545
Ramaty, R., Scully, S., Lingenfelter, R., Kozlovsky, B., 2000, 534, 747
Ramaty, R., Kozlovsky, B., Lingenfelter, R., Reeves, H., 1997, 488, 730
Reeves, H., 1994, RevModPhys, 66, 193
Reeves, H., 2005, astro-ph/0509380
Reeves, H., Fowler, W., Hoyle, F., 1970, Nature 226, 727
Rollinde, E., Vangioni, E., Olive, K., 2005, ApJ 627, 666
Richard, O, Michaud, G, Richer, J., 2005, ApJ 619, 538
Ryan, S., Norris, J., Beers, T., 1999, ApJ 523, 654
Serpico, P. D., Esposito, S., Iocco, F., et al., 2004, JCAP, 04120, 010
Spite, F., Spite, M., 1982, 115, 357
Steigman, G., Walker, T., 1992, 385, L13
Suzuki, T. K., Inoue, S., 2002, 573, 128
Thielemann, K.-F., Hashimoto, M., Nomoto, K., 1996, 460, 408
Umeda, H., Nomoto, K., 2005, 619, 427
van den Bosch, F., 2002, MNRAS 331, 98
Wiedenbeck, M., Yanasak, N., Cummings, A., et al., 2001, SpSciRev 99, 15
Woosley, S., Weaver, T., 1995, 101, 181
[^1]: See e.g. Fig. 8c in Ramaty et al. 1997, displaying clearly the effects of slope $s$ and cut-off energy $E_0$ on the energetics of Be; increasing $s$ (steeper spectra) increases dramatically the energetic requirements. A similar, albeit not so dramatic effect (in view of the large low energy $\alpha + \alpha$ cross sections) holds for .
[^2]: The large energy value of SN1998bw is derived by Iwamoto et al. (1998) under the assumption of spherical symmetry; however, it could be as low as only 3 10$^{51}$ erg, if that assumption is dropped and an ellipsoidal geometry for the ejecta is assumed, as suggested in Hoeflich et al. (1998)
[^3]: A recent work (Bullock and Johnston 2005), that studies the formation of Milky Way’s halo in a cosmological context confirms that the halo was indeed formed by the merging and tidal disruption of dwarf galaxies, mostly embedded in dark haloes of $\sim$10$^{10}$ .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cross-lingual representations of words enable us to reason about word meaning in multilingual contexts and are a key facilitator of cross-lingual transfer when developing natural language processing models for low-resource languages. In this survey, we provide a comprehensive typology of cross-lingual word embedding models. We compare their data requirements and objective functions. The recurring theme of the survey is that many of the models presented in the literature optimize for the same objectives, and that seemingly different models are often equivalent, [*modulo*]{} optimization strategies, hyper-parameters, and such. We also discuss the different ways cross-lingual word embeddings are evaluated, as well as future challenges and research horizons.'
author:
- |
Sebastian Ruder$^*$ [email protected]\
Insight Research Centre, National University of Ireland, Galway, Ireland\
Aylien Ltd., Dublin, Ireland Ivan Vulić [email protected]\
Language Technology Lab, University of Cambridge, UK Anders Søgaard [email protected]\
University of Copenhagen, Copenhagen, Denmark
bibliography:
- 'xlingual\_survey.bib'
title: 'A Survey of Cross-lingual Word Embedding Models'
---
Conclusion {#sec:conclusion}
==========
This survey has focused on providing an overview of cross-lingual word embedding models. It has introduced standardized notation and a typology that demonstrated the similarity of many of these models. It provided proofs that connect different word-level embedding models and has described ways to evaluate cross-lingual word embeddings as well as how to extend them to the multilingual setting. It finally outlined challenges and future directions.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the anonymous reviewers for their valuable and comprehensive feedback. Sebastian is supported by Irish Research Council Grant Number EBPPG/2014/30 and Science Foundation Ireland Grant Number SFI/12/RC/2289. Ivan’s work is supported by the ERC Consolidator Grant LEXICAL: Lexical Acquisition Across Languages (no 648909).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the possibility to induce an effective gauge field for light confined to a Kagomé lattice of identical optical resonators using an on-site modulation of the resonant frequencies. We find that the Haldane Quantum Hall effect arises simply through a site-dependent phase (but constant amplitude) of the dynamic modulation. Within this scheme, we further demonstrate the existence of topological one-directional edge states immune to back-scattering losses, and discuss the possibilities for a practical implementation, which would enable slow-light devices of unprecedented quality.'
author:
- Momchil Minkov
- Vincenzo Savona
title: Haldane Quantum Hall Effect for Light in a Dynamically Modulated Array of Resonators
---
Topological order has opened a new frontier in the classification of distinctive phases of matter, and is thus a center of attention of theoretical and condensed matter physics [@Hasan2010]. Its study has also reached the field of photonics [@Lu2014], for two main reasons. First, photonic analogues of topological systems are a promising route to bridging theory and experiment. Second, a signature of a topologically non-trivial material is the presence of one-directional edge states providing energy transport immune to disorder. This could prove extremely valuable for slow-light photonic devices, which find a variety of applications [@Krauss2008; @Baba2008], but whose performance is severely limited by back-scattering due to fabrication imperfections [@John1987; @Patterson2009; @Mazoyer2009].
Historically, topological order was first recognized in relation to the Quantum Hall effect. In that area, Haldane had a ground-breaking contribution in demonstrating that the effect can arise even with zero magnetic field averaged over a primitive cell. The research into topological photonics was also started by Haldane in two theoretical studies [@Haldane2008; @Raghu2008], which were quickly followed by an experimental realization of a photonic topological insulator using gyromagnetic media [@Wang2009]. This result was however obtained in the GHz frequency range. Due to the lack of suitable materials, reproducing this scheme in the visible or the near-infrared spectrum – which are the most interesting for applications – is still a major challenge. The milestone of an experimental realization of topological edge states for light in the near-infrared has been reached using coupled microring resonators [@Hafezi2011] or coupled waveguides [@Rechtsman2013] by taking advantage of the symmetry-induced degeneracy of rotating and counter-rotating modes. More specifically, these systems are characterized by a preserved time-reversal symmetry (TRS), which leads to an important limitation of the topological protection. The ground-breaking result (which is now known as the Spin Quantum Hall effect) of Kane and Mele [@Kane2005] that, for electrons, this protection is still present in TRS systems, relies on the anti-unitarity of the time-reversal operator ($T^2 = -1$). For photons, this operator is unitary, and the result no longer holds [@Lu2014], at least not in its full strength. Instead, the protection relies on the symmetry that prevents the mixing of propagating and counter-propagating modes in a waveguide, which in practice may be broken by disorder. This suggests the need for systems where TRS is broken [@Haldane2008; @Raghu2008; @Koch2010; @Umucallar2011; @Umucallar2012; @Fang2012a; @Schmidt2015; @Peano2015]. Recently, the possibility to use a fine-tuned dynamic modulation of a system to engineer a gauge field for photons has been shown both theoretically [@Fang2012a] and experimentally [@Fang2012; @Tzuang2014]. This scheme is employed here to induce a Haldane-like magnetic flux for photons on a lattice of optical resonators.
The seminal work by Haldane [@Haldane1988] considered a honeycomb lattice (see the Supplemental material) with real first-neighbor and complex second-neighbor couplings. In the absence of the latter, the band structure of the lattice has six Dirac points, and no band gap. Haldane showed that the complex second-neighbor hopping terms, which result in zero average magnetic field over the unit cell, but non-zero magnetic flux through a triangle enclosed by second-neighbor hopping, break the TRS and open a topological band gap. Recently, this was successfully observed in a system of cold atoms in a ‘shaken’ optical lattice [@Jotzu2014], which, together with previous research in that field [@Hauke2012; @Struck2012; @Goldman2014], inspired the results presented here. In this work, we show how an analogue of the Haldane model can be achieved in a Kagomé lattice of photonic resonators using a time-periodic modulation of the resonant frequencies, where *only the phase of the modulation varies among different sites, in a spatially periodic manner*. We further show the existence of back-scattering-immune edge states, and discuss the possibilities for a practical implementation of the system.
*Model–* We consider a lattice of optical resonators, in which the resonant frequencies $\omega_i$ are subject to a periodic modulation in time. The linear photonic Hamiltonian, most generally, reads $$H = \sum_i (\omega_i + A_i \cos(\Omega t + \phi_i)) a^{\dagger}_i a_i - \sum_{i j} J_{ij} a^{\dagger}_i a_j,
\label{ham_photon}$$ where $a^\dagger$ is the photon creation operator, $J_{ij}$ are the hopping coefficients, and $A_i$ and $\phi_i$ denote the site-dependent amplitude and phase of the dynamic modulation, which can be achieved for example through electro-optic modulation [@Xu2005; @Kuo2005], optically-induced material non-linearities [@Yuce2013], or optomechanical interaction with phonon modes [@Aspelmeyer2014]. The Hamiltonian is particle-number preserving, thus eq. (\[ham\_photon\]) describes the system with any fixed number of photons (sub-spaces of different photon numbers are decoupled). The equation also applies to classical light, since it is a concise way to write the coupled-mode theory that can be used for an array of optical resonators. In the Supplemental material, we outline the theoretical details of the Floquet theory [@Shirley1965; @Sambe1973; @Eckardt2005] that we employ to solve the time-periodic Hamiltonian of eq. (\[ham\_photon\]). The way the Haldane model arises is best revealed through Floquet perturbation theory. There, to first order in $1/\Omega$, new effective couplings can be derived: $J_{ij} \rightarrow J'_{ij} + iJ''_{ij}$, with $J'_{ij}, J''_{ij}$ real constants (see the Supplemental material). In our system, the imaginary $iJ''_{ij}$ thus introduces the magnetic flux required for the Haldane effect.
The most straightforward way to achieve this effect would be by replicating the system of Ref. [@Jotzu2014] through an appropriate modulation of a honeycomb lattice of resonators. However, as discussed in the Supplemental material, this would require a spatial gradient of the modulation amplitude $A_i$. This breaks the spatial periodicity and makes it impossible to analyze the system in momentum space, which is a significant theoretical disadvantage. In addition, in view of potential experimental realizations, this feature introduces an extra challenge, since the maximum amplitude of the modulation is inevitably limited, which in turn would limit the maximum system size. Fortunately, this can be easily overcome through a modification of the lattice geometry – namely, by considering the Kagomé lattice illustrated in Fig. \[fig1\](a). This lattice has three lattice sites per elementary cell, and the band structure (Fig. \[fig1\](b)) is similar to the one of the honeycomb lattice in that there are six Dirac cones. The main difference comes from the additional flat band. Importantly, in the presence of a flux similar to the one of the Haldane model, topologically non-trivial band gaps can be opened between the first and the second and/or the second and the third bands [@Ohgushi2000; @Guo2009].
The Kagomé lattice studied here has identical resonators of frequency $\omega_0$ on all sites, and first-neighbor couplings only (along the black lines of Fig. \[fig1\](a)), with a hopping coefficient $J$. The dynamic modulation that we assume has the form $\omega_A = \omega_0 + A_0 \cos(\Omega t + \varphi)$, $\omega_B = \omega_0 + A_0 \cos(\Omega t + 2\varphi)$, $\omega_C = \omega_0 + A_0 \cos(\Omega t + 3\varphi)$, where A, B, and C refer to the three sites of the primitive cell. Since the modulation is time-periodic, for time-scales larger than the period $T = 2\pi/\Omega$, we can apply the Floquet theory of quasi-energies [@Shirley1965; @Sambe1973; @Eckardt2005], which is outlined in detail in the Supplemental material. There, we give two possible ways to compute the Floquet spectrum of the system. This spectrum has a Brillouin-zone-like structure in the sense that, if $\omega$ is an eigenstate of the time-dependent Hamiltonian, so are $\omega + m\Omega$, for all integer $m$. The eigenstates are time-periodic functions with period $T$, and can be expanded on the Floquet basis $| \{n_i\}, m\rangle$, where $n_i$ is the occupation number of site $i$. This results in an infinite-dimensional matrix for diagonalization with matrix elements given in the Supplemental material. We consider the sub-space of a single excitation only, i.e. $n_i = 1, n_{j\neq i} = 0$, and truncate the orders of $m$ at $m_{\mathrm{max}} = 10$ (convergence is always checked), which yields a finite matrix that can be diagonalized numerically.
The Floquet two-dimensional band structure of the lattice can then be computed in this way, with time- and space-periodic solutions $$u_n(\mathbf{k}, t) = \sum_{i,m} v_{i, m}(\mathbf{k}, n)e^{-i\mathbf{k}\mathbf{R}_i}e^{i m \Omega t},$$ with $v_{i, m}(\mathbf{k}, n)$ the eigenvectors from the diagonalization, and $\mathbf{R}_i$ the position of site $i$. The Floquet band diagram is shown in Fig. \[fig2\](a)-(b) for $J = 0.1\Omega$, $A_0 = 0.9\Omega$, $\varphi = 2.1$. As discussed and displayed in panel (a), the bands are repeated in frequency space at an interval of $\Omega$. In panel (b), which shows a close-up of the zero-th order bands of panel (a), we see that band gaps are opened due to the dynamic modulation. To quantify their topological properties, we compute the Chern number for all bands by integrating the Berry curvature $\mathcal{F}(\mathbf{k})$ [@Berry1984; @Zak1989] over the Brillouin zone. Numerically, we compute $\mathcal{F}(\mathbf{k})$ on a discrete mesh in $\mathbf{k}$-space using the eigenvectors $v_{m, i}(\mathbf{k}, n)$ [@Resta2000; @Soluyanov2012]. The non-zero Chern numbers (1, -2 and 1 for the three bands, respectively) confirm the non-trivial nature of the band gaps.
The second way to handle eq. (\[ham\_photon\]) (see the Supplemental material) is through a perturbative expansion for an effective time-independent Hamiltonian. As mentioned above, this has the advantage of making the connection between this system and the Haldane model manifest, since the first-order terms in the expansion are imaginary couplings that introduce a flux in the red triangle of Fig. \[fig1\](b). In Fig. \[fig2\](c), we show the bands computed by diagonalizing this effective Hamiltonian, which agree very well with the exact solution of panel (b), and the computed Chern numbers are the same.
*Edge states–* Topological invariants like the Chern number cannot change as long as the band gap remains open. Hence, the width of the band gap is an important parameter, giving an energy scale to the topological protection against disorder (only fluctuations on a larger scale can destroy the topological properties). Thus, in Fig. \[fig3\], we plot maps of the gap width $\Delta_T$ (if two gaps are present, the largest value is taken), versus the parameters $A_0$ and $\varphi$. The data in panels (a) and (b) are computed for $J = 0.1\Omega$, with the perturbation theory Hamiltonian in (a), and the full diagonalization in (b), and show very good agreement. In panels (c) and (d), $J = 0.5\Omega$ was used, and the agreement is no longer present. It is natural that the perturbative expansion works well for small $J/\Omega$ when the Floquet bands of different orders are well-separated (Fig \[fig2\](a)), but has limited reliability as $J$ increases. Importantly, however, the topological effect is present even beyond perturbation theory: a gap of width larger than $0.2\Omega$ is opened for $J = 0.5\Omega$, $A_0 = 1.6\Omega$, $\varphi = 2.1$. Notice that for any value of the parameters in this system, the band gap is inevitably limited to a fraction of $\Omega$ due to the higher-order Floquet bands.
In Fig. \[fig4\], we show the band structures with the largest band gaps for $J = 0.3\Omega$, $J = 0.5\Omega$ and $J = 0.7\Omega$, with parameters $A_0$ and $\varphi$ chosen for the largest $\Delta_T$ (see Fig. 2 of the Supplemental material). Topologically, there is a difference between the bands in Fig. \[fig2\](b) and \[fig4\](a), with Chern numbers 1, -2, and 1, and those of Fig. \[fig4\](b)-(c) with Chern numbers 1, 0, and -1. What is important, however, is that in both cases there are bands with a non-zero topological invariant. The bulk-boundary correspondence principle [@Hasan2010; @Lu2014] then applies, guaranteeing the existence of gapless edge states at an interface between the topological material and a topologically trivial one (e.g. empty space). In terms of practical applications, propagating modes robust to disorder are thus expected to appear in a finite system.
The existence of the topological edge modes is illustrated in Fig. \[fig5\] for a ribbon geometry, with a finite number of sites in one direction, and periodic boundary conditions in the other. The one-dimensional Floquet band structure can again be computed by expanding on the Floquet basis, and is shown in panel (a) and (d) for $J = 0.5\Omega$, $A_0 = 1.6\Omega$, $\varphi = 2.1$. The difference between the two panels comes from the truncation at the edges – compare panels (b) and (e). Regardless of how we truncate, there is a band that closes the band gap of the bulk structure, due to the non-zero topological invariants. Modes belonging to that band are localized close to the boundaries of the ribbon; the important point, however, is that the modes at $k_x$ and $-k_x$ are localized at opposite edges. This is illustrated in panels (c) and (f), where we plot the position dependence of the magnitude of the eigenvectors of the two states indicated by a blue and a red dot in panels (a) and (d), respectively. The amplitude on the x-axis is the quantity $\sum |v_{m, i}(\mathbf{k}, n)|^2$, where the sum is over all $m$, and over all sites at the same position along y. The edge modes are exponentially localized at the boundaries (notice the logarithmic scale on the x-axes of panels (c) and (f)), thus the overlap between the forward and backward-propagating modes decreases exponentially with the width of the ribbon in the y-direction. This is only possible due to the broken TRS, and ensures protection against back-scattering in the presence of disorder.
*Discussion–* Several considerations have to be made for the results presented here to have practical implications. We have not considered the loss rate $\kappa$ of the optical resonators, which is in practice always non-zero. To be able to meaningfully talk about light transport, this must be smaller than the coupling constant $J$. In addition, $\kappa$ must also be smaller than the band gap $\Delta_T$, so that the latter can be resolved. By extension, this also implies $\kappa \ll \Omega$. In state-of-the-art photonic crystal cavities, $\kappa/\omega_0$ of the order of $10^{-6}$ can now be routinely achieved [@Notomi2008; @Lai2014; @Sekoguchi2014] at telecommunication frequencies $\omega_0/2\pi \approx 200\mathrm{THz}$, thus $\kappa/2\pi = 0.5\mathrm{GHz}$ is a reasonable and conservative assumption. The coupling constant $J$ is the easiest parameter to control by varying the distance between resonators. Thus the more important challenge is to have a sufficiently high $\Delta_T$. In fact, independently of $\kappa$, $\Delta_T$ is a general figure of merit for the magnitude of the topological protection that should be maximized.
In Ref. [@Fang2012a], electro-optic modulation was suggested as the practical tool for driving the resonant-frequency oscillation. This offers sufficient control over the phase, and has been shown to be scalable [@Fang2012a; @Lira2012]. The maximum achievable modulation frequency $\Omega/2\pi$ is of the order of several GHz. A band-gap $\Delta_T$ of the order of $1\mathrm{GHz}$ could thus be achieved, which lies just above the limit set by $\kappa$. We note that this challenge holds both for our proposal and for that of Ref. [@Fang2012a]. Very recently [@Schmidt2015], it was suggested to use the coupling of the optical resonators to localized phonon modes to induce the frequency modulation. In this scheme, $\Omega$ is fixed by the phonon resonant frequency, which can be as high as $\Omega/2\pi = 10\mathrm{GHz}$ in two-dimensional optomechanical crystals [@Safavi-Naeini2014]. This is sufficiently large for our scheme, and the required phase control can be easily implemented through the phase of the lasers driving the mechanical oscillations [@Aspelmeyer2014; @Schmidt2015]. We note that, when compared to Refs. [@Fang2012a] and [@Schmidt2015], our proposal has a significant structural advantage, as it involves identical resonators with no intermediate (link) elements. Another recent optomechanical scheme [@Peano2015] investigated the Kagomé lattice of resonators, focusing on creating and probing topological states for sound (i.e. phonons). Within that proposal, it is also possible to create topological states of light, but the size of the band gap is shown to be proportional to the phonon hopping coefficient. This is typically orders of magnitude smaller than the phonon resonant frequency, and thus also smaller than the best optical loss rate $\kappa$ that could possibly be achieved in state-of-the-art photonic devices.
While both of the modulation schemes mentioned above could be employed for an experimental realization of our system, a third option is also worth mentioning. Using the optically-induced Kerr nonlinearity, repeated switching at a $\mathrm{THz}$ rate has been recently demonstrated in a micropillar cavity [@Yuce2013]. The maximum amplitude in such a scheme is limited to only a fraction of $\Omega$, but assuming $\Omega/2\pi = 1\mathrm{THz}$, $A_0 = 0.05 \Omega$ (which can be read out of the sine-like dependence of the cavity resonant frequency measured in Ref. [@Yuce2013]), $J = 0.2\Omega$, and $\varphi = 2.1$, we obtain for our Kagomé lattice a topological band gap of width $0.033\Omega$, i.e. $\Delta_T/2\pi = 33\mathrm{GHz}$. This value is very similar to the magnitude of the disorder-induced fluctuations in the resonant frequencies of nominally identical photonic crystal cavities [@Hagino2009; @Minkov2013], and, furthermore, the latter can in principle be reduced by post-processing techniques [@Intonti2012; @Piggott2014]. The predicted $\Delta_T$ is thus, on one hand, two orders of magnitude larger than the loss rate of state-of-the-art cavities, and on the other high enough to ensure a truly sizable protection against disorder.
*Conclusion–* In conclusion, we have described a straightforward implementation of the Haldane-like Quantum Hall effect for light in a lattice of optical cavities, with an effective gauge field produced through a time-periodic modulation of the resonant frequencies. The site-dependence of the phase of the modulation breaks time-reversal symmetry and opens topologically non-trivial band gaps, which, in a finite geometry, yields propagating, back-scattering-free edge states. These can find applications for high bit-rate storage [@Baba2008], for enhanced non-linear effects e.g. for frequency conversion or generation of non-classical light for quantum information processing [@Corcoran2009; @Monat2010; @Azzini2012; @Takesue2014], and for enhanced radiative coupling between distant quantum dots for on-chip quantum computation [@Hughes2007; @Minkov2013a].
This work was supported by the Swiss National Science Foundation through Project N^^ 200020\_149537. We thank Hugo Flayac for discussions, and Willem Vos for highlighting the relevance of Ref. [@Yuce2013].
[54]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](\doibase 10.1103/RevModPhys.82.3045) [****, ()](\doibase 10.1038/nphoton.2014.248) [****, ()](\doibase 10.1038/nphoton.2008.139) [****, ()](\doibase 10.1038/nphoton.2008.146), [****, ()](\doibase 10.1103/PhysRevLett.58.2486) [****, ()](\doibase
10.1103/PhysRevLett.102.253903), [****, ()](\doibase 10.1103/PhysRevLett.103.063903) [****, ()](\doibase 10.1103/PhysRevLett.100.013904) [****, ()](\doibase 10.1103/PhysRevA.78.033834) [****, ()](\doibase 10.1038/nature08293) [****, ()](\doibase 10.1038/nphys2063) [****, ()](\doibase 10.1038/nature12066) [****, ()](\doibase 10.1103/PhysRevLett.95.226801) [****, ()](\doibase 10.1103/PhysRevA.82.043811), [****, ()](\doibase 10.1103/PhysRevA.84.043804) [****, ()](\doibase 10.1103/PhysRevLett.108.206809) [**** (), 10.1038/nphoton.2012.236](\doibase
10.1038/nphoton.2012.236) [****, ()](\doibase
10.1364/OPTICA.2.000635) [ ****, ()](\doibase 10.1103/PhysRevX.5.031011) [****, ()](\doibase
10.1103/PhysRevLett.108.153901) [****, ()](\doibase
10.1038/nphoton.2014.177) [****, ()](\doibase 10.1103/PhysRevLett.61.2015) [****, ()](\doibase
10.1038/nature13915) [****, ()](\doibase
10.1103/PhysRevLett.109.145301), [****, ()](\doibase
10.1103/PhysRevLett.108.225304), [****, ()](\doibase 10.1103/PhysRevX.4.031027) [****, ()](\doibase 10.1038/nature03569) [****, ()](\doibase 10.1038/nature04204) [****, ()](\doibase
10.1364/OL.38.000374), [****, ()](\doibase 10.1103/RevModPhys.86.1391) [****, ()](\doibase 10.1103/PhysRev.138.B979) [****, ()](\doibase 10.1103/PhysRevA.7.2203) [****, ()](\doibase 10.1103/PhysRevLett.95.260404), [****, ()](\doibase 10.1103/PhysRevB.62.R6065) [****, ()](\doibase 10.1103/PhysRevB.80.113102), [****, ()](\doibase 10.1098/rspa.1984.0023) @noop [****, ()]{} @noop [****, ()]{} **, @noop [Ph.D. thesis]{}, () [****, ()](\doibase 10.1038/nphoton.2008.226) [****, ()](\doibase 10.1063/1.4882860) [****, ()](\doibase 10.1364/OE.22.000916) [****, ()](\doibase
10.1103/PhysRevLett.109.033901) [****, ()](\doibase
10.1103/PhysRevLett.112.153603) [****, ()](\doibase
10.1103/PhysRevB.79.085112) @noop [****, ()]{} [**** (), 10.1063/1.3678036](\doibase
10.1063/1.3678036) [****, ()](\doibase 10.1364/OE.22.015017) [****, ()](\doibase
10.1038/nphoton.2009.28) [****, ()](\doibase
10.1364/OE.18.022915) [****, ()](http://www.ncbi.nlm.nih.gov/pubmed/23188274) [****, ()](\doibase 10.1038/srep03913), [****, ()](\doibase 10.1103/PhysRevLett.98.083603) [****, ()](\doibase 10.1103/PhysRevB.87.125306)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recent advances in deep convolutional neural networks (CNNs) have motivated researchers to adapt CNNs to directly model points in 3D point clouds. Modeling local structure has been proven to be important for the success of convolutional architectures, and researchers exploited the modeling of local point sets in the feature extraction hierarchy. However, limited attention has been paid to explicitly model the geometric structure amongst points in a local region. To address this problem, we propose Geo-CNN, which applies a generic convolution-like operation dubbed as GeoConv to each point and its local neighborhood. Local geometric relationships among points are captured when extracting edge features between the center and its neighboring points. We first decompose the edge feature extraction process onto three orthogonal bases, and then aggregate the extracted features based on the angles between the edge vector and the bases. This encourages the network to preserve the geometric structure in Euclidean space throughout the feature extraction hierarchy. GeoConv is a generic and efficient operation that can be easily integrated into 3D point cloud analysis pipelines for multiple applications. We evaluate Geo-CNN on ModelNet40 and KITTI and achieve state-of-the-art performance.'
author:
- |
Shiyi Lan$^{1}$ Ruichi Yu$^{1}$ Gang Yu$^{2}$ Larry S. Davis$^1$\
$^1$University of Maryland, College Park $^2$Megvii Inc (Face++)\
[, [email protected] ]{}
bibliography:
- 'egbib.bib'
title: |
Modeling Local Geometric Structure of\
3D Point Clouds using Geo-CNN
---
Conclusion
==========
We address the problem of modeling local geometric structure amongst points with GeoConv operation and a hierarchical feature extraction framework dubbed Geo-CNN. Inspired by the success of exploiting local structure using CNNs on 2D image analysis task, we propose to extract features from each point and its local neighborhood with a convolutional-like operation. GeoConv explicitly models the geometric structure between two points by decomposing the feature extraction process onto three orthogonal directions, and aggregating the features based on the angles between the edge vector and the bases. The Geo-CNN with GeoConv operation achieves state-of-the-art performance on the challenging ModelNet40 and KITTI datasets.
Acknowledgement {#acknowledgement .unnumbered}
===============
The research was partially supported by the Office of Naval Research under Grant N000141612713: Visual Common Sense Reasoning for Multi-agent Activity Prediction and Recognition.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We reproduced the results of CheXNet with fixed hyperparameters and 50 different random seeds to identify 14 finding in chest radiographs (x-rays). Because CheXNet fine-tunes a pre-trained DenseNet, the random seed affects the ordering of the batches of training data but not the initialized model weights. We found substantial variability in predictions for the same radiograph across model runs (mean $\ln(P_{\text{max}}/P_{\text{min}})$ 2.45, coefficient of variation 0.543). This individual radiograph-level variability was not fully reflected in the variability of AUC on a large test set. Averaging predictions from 10 models reduced variability by nearly 70% (mean coefficient of variation from 0.543 to 0.169, t-test 15.96, p-value < 0.0001). We encourage researchers to be aware of the potential variability of CNNs and ensemble predictions from multiple models to minimize the effect this variability may have on the care of individual patients when these models are deployed clinically.'
author:
- |
John R. Zech[^1]\
Department of Radiology\
Columbia University Irving Medical Center\
New York, NY\
`[email protected]`\
Jessica Zosa Forde$^*$, Michael L. Littman\
Department of Computer Science\
Brown University\
Providence, RI\
`[email protected]`\
`[email protected]`
bibliography:
- 'refs.bib'
title: 'Individual predictions matter: Assessing the effect of data ordering in training fine-tuned CNNs for medical imaging'
---
Introduction
============
While there is interest in using convolutional neural networks (CNNs) to identify findings in medical imaging [@Rajpurkar2017-nh; @Irvin2019-lc; @Johnson2019-as; @Wang2017-py; @Wang2018-ry; @Zech2018-tg], some researchers have questioned the reliability [@Sculley2015-da; @Papernot2016-lz; @Rahimi2017-xb; @Finlayson2019-ed; @Forde2019-fv] and reproducibility [@Vaswani2017-iu; @Melis2018-mv; @Lucic2018-bp; @Riquelme2018-cn; @Henderson2018-yx] of deep learning methods. Often, researchers in medical imaging evaluate a single model’s predictions to measure performance [@gale; @mabal; @Rajpurkar2017-nh; @tienet]. However, the loss surface of a CNN is non-convex, and differences in training such as random seed [@Henderson2018-yx] and optimization method [@Wilson2017-ca; @Choi2019-od] can affect the learned model weights and, consequently, the predictions of the model. Within the machine learning community, there are concerns that evaluation of a single trained model does not provide sufficient measurement of its variability [@Henderson2018-yx; @Forde2019-fv]. To mitigate this variability, some researchers in medical imaging have used cross-validation [@thrun; @Chang2018-ch], which varies the data used for both training and testing, but still uses a single model to make predictions during each cross-validation run. Others have used ensembling [@Sollich1996-wj], combining predictions from 3 [@titano], 5 [@wu; @mura], 10 [@Gulshan2016-mb; @Rajpurkar2018-vn; @Pan2019-ux], and 30 [@Irvin2019-lc] different trained models to optimize classification performance. While ensembles have been proposed as a simple method for estimating the uncertainty of the predictions of a CNN [@Lakshminarayanan2017-en; @Ovadia2019-qr], such variability measurements, as seen in @wu, are not common in medical imaging research.
If models generate different predictions when they are retrained, they may make inconsistent predictions for the same patient. The AUC of a single trained model will not give a direct indication of this inconsistency. AUC itself will likely vary between retrained models, which could complicate efforts to compare CNN performance to other models or human experts using statistical testing [@Silva-Aycaguer2010-fu]. Researchers have attempted to quantify the uncertainty of predictions for individual patients by statistical estimation [@Schulam2019-ui] and direct prediction [@Raghu2019-uk]. @Raghu2019-uk proposed a machine learning method which identifies which retinal fundus photographs [@Gulshan2016-mb] would be likely to have human expert disagreement in diagnosing diabetic retinopathy; such a model could be used to identify high uncertainty cases likely to benefit from a second opinion. @Dusenberry2019-xh examined the variability of RNN-based mortality prediction using the medical records of ICU patients in MIMIC-III [@Johnson2019-as] and recommended the use of Bayesian RNNs [@Fortunato2017-fo] with stochastic embedding over ensembling as a way to estimate the variability of predictions in clinical time-series data.
In this study, we explicitly characterized the variability in individual predicted findings and overall AUC of a CNN that was trained multiple times to predict findings on chest radiographs. Like @Dusenberry2019-xh, we found notable variability of predictions on individual patients with similar aggregate performance metrics. Because many real-world clinical decision support systems rely on single values of predicted probability rather than statistical distributions incorporating uncertainty, we focused our analysis on the use of ensembling for the purposes of robust prediction. We found that, in the case of chest radiographs, simple ensembling can reduce the variability of these probability estimates; ensembles of as few as ten models were found to reduce the variability of predictions by 70% (mean coefficient of variation from 0.543 to 0.169, t-test 15.96, p-value < 0.0001).
Methods
=======
Using an open source implementation [@Zech2018-gf], we replicated the model described in @Rajpurkar2017-nh 50 times, varying the random seed with each fine-tuning [@Lakshminarayanan2017-en]. Per @Rajpurkar2017-nh, a DenseNet-121 [@Huang2017-hi] pre-trained on ImageNet [@Russakovsky2015-ib] was fine-tuned to identify 14 findings in the NIH chest radiography dataset ($n$=112,120) [@Wang2017-py]. The dataset was partitioned into 70% train, 10% tune, and 20% test data (train $n$=78,468, tune $n$=11,219, test $n$=22,433). These 50 models were fine-tuned using SGD with identical hyperparameters on the same train and tune datasets. Because the CNN was consistently initialized with parameters from a DenseNet-121 pre-trained to ImageNet, the only difference in the training procedure across model runs was the order in which data was batched and presented to the model during each epoch of fine-tuning. Each model’s performance was assessed on the full test partition ($n$=22,433). To replicate the test set used for labeling by radiologists [@Rajpurkar2017-nh], a smaller test partition ($n$=792) was created by randomly sampling 100 normal radiographs and 50 positive examples for each finding except for hernia ($n$=42 in the test set).
We calculated various statistical measurements for each finding on each radiograph across our 50 models in the full test set ($n$=22,433). Table \[table-metrics\] describes each metric in detail. In addition to mean, standard deviation, and coefficient of variation, we calculated $\ln(\frac{P_{\text{max}}}{P_{\text{min}}})$, where $P_{\text{max}}$ is the greatest and $P_{\text{min}}$ the least probability predicted by the 50 trained models for a given finding on a given radiograph. This ratio provides a scaled measurement of the variability of the predicted probability of a finding on a radiograph. To contextualize predictions within the population of predictions for that finding, we calculated the percentile rank range of the predictions, $R(P_{\text{max}}) - R(P_{\text{min}})$, where $R(P)$ is the percentile rank of a prediction relative to all predictions for that finding in the test set.
To evaluate the effectiveness of ensembling in reducing variance, we averaged predictions over disjoint groups of 10 models to yield 5 separate averaged predictions for each finding for each radiograph ($n$=5 groups $\times$ 10 models per group = 50 total models). We reported the standard deviation and coefficient of variation across these averaged groups. A paired t-test was used to compare coefficients of variation across the raw ($n$=50) and averaged ($n$=5) predictions.
We examined the variance in overall AUC for the 14 possible targets. For each of the 50 models, AUC was calculated on both the full and limited test sets using the pROC package in R [@Robin2011-lg; @Ihaka1996-kl]. This calculation provided an empirical distribution of the test AUC relative to the order of samples in the training data. For both the full and limited test sets, we calculated the 95% confidence interval of this distribution by subtracting the second smallest AUC from the second largest of the 50 AUCs. For the limited test sets, the average width of the 95% confidence interval was also estimated using DeLong’s method [@DeLong1988-je] and bootstrapping [@Carpenter2000-xn]. DeLong expresses AUC in terms of the Mann-Whitney U statistic [@Mann1947-kd], a non-parametric test statistic that is approximately normally distributed for large sample size, and thus can used to calculate confidence intervals.
Results
=======
{width="60.00000%"}
The variability in predictions for a given radiograph was substantial across models. An example radiograph classification is shown in Figure \[fig:lnp\_scatter\_pna\_only\]. Figure \[fig:lnp\_scatter\_pna\_only\] compares the variability across models ($n$=50) in predicted probability of pneumonia for this radiograph to the variability of predictions for pneumonia in the full test set ($n$=22,433 cases $\times$ 50 models = 1,121,650 total predictions). In this example, the percentile rank range was 95.3% $-$ 48.2% = 47.1%. The variability of each finding for this radiograph relative to all predictions for a given finding is shown in Figure \[fig:lnp\_scatter\].
In the full test set ($n$=22,433), the mean coefficient of variation for an individual radiograph over 50 retrainings was 0.543, and mean $\ln(\frac{P_{\text{max}}}{P_{\text{min}}})$ was 2.45 (Table \[table-pmax-pmin\], Figure \[fig:ln-pmax-pmin\]); for a model with unvarying predictions, $\ln(\frac{P_{\text{max}}}{P_{\text{min}}})$ would equal zero. The radiographs had a mean percentile rank range of 43.0%. In other words, the average difference between the percentile rank of a radiograph’s highest prediction, relative to all predictions for that finding in the test set, and the radiograph’s lowest prediction of that finding, was 43.0%—nearly half the available range.
Averaging model predictions significantly reduced the mean coefficient of variability from 0.543 to 0.169 (t-test 15.96, p-value < 0.0001). The distribution over AUC across models showed a degree of variability in both the full and limited test sets (Figure \[fig:auc\], Table \[table:auc\]). In the limited test set, the empirical variability in predictions did not exceed the average DeLong or bootstrap confidence interval for each model (Table \[table:auc\]). The DeLong and bootstrap 95% confidence intervals for AUC contained the mean AUC across models in 99.7% of cases ($n$=698/700).
{width="100.00000%"}
Discussion
==========
We found substantial variation among the predicted probability of findings when varying the sampling of batches in the training set (mean coefficient of variation across all findings 0.543, mean $\ln(P_{\text{max}}/P_{\text{min}})$ of 2.45; Figure \[fig:ln-pmax-pmin\], Table \[table-pmax-pmin\]). We highlighted a case that demonstrated how predicted probabilities could vary across models (Figure \[fig:lnp\_scatter\]), shifting its estimated risk relative to the test set population based on the random seed used to train the model. The average case had a 43.0% percentile range between its highest and lowest estimated probability of disease across all 50 models.
We found that there was variability across models in AUC for all findings. The overall AUC for each finding in the full test set of over 20,000 cases was much more stable than the substantial variability in predictions for individual radiographs. As explained by @delong-article, “the area under the population ROC curve represents the probability that, when the variable is observed for a randomly selected individual from the abnormal population and a randomly selected individual from the normal population, the resulting values will be in the correct order (e.g., abnormal value higher than the normal value).” In our case, AUC represents the probability that a randomly selected radiograph that is ground-truth positive for pathology will be assigned a higher score by the CNN than a randomly selected radiograph that is ground-truth negative for pathology. Calculating AUC is thus identical to estimating $p$ for a Bernoulli random variable (i.e., a weighted coin-flip) by repeatedly sampling from this distribution and calculating the average $\hat{p}$ over all draws. As our sample size grows larger and larger, our uncertainty interval over the true value of $p$ (and, equivalently, our uncertainty over AUC) grows progressively narrower. AUC can thus be relatively consistent across CNNs that make variable radiograph-level predictions, provided that these variable predictions are similar overall in their ability to classify positive and negative cases.
[.8]{} ![Boxplots of AUC across all 50 models in the full test set ($n$=22,433; top panel) and limited test set, representative of test sets used for expert labeling ($n$=792; bottom panel). Despite substantial individual radiograph-level variability, variability in AUC was low for most findings on the full test set due to the large sample size. The limited test set had expectedly wider distributions over AUC.[]{data-label="fig:auc"}](output_big.pdf "fig:"){width="\linewidth"} \[fig:fig\_a\]
[.8]{} ![Boxplots of AUC across all 50 models in the full test set ($n$=22,433; top panel) and limited test set, representative of test sets used for expert labeling ($n$=792; bottom panel). Despite substantial individual radiograph-level variability, variability in AUC was low for most findings on the full test set due to the large sample size. The limited test set had expectedly wider distributions over AUC.[]{data-label="fig:auc"}](output_small.pdf "fig:"){width="\linewidth"} \[fig:fig\_c\]
The variability in AUC was expectedly wider in the limited test set compared to the full test set. We compared realized variability across models to 95% confidence intervals estimated by two commonly used methods, DeLong and bootstrapping, on the limited test set and found that the realized variability did not exceed these estimated bounds. We note that this comparison is limited and not fully powered; we use sample mean instead of unknown population mean, limiting our ability to detect true differences. Nevertheless, it provides evidence that variability in AUC does not grossly exceed the estimates of common statistical tests, and that these tests can be used to compare the performance of different CNNs, provided researchers are aware of their variability.
{width="100.00000%"}
Stability in AUC across trained models can mask the wide variation in predictions for a single radiograph, and should not reassure researchers that predictions will remain consistent. In our experiments, each DenseNet-121 [@Huang2017-hi] model was initialized with the same pre-trained weights from ImageNet [@Russakovsky2015-ib] and trained with the same train/tune/test data, optimizer, and hyperparameters. From this consistent configuration, we fine-tuned each model on the NIH chest radiograph dataset [@Wang2017-py], varying only the order in which training data was batched and presented to the model. The substantial variability we observed in predictions for individual radiographs might have been even wider had we allowed the model’s initialization parameters, choice of optimizer, or hyperparameters to vary. [@Wilson2017-ca; @Choi2019-od]. @Raghu2019-js suggested that pre-training may not be necessary to achieve competitive performance on medical imaging tasks. Our results call to question whether the absence of pre-training may induce additional variability in predictions.
In the context of healthcare, it is particularly important to remain aware of the variability in individual predictions. If deep learning-based decision support will be deployed in clinical settings, their predictions will alter the diagnoses and treatments given to some patients. Justice, beneficence, and respect for persons are the three ethical principles proposed by the Belmont Report [@belmont], which guides discussion of ethical considerations in medical research. An algorithm that treats identical patients differently challenges the value of justice and potentially leaves the care of patients up to a multi-dimensional coin flip. At the same time, radiologists are also far from perfectly consistent [@Bruno2015-cs]. @Rajpurkar2017-nh observed relatively low inter-rater agreement between the radiologists who contributed the expert labels for pneumonia (0.387 average F1 score comparing each individual radiologist to the majority vote of three other radiologists). Similarly, a study of radiologists at Massachusetts General Hospital found 30% disagreement between colleagues’ interpretations of abdominopelvic CTs and 25% disagreement for the same radiologist viewing the CT at different times [@Abujudeh2010-re; @Bruno2015-cs]. Machine learning algorithms may offer an opportunity to improve the consistency of medical decisions, but only if we are attentive to the inconsistency of which they, too, are capable.
Straightforward workarounds, such as averaging predictions across models [@Sollich1996-wj; @Lakshminarayanan2017-en], can substantially mitigate the effect of this individual-level variability (coefficient of variation reduced from 0.543 to 0.169, p-value < 0.0001). Reducing the variability in individual predictions is also likely to improve performance metrics such as AUC; ensembling of CNN predictions has been successfully demonstrated in medical imaging literature [@titano; @wu; @mura; @Gulshan2016-mb; @Rajpurkar2018-vn; @Irvin2019-lc; @Pan2019-ux], primarily to optimize model performance. No prior work to our knowledge examines how variability in the predictions of CNNs for radiologic imaging may translate to the care of individual patients. We encourage researchers to be vigilant of the variability of deep learning models and to provide some measure of how consistently their final (possibly ensembled) model performs in predicting findings for individual patients to assure readers that end users of the model will not be fooled by randomness.
### Acknowledgments {#acknowledgments .unnumbered}
We would like to thank the Internal Medicine residency program at California Pacific Medical Center (San Francisco, CA) for giving one of the authors (J.Z.) dedicated time to work on this research while he was a preliminary medicine resident.
[^1]: equal contribution
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Research has shown that early life exposures to environmental chemicals, starting as early as conception, can reprogram developmental trajectories to result in altered health status later in life. These principles likely apply to complex mixtures as well as individual chemicals. We thus consider statistical methods for analyzing data on multiple time-varying exposures and a future health outcome when interest simultaneously focuses on both identifying windows of susceptibility to exposure and estimating complex effects of multiple pollutants. First, we apply traditional distributed lag models, distributed lag nonlinear models, and Bayesian kernel machine regression to demonstrate that each can estimate certain features of the exposure-response relationship well but cannot fully characterize associations between an outcome and multiple time-varying exposures. Second, we propose a novel method, a Bayesian kernel machine regression distributed lag model (BKMR-DLM), that simultaneously accounts for nonlinear associations and interactions between time-varying exposures. BKMR-DLM uses a functional weight for each exposure that parameterizes the window of susceptibility corresponding to that exposure within a kernel machine framework that captures non-linear and interaction effects of the multivariate exposure on the outcome. We use the methods to analyze the association between exposure to four ambient pollutants and birth weight in a Boston-area perinatal cohort. We found evidence of a negative association between OC and BWGAz and that nitrate modifies the OC, EC, and sulfate exposure-response functions.'
author:
- 'Ander Wilson[^1]'
- 'Hsiao-Hsien Leon Hsu'
- 'Yueh-Hsiu Mathilda Chiu'
- 'Robert O. Wright'
- 'Rosalind J. Wright'
- 'Brent A. Coull'
bibliography:
- 'references.bib'
title: '**Kernel Machine and Distributed Lag Models for Assessing Windows of Susceptibility to Mixtures of Time-Varying Environmental Exposures in Children’s Health Studies**'
---
Introduction {#sec:intro}
============
Humans are inevitably exposed to a complex mixture of chemicals and other pollutants throughout the life course beginning with conception [@Woodruff2011; @Wright2017]. Epidemiological evidence about the toxicity of environmental chemicals has traditionally come from studies of a single exposure observed during a single time window, such as averaged over a pre-specified time period. The one-chemical-at-a-time and one-exposure-window-at-a-time approaches can result in misleading estimates by failing to distinguish between the effects of multiple highly correlated chemical exposures [@Braun2016] or by incorrectly identifying the time window during which someone is vulnerable to a chemical exposure [@Wilson2017], respectively. In the study of the risks associated with maternal exposures to air pollution and children’s health, there is particular interest in exposure timing. There exist windows of susceptibility in time during which the developing fetus has increased vulnerability to chemical exposures. These windows correspond to specific developmental processes that may be altered by environmental insults and likely vary depending on the mechanisms of the chemical exposures which may impact specific aspects of development. Recent research has identified windows of susceptibility in the association between air pollution exposure and lower birth weight, increased risk of preterm birth, and decreased childhood respiratory health, among other outcomes [@Chang2012; @Warren2013; @Hsu2015; @Bose2017a; @Lee2017].
Partly due to a dearth of available methods, all of these studies estimate the association between time-varying exposure to a single environmental chemical and a health outcome; thus a gap exists with regards to time-varying exposure and chemical mixtures. Assessing the relationship between mixtures of time-varying exposures and a health outcome is complicated by several factors. These include: 1) high correlation between exposure levels at each time point; 2) high temporal correlation within each exposure; 3) potential nonlinear associations; and 4) potential interactions between both simultaneous and sequential exposures. Statistical approaches have been proposed to address each of these challenges individually. For analyzing mixtures observed at a single point in time the proposed methods include Bayesian nonparametric shrinkage and selection priors [@Herring2010], clustering approaches [@Molitor2010; @Molitor2011; @Austin2012; @Zanobetti2014; @Pearce2014], exposure index methods [@Carrico2014], and exposure-response surface methodology [@Bobb2015]. Of particular relevance to the current paper is Bayesian kernel machine regression (BKMR) which estimates a flexible, high-dimensional exposure-response surface [@Bobb2015]. For recent reviews if statistical methods for chemicals mixtures see [@Billionnet2012; @Taylor2016b; @Davalos2017] and [@Hamra2018]. For estimation of windows of susceptibility using time-varying exposures, several distributed lag model (DLM) approaches have been proposed to regress a birth or children’s health outcome on exposures during the gestational period [@Warren2012; @Chang2015; @Warren2016; @Wilson2017a] including extensions to nonlinear associations [@Gasparrini2010; @Gasparrini2011; @Gasparrini2017]. However, no single approach has been proposed to address all of these challenges simultaneously.
To extend DLM methods beyond time-varying measures of a single chemical exposure [@Warren2013] consider an additive DLM with two pollutants and [@Chen2018] proposed a two-pollutant DLM with interaction. However, DLMs have not been extended to more than two pollutants with or without interactions. For a linear association between a continuous outcome and multiple time-varying exposures [@Bello2017] proposed a lagged weighted quantile sums model. The approach regresses the time-varying exposures on the outcome to estimate the association but does not account for nonlinearities or interactions. [@Liu2018a] and [@Liu2018b] developed lagged kernel machine regression to estimate nonlinear associations between exposures observed at multiple points in time and interactions between simultaneous exposures (i.e. those recorded at the same time point). The approach is appropriate for exposures observed at a small number of times that are common to all individuals in the study (e.g. blood biomarkers measured once per trimester) and estimates only interactions between exposures at a single time point and not between exposures at different times. No methods have been proposed that fully integrate methods for mixtures with those for time-varying exposures. None of the proposed approaches estimates the effects of multiple time-varying exposures, including nonlinearities and interactions between both simultaneous and subsequent exposures.
In this paper we make two contributions to the literature on statistical methods for analyzing data on multiple time-varying exposures when interest simultaneously focuses on both identifying windows of susceptibility to exposure and estimating complex effects of multiple pollutants. First, as a baseline of comparison, we adapt existing methods that accommodate either a single time-varying exposure or exposures to chemical mixtures assessed at a single time point to handle time-varying measures of the mixture. Specifically, we consider a straightforward additive extension of traditional DLMs and DLNMs to multi-pollutant models. While previous studies have considered additive DLMs for two exposures, we are unaware of any instances of additive DLMs or DLNMs being applied to mixtures of three or more chemicals. Additive DLMs and DLNMs accommodate data on multiple time-varying exposures, but do not allow for interactions between pollutants. We also evaluate the performance of BKMR applied using exposure averaged over pregnancy for multiple pollutants. This strategy allows for nonlinear associations and interactions but does not account for exposure timing. Our second contribution is to propose a novel Bayesian kernel machine regression distributed lag model (BKMR-DLM) framework. This new framework integrates the DLM methodology for high temporal-resolution time-varying exposures and the BKMR framework for mixtures which have been developed, until now, with independent research threads. The approach uses the concept of time-weighted exposures [@Wilson2017a] to reduce the dimension of the time-varying exposure and to identify windows of susceptibility. The potentially nonlinear and non-additive association between these time-weighted exposures and a health outcome is then modeled using kernel machine regression. To handle the larger parameter space required to account for exposure timing we propose a new MCMC algorithm for the BKMR framework that allows for block updates of the parameters in the kernel model. We evaluate the relative strengths and weaknesses of the different approaches both via simulation and by applying them to analyze data on the association between maternal exposure to four air pollutants during pregnancy and birth weight in a pre-birth cohort study.
Data Description {#s:data}
================
Asthma Coalition on Community, Environment, and Social Stress (ACCESS) cohort [@Wright2008] is a prospective, longitudinal study designed to examine the effects of psychosocial stressors and chemical stressors (e.g., air pollution and other environmental influences) on children’s birth and health outcomes. ACCESS includes 955 mother-child dyads recruited between August 2002 and January 2007 who continued active follow-up after birth in the Boston, MA area. Procedures were approved by the human studies committees at Brigham and Women’s Hospital and Boston Medical Center.
Previous analyses of the ACCESS cohort identified an association between increased maternal exposure to fine particulate matter (PM$_{2.5}$) averaged over pregnancy and decreased birth weight for gestational age $z$-score (BWGAz) particularly among boys born to obese mothers [@Lakshmanan2015]. In this paper we consider the association between weekly levels of exposure to four components of particulate matter–elemental carbon (EC), organic carbon (OC), nitrate, and sulfate–and BWGAz among the same population of boy babies with obese mothers. We include as covariates maternal age at enrollment, an indicator of maternal education at high school level or above, maternal pre-pregnancy body mass index (BMI), indicators of black and Hispanic race/ethnicity, parity, and an indicator of season on birth.
Maternal exposures of EC, OC, nitrate, and sulfate were previously estimated with a hybrid land use regression model that incorporates satellite-derived aerosol optical depth measures and a chemical-transport model GEOS-Chem [@Di2016]. Each mother was assigned an average exposure level for each pollutant for each week of pregnancy based on the predicted value at her address of residence. We limit our analysis to full-term infants (born at $\ge 37$ weeks gestation) and their exposure during the first 37 weeks of pregnancy. A total of 109 children had complete exposure, outcome, and covariate data.
BKMR and DLMs for time-varying measures of an environmental mixture {#sec:meth}
===================================================================
Objectives and notation
-----------------------
Interest focuses on estimating the association between time-varying exposure to $M$ pollutants $X_{1}(t),\dots,X_{M}(t)$ and a scalar outcome $Y$, while controlling for a $p$-vector of baseline covariates $\mathbf{Z}$. We assume these quantities are observed for a sample of size $n$, with subject indexed by $i$, and the exposures are observed over the time domain $\mathcal{T}$, in our case week of gestation. There are two primary objectives: 1) to identify windows of susceptibility during which the time-varying exposures are associated with a future health outcome and 2) estimate the exposure-response relationship while allowing for a nonlinear and non-additive relationship between the multiple exposures and the outcome. In this section we introduce methods that can address only one of these goals–DLM, DLNM and BKMR–and show how these approaches can be readily applied to time-varying measures of a multi-pollutant mixture. We then introduce our proposed methods that can simultaneously address both goals in Section \[s:kmrwithtime\].
Additive DLMs for time-varying measures of an environmental mixture {#sub:dlm}
-------------------------------------------------------------------
The Gaussian discrete-time DLM for a single exposure is $$Y_i=\alpha + \sum_{t=1}^T X_{it}\delta(t)+\mathbf{Z}_i^T\boldsymbol{\gamma}+\epsilon_i,
\label{eq:DLM}$$ where $\delta(t)$ parameterizes the time-varying association between exposure and outcome, $\boldsymbol{\gamma}$ is a $p$-vector of unknown regression coefficients for the confounders, and $\epsilon_i$ are independent $\text{N}(0,\sigma^2)$ random variables. Constrained DLMs impose smoothness on the distrusted lag function. The smoothness constraint can be imposed by modeling $\delta=\delta(t)$ using splines, Bayesian priors, Gaussian processes, or other penalization approaches [@Zanobetti2000; @Peng2009; @Heaton2012; @Chen2017]. For multiple time-varying exposures, an additive DLM is $$Y_i=\alpha + \sum_{m=1}^M\sum_{t=1}^T X_{mit}\delta_{m}(t)+\mathbf{Z}_i^T\boldsymbol{\gamma}+\epsilon_i.
\label{eq:additiveDLM}$$
In prior work, researchers typically identify a window of susceptibility as a period of time during which the pointwise $95\%$ interval for $\delta_{m}(t)$ does not contain 0. We take that approach here. Additionally, the cumulative effect is the expected change in outcome associated with a one unit increase in exposure occurring simultaneously at every exposure time point, or $\text{CE}=\sum_{t=1}^T \delta_{m}(t)$.
In this paper, we use natural splines impose smoothness on $\delta_{m}(t)$ and select the degrees of freedom as the value that minimizes the Akaike Information Criterion (AIC). The model in is then a linear model. An additive DLM is appealing because it can easily be implemented with existing software and is easy to visualize and interpret. However, DLMs do not allow for nonlinear associations or interactions between exposures.
Additive DLMs for mixtures of time-varying exposures {#sub:dlm}
----------------------------------------------------
In their distributed lag nonlinear model (DLNM) framework, [@Gasparrini2010] extended DLMs to accommodate a nonlinear association between outcome and exposure at any given time; that is, they replaced linear association $X_{it}\delta(t)$ in with nonlinear association $f(X_{it},t)$. They model the function $f$ as the outer-product of two basis expansions, which allows the nonlinear association between outcome and exposure at given time point to vary smoothly over time. The additive DLNM for time-varying measures of a mixture is then $$Y_i=\alpha + \sum_{m=1}^M\sum_{t=1}^T f_m(X_{mit},t)+\mathbf{Z}_i^T\boldsymbol{\gamma}+\epsilon_i.
\label{eq:additiveDLNM}$$
In this paper we estimate additive DLNMs using penalized splines and the [dlnm]{} package [@Gasparrini2011]. Using penalized splines for $f_m$ is particularly appealing for mixtures because it facilities simultaneous tuning of each $f_m$ without comparing an unwieldy number of degrees of freedom for the splines bases.
DLNMs allow for nonlinear association between each exposure and the outcome but do not allow for interactions. The identification of a window of susceptibility is more challenging within the DLNM framework because windows may appear at some levels of exposure but not others.
BKMR applied to pregnancy-averaged exposures {#sub:kmr}
--------------------------------------------
BKMR is a popular approach to estimating the association between multiple scalar exposures and a health outcome. Unlike DLM and DLNM, BKMR allows for nonlinear associations and interactions, but does not account for exposure timing. For $M$ scalar exposures $\mathbf{E}_i=(E_{i1},\dots,E_{iM})^T$, a BKMR model takes the form $$Y_i = h(E_{1i},\dots,E_{Mi}) + \mathbf{Z}_i^T\boldsymbol{\gamma} + \epsilon_i.
\label{eq:bkmrsingle}$$ The function $h(\cdot)$ is a potentially nonlinear and non-additive exposure-response function. In the context of maternal exposures during pregnancy, we take $\mathbf{E}_{mi}$ to be the levels of the $m^{th}$ exposure averaged over 37 weeks of pregnancy.
BKMR assumes that the exposure-response function $h:\mathbb{R}^M\rightarrow\mathbb{R}$ resides in the functional space $\mathcal{H}_K$ that is uniquely defined by the positive semidefinite reproducing kernel $K:\mathbb{R}^M\times\mathbb{R}^M\rightarrow\mathbb{R}$. The function $h(\cdot)$ can be represented with a positive-definite kernel function $K(\cdot,\cdot)$ and coefficients $\{\alpha_i\}_{i=1}^n$ as $h(\mathbf{E})=\sum_{i=1}^n K(\mathbf{E},\mathbf{E}_i)\alpha_i$. According to Mercer’s Theorem [@Cristianini2000], the kernel $K(\cdot,\cdot)$ implicitly specifies a basis expansion. For example, the Gaussian kernel $K(\mathbf{E},\mathbf{E}')=\exp\{-\sum_{m=1}^M \rho_m^{-1}(E_m-E_m' )^2\}$ corresponds to the set of Gaussian radial basis functions. The polynomial kernel $K(\mathbf{E},\mathbf{E}')=\left(1+\mathbf{E}^T\mathbf{E}'\right)^d$ corresponds to a polynomial of order up to $d$.
Using the kernel representation of $h(\cdot)$, [@Liu2007] showed that the regression model in is equivalent to the hierarchical model $$\begin{aligned}
\label{eq:hiermodel}
Y_i &\sim&\text{N}(h_i+\mathbf{Z}_i^T\boldsymbol{\gamma}, \sigma^2)\\
\mathbf{h}=\left(h_1,\dots,h_n\right)^T &\sim&\text{N}(\mathbf{0},\sigma^2\tau^2\mathbf{K}),\nonumber\end{aligned}$$ where $\mathbf{K}$ is an $n\times n$ matrix with $i,j$ element $K_{ij}=K(\mathbf{E}_i,\mathbf{E}_j)$. In this work we consider both Gaussian and polynomial kernel functions within BKMR. The Gaussian kernel allows for flexible estimation of $h()$, while the a polynomial kernel potentially yields increased power to detect associations together with a concomitant reduction in flexibility. For the Gaussian kernel, the $i$, $j$ element of $\mathbf{K}$ is $$K_{ij}=\exp\left[-\sum_{m=1}^M \rho_m(E_{im}-E_{jm})^2\right].
\label{eq:kernelscalar}$$ In this paper, for BKMR analyses using exposures averaged over the first 37 weeks of gestation, we fit the Gaussian kernel model as implemented in the R package [bkmr]{} [@Bobb2017; @Bobb2018].
For the BKMR-models specified in the next section, we also considered models using a polynomial kernel with $d=2$ (i.e. a quadratic kernel). The polynomial kernel of order $d$ is $$K_{ij}=\left(1+\sum_{m=1}^M \rho_m E_{im}E_{jm}\right)^d.
\label{eq:kernelscalarquad}$$ Note both the Gaussian and polynomial kernels include feature weights $\rho_m>0$ [@allen2013; @Bobb2015].
BKMR-DLM to account for exposure timing, interactions, and nonlinearity {#s:kmrwithtime}
=======================================================================
BKMR-DLM model specification
----------------------------
The methods presented in Section \[sub:kmr\] are well established for scalar exposures. The proposed BKMR-DLM approach simultaneously identifies windows of susceptibility using pollutant-specific weight functions and estimates the potentially nonlinear and non-additive associations between a health outcome and time-weighted exposures.
Time-weighted exposure are closely related to DLMs and functional regression methods. [@Wilson2017a] proposed estimating the linear association between a single time-varying exposure and health outcome using the model $$Y_i=\alpha + \beta\int_\mathcal{T} X_i(t)w(t)dt+\mathbf{Z}_i^T\boldsymbol{\gamma}+\epsilon_i,
\label{eq:bdlim}$$ where $X_i(t)$ is a functional representation of the exposure and $w(t)$ is a functional weight parameter defined over $\mathcal{T}$. The approach constraints $\int_\mathcal{T} w(t)dt>0$ and $\int_\mathcal{T} \left[w(t)\right]^2dt=1$ to the weight function for identifiability.
The model can be viewed as a linear regression model using the weighted exposure $E_i=\int_\mathcal{T} X_i(t)w(t)dt$ as a scalar covariate: $Y_i=\alpha + \beta E_i +\mathbf{Z}_i^T\boldsymbol{\gamma}+\epsilon_i$. It is equivalent to the functional DLM $Y_i=\alpha + \int_\mathcal{T} X_i(t)\delta(t) dt+\mathbf{Z}_i^T\boldsymbol{\gamma}+\epsilon_i$ with functional predictor $\delta(t)=\beta w(t)$.
For BKMR-DLM we insert the weighted exposure $E_{mi}=\int_\mathcal{T} X_{mi}(t)w_m(t)dt$ into the kernel function or . The Gaussian kernel is then $$K_{ij}=\exp\left[-\sum_{m=1}^M \rho_m\left\{\int_\mathcal{T} X_i(t)w(t)dt-\int_\mathcal{T} X_j(t)w(t)dt\right\}^2\right]
\label{eq:kernelfunc}$$ and the polynomial kernel is $$K_{ij}=\left[1+\sum_{m=1}^M \rho_m \left\{\int_\mathcal{T} X_i(t)w(t)dt\right\}\left\{\int_\mathcal{T} X_j(t)w(t)dt\right\}\right]^d.
\label{eq:kernelfuncquad}$$ Under the constraints $\int_\mathcal{T} w(t)dt>0$ and $\int_\mathcal{T} \left[w(t)\right]^2dt=1$, both $\rho_m$ and $w_m(t)$ are identifiable.
When $w(t)$ is constant in time, then BKMR-DLM is equivalent to a BKMR model that uses pregnancy-averaged exposure as a predictor. When $w(t)$ varies in time, BKMR-DLM up- or down-weights exposures during certain time periods. The up-weighted time periods represent windows of susceptibility to exposure.
For each time-varying exposure, we parameterize both $X_{mi}(t)$ and $w_m(t)$ using a basis function representation [@Morris2015a]. We assume $X_{mi}(t)=\sum_{L=1}^{L_{m}}\xi_{mli}\psi_{mli}(t)$ and $w_m(t)=\sum_{l=1}^{L_m}\theta_{ml}\psi_{ml}(t)$, where $\{\psi_{ml}(t)\}_{l=1}^{L_m}$ is an exposure-specific orthogonal basis expansion and $\{\xi_{mli}\}_{l=1}^{L_m}$ and $\{\theta_{mli}\}_{l=1}^{L_m}$ are regression coefficients. The $m^\text{th}$ weighted exposure for individual $i$ can then be rewritten as $E_{mi}=\boldsymbol{\xi}_{im}^T\boldsymbol{\theta}_m$ where $\boldsymbol{\theta}_m=\left[\theta_1,\dots,\theta_{L_m}\right]^T$ and $\boldsymbol{\xi}_{mi}=\left[\xi_{1i},\dots,\xi_{L_mi}\right]^T$. We estimate $\boldsymbol{\xi}_{mi}$ using ordinary least squares which gives $E_{mi}=\mathbf{X}^T_{mi}\boldsymbol{\Psi}_m\boldsymbol{\theta}_m$, where $\mathbf{X}^T_{mi}$ is the row-vector of observed exposures for chemical $m$ and person $i$ and $\boldsymbol{\Psi}_m=\left[\psi_{1}(t),\dots,\psi_{L_m}(t)\right]$ is design matrix of orthonormal basis functions with columns $\psi_{1}(t)$ being the basis expansion at time point $t$.
Using an orthonormal basis, the constraint $\int [w_m(t)]^2dt=1$ is satisfied if and only if $\|\boldsymbol{\theta}_m\|=1$. The constraint $\int w_m(t)dt\ge0$ is satisfied for a set of observed times if and only if $\mathbf{1}^T\boldsymbol{\Psi}_m\boldsymbol{\theta}_m\ge0$, where $\mathbf{1}_{L_m}$ is a $L_m$-vector of ones. As such, the constraints on $w_m(t)$ are now constraints on $\boldsymbol{\theta}_m$. The constrained parameter space is half of a unit $K_m$-ball on one side of a hyperplane defined by $\mathbf{1}^T\boldsymbol{\Psi}_m\boldsymbol{\theta}_m=0$.
Using the weighted exposures as inputs, the Gaussian kernel function in is $$K_{ij} = \exp\left[-\sum_{m=1}^M\rho_m\left\{(\mathbf{X}_{mi}-\mathbf{X}_{mj})^T\boldsymbol{\Psi}_m\boldsymbol{\theta}_m\right\}^2\right],
\label{eq:K1}$$ and the polynomial kernel in is $$K_{ij} = \left[1+\sum_{m=1}^M\rho_m(\mathbf{X}_{mi}\boldsymbol{\Psi}_m\boldsymbol{\theta}_m)(\mathbf{X}_{mj}\boldsymbol{\Psi}_m\boldsymbol{\theta}_m)\right]^d.
\label{eq:K2}$$ The parameters $\rho_m$ and $\boldsymbol{\theta}_m$ represent the importance and the timing, respectively, of exposure $m$. Because these parameters are only jointly identifiable and to ease computation, we reparameterize the model in terms of $\boldsymbol{\theta}^*_m=\boldsymbol{\theta}_m\rho_m^{-1/2}$. The Gaussian kernel in is then $$K_{ij} = \exp\left[-\sum_{m=1}^M\left\{(\mathbf{X}_{mi}-\mathbf{X}_{mj})^T\boldsymbol{\Psi}_m\boldsymbol{\theta}^*_m\right\}^2\right],
\label{eq:newkernel}$$ and the polynomial kernel in can be written as $$K_{ij} = \left[1+\sum_{m=1}^M(\mathbf{X}_{mi}\boldsymbol{\Psi}_m\boldsymbol{\theta}^*_m)(\mathbf{X}_{mj}\boldsymbol{\Psi}_m\boldsymbol{\theta}^*_m)\right]^d.
\label{eq:newkernelquad}$$
In the above model written in terms of $\boldsymbol{\theta}_m^*$, both $\rho_m$ and $\boldsymbol{\theta}_m$ are uniquely identified by $\boldsymbol{\theta}_m^*$ as $\rho_m^{-1/2}=||\boldsymbol{\theta}_m^*||$ and $\boldsymbol{\theta}_m=\rho_m^{1/2}\boldsymbol{\theta}_m^*\text{sign}\{\int_\mathcal{T} w_m(t) dt \}$. Hence, we can estimate the full model parameterized in terms of $\boldsymbol{\theta}^*$ and then partition the posterior sample of $\boldsymbol{\theta}^*_m$ into $\rho_m$ and $\boldsymbol{\theta}_m$, where $\boldsymbol{\theta}_m$ uniquely describes the weight function $w_m(t)$.
To induce smoothness in the weight function we use the eigenfunctions of the covariance matrix of smoothed exposures. Specifically, we pre-smooth each exposure with a parsimonious natural spline bases and then use the eigenfunctions of the covariance matrix of the smoothed exposures in the model as specified above. An alternative is to use the eigenfunctions of the smoothed covariance matrix of the exposures obtained by the fast covariance estimation method proposed by [@Xiao2016] and implemented in the [R]{} package [refund]{} [@Goldsmith2018], as done in [@Wilson2017a]. However, we find that pre-smoothing the exposures with a parsimonious basis is more reliable in practice, particularly for moderate sample sizes.
Prior specification and posterior computation
---------------------------------------------
We use as a prior for $\boldsymbol{\theta}_m$, $m=1,\dots,M$ a uniform distribution over its parameter space, which can be written as $p(\boldsymbol{\theta}_m)\propto\exp\left(-\boldsymbol{\theta}_m^T\boldsymbol{\theta}_m/2\right)\mathbbm{1}(\boldsymbol{\theta}_m^T\boldsymbol{\theta}_m=1)\mathbbm{1}(\mathbf{1}_{L_m}^T\boldsymbol{\theta}_m>0)$, where $\mathbbm{1}(\cdot)$ is an indicator function. We then let $\rho_m/\kappa_m\sim\chi^2_1$ for fixed value $\kappa_m$. It follows that $\boldsymbol{\theta}_m^*=\rho_m^{1/2}\boldsymbol{\theta}_m\sim\text{N}(0,\nu_m\kappa\mathbf{I}_{L_m})$, with $\nu_m\sim\chi^2_1$. We complete the prior specification by assuming a flat prior on $\boldsymbol{\gamma}$, $\sigma^{-2}\sim\text{gamma}(a_1,b_1)$, and $\log(\tau^2)\sim\text{N}(0,b)$.
To estimate the model parameters, we first integrate out $\mathbf{h}$ from . This yields $\mathbf{Y} \sim\text{N}(\mathbf{Z}\boldsymbol{\gamma}, \sigma^2\widetilde{\mathbf{K}})$, where $\widetilde{\mathbf{K}}= \mathbf{I}_n+\tau^2\mathbf{K}$. The posterior can be estimated using the decomposition $$\begin{aligned}
\label{eq:postdecomp}
p(\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*,\tau^2,\nu_1,\dots,\nu_M,\boldsymbol{\gamma},\sigma^{-2}|\mathbf{Y}) &=& \\
&&\hspace{-15em} p(\boldsymbol{\gamma}|\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*,\tau^2,\nu_1,\dots,\nu_M,\sigma^{-2},\mathbf{Y}) \nonumber\\
&&\hspace{-15em} p(\sigma^{-2}|\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*,\tau^2,\nu_1,\dots,\tau_M,\mathbf{Y}) \nonumber\\
&&\hspace{-15em} p(\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*,\tau^2,\nu_1,\dots,\nu_M|\mathbf{Y}). \nonumber\end{aligned}$$ In , $p(\boldsymbol{\gamma}|\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*,\tau^2,\nu_1,\dots,\nu_M,\sigma^{-2},\mathbf{Y})$ and $p(\sigma^{-2}|\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*,\tau^2,\nu_1,\dots,\tau_M,\mathbf{Y})$ are the typical multivariate normal and gamma full conditionals. The final term in takes the form $$\begin{aligned}
\label{eq:integrated}
p(\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*,\nu_1,\dots,\nu_M,\tau^2|\mathbf{Y}) &\propto&\\\nonumber
&&\hspace{-30ex} (\tau^{-2})^{b_1-1}\left|\widetilde{\mathbf{K}} \right|^{-1/2}\left|\mathbf{Z}^T\widetilde{\mathbf{K}}^{-1}\mathbf{Z} \right|^{-1/2} \exp\left(-b_2\tau^{-2}\right) \\\nonumber
&&\hspace{-30ex}\times \bigg[a_2+\mathbf{Y}^T\widetilde{\mathbf{K}}^{-1} \mathbf{Y}/2 \\\nonumber
&&\hspace{-27ex} + \mathbf{Y}^T\widetilde{\mathbf{K}}^{-1}\mathbf{Z}\left(\mathbf{Z}^T\widetilde{\mathbf{K}}^{-1}\mathbf{Z}\right)^{-1} \mathbf{Z}^T\widetilde{\mathbf{K}}^{-1}\mathbf{Y}/2 \bigg]^{-[(n-p)/2+a_1]}\\\nonumber
&& \hspace{-30ex} \times p\left(\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*
|\nu_1,\dots,\nu_M\right)p(\nu_1,\dots,\nu_M)p(\tau).\end{aligned}$$
[@Bobb2015] updated each of the $M$ parameters in the kernel function independently with Metropolis-Hastings. This approach is unappealing for our model as we have $\sum_{m=1}^M L_m$ parameters in the kernel function, increased from $M$ with BKMR, and potentially increased correlation between parameters due to the temporal correlation in the exposures. Our Markov chain Monte Carlo (MCMC) algorithm instead iteratively samples each $\boldsymbol{\theta}_m^*$ using an elliptical slice sampler [@Murray2009] and the kernel of . Then we sample $\tau^{-2}$ using random walk Metropolis-Hastings and the same integrated kernel in . Finally, we use a Gibbs sampler to sample $\sigma^{-2}$, $\boldsymbol{\gamma}$, and $\nu_1,\dots,\nu_M$ from their respective full conditionals. The full conditional for $\nu_m$ is a generalized inverse-Gaussian distribution with density function $f(\nu_m;\lambda,\chi,\psi)\propto\nu_m^{\lambda-1} \exp\{-(\chi/\nu_m+\psi\nu_m)/2\}$, where $\lambda=-(L_m-1)/2$, $\psi=1$, and $\chi=\kappa_m^{-1}\boldsymbol{\theta}^{*T}_m\boldsymbol{\theta}^*_m$. Algorithm 1 in the supplemental material shows the full MCMC approach.
Posterior inference for $w(t)$
------------------------------
Windows of susceptibility during which there is an increased association between exposure and outcome are identified using the estimated weight function. Let $\boldsymbol{\theta}^{*(r)}_m$ for $r=1,\dots,R$ be the posterior sample of size $R$. We can identify $\boldsymbol{\theta}^{(r)}_m=\boldsymbol{\theta}^{*(r)}_m\|\boldsymbol{\theta}^{*(r)}_m\|^{-1}\text{sign}(\mathbf{1}_{L_m}^T\boldsymbol{\theta}^{*(r)}_m)$ and $w_m^{(r)}(t)=\sum_{l=1}^{L_m} \theta^{(r)}_{ml}\psi_{ml}(t)$. It is then straightforward to compute the pointwise posterior mean and credible interval for $w_m(t)$. While the credible interval provides valid pointwise posterior inference for w(t) and can be used to identify windows of susceptibility, the posterior mean does not satisfy the constraint $\int_\mathcal{T}[w(t)]^2dt=1$. We use the point estimate projected onto the parameter space of $\boldsymbol{\theta}_m$: $\widehat{w}_m(t)=\mathbf{X}^{*T}_{mi}\widehat{\boldsymbol{\theta}}_m$ with $\widehat{\boldsymbol{\theta}}_m=\bar{\boldsymbol{\theta}}_m\|\bar{\boldsymbol{\theta}}_m\|^{-2}$ and $\bar{\boldsymbol{\theta}}_m$ is the posterior mean. The resulting estimator, equivalent to the Bayes estimate with respect to the loss function $L(\boldsymbol{\theta}_m,\widehat{\boldsymbol{\theta}}_m) = [(\boldsymbol{\theta}_m-\widehat{\boldsymbol{\theta}}_m)^T(\boldsymbol{\theta}_m-\widehat{\boldsymbol{\theta}}_m)]/\mathbbm{1}\{\|\widehat{\boldsymbol{\theta}}_m\|=1\}$, is a central estimate in the parameter space of $\boldsymbol{\theta}_m$.
Posterior inference for $h(\cdot)$
----------------------------------
Estimates of $h(\cdot)$ for the observed exposure levels can be obtained by sampling from conditional distribution of $\mathbf{h}$ from . Specifically, for each MCMC iteration, we sample $\mathbf{h}$ from $$\mathbf{h}|\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*, \sigma^2,\tau^2,\boldsymbol{\gamma} ,\mathbf{Y}
\sim\text{N}\left[\tau^2\mathbf{K}\widetilde{\mathbf{K}}^{-1}(\mathbf{Y}-\mathbf{Z}\boldsymbol{\gamma}),\sigma^2\tau^2\mathbf{K}\widetilde{\mathbf{K}}^{-1}
\right].$$ To predict $h(\cdot)$ at new values, including over a regularly spaced grid, we predict $\mathbf{h}_{new}$ by considering the joint distribution $$\left(\begin{array}{c}\mathbf{h}\\\mathbf{h}_{new} \end{array}\right)\sim\text{N}\left[\mathbf{0},\sigma^2\tau^2\left(\begin{array}{cc}\mathbf{K}&{\mathbf{K}''}^T\\\mathbf{K}'' & \mathbf{K}' \end{array}\right)\right]$$ and the subsequent posterior predictive distribution $$\begin{aligned}
\mathbf{h}_{new}|\boldsymbol{\theta}_1^*,\dots,\boldsymbol{\theta}_M^*, \sigma^2,\tau^2,\boldsymbol{\gamma} ,\mathbf{Y}
&\sim& \\\nonumber
&& \hspace{-25ex} \text{N}\left[\tau^2\mathbf{K}''\widetilde{\mathbf{K}}^{-1}(\mathbf{Y}-\mathbf{Z}\boldsymbol{\gamma}),\sigma^2\tau^2\left\{\mathbf{K}' - \tau^2\mathbf{K}''\widetilde{\mathbf{K}}^{-1} {\mathbf{K}''}^T\right\}
\right].\end{aligned}$$
Simulation Study {#s:sim}
================
We evaluate the effectiveness of the proposed BKMR-DLM and other methods for mixtures of time-varying exposures in two scenarios: one with two exposures and the second with five exposures. The study is design to both evaluate the effectiveness of BKMR-DLM and to identify the best method to estimate the multi-pollutant exposure-response relation and to identify windows of susceptibility in the setting of mixtures of time-varying exposures.
Scenario A: two pollutants {#sub:sim2polls}
--------------------------
In scenario A we considered two exposures and compared BKMR-DLM with: 1) BKMR using mean exposure over pregnancy; 2) an additive DLM; and 3) an additive DLNM. To compare the results of DLM and DLNM to the true simulated weight functions, we normalized the estimates to match the constraints imposed on the true weight functions. For DLNM, where the distributed lag function varies smoothly with concentration and windows may only be identified at some concentrations, we made this comparison using the cross-section of the DLNM that shows the time-varying association at the mean exposure level for each pollutant. We fit BKMR-DLM using both a Gaussian kernel and a polynomial kernel of degree two.
For the simulation, we used real exposure data for two pollutant, PM$_{2.5}$ and nitrogen dioxide (NO$_2$), taken from one Boston, MA, USA monitor and created weekly mean exposure values for births simulated at randomly selected birth dates. This simulation strategy yields a realistic correlation structure among weekly exposures within a pregnancy, as well as seasonal variation in exposure across multiple pregnancies. The simulated weight functions were a normal density function peaking mid-pregnancy ($w_1$) and a logistic link function identifying a window in the second half of gestation ($w_2$). Both weight functions were truncated to span 37 weeks and scaled to meet the $\int_\mathcal{T}[w(t)]^2dt=1$ constraint (see Figure \[fig:simweights\] for a visualization).
We simulated the outcomes using the model $y_i = h_i + \mathbf{Z}_i^T\boldsymbol{\gamma} + \epsilon_i$. The exposure-response function, $h$, was $H(E_1,E_2)=h_1(E_1)+h_2(E_2) + h_{12}(E_1,E_2)$ with $h_1(E_1)=3/(1+\exp(-2E_1^s))$, $h_2(E_2)=2E_2^s\mathbbm{1}_{(E_2^s>0)}$, and $h_{12}(E_1,E_2)= - E_1^sE_2^s$ and $E_1^s$ and $E_2^s$ are scaled and centered versions of the weighted exposures. Hence, $h$ was nonlinear in both $E_1$ and $E_2$ and there was a multiplicative interaction between $E_1$ and $E_2$. We assumed five covariates, $\mathbf{Z}_i^T=(Z_{1i},\dots,Z_{5i})$, each simulated independent standard normal and the covariate regression coefficients, $\boldsymbol{\gamma}=(\gamma_1,\dots,\gamma_5)^T$, was also simulated independent standard normal. The random error $\epsilon_i$ was simulated normal mean zero and standard deviation 3, 7.5, and 15 which represent approximately a 1:2, 1:5, and 1:10 ratio between the standard deviation of $h$ and $\epsilon$. Finally, we considered sample sizes of $n=100$ and $n=500$ and evaluated model permanence based on 200 simulated data sets.
We note that, because the multi-dimensionality of the problem (exposure timing, multiple pollutants, potential nonlinearity and non-additivity of the exposure-response function), we purposely simulate under a scenario in which all models are misspecified. BKMR-DLM is the only model that accounts for exposure-timing, nonlinearity, and interactions. However, the natural spline basis that we employ (4 degrees of freedom) is not sufficiently flexible to model the simulated weight functions and the polynomial kernel is not sufficiently flexible to accurately represent the true exposure-response function. We include additional scenarios C and D in the supplemental material for which BKMR-DLM is perfectly specified.
Figure \[fig:simweights\] shows the true weight functions and the estimated weight functions for the first 100 simulated data sets from for $n=500$ and error standard deviation $7.5$ using both the Gaussian and polynomial kernel. Overall both approaches estimated the general pattern but over-smooth the weight function for the first exposure due to the fact that the spline basis used is not sufficiently flexible to match the peak of the window in the middle of pregnancy. Estimated weight functions for other settings are shown in the supplemental material. Additional simulation scenarios C and D with smoother weight functions are also shown in supplement. In those scenarios BKMR-DLM and BKMR-DLM-poly are better able to estimate the true weight function.
Table \[tab:simA\] compares the simulated performance of the five approaches. Overall, differences in root mean square error (RMSE) for $h$ show that BKMR-DLM with a polynomial kernel estimated the exposure-response function characterizing the association between exposure and health the best. This is despite the fact that the polynomial kernel is not sufficiently flexible to match the true exposure-response function. For the larger sample size and smaller error variance scenario, the Gaussian kernel also performed well but lacked power to estimate the exposure-response function in the more challenging scenarios. For a small sample size and larger error variance, BKMR applied to the 37-week average exposures performed slightly better. While BKMR using 37-week average exposures had lower RMSE in some cases, interval coverage was well below the nominal level. BKMR-DLM with a polynomial kernel did not have 95% coverage for $h$ in some scenarios due the insufficient flexibility of the quadratic kernel. Direct comparison of BKMR-DLM with a Gaussian kernel and BKMR with averaged exposures, which also uses the same Gaussian kernel, highlights the importance accounting for exposure timing. When there exists a sufficiently strong signal that provides information about exposure timing, the BKMR-DLM model better estimates the exposure-response function. DLM and DLNM also had low interval coverage and had larger RMSE for $h$ than BKMR-DLM with a polynomial kernel, illustrating the importance of accounting for interactions.
Similarly, BKMR-DLM had the lowest RMSE for estimating the weight functions in all scenarios except the lower signal-to-noise ones. However, The intervals for the weight functions from BKMR-DLM did not have 95% coverage. DLM and DLNM, which do not account for interactions, had larger RMSE for the weight functions and did not have adequate interval coverage. However, DLM was the most likely to identify a window of susceptibility due to increased power resulting from parsimony.
[lccccc]{}Model & RMSE $h$ & Coverage $h$ & RMSE $w(t)$ & Coverage $w(t)$ & Pr(window)\
\
\
\
\
\
\
Scenario B: five pollutants
---------------------------
Scenario B was largely the same as scenario A except we included five exposures (PM$_{2.5}$, NO$_2$, CO, O$_3$, and SO$_2$), all taken from the same Boston monitor. The first two exposures had the same weights and exposure-response functions as described above (same $w_1$, $w_2$, $h_1$, $h_2$, and $h_{12}$). We added a third active exposure, CO, and let $w_3=w_2$ and $h_3(x)=-h_2(x)$. We added two additional exposures, O$_3$ and SO$_2$, that had no association with the outcome. All other details remained the same as in scenario A.
Table \[tab:simB\] shows results for scenario B. For the purpose of estimating the exposure-response and weight functions, the pattern of relative performance of the different methods was similar to that in scenario A. For a larger sample size and lower error variance, BLMR-DLM with a polynomial kernel had the lowest RMSE for $h$ and for the weight functions. For smaller sample sizes and larger error variance, BKMR using 37-week average exposure yielded lower RMSE for $h$ but also low interval coverage. BKMR-DLM did not have sufficient power to identify windows of susceptibility; however, the improvement in estimation of $h$ and the weight functions indicate that, despite not formally identifying windows associated with exposure, some information can be learned about the timing, and this information can improve exposure-response estimation. While DLM and DLNM did not estimate the exposure-response function as well as BKMD-DLM with the polynomial kernel based on RMSE, these two simpler models had the best power to identify windows of susceptibility.
Data Analysis {#s:da}
=============
We applied BKMR-DLM with a quadratic kernel to birth weight z-score in the ACCESS cohort. For comparison, we considered an additive DLM model and BKMR with 37-week averaged exposure. In total, 109 mother-child dyads were analyzed. The analysis included four pollutant indicators: nitrate, OC, EC, and sulfate. All covariates listed in Section \[s:data\] were included in each model.
Figure \[subfig:dlmwindows\] shows results from the additive DLM analysis. We identified a susceptibility window in weeks 29-33 for nitrate, in weeks 26-27 for OC, and weeks 9-13 in sulfate. There was moderate evidence of a cumulative effect, representing the change in birth weight associated with a one unit increase in exposure at every time point, for sulfate ($p$-value$=0.07$) but not for any other pollutant.
Figure \[subfig:bkmrdlmwindows\] shows the estimated weight functions from BKMR-DLM. No windows were identified. Recall that the weight functions are constrained. As such, the magnitude and sign do not reflect that of the association. They do identify periods of time with increased importance as periods where there is a “bump” in the weight function. With that in mind, the weight functions for OC and sulfate show similarities in shape to the estimated distributed lag functions from the additive DLM but with increased uncertainty.
. \[fig:windows\]
Figure \[fig:bkmrdlmcross\] shows estimates of the exposure-response function from BKMR-DLM. The diagonal shows $h$ as a function of weighted exposure for one pollutant at the median value of the weighted exposures for the other pollutants. We found some evidence of a negative association between OC and BWGAz. Supplemental Figures S9 and S10 shows a similar estimated association using BKMR with 37-week averaged exposures. The off diagonals show the posterior mean of $h$ at different quantiles of one co-pollutant and the median of the other two co-pollutants. There was slight evidence that nitrate modifies the OC, EC, and sulfate exposure-response functions and that sulfate may modify the OC exposure response. Supplement Figures S11 and S12 and Table S3 show results for univariate DLM analyses of OC, EC, and sulfate stratified by nitrate and OC stratified by sulfate. This simpler analysis supports the conclusion that there may be an association between OC, EC, and sulfate at higher levels of nitrate. This follow-up analysis showed that the windows of susceptibility associated with OC, EC, and sulfate only exist at higher levels of co-exposure to nitrate.
![Cross sections of the estimated exposure-response function ($\hat{h}$). Each column shows the exposure-response function as the weighted exposure for a single pollutant varies. The panels on the diagonal show the main effect, the association between a weighted exposure and the outcome at the median level of all other weighted exposures. The off-diagonals show the exposure-response function at different quantiles of the co-exposure. []{data-label="fig:bkmrdlmcross"}](bkmrdlm_crosssections_greyscale.pdf){width="\textwidth"}
Discussion {#s:discussion}
==========
In this paper we consider multiple strategies for quantifying the association between time-varying measures of an environmental mixture and a prospectively assessed health outcome, and compared the relative advantages and disadvantages of each approach through simulation and application of each to a case study involving prenatal exposures to multiple air pollutants and birth weight. In this setting there are three key challenges: accounting for exposure timing, accounting for nonlinear associations, and accounting for interactions. Table \[tab:summary\] summarizes methods that can be applied in this setting and the advantages and disadvantages of each approach.
We proposed BKMR-DLM to estimate the association between multiple time-varying exposures and an outcome. To our knowledge, this is the first approach that accounts for exposure timing, interactions between exposures, and nonlinear associations–thereby more comprehensively modeling the underlying complexity of the relationships. The approach uses time-weighted exposures in a kernel machine regression framework. The weight-functions identify windows of susceptibility during which there is an increased association between exposure and outcome. Such information will be important as developmental processes are both timed and linked to windows of susceptibility, thus exposure timing provides hints to the underlying mechanisms. By using kernel machine regression we allow for nonlinear associations and interactions among the multiple weighted exposures.
----------------------- -------------- -------------- -------------- -------------- --------------- ---------------
(l)[2-4]{} (l)[5-7]{}
exposure inter- non- detecting wk signal$^*$ st signal$^*$
Method timing actions linearity windows modest $n$ large $n$
Add. DLM $\checkmark$ $\checkmark$
Add. DLNM $\checkmark$ $\checkmark$
BKMR $\checkmark$ $\checkmark$ $\checkmark$
BKMR-DLM $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$
----------------------- -------------- -------------- -------------- -------------- --------------- ---------------
: Summary of methods and performance. The first three columns show what data features the method accounts for by design. The last three columns summarize relative performance of the methods and indicates the recommended method based on study objectives of estimating the exposure-response function (ER) and detecting windows of susceptibility. These recommendations are based on the the simulation study results.[]{data-label="tab:summary"}
$^*$ denotes weaker signal and stronger signal, respectively.
ER: exposure-response
In addition to this novel method we considered BKMR using averaged exposures. BKMR allows for nonlinear associations and interactions but does not account for exposure timing. Further, we used additive versions of both DLMs and DLNMs to estimate exposure-response relationships. DLM accounts for exposure timing but limits the analysis to linear associations and no interactions. DLNM accounts for exposure-timing and nonlinear associations, but also limits the analysis to main effects only.
In a simulation study, we showed that BKMR-DLM with a polynomial kernel function was best able to estimate the exposure-response relation in most situations. When the sample size is small or the signal-to-noise ratio is small, BKMR, which is a simpler model because exposure timing is not estimated, performed slight better. While DLNM and DLM were not as accurate in estimating the exposure-response relationship, these approaches had greater power to identify windows of susceptibility. DLM, which is the simplest model, was the most effective at identifying windows of susceptibility. Hence, while BKMR-DLM has the advantage of accounting for exposure timing, nonlinear associations, and interactions in a single model and performed best with estimation of the exposure-response function, the preferred method for time-varying measures of a multi-pollutant mixture may depend on both the data, the strength of the exposure-response relationship, and the primary study objectives.
We applied this strategy of combining the use of BKMR-DLM to identify potential interactions and DLM to identify windows of susceptibility to analyze data on prenatal exposure to an air pollution mixture and birth weight in the ACCESS cohort. In a sample of 109 boys born to obese mothers, we estimated the association between four ambient pollutants and birth weight using BKMR-DLM as a screening method to identify potential interactions between time-varying exposures. We then performed a stratified DLM analysis to confirm that the findings from the primary BKMR-DLM analysis were not driven solely by modeling assumptions. The fact that the simpler stratified DLM analysis yielded similar results does not diminish the utility of BKMR-DLM as BKMR-DLM was used to identify which specific pairs of pollutants to further investigate with the stratified DLM. Using a stratified DLM alone would require estimating an unacceptable large number of models when the number of pollutants is large.
The analysis in this paper included 109 dyads and four pollutants each estimated at 37 time points. While this analysis may have been slightly under-powered there are several larger studies that will be yielding larger data in the near future. In particular the National Institutes of Health Environmental influences on Child Health Outcomes (ECHO) program is pooling similarly designed birth cohorts and will provide the potential and power to investigate this complex interplay of multiple pollutants, complex exposure-response functions, and timing. Therefore the knowledge gained from this investigation will provide guidelines for the statistical analysis of these future studies
Both the estimation of health effects associated with multi-pollutant mixtures and the identification of windows of susceptibility are important areas of environmental health research. BKMR-DLM and other related distributed lag methods are tools that can be used to combine these two important areas of research and simultaneously estimate windows of susceptibility and multi-pollutant exposure-response functions, while properly quantifying uncertainty arising from uncertain exposure timing.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported in part by NIH grants R01ES028811, R01ES013744, P30ES000002, P30ES023515, and UH3OD023337 and US EPA grant RD-83587201. Its contents are solely the responsibility of the grantee and do not necessarily represent the official views of the US EPA. Further, US EPA does not endorse the purchase of any commercial products or services mentioned in the publication. The ACCESS cohort has been supported by NIH grants R01ES010932, U01HL072494, and R01HL080674. This work utilized the RMACC Summit supercomputer, which is supported by the NSF (awards ACI-1532235 and ACI-1532236), the University of Colorado Boulder and Colorado State University. The RMACC Summit supercomputer is a joint effort of the University of Colorado Boulder and Colorado State University.
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[Microsoft]{}’s internal big data analytics platform is comprised of hundreds of thousands of machines, serving over half a million jobs daily, from thousands of users. The majority of these jobs are recurring and are crucial for the company’s operation. Although administrators spend significant effort tuning system performance, some jobs inevitably experience slowdowns, i.e., their execution time degrades over previous runs. Currently, the investigation of such slowdowns is a labor-intensive and error-prone process, which costs [Microsoft]{}significant human and machine resources, and negatively impacts several lines of businesses. In this work, we present [Griffin]{}, a system we built and have deployed in production last year to automatically discover the root cause of job slowdowns. Existing solutions either rely on labeled data (i.e., resolved incidents with labeled reasons for job slowdowns), which is in most cases non-existent or non-trivial to acquire, or on time-series analysis of individual metrics that do not target specific jobs holistically. In contrast, in [Griffin]{}we cast the problem to a corresponding regression one that predicts the runtime of a job, and show how the relative contributions of the features used to train our interpretable model can be exploited to rank the potential causes of job slowdowns. Evaluated over historical incidents, we show that [Griffin]{}discovers slowdown causes that are consistent with the ones validated by domain-expert engineers, in a fraction of the time required by them.'
author:
- |
Liqun Shao$^1$, Yiwen Zhu$^1$, Abhiram Eswaran$^1$, Kristin Lieber$^1$,\
Janhavi Mahajan$^1$, Minsoo Thigpen$^1$, Sudhir Darbha$^1$,\
Siqi Liu$^2$, Subru Krishnan$^1$, Soundar Srinivasan$^1$,\
Carlo Curino$^1$ and Konstantinos Karanasos$^1$\
\
\[2mm\]
bibliography:
- 'soccbib.bib'
title: |
[Griffin]{}: Reasoning about Job Anomalies with\
Unlabeled Data in Cloud-based Platforms
---
Introduction
============
[Microsoft]{}operates one of the biggest data lakes worldwide for its big data analytics needs [@hydra]. It is comprised of several clusters for a total of over 250k machines and receives approximately half a million jobs daily that process exabytes of data on behalf of thousands of users across the organization. The majority of these jobs are recurring and several of them are critical services for the company. Hence, administrators and users put significant effort in tuning the system and the jobs to optimize their performance. Nevertheless, some jobs inevitably experience slowdowns in their execution time (i.e., they take longer to complete than their previous occurrences) due to either system-induced (e.g., upgrades in the execution environment, network issues, hotspots in the cluster) or user-induced reasons (e.g., changes in job scripts, increase in data consumed).
Such *job slowdowns* can have a catastrophic impact to the company. In fact, runtime predictability is often considered more important than pure job performance in recurring production jobs [@morpheus]. First, several jobs are interdependent, that is, the output of a job might be consumed by multiple other jobs [@owl]. Thus, the slowdown of the first job can have a cascading effect on all other dependent jobs, impacting vital services across the company. Second, some business-critical jobs are associated with deadlines in the form of service-level objectives (SLOs). Missing those SLOs can result in substantial financial penalties in the order of millions of dollars.
Despite the importance of promptly resolving such incidents, the current approach remains largely manual. Job slowdowns are signaled either through tickets raised by customers or by missed deadlines (for jobs with SLOs). In either case, a slow, labor-intensive process of error triaging and root-cause analysis must be initiated. In particular, on-call engineers manually investigate causes of job slowdowns by analyzing hundreds of logs and system traces through a complex monitoring dashboard. Despite the existence of detailed metrics, it can sometimes take several hours to resolve an incident. This bottleneck costs millions of dollars in engineering time wasted on investigation and in job SLO violations, and results in degraded user experience.
In this work, we present [Griffin]{}, the system we built and have deployed in our production big data analytics infrastructure to automatically discover the main factors causing a job’s runtime deviation through the use of machine learning. [Griffin]{}greatly improves the situation described above. First, it helps users find user-induced causes of their job slowdowns and prevents them from raising tickets that are “false alarms” to system administrators. Second, in case of actual infrastructure issues, it directs administrators towards the most probable causes for a job slowdown and allows early elimination of factors unrelated to the slowdown. Third, by observing slowdowns in jobs submitted for testing purposes, administrators can resolve system issues before they affect user jobs.
Existing related works have used detection methods such as classification and clustering to perform analysis of anomalies in cloud computing [@agrawal2015survey; @modi2013survey]. However, to analyze anomalies, these methods rely on labeled data, e.g., data from existing incidents that associate jobs with their slowdown causes. Such labeled training data in production cloud systems are extremely hard to obtain and can also be erroneous. A few approaches do consider unlabeled data, but rely either on time-series analysis or restrict their focus to machine or VM behavior [@cherkasova2009automated; @cohen2004correlating; @dean2012ubl; @gu2009online; @tan2010adaptive; @tan2012prepare; @zhang2013intelligent]. In contrast, we focus on job instances that span several hundreds of machines but only during the lifetime of the job. As a result, existing techniques are frequently not applicable to identify root causes of job slowdowns.
Unlike existing works, [Griffin]{}employs an interpretable regression model to predict job runtime and then suggest reasons for runtime deviations. [Griffin]{}exploits two characteristics of available data. First, in our clusters we collect valuable telemetry at various levels of abstraction (at the job, machine, and cluster level). Second, the majority of our jobs are recurrent, meaning similar jobs which run regularly, for example every day or several times a day, allowing [Griffin]{}to leverage this historic data. Then, based on the relative contribution of each metric/feature to the runtime of a job that experienced a slowdown, we emit a list of possible causes for the slowdown, ranked by their importance.
Our contributions are the following:
1. We present an end-to-end ranking system to identify the root causes of job slowdowns without human-labeled data.
2. We show how an interpretable regression model can be used to reason about job slowdowns.
3. We experimentally compare various models in terms of accuracy, scalability in the model size and number of jobs, and generalizability to jobs not seen before by the system.
4. [Griffin]{}is deployed in our clusters and is used by our engineers. Early indications show that slowdown causes generated by [Griffin]{}are closely correlated to causes validated by domain experts. At the same time, [Griffin]{}drops the time of investigation by orders of magnitude compared to the existing manual process.
The rest of the paper is organized as follows. Section \[sec:background\] provides details on our production environment and the relevance of the problem we focus on. Section \[sec:architecture\] gives an overview of [Griffin]{}. Section \[sec:reasoning\] describes our anomaly reasoning algorithm. Section \[sec:data collection\] discusses feature engineering and data collection, whereas Section \[sec:deployment\] provides details on [Griffin]{}’s deployment. Section \[sec:experiment\] presents the results of our experimental evaluation. Section \[sec:related\] discusses related work, and Section \[sec:conclusion\] provides our concluding remarks.
Background on our Environment {#sec:background}
=============================
In this section, we provide background on the characteristics of our analytics clusters to give a sense of the scale of the problem we target (Section \[sec:bg:env\]). Then, we describe the current state of affairs in finding the reasons for a job’s slowdown (Section \[sec:bg:debug\]).
Cluster Characteristics {#sec:bg:env}
-----------------------
At [Microsoft]{}we operate a massive data infrastructure, powering our internal analytics processing. This infrastructure consists of several clusters, each comprised of tens of thousands of machines—Table \[tab:req3\] highlights some details of our clusters’ scale. To make the situation even more challenging, our cluster environments are also heterogeneous, including several generations of machines.
![Running tasks in one of [Microsoft]{}’s production analytics clusters, comprised of tens of thousands of machines.\[fig:tasks\]](imgs/running-tasks){width="\columnwidth"}
{width="99.00000%"}
Tens of thousands of users submit hundreds of thousands of jobs to these clusters daily. Each job is a directed-acyclic graph (DAG) of operators (which we term *stages*), and each stage consists of several tasks [@scope]. Each task gets executed in a cluster’s machine (and each machine runs several tasks in parallel). Figure \[fig:tasks\] depicts the number of running tasks in one of our clusters over the course of a week. At each moment in time there are between 200k–300k tasks running.
Given this extreme scale and complexity, job slowdowns are quite common. Manually investigating such slowdowns, as we explain in the following section is a painful and time-consuming effort.
Manual Job Slow-down Investigation {#sec:bg:debug}
----------------------------------
We now describe how [Microsoft]{}engineers used to approach job slowdowns before [Griffin]{}got deployed in our clusters. Figure \[fig:manual\] shows two occurrences of the same job. The top one corresponds to its regular execution, taking 44 mins to complete. The bottom one experiences a slow down with a completion time of 88 mins. The figure visualizes how long the various stages of the job take to execute (although several stages might run in parallel, this tool shows the ones in the critical path, as those determine the job’s execution time).
To investigate this slowdown, an engineer will typically start by looking at the visualization tool of Figure \[fig:manual\], trying to detect the stages that seem abnormal. Note, however, that in this example, the “regular” top occurrence is the one that seems to have longer stages. Therefore, this tool is of limited use. Next, the engineer will have to manually combine several other tools and system files to get more information about the job and the system during the time this job was executed. Given the scale of the system and the amount of metrics collected, this process can take a considerable amount of time to complete. Multiply this by the number of slowdowns and one can easily see the significant opportunity in saving engineering time and improving user experience by speeding up this process. Note also that only a few engineers have the knowledge to perform this manual analysis.
System Overview {#sec:architecture}
===============
{width="85.00000%"}
[Griffin]{}’s goal is to find the causes for job runtime degradations in our big data analytics clusters. A central requirement is to not rely on labeled data, i.e., there should be no need for existing slowdown instances associated with their causes. Each job in our clusters is associated with a set of telemetry data that we already collect for monitoring and debugging purposes (e.g., number of tasks, size of input data, load of machines the job was executed on—see Section \[sec:data collection\] for details), some of which contributed to the job’s slowdown. Instead of finding a subset of the slowdown causes (as a system that relies on labeled data would do), [Griffin]{}*ranks* the causes (i.e., the features) in the order they affected the deviation of the job’s runtime from its expected runtime, and then suggests the top causes to the users. A formal description of the problem and our approach for solving it is presented in Section \[sec:reasoning\].
[Griffin]{}’s architecture is depicted in Figure \[fig:frm\]. It consists of two pipelines: the offline training and the online prediction, which we detail below.
[**Training**]{} The offline training process uses various metrics that we collect at the job, machine, and cluster level to generate a model that will be able to predict the runtime of a job given these features. Our anomaly reasoning algorithm will use this model to rank the features that contributed to the job’s slowdown. The training process involves the following steps: (1) data preparation, i.e., collect and clean the data from different sources in the cluster at regular intervals; (2) feature engineering, i.e., extract raw features from the collected data, create new ones, and choose the ones that we will be using to train the model; and (3) model generation, i.e., the creation of the model, including hyper-parameter tuning, training, and model evaluation. The data preparation is detailed in Section \[sec:data collection\], whereas the model details are in Section \[sec:reasoning\].
![[Griffin]{}’s output for the slowdown of the job depicted in Figure \[fig:manual\].[]{data-label="fig:griffon-output"}](imgs/piechart){width="0.9\columnwidth"}
The generated model is stored in the Tracking Server, which tracks model runs and stores performance metrics and hyper-parameter values (see Section \[sec:deployment\]).
[**Prediction**]{} The online prediction pipeline provides an API that takes as input the ID of a job that was executed and experienced a slowdown. This API is exposed to the users through a web application. Then the Online Feature Building component gathers the metrics associated with that job and provides the data to the Model Server where the prediction model is deployed. The output is a report with ranked reasons for job slowdown. In addition, the system provides a confidence level in the results (see Section \[sec:confidence\]).
High confidence means that users can rely on the output of the system. On the other hand, low confidence means that the metrics used are not sufficient to explain the job slowdown. The output of the system is useful even in the latter case, because users can rule out these metrics and focus their investigation in other areas.
Figure \[fig:griffon-output\] shows [Griffin]{}’s output for the job slowdown of Figure \[fig:manual\]. In this case, [Griffin]{}suggests with high confidence that the increase in data written by the job is the main reason for its slowdown. As this is a user-induced reason and not a problem with the system, the corresponding ticket can be closed without further investigation.
Anomaly Reasoning Algorithm {#sec:reasoning}
===========================
In this section, we describe our algorithm for reasoning about anomalies. First, we formally define the problem (Section \[sec:propstate\]) and discuss model interpretability (Section \[sec:model\]). Then we describe the interpretable tree-based model that [Griffin]{}uses for determining the reasons for a job’s slowdown (Section \[sec:treemodel\]) and its associated confidence level (Section \[sec:confidence\]).
Problem Statement {#sec:propstate}
-----------------
We consider a set of jobs that have already been executed. Hence, we know the runtime of each job. Through our collected metrics, we also know the values of the features that we are interested in (see Section \[exploreselection\] for feature selection).
The majority of the jobs submitted in our clusters are analytics jobs[^1] that are recurring, i.e., they are submitted at regular intervals (typically hourly, daily, or weekly) [@hydra]. We use the notion of **job template** to refer to each of these recurring jobs. Jobs belonging to the same template have very similar scripts with minor differences, e.g., to access the latest data.
\[t\] ![Example of baseline selection for a job template.[]{data-label="fig:baselines"}](base1.png "fig:"){width="0.8\columnwidth"}
We also define the **baseline** of a job to be its “expected” runtime, given the runtime of the other jobs that belong to the same template. In practice, we use the mean runtime of the jobs whose runtime falls between the 45$^{th}$ and 55$^{th}$ percentile for that template. A benefit of using a percentile measure is that we avoid outliers. Therefore “slow jobs" in our training data will not affect the baseline set. Figure \[fig:baselines\] shows the runtime distribution for one job template. The data we use for baseline selection falls between the two orange lines.
For jobs that belong to templates with no previous occurrences, we use the baseline of jobs with similar characteristics (in data size and performed operations). Similarly, we define the baseline of various features of a job to be their expected value, given the jobs of the corresponding template.
Let ${\bm{x}} \in \mathbb{R}^p$ be the $p$-dimensional features, and $y \in \mathbb{R}$ be the job runtime. Let [${y}^\beta$]{} and [${{\bm{x}}}^\beta$]{} be the baseline of the runtime and the features, respectively. We define the problem as follows: for each job $i$, lacking human labeled reasons, with features ${\bm{x}}^i$ and runtime $y^i$, predict the rank of different features based on their influence on the deviation of $y^i$ from [${y^i}^\beta$]{}.
Interpretable Model {#sec:model}
-------------------
Consider a machine learning model that is trained to predict the runtime $y$ of a job using a set of features ${\bm{x}}$. That alone would be a standard regression problem. However, in our setting we want to use such a runtime prediction to find the features that contribute the most to a job’s slowdown, that is, to the runtime’s deviation from the job’s baseline [${y}^\beta$]{}. To this end, we need an *interpretable* regression model for the job’s runtime.
We define a regression model to be interpretable, if the output of the model can be expressed as the sum of contributions of each of the model’s features:
$$\label{eq:linear_contrib}
y = c + \sum_{k=1}^K fc_{k}$$
where $c$ is a constant and $fc_{k}$ is the contribution of feature $x_k$ to the prediction.
Similarly, for a baseline job, let $y'^\beta$ be the predicted runtime based on the same model using the baseline features $x_k^\beta$, we can decompose the model prediction as: $$\label{eq:linear_contrib_baseline}
y'^{\beta} = c + \sum_{k=1}^K fc_{k}^\beta$$ where $fc_{k}^\beta$ is the contribution of feature $x_k^\beta$ to the prediction.
In our setting, we can quantify the **delta feature contribution** ${\bm{\mathit{\Delta}}}fc_{k}$ of each feature $x_k$ to the deviation of $y$ from $y'^\beta$:
$$\label{eq:delta_contrib0}
y - y'^\beta = \sum_{k=1}^K (fc_{k}-fc_{k}^\beta) = \sum_{k=1}^K {\bm{\mathit{\Delta}}}fc_{k}$$
In our case, for the baseline jobs of all templates, the model prediction is very accurate (with Mean Absolute Ratio Error as 2.2%). Thus the sum of the delta feature contribution approximately equals to the deviation from the baseline job runtime $y^\beta$. If the predicted baseline runtime is not accurate, i.e. the difference between $y^\beta$ and $y'^\beta$ is large, we can raise a flag about our confidence of the model result, as will be discussed in Section \[sec:confidence\].
Being able to quantify the contributions of each feature to a job’s slowdown allows us to rank the features in order of importance, which is the goal of [Griffin]{}.
[**Model choice**]{} We considered various model categories to predict the runtime of a job, namely Linear Regression (LR), Random Forest (RF), Gradient Boosted Trees (GBT), and Deep Neural Network (DNN). Our main requirements were that the model be interpretable and that it offers good accuracy.
A linear model can be expressed as $y = \alpha + \sum_{k=1}^K \beta_k x_k$, where $\alpha, \beta_k \in \mathbb{R}$. It is trivial to show that it satisfies the interpretability criterion of Eq. \[eq:linear\_contrib\]. However, as we show in Section \[choosemodel\], the accuracy is worse than that of the other models. GBT and DNN have acceptable accuracy but their interpretability is hard to establish. Lastly, the RF model exhibited the best accuracy in our experiments and therefore, is our model of choice in [Griffin]{}. In the next section, we describe an appropriate tree interpreter that reformulates a tree-based model to a linear form, so that we can use it to rank feature contributions to job slowdowns.
Note that when training our models, we considered both a *global* and *per-template* models. In the former case, we train a single unified model to predict runtime using jobs of all templates together in the training set. In the latter, we train one model per job template utilizing training data drawn exclusively from jobs that belong to the particular template. In Section \[sec:experiment\], we compare the two approaches in terms of accuracy, scalability, and generalizability.
Interpretable Random Forest {#sec:treemodel}
---------------------------
In a Random Forest (RF) model, for each tree, in order to make a prediction, we traverse a path from the root of the tree to a leaf. This path consists of a series of decisions based on the model’s features. Assuming there are $M$ nodes on the path, each node separates the feature space into two, given a feature $x_k$ and a threshold $t_k$: the one child node corresponds to $x_k \leq t_k$, the other to $x_k > t_k$. In other words, from the root node where all the samples reside, a partition based on feature $x_k$ and threshold $t_k$ thus separates the data samples to the two children that correspond to smaller feature spaces.
Consider a tree $j$ of the model and a node $m \in j$ that is partitioned from its sibling based on feature $x_k$. Let $\bar{y}_{m,j}$ be the mean target value for all samples that reside on node $m$. Then the contribution of feature $x_k$ to the final prediction due to this partitioning is calculated as: $$\begin{aligned}
\label{eq:tree1}
\Delta_{m}\text{contrib}_j(x,k)=(\bar{y}_{m,j} - \bar{y}_{m-1,j})I_j(m,k)\end{aligned}$$
for $2<m\leq M$, where node $m$$-$$1$ is $m$’s parent. $I_j(m,k)$ equals to 1 if the partitioning at node $m$$-$$1$ involves feature $x_k$ for tree $j$ or 0 otherwise. The number of samples that reside on each node becomes smaller and smaller by traversing the path, as the feature space gets smaller. The contribution of $x_k$ to the final prediction can be calculated as the sum of all $\Delta_{m}\text{contrib}_j(x,k)$:
$$\begin{aligned}
\label{eq:tree2}
\text{contrib}_j(x,k)=\sum_{m=2}^{M}\Delta_{m}\text{contrib}_j(x,k) \\ =\sum_{m=2}^{M}(\bar{y}_{m,j} - \bar{y}_{m-1,j})I_j(m,k)
$$
The prediction of the target value from this tree is $\bar{y}_{M,j}$ and can be expressed using the sum of all features’ contributions along the path: $$\begin{aligned}
\label{eq:tree3}
y_j = \bar{y}_{M,j} = \bar{y}_{1,j} + \sum_{m=2}^{M}\bar{y}_{m,j} - \bar{y}_{m-1,j} \\ = c_{j} + \sum_{m=2}^{M}\sum_{k=1}^{K}\Delta_{m}\text{contrib}_j(x,k)\end{aligned}$$
where $c_{j}$ is the full sample mean. *TreeInterpreter* [@TreeInterpreter] combines the results of all trees in our Random Forest by taking the sum of the contribution from each tree. Thus, each prediction is decomposed into a sum of contributions from the features, as follows: $$\label{eq:tree_contrib}
y = \dfrac{1}{J}\sum_{j=1}^J c_{j} + \sum_{k=1}^K (\dfrac{1}{J}\sum_{j=1}^J \text{contrib}_j(x, k))$$ where $J$ is the number of trees, $c_{j}$ is the full sample mean for each $j^{th}$ tree, and $K$ is the number of features involved. Using $c = \dfrac{1}{J}\sum_{j=1}^J c_{j}$ for the average runtime across the whole training set and $fc_{k} = \dfrac{1}{J}\sum_{j=1}^J \text{contrib}_j(x, k)$ for the contribution of feature $x_k$ to the predicted runtime, we get to Eq. \[eq:delta\_contrib0\], which shows that our Random Forest model meets the interpretability criterion. Therefore, it can be used to detect reasons for job slowdowns in [Griffin]{}.
Confidence Level {#sec:confidence}
----------------
The confidence level shows how reliable is the prediction made by our model for the contribution of each feature to a job’s slowdown. We consider two factors that affect our model confidence: (1) the relative error in predicting the runtime of the job (by comparing the model prediction with the actual runtime of the job); (2) the confidence intervals estimated by the random forest [@Mein2006].
The relative error is defined as following: $$\label{eq:error}
\text{error\_rate} = \frac{|\text{predicted\_runtime} - \text{actual\_runtime}|}{\text{actual\_runtime}}$$
We use two thresholds, $t_1$ and $t_2$, for the relative error, as explained below.
The confidence interval of the random forest method is estimated based on the prediction of each decision tree, $y_j, \forall j \in \{1,2,3,\cdots, J\}$. We take the $p^{\text{th} }$ and $(100-p)^{\text{th} }$ percentile of the distribution of $y_j$. If the final prediction $y$ is within this range, we consider the prediction to have low variance, since the predictions from all trees are consistent. We define three confidence levels as follows:
High
: The prediction $y$ is within the range of $p^{\text{th} }$ and $(100-p)^{\text{th} }$ percentile of $y_j$, and the relative error is lower than threshold $t_1$;
Medium
: The relative error is between $t_1$ and $t_2$;
Low
: Other scenarios.
Parameters $p$, $t_1$ and $t_2$ are tuned as hyper parameters using validation data. High confidence means our model can reliably predict the slowdown reasons. Low confidence means the reasons are likely to fall outside the metrics we used. In the API, the user will be presented the level of confidence. A low confidence indicates more investigation will be needed. However, even in this case, as the model has examined many metrics, the DRI can focus their investigation in other areas.
Data Preparation and Feature Engineering {#sec:data collection}
========================================
Data Preparation {#dataprep}
----------------
In our clusters, we keep hundreds of metrics for monitoring, reporting, and troubleshooting purposes, which result in petabytes of logs and metrics per day. Moreover, the features we are interested in are scattered both *physically* (in different files across our hundreds of thousands of machines) and *logically* (we need to process and combine different files to generate features). To perform the required data preparation and extract features out of the data, we use Scope [@scope], which provides a SQL-like language and can support our scale. Feature extraction occurs both during the offline training phase and the online prediction (see Figure \[fig:frm\]). Given that data freshness is not an issue for training (we do not need data of the current day), we use data that becomes available daily in our clusters and includes years-worth of historical data. For prediction, we need to collect features only for a single job (or a group of jobs), but latest data is required, as a user might want to debug their job that just finished. Thus, we use different data sources that allow us to access data within minutes from when they are produced (but that do not allow access to historical data, so cannot be used for training).
Feature Engineering and Selection {#exploreselection}
---------------------------------
In collaboration with domain experts, we selected a subset of the features we collect to train our models in [Griffin]{}, based on what could potentially impact the runtime of a job. As already discussed, a job can experience a slowdown compared to its previous occurrences, due to either user-induced or system-induced reasons. User-induced reasons can be captured by metrics collected at the job-level, whereas metrics related to system-induced reasons can be split to either machine-level or cluster-level, as detailed below.
Job-level
: These are metrics collected for each job. [Griffin]{}currently uses approximately 15 such features, including data read within and across racks, data written, data skewness metrics, job priority, execution DAG features (e.g., number of stages and tasks), and user information.
Machine-level
: These are metrics collected at the machines used during the execution of a job, such as CPU load, allocation delays, I/O reads/writes. Due to the challenges of collecting such features and correlating them with each job, in production [Griffin]{}currently uses only a few of them, but we are working on adding more.
Cluster-level
: These relate to the cluster environment when a job was executed. Examples of the ones we use are job queuing times, number of failed and revoked vertices, and execution environment version.
[**Challenges**]{} \[challenges\] One of the biggest challenges with feature engineering is the correlation between features, which often results in high variance in the model prediction. While certain highly correlated features ($>0.95$) were removed, feature data was preserved to the greatest extent, because correlated features may indicate different problems with a job. For instance, input size and input size per task have a correlation of $0.9$. However, the former might indicate that the slowness reason was more data; the latter might indicate data skew. Fortunately, [Griffin]{}’s tree-based models have an innate feature of being robust to correlated features.
Tracking and Deployment {#sec:deployment}
=======================
[**Model Tracking**]{} Machine learning is an iterative process, and reproducibility and versioning are crucial to productionalize machine learning models. To track model history in [Griffin]{}, we use MLflow [@MLflow] in our Tracking Server (see Figure \[fig:frm\]), which is deployed as an Azure Linux VM. For each model, we track logging parameters, code versions, metrics, and model artifacts. [**Model Serving**]{} In order to make available to end users a model that we trained and stored in the Tracking Server, we use one of MLflow’s “flavors”[^2] to build an Azure Machine Learning image out of that model. Then we deploy this image as an Azure Container Instance, using the Azure ML Service’s [@AML] in the Model Server (see Figure \[fig:frm\]).
The models that are deployed in the Model Server are made available to end users through a web application, which exposes a scoring API. The web application runs on a flask web server [@Flask]. [**Model Monitoring**]{} \[monitoring\] It is critical to monitor the model performance and retrain models if they go stale. To ensure this, we have provisioned our pipeline to allow single click retraining. Decoupling data sources from the training pipeline helps to easily refresh our data, retrain, and deploy the model with minimal impact to end users.
Evaluation {#sec:experiment}
==========
We now present our experimental evaluation for [Griffin]{}. In Section \[sec:validation\] we discuss [Griffin]{}’s effectiveness in finding the actual causes of job slowdowns. In Section \[choosemodel\] we compare different machine learning models for training, whereas Section \[scale\] studies the scalability of different models as the number of job templates increases. In Section \[sec:size\] we compare the model performance when training our model with an increasing number of jobs per template. In Section \[modgeneralize\] we show early evidence that [Griffin]{}can be applicable in different domains.
We carried out our experiments on Windows using Python 3.7. We used a machine with eight 2.90 GHz processors and 64 GB RAM for the experiments in Sections \[sec:validation\] and \[choosemodel\], and a high-memory Virtual Machine (VM) for the scalability experiments of Section \[scale\]. The job and feature data is obtained from the [Microsoft]{}production clusters, as described in Section \[dataprep\]. We use historical job data with different job templates, as described in Section \[exploreselection\], over a period of three months.
![Runtime distribution for the jobs of different templates.[]{data-label="fig:template-runtime"}](template-runtime.png){width="0.8\columnwidth"}
Validation Results {#sec:validation}
------------------
Working with domain experts at [Microsoft]{}, we picked a set of job templates that are considered important for our production clusters (SLO critical), and trained [Griffin]{}based on those. Note that the runtime distribution of the jobs of different templates varies significantly, which poses extra challenges for the runtime prediction based on machine learning models. Figure \[fig:template-runtime\] shows the runtime distributions for five of the templates that we used.
From these templates, we then randomly picked seven jobs that experienced slowdowns (five from these templates and two from different templates), and compared the causes for slowdowns that were identified by the experts with those suggested by [Griffin]{}. For these jobs, Table \[validation\] shows the reasons identified by the experts and [Griffin]{}(with their ranking), [Griffin]{}’s confidence level, and if the job belonged to one of the templates used for training the model (in-t). We use $R_x$ to denote the reason that [Griffin]{}predicted for a job’s slowdown with rank $x$. For readability, we show only the reasons that were common between [Griffin]{}and experts and use $R_x$ variables for the rest.
--------------------------------------------------------------------------------------------------------------------- -- -- -- --
**Job & **[Griffin]{}’s Predicted List ofRanked Reasons & **Engineer ValidatedReason & **Confidence Level & **in-t\
1 & \[Input Size, $R_2$, $R_3$\] & Input size & High & Yes\
2 & \[$R_1$, $R_2$, $R_3$, Revocation, $R_5$\] & Revocation & Medium & Yes\
3 & \[$R_1$, $R_2$, $R_3$, $R_4$, $R_5$, $R_6$\] & Framework issue & Low & Yes\
4 & \[$R_1$, $R_2$, $R_3$, $R_4$, High compute hours\] & High compute hours & Medium & Yes\
5 & \[Time skew, $R_2$, $R_3$, $R_4$\] & Time skew & High & Yes\
6 & \[High compute hours, $R_2$, $R_3$\] & High compute hours & High & No\
7 & \[$R_1$, Usable machine count, $R_3$, $R_4$\] & Usable machine count & High & No\
8 & \[High compute hours, $R_2$\] & High compute hours & High & No\
**********
--------------------------------------------------------------------------------------------------------------------- -- -- -- --
The results in Table \[validation\] show that the reasons generated by [Griffin]{}are highly correlated with the reasons manually validated by our domain expert engineers. For job 1, the top predicted reason is the same as the manually validated reason with high confidence. For jobs 2 and 4, our system predicted the validated reason in the top 5 slowdown reasons, which is consistent with the confidence level medium. For job 3, our model does not identify the same reasons as the experts, also consistent with low confidence. Adding more features as planned (e.g., additional machine-level features) will allow us to improve the model’s prediction capability and minimize such low-confidence cases. Jobs 6, 7 and 8 show the robustness of our model: although [Griffin]{}was not trained using these job templates, it can still find the correct reasons with high confidence by using knowledge gathered by other similar job templates. Importantly, we observed no misleading predictions, i.e., there were no cases where [Griffin]{}predicted wrong slowdown reasons with high confidence. This means that even predictions with low confidence can be useful in ruling out the currently used features from the investigation.
Picking the Right Model {#choosemodel}
-----------------------
As discussed in Section \[sec:reasoning\], we experimented with various categories of models for the job runtime prediction, including Linear Regression (LR), Random Forest (RF), Gradient Boosted Trees (GBT) and Deep Neural Networks (DNN) with two hidden layers without hyper parameter tuning. For each of these categories, we consider both a global and per-template models. We use Mean Absolute Ratio Error (MARE) as a metric to evaluate each model’s accuracy. As the runtime distribution varies significantly across different job templates, we normalize the estimation error by the baseline runtime of each job template (see Section \[sec:propstate\]), calculating the average runtime per job template in the training data: $$\label{eq:mare}
\text{MARE} = \dfrac{1}{n}\sum_{i=1}^n |\dfrac{y_{\text{test}_i}-\hat{y}_{\text{test}_i}}{y_{\text{test}_i}^{\beta}}|$$
where $n$ is the number of jobs in the testing data, $y_{\text{test}_i}$, $\hat{y}_{\text{test}_i}$, and $y_{\text{test}_i}^{\beta}$ are the predicted, actual, and baseline runtime from testing data, respectively.
[0.4]{} {width="\textwidth"}
[0.4]{} {width="\textwidth"}
Table \[tab:model\] shows the results of MARE scores for the four model categories. Random forest performs best in terms of accuracy, both for the global and per-template model. Given its high accuracy and interpretability (as discussed in Section \[sec:model\]), RF is the approach we use in our production [Griffin]{}deployments. Moreover, we observe that the global model tracks closely the performance of the per-template model, while allowing to reason about jobs that we have not sufficiently encountered previously. In the next section, we demonstrate that the global model scales much better than the per-template models with an increasing number of templates. Hence, we use the global model in production.
-------------------------------------------------------- -- -- -- --
**& **LR & **RF & **GBT & **DNN\
Per-Template Model & 0.186 & **0.116 & 0.124 & 0.146\
Global Model & 0.235 & **0.121 & 0.277 & 0.353\
**************
-------------------------------------------------------- -- -- -- --
: MARE scores on runtime prediction by LR, RF, GBT, DNN with global and per-template models. Lower MARE is better.[]{data-label="tab:model"}
![Model performance with different training sample sizes[]{data-label="fig:sizeb"}](sample_size.pdf){width="0.9\columnwidth"}
Scalability of Global vs. Per-template Models {#scale}
---------------------------------------------
We assess the scalability of the global and per-template models by training them with an increasing number of job templates, as shown in Figure \[fig:more\_temp\]. We added job templates incrementally to the training set and repeated each experiment 10 times. We use 5-fold cross validation for hyper parameter tuning with random grid search. With more job templates, the training time (TT) and the model size (MS) for both global and per-template models increased. We observe that the MARE for global model is better than per-template models when training with a larger number of job templates. This can be attributed for the most part to the larger sample size for the global model, as a result of the unification of the training set. In contrast, when training the per-template models with 240 job templates, many templates only have a small number of samples and the MARE is high. We also report the prediction time on a single job, i.e. the single inference time (SIT). The SIT didn’t increase with the model size, which is important to deliver a real-time experience to [Griffin]{}’s users. Overall, the above experiments demonstrate the scalability of the global model for cloud-scale training. At the other extreme, a template-specific model suffers from lack of training data and the ability to generalize to new (unseen) templates. As part of our future work, we plan to cluster job templates using unsupervised machine learning methods and train “semi-global” models that take into account multiple job templates that share similar characteristics to strike a balance between the two approaches.
Varying Size of Training Data {#sec:size}
-----------------------------
In order to determine the impact of the number of jobs per template on our model performance, we retrain our global model assuming that we only have a limited number of observations for each job template. In particular, we train our model with $n$ observations per job template, where $n \in \{1, 10, 100, 1000,10000\}$. The MARE and the training time are reported in Figure \[fig:sizeb\]. We used 22 job templates and, for each run of the experiment, a random sample of $n$ observations were selected as the training data. When $n\geq 10$, the MARE dropped below 20%, and the prediction accuracy continues to improve for larger sample sizes, although less significantly. Note that the training time increased exponentially to the sample size.
Model Generalization {#modgeneralize}
--------------------
The baseline approach for interpretation described in this paper allows job runtime prediction results to be interpreted and compared to a set of similar jobs. Data in the real world frequently looks similar to our dataset: Gaussian mixture distributions of a target variable are commonly encountered.
This section presents an example of employing [Griffin]{}on another dataset in another field. A classic dataset from statistics is the “Auto” dataset of gas mileage, which is represented well by a linear superposition of three Gaussians by region of origin: American, European and Japanese. Manufacturers might be interested in understanding what factors drive higher gas mileage in American cars relative to other American cars. Here gas mileage is the equivalent to job runtime.
Table \[tab:granada\] summarizes the delta contribution of each feature (FC) to the high gas mileage of American-made Ford Granada compared to other American cars based on our anomaly reasoning algorithm, described in Section \[sec:reasoning\]. We observe that “Year" and “Horsepower" contribute the most to high gas mileage, while “Weight" and “Displacement" make marginal contributions. “Acceleration" and “Cylinders" contribute to low gas mileage.
Related Work {#sec:related}
============
Anomaly detection [@Chandola2009AnomalyDA] refers to the problem of finding patterns in data that do not conform to expected behavior. In contrast, *anomaly reasoning*, which is the purpose of this work, encompasses recognizing, interpreting, and reacting to unfamiliar objects or familiar objects appearing in unexpected contexts. Anomaly reasoning is particularly important for large-scale systems, as it is not possible to manually track all machines and applications at scale. Below we discuss the main categories of works that are related to [Griffin]{}.
[**Interpretable models**]{} Building an effective anomaly system requires both interpretable models and reasoning algorithms. Several efforts have focused on interpreting the results of machine learning models. Their goal is to provide proper explanation about how or why the algorithm produces a specific prediction and to identify interactions between features and estimation results [@dovsilovic2018explainable; @gunning2017explainable; @samek2017explainable]. In [Griffin]{}, we use an interpretable random forest model to rank a job’s slowdown reasons given the contribution of various features. Similar methods to interpret a tree model can be generalized to boosting algorithms [@welling2016forest].
[**Anomaly reasoning with labeled data**]{} Existing related works have used detection methods such as classification and clustering to perform analysis of anomalies in cloud computing. A detailed survey of those work can be seen in [@agrawal2015survey; @modi2013survey]. For example, a fault detection and isolation system based on k-nearest neighbor has been proposed to rank machines in order of their anomalous behavior [@bhaduri2011detecting]. Other works have used a hybrid of SVM classification and k-medroids clustering to detect intrusions of the network [@chitrakar2012anomaly]. An anomaly-based clustering method has also been suggested to detect failures in general production systems [@duan2009fa].
The downside of these approaches is that they require labeled data. Such data is hard to acquire in many settings, including ours. In the context of an infrastructure that has been operating for many years, labeling data requires infrastructure support and, most importantly, training a large number of engineers to add labels when resolving anomalies. Moreover, it is almost impossible to perform labeling for the years-worth of historical data.
[**Anomaly reasoning with unlabeled data**]{} A few approaches have considered unlabeled data, but focus either on time-series analysis or are restricted to machine or VM metrics [@cherkasova2009automated; @cohen2004correlating; @dean2012ubl; @gu2009online; @tan2010adaptive; @tan2012prepare; @zhang2013intelligent]. For instance, an anomaly detection and reasoning system has been proposed to detect security problems during VM Live Migration [@zhang2013intelligent]. This work is based on time series data related to resource utilization statistics, e.g., file read/write, system call, CPU usage. PREdictive Performance Anomaly pREvention (PREPARE) [@tan2012prepare] predicts performance anomalies using a 2-dependent Markov model and a classifier based on system-level metrics, such as CPU, memory, network traffic. Another related work developed an Unsupervised Behavior Learning (UBL) system to capture the anomalies and infer their causes [@dean2012ubl]. To circumvent the need for labeling data, Self Organizing Map (SOM) has been suggested to model the system behavior, and deviations are used for the anomaly detection [@kohonen2012self]. Similar to [Griffin]{}, those methods estimate the contribution of each attribute to the anomaly and provides information about which system-level metrics to look into. However, they require time-dependent series of data to capture the anomalies. In contrast, we focus on job instances that span several hundreds of machines but only during the lifetime of the job. Thus, our features are neither time series nor machine-centric (although we do employ some system-level data to examine the system’s impact on a particular job’s execution). Other works in anomaly reasoning aim to pinpoint the faulty components of a system by tracing the system’s activities [@aguilera2003performance; @mi2012performance; @mi2011magnifier; @nguyen2011pal]. The methods rely more on the estimation of the time series’ change point and the propagation pattern or the execution graph. However, those methods require significant domain knowledge and are hard to generalize.
To the best of our knowledge, [Griffin]{}is the first anomaly reasoning system to be deployed at this scale in production to identify the causes of job slowdowns in analytics clusters. Unlike existing approaches, it follows a job-centric approach and does not rely neither on labeled data nor on time series analysis.
Conclusion & Future Work {#sec:conclusion}
========================
We presented [Griffin]{}, a system that we built and have deployed in production to detect the causes of job slowdowns in [Microsoft]{}’s big data analytics clusters, consisting of hundreds of thousands of machines. [Griffin]{}does not require labeled data to perform anomaly reasoning. Instead, it uses an interpretable machine learning model to predict the runtime of a job that has experienced a slowdown. Using this model, we can determine the contribution of each feature in the deviation of the job’s runtime compared to previous normal executions of the job (or of jobs with similar characteristics).
Our evaluation results using historical incidents showed that [Griffin]{}discovers the same slowdown reasons that were detected by domain expert engineers. We also compared various categories of models and showed that a global (i.e., trained over all jobs) random forest model strikes a good balance between accuracy, training time, model size, and generalization capabilities. [**Towards data-driven decisions**]{} [Griffin]{}is part of our bigger vision towards employing data-driven decisions to optimize various aspects of our systems. Taking [Griffin]{}’s capabilities a step further, knowing the job slowdown reasons allows us to automatically tune the system to avoid such slowdowns in the future. This may include both system parameters, such as dynamically setting the number of running tasks per machine, and application parameters, such as the degree of parallelism for each stage of a job. Moreover, such parameter autotuning does not have to be constrained to job slowdowns—we can use it to automatically and dynamically set various parameters in our systems to improve their performance.
Furthermore, although [Griffin]{}currently targets our internal analytics clusters, the above techniques can be applied to other environments, such as various public Azure services, including the Azure SQL and HDInsight offerings. Similar data-driven decisions are increasingly applied in various companies [@uberds; @linkedinds].
[^1]: Analytics jobs are executed using Scope, an internal SQL-like distributed query engine that enables processing of petabytes of data per job [@scope].
[^2]: An MLflow flavor is a convention that deployment tools use to understand the model.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Consider the rational map $\phi: \mathbb{P}^{n-1}_{{\mathbf k}}{{\ext@arrow 0359\rightarrowfill@@{}{[f_0:\cdots: f_n]}}} \mathbb{P}^{n}_{{\mathbf k}}$ defined by homogeneous polynomials $f_0,\dots,f_n$ of the same degree $d$ in a polynomial ring $R={{\mathbf k}}[x_1,\dots,x_n]$ over a field ${{\mathbf k}}$. Suppose $I=(f_0,\dots,f_n)$ is a height two perfect ideal satisfying $\mu(I_p)\leq\dim R_p$ for $p\in \operatorname{Spec}(R) \setminus V(x_1,\dots, x_n)$. We study the equations defining the graph of $\phi$ whose coordinate ring is the Rees algebra $R[It]$. We provide new methods to construct these equations using work of Buchsbaum and Eisenbud. Furthermore, for certain classes of ideals satisfying the conditions above, our methods lead to explicit equations defining Rees algebras of the ideals in these classes. These classes of examples are interesting, in that, there are no known methods to compute the defining ideal of the Rees algebra of such ideals. These new methods also give rise to effective criteria to check that $\phi$ is birational onto its image.'
address:
- 'Department of Mathematical Sciences, University of Arkansas, Fayetteville, Arkansas 72701 U.S.A'
- 'Department of Mathematics, University of Virginia, Charlottesville, Virginia 22902 U.S.A'
author:
- Youngsu Kim
- Vivek Mukundan
title: Equations defining certain graphs
---
[*Dedicated to Professor Bernd Ulrich on the occasion of his 65th birthday*]{}
Introduction
============
Our primary goal in this paper is to understand certain rational maps from the projective $n-1$ space to the projective $n$ space. In particular, we will provide an explicit description of the equations defining the image and graph of certain rational maps. Fix a field ${{\mathbf k}}$. Let $X = \mathbb{P}^{n-1}_{{\mathbf k}}$ and $W$ a linear system in $
\operatorname{H}^0 (X, \mathcal{O}_X (d))$ for some positive integer $d$. In general, $W$ does not induce a morphism. However, it defines a rational map from $X$ to $Y := \mathbb{P}^{\dim W - 1}_{{\mathbf k}}$ with the base locus defined by $W$. We denote such a rational map by $\phi: X \stackrel{W}{\dashrightarrow} Y$. The image of $\phi$ may not be closed in $Y$. To study the algebraic properties of the image, we take the closure of the image $\phi(X)$ in $Y$. The questions we are interested in are the following:
\[intQuestion\]
1. What are the equations defining $\overline{\phi(X)}$ in $Y$. More generally, let $\Gamma_\phi$ denotes the closure of the graph of $\phi$. What are the equations defining $\Gamma_\phi$ in $X \times Y$?
2. Is the rational map $\phi$ birational onto its image?
These questions have a tight connection with commutative algebra as their coordinate rings are well-studied rings among algebraists. In the coordinate ring $R = {{\mathbf k}}[x_1,\dots, x_n]$ of $X = \mathbb{P}^{n-1}_{{\mathbf k}}$, let $I$ be the ideal generated by the elements in $W$. The ideal $I$ is a homogeneous ideal generated in degree $d$. The *Rees algebra* $$R[It] := R \oplus I t \oplus I^2 t^2 \oplus \cdots$$ and the *special fiber ring* $$\mathcal{F}(I) := R[It] \otimes {{\mathbf k}}\cong R/ {\mathfrak{m}}\cong R[It]/{\mathfrak{m}}R[It],$$ where ${\mathfrak{m}}$ is the maximal ideal $(x_1,\dots, x_n)$ are the coordinate rings of the projective varieties $\Gamma_\phi$ in $X \times Y$ and $\overline{\phi(X)}$ in $Y$, respectively. Hence, the answer to (a) is nothing but the equations in the defining ideals of these rings (see for details). In the case where ideal $I$ is *nice*, for instance, if $W = \mathcal{O}_X (d)$ the d-th Veronese embedding or $I$ is a complete intersection ideal, the defining ideal of the Rees algebra of $I$ is well-understood, cf. [@SH Sec. 5.5]. However, determining defining ideals of Rees algebras is a challenging task, and there is a series of work on this topic.\
Our main case concerns perfect ideals of codimension (equivalently height) two. Such ideals satisfy the structure theorem of Burch, and a number of papers has been devoted to understanding this class of ideals; for instance [@HSV83; @Mor96; @MU96; @Ha02; @HSV08; @CHW08; @Bus09; @KPU11; @CDA13; @CDA14; @Lan14; @Mad15; @BM16; @KPU17]. The novelty in our approach is to incorporate the work of Buchsbaum and Eisenbud [@BE74], in particular, Buchsbaum-Eisenbud multipliers.\
We work on in the following setup. Let $W \subset \operatorname{H}^0(\mathbb{P}^{n-1}, \mathcal{O}_{\mathbb{P}^{n-1}} (d))$ be a subsystem. Assume that $\dim_{{{\mathbf k}}} W = n+1$ (so $\phi: \mathbb{P}^{n-1}_{{\mathbf k}}\stackrel{W}{\dashrightarrow} \mathbb{P}^{n}_{{\mathbf k}}$), $\operatorname{codim} \operatorname{Fitt}_i (I) \ge i+1$ for $1 \le i \le n-1$, and $\operatorname{Proj}R/I$ is arithmetically Cohen-Macaulay of codimension $2$. Here, $\operatorname{Fitt}_i (I)$ denotes the $i$th Fitting ideal of $I$. The condition that $\operatorname{Proj}R/I$ is arithmetically Cohen-Macaulay of codimension 2 is equivalent to the condition of the base locus being defined by a height $2$ perfect ideal $I$. The *defining ideal* of the Rees algebra $R[It]$ is the kernel of the surjective $R$-algebra homomorphism $$\label{introPresentationB}
B := R \otimes_{{{\mathbf k}}} S = R[T_0,\dots, T_n] \to R[It],$$ where $S = {{\mathbf k}}[T_0,\dots, T_n]$. This kernel a bigraded ideal in $B$. We often study the defining ideal of $R[It]$ in $\operatorname{Sym}_R(I)$, the symmetric algebra of $I$, since the presentation in \[introPresentationB\] factors through $\operatorname{Sym}_R(I)$, and the kernel of this induced presentation is well-understood, see . By abuse of notation, we often call the kernel of the induced map $\operatorname{Sym}_R(I) \to R[It]$ the *defining ideal* of the Rees algebra $R[It]$.\
Under this setup, $\operatorname{Sym}_R(I)$ is a complete intersection ring, and the defining ideal of the Rees algebra $R[It]$ in $\operatorname{Sym}_R(I)$ is equal to $\mathcal{A} := H^0_{\mathfrak{m}}(\operatorname{Sym}_R(I))$, the zeroth local cohomology module supported in the maximal ideal ${\mathfrak{m}}= (x_1, \dots, x_n)R$. These facts allow us to take the advantage of the bigraded structure of $\mathcal{A}$ and the duality theorem (). In the sequel, for a bigraded module $M = \oplus_{i,j \in \mathbb{Z}} M_{i,j}$ in $B$, we set $M_i := \oplus_{j \in \mathbb{Z}} M_{i,j}$.
\[introDualThm\] For each $i \in \{ 0,\dots, d-n \}$, there exists an isomorphism of finitely generated graded $S$-modules $$\mathcal{A}_i\cong \operatorname{Hom}_S (\operatorname{Sym}(I)_{d-n-i},S(-n)).$$
In [@Jou97], Jouanolou constructed the isomorphisms in explicitly. Hence to understand $\mathcal{A}_i$ it suffices to study its dual $\operatorname{Hom}_S (\operatorname{Sym}(I)_{d-n-i},S(-n))$. Our first result concerns this Hom module. Fix $i \in \{0,\dots, d-n \}$ and let $$\label{intPres}
F_1 \stackrel{\alpha}{\to} F_0 \to [\operatorname{Sym}_R(I)]_{d-n-i}$$ be a graded presentation of $[\operatorname{Sym}_R(I)]_{d-n-i}$. Then $\operatorname{Hom}_S (\operatorname{Sym}(I)_{d-n-i}, S(-n))$ corresponds to the kernel of $\alpha^*$. In , we show that there exists a complex which induces elements in $\ker \alpha^*$.
We have a complex of graded free $S$-modules $$\label{eqIntro}
\wedge^{r-1}F_0\otimes S(-s_1)\stackrel{\partial}\to F_0^*\xrightarrow{\alpha^*}F_1^*,$$ where $r = \operatorname{rank}[\operatorname{Sym}_R (I)]_{d-n-i}$, i.e., $\operatorname{Im} (\partial)(-n) \subset \ker \alpha^*(-n) \cong \mathcal{A}_{i}$.
We note that the above is a complex for any degree of $\mathcal{A}_i$ (equivalently, $[\operatorname{Sym}_R (I)]_{d-n-i}$). It is natural to ask under what conditions the complex in \[eqIntro\] is exact. We show that for $r \le 2$, this complex is exact provided that $[\operatorname{Sym}_R(I)]_{d-n-i}$ satisfies Serre’s condition $(S_{r})$ (), and we ask that whether this holds true in general (). Our theorems provide the differential map $\partial$ and the shift $s_1$ in \[eqIntro\] explicitly. To do this we use a structure theorem of Buchsbaum-Eisenbud [@BE74], and their main theorem and lemmas are our main technical tools in this paper.\
Our next theorem concerns (b). It is a well-known fact that the closed image $\overline{\phi(X)}$ is defined by a single equation, for instance, see [@UV Proposition 2.4]. Hence, it suffices to study $$\mathcal{A}_0 \cong \operatorname{Hom}_S (\operatorname{Sym}(I)_{d-n},S(-n)) = \ker \alpha^*(-n),$$ where $\alpha^*$ as in \[eqIntro\]. We provide equivalent conditions that the complex in Theorem A is exact. The equivalence $(1)$ and $(4)$ in Theorem B below was first established in [[@KPU16 Corollary 3.7]]{}.
Assume ${{\mathbf k}}$ is a field of characteristic zero and $n \ge 3$. Then with $s_1$ as in \[eqIntro\], the generating degree of $\mathcal{A}_0$ is at most $s_1 + n$. Furthermore, the following statements are equivalent:
1. The rational map $\phi$ is birational to its image.
2. $\mathcal{A}_0$ is generated in degree $s_1 + n$.
3. The greatest common divisor of the entries of $\ker \alpha^*$ is 1.
4. $e(\mathcal{F}(I))= e ({R}/{(g_0,\dots,g_{n-2}):I})$, where $g_0, \dots, g_{n-2}$ are general ${{\mathbf k}}$-linear combinations of the $f_i'$s (see for the definition of the term general). Here, $e(-)$ denotes the Hilbert-Samuel multiplicity.
The explicit isomorphism of Jouanolou provides an explicit form of the defining equation from $\ker \alpha^*$, and there are a few other ways to treat this case. In [@BuseChardinJouanolou], Busé, Chardin, and Jouanolou achieved this by analyzing $[\operatorname{Sym}_R(I)]_q$ for $q \gg 0$ and using the determinant of a free resolution of $[\operatorname{Sym}_R(I)]_q$. We note that as $q$ becomes larger, the size of the presentation matrix of $[\operatorname{Sym}_R(I)]_q$ grows in the binomial order of $d = \dim R$. With our approach, one only needs to analyze the presentation matrix of $[\operatorname{Sym}_R(I)]_{d-n}$. As far as item (1) in the above theorem is concerned, one may apply a theorem of Doria, Hassanzadeh, and Simis [@DHS Section 2.3]. Their approach is more general, but it requires understanding an additional graded piece $\mathcal{A}_{1}$ of the defining ideal and its Jacobian dual. Lastly, we compare ours with a result by Boswell and Mukundan [@BM16], where the authors use an iterative Jacobian dual. This iteration involves computing large size matrices, and it is computationally not as efficient as our approach. However, their statement provides a closed formula for $\mathcal{A}$ in terms of colon ideals.\
One of the advantages of our approach is that every condition we impose is general, for instance see [@BJ03 p. 316] and . By the semi-continuity theorem [@Eis Thm 14.8b], one can see that the exactness of the complex in \[eqIntro\] is a general condition. That is, there exists an open subset $U$, which may be empty, in a parameter space such that the fiber of each point of $U$ satisfies this condition. The difficult part is to show that such an open subset $U$ is non-empty. In other words, one needs to exhibit an example whose corresponding point belongs to $U$. This turned out to be the most challenging part of our paper. We believe that for $d_1, \dots, d_n \in \mathbb{N}$ and a presentation matrix $$\varphi = \begin{bmatrix}
x_1^{d_1} & x_1^{d_2} & \cdots & x_1^{d_n} \\
x_2^{d_1} & x_2^{d_2} & \cdots & x_2^{d_n} \\
\vdots & \vdots & \ddots & \vdots \\
x_n^{d_1} & x_n^{d_n} & \cdots & x_n^{d_n} \\
g_1 & g_2 & \cdots & g_n
\end{bmatrix},$$ where $g_i$’s are symmetric polynomials of degree $d_i$, the ideal generated by the $n \times n$ minors of $\varphi$ would provide a family of examples. As an evidence, we show that a variant of this with $d_1 = 1, d_2 = 2, d_3 \ge 3$ provides a necessary example (). This allows us to state our last main result in this paper.
\[thmC\] Let $R = {{\mathbf k}}[x,y,z]$, where ${{\mathbf k}}$ is a field of characteristic $0$. Let $M$ be a $4 \times 3$ matrix whose entries are in $R$ and let $I = I_3(M)$. If $M$ is general of type $(1,2, {{\mathbf q}})$, where ${{\mathbf q}}> 2$ (see Case 2 in ), then the defining ideal of the Rees algebra $R[It]$ is minimally generated by $$\def\arraystretch{1.2}
\begin{array}{c;{2pt/2pt}c|c}
\text{~} & \text{bidegree} & \text{number of elements} \\
\hline
l_1 & (1,1) & 1 \\ \hline
l_2 & (2,1) & 1 \\ \hline
l_3 & ({{\mathbf q}},1) & 1 \\ \hline
\mathcal{A}_0 & (0, 3{{\mathbf q}}+2 ) & 1 \\ \hline
\mathcal{A}_{{{\mathbf q}}- i} & ({{\mathbf q}}- i, 3i+1 )& {2 + i \choose 2} \\ \hline
\mathcal{A}_{{\mathbf q}}& ({{\mathbf q}}, 3) & 1
\end{array}$$ for $1 \le i < {{\mathbf q}}-1$. In particular, the defining ideal is minimally generated by ${{{\mathbf q}}+2 \choose 3} + 4$ elements. Here, $l_1,l_2,l_3$ denote the equations defining $\operatorname{Sym}_R(I)$ in $B$, see .
The paper is organized as follows. In , we set up the notation and provide preliminaries. In , we explain our setup mentioned in the introduction, the work of Buchsbaum and Eisenbud, and the duality theorem in detail. In , we prove Theorem A and related statements. is devoted to the equivalence in Theorem B. In , we present Theorem C and the promised example.\
[**Acknowledgment:**]{} We would like to thank professors Bernd Ulrich and Craig Huneke for their helpful comments and suggestions. Also, we owe a lot to the anonymous referee for his/her invaluable comments. The earlier version of the paper had an erroneous definition of the notion of general elements which was kindly pointed out to us by the referee.
Preliminaries {#secPrel}
=============
In this section, we will setup the notation and review some background materials. We refer the reader to [@Eis; @Har] for basic definitions and notations for algebraic geometry and commutative algebra.
Morphisms between projective spaces and Rees algebras
-----------------------------------------------------
Let ${{\mathbf k}}$ be a field, ${{\mathbf k}}[x_1,\dots, x_n]$, the coordinate ring for $\mathbb{P}^{n-1}_{{\mathbf k}}$, and $f_0,\dots,f_n$ homogeneous polynomials of degree $d$ in ${{\mathbf k}}[x_1,\dots, x_n]$. Then we have a rational map between projective spaces $$\label{rationalMap}
\phi: \mathbb{P}^{n-1}_{{\mathbf k}}{{\ext@arrow 0359\rightarrowfill@@{}{[f_0:\cdots: f_n]}}} \mathbb{P}^{n}_{{\mathbf k}}$$ defined by the polynomials $f_0,\dots, f_n$. This map is defined on $\mathbb{P}^{n-1}_{{\mathbf k}}\setminus V(I)$ (equivalently, the base locus is $V(I)$). The image $\operatorname{Im}\phi$ is not a closed subscheme of $ \mathbb{P}^{n}_{{\mathbf k}}$ in general. In this article, we study the closed subscheme $\overline{\operatorname{Im}\phi}$ in $\mathbb{P}^{n}_{{\mathbf k}}$. Let ${{\mathbf k}}[y_0,\dots, y_n]$ be the coordinate ring of $\mathbb{P}^{n}_{{\mathbf k}}$. Then $\overline{\operatorname{Im}\phi} = \operatorname{Proj}{{\mathbf k}}[f_0,\dots, f_n]$, and the (rational) maps between projective schemes $$\mathbb{P}^{n-1}_{{\mathbf k}}\stackrel{\phi}{\dashrightarrow} \overline{\operatorname{Im}\phi} \subset \mathbb{P}^{n}_{{\mathbf k}}$$ corresponds to the maps between $k$-algebras $${{\mathbf k}}[y_0,\dots,y_n] \twoheadrightarrow {{\mathbf k}}[f_0,\dots,f_n] \subset {{\mathbf k}}[x_1,\dots,x_n],$$ where the first map is defined by $y_i \mapsto f_i$ for $i = 0, \dots, n$. Observe that $\overline{\operatorname{Im}\phi} = \operatorname{Proj}{{\mathbf k}}[f_0,\dots, f_n] = V(J)$ for some homogeneous ideal $J$ of ${{\mathbf k}}[y_0,\dots, y_n]$[^1].\
Let $R = {{\mathbf k}}[x_1,\dots, x_n]$, ${\mathfrak{m}}= (x_1,\dots, x_n)R$, and $I = (f_0,\dots, f_n)$. The *Rees algebra* of $I$ is the graded ring $$R[It] = R \oplus I t \oplus I^2 t^2 \oplus \cdots \subset R[t],$$ and *the special fiber ring* of $I$ is the graded ring $$\mathcal{F}(I) = R[It] \otimes_R R/ {\mathfrak{m}}= R[It]/{\mathfrak{m}}R[It].$$
The Rees algebra $R[It]$ is the coordinate ring of the closure of the graph of the rational map in \[rationalMap\]. Here, the *graph* of a rational map $\phi: X \dashrightarrow Y$ with base locus $W$ between projective schemes is $\Gamma_{\phi} := \{ (x,y) \in X \times Y \mid y = \phi(x), x \not\in W \}$. In addition, if $\deg f_i = d$ for $i = 0,\dots, n$, then the special fiber ring $\mathcal{F}(I)$ is an integral domain and is isomorphic to the subring ${{\mathbf k}}[f_0,\dots, f_n]$ of ${{\mathbf k}}[x_1,\dots, x_n]$.\
Let $\operatorname{Quot}(A)$ denote the total ring of fractions of a ring $A$. In the case where $A$ is an integral domain, $\operatorname{Quot}(A)$ is the field of fractions. For fields $F \subset K$, let $[K : F]$ denote the field extension degree.
\[birationalprelim\] The rational map in \[rationalMap\] is birational to the image if and only if ${{\mathbf k}}[f_0,\dots, f_n]$ and ${{\mathbf k}}[x_1,\dots, x_n]$ have the same field of fractions.
Free resolutions and minors of matrices
---------------------------------------
Let $R$ be a Noetherian ring. In this subsection, we review two theorems of Buchsbaum-Eisenbud on finite free complexes. One provides a characterization of the acyclicity of a finite free (graded) $R$-complex, and the other one provides a structure theorem for a finite free (graded) acyclic $R$-resolution.\
Let $\varphi: F \to G$ be a map between finite free $R$-modules of rank $f$ and $g$, respectively. Once we fix ordered bases for $F$ and $G$, we obtain a matrix representation $M$ of $\varphi$, which is an $g \times f$ matrix with entries in $R$. For an $m \times n$ matrix $N$ with entries in $R$, let $I_t(N)$ be the ideal generated by $t \times t$ minors of $N$ if $1 \le t \le \min \{ m, n\}$, and we set $I_0(N) = R$ and $I_t (N) = 0$ if $t > \min \{ m,n\}$. By abuse of notation, let $I_t(\varphi)$ denote $I_t(M)$, where $M$ is a matrix representation for $\varphi$. Matrix representations depend on the choice of bases. However, the ideal $I_t(M)$ does not depend on the choice of bases. The *rank* of $\varphi$, denoted $\operatorname{rk}\varphi$, is the number $t$ where $I_{t+1}(\varphi) = 0$, but $I_t(\varphi) \neq 0$, and we set $I (\varphi) = I_{\operatorname{rk}\varphi} (\varphi)$. For a proper ideal $I$ of $R$, the grade of $I$, denoted $\operatorname{grade}I$ (equivalently, the depth of $I$ in $R$ denoted by $\operatorname{depth}_I R$), is the length of a maximal $R$ regular sequence contained in $I$. It is well-known that the maximal length is independent of regular sequences, cf. [@BH Def. 1.2.11].
\[thmBE73\] Let $\mathcal{C}_\bullet$ be a finite complex of free $R$-modules of finite rank $$0 \to F_n \stackrel{\varphi_n}{\to} \cdots \stackrel{\varphi_2}{\to} F_1 \stackrel{\varphi_1}{\to} F_0.$$ Then $\mathcal{C}_\bullet$ is acyclic if and only if for $k = 1,\dots, n$,
1. $\operatorname{rk}F_k = \operatorname{rk}\varphi_{k+1} + \operatorname{rk}\varphi_{k}$ and
2. $\operatorname{grade}I(\varphi_k) \ge k$ or $I(\varphi_k) = R$.
The map $\varphi : F \to G$ induces maps between exterior powers $\wedge^k \varphi: \wedge^k F \to \wedge^k G$ for any $k$. We also denote the image of $\wedge^k \varphi$ by $I_k(\varphi)$. (Once we fix bases for $F,G$ and a matrix representation $M$ for $\varphi$, the matrix representation of $\wedge^k \varphi$ is the $k$-minors of $M$ (up to sign). Therefore, the image of $\wedge^k \varphi$ in $R$ is $I_k(\varphi)$.) For $F$ a free module of rank $f$, an isomorphism $\eta: \wedge^f F \to R$ is called an *orientation* of $F$. We say a finite free module is *oriented* if it is equipped with an orientation. For an oriented finite free module $F$, we have $ \wedge^k F \otimes \wedge^{f-k} F \to \wedge^f F \stackrel{\eta}{\to} R$. Hence, we identify $(\wedge^k F)^*$ with $\wedge^{f-k} F$ for oriented free modules. Here, for any $R$-module $L$, $L^* := \operatorname{Hom}_R(L,R)$ denotes the $R$-dual of $L$.
\[BEstatementprelim\] Consider a finite free acyclic $R$-complex\[BEprelim\] $$0 \to F_n \stackrel{\varphi_n}{\to} \cdots \stackrel{\varphi_2}{\to} F_1 \stackrel{\varphi_1}{\to} F_0.$$ Write $r_i = \operatorname{rk}\varphi_i$. For $k = 1, \dots, n$, there exists unique $R$-homomorphism $a_k : R \to \wedge^{r_k} F_{k-1}$ such that
1. $a_n := \wedge^{r_n}\varphi_n: R = \wedge^{r_n} F_n \to \wedge^{r_n} F_{n-1}$, and
2. for $k < n$, the diagram $$\xymatrix{
\wedge^{r_k} F_k \ar[rr]^{\wedge^{r_k} \varphi_k} \ar[dr]_{a_{k+1}^*}& & \wedge^{r_{k}} F_{k-1} \\
& R \ar[ur]_{a_k}
}$$ commutes.
3. For all $k>1$, $\sqrt{I(a_k)}=\sqrt{I(\varphi_k)}$.
General property {#secGeneral}
----------------
In this subsection, we recall the notion of a general property. We will follow the section “general object” in [@Har92]. Let $X$ be a variety (or a scheme) parametrized by (closed) points in an (irreducible and reduced) variety $Y$ and $\mathcal P$ a property on $X$. We say that $\mathcal P$ is *general* or a *general property* with respect the pair $X$ and $Y$, if the set $\{ p \in Y \mid \text{object parametrized by} p~\text{satisfies}~\mathcal{P} \} \subset Y$ is a dense open subset. If $\mathcal{P}$ is a general property and $y \in Y$ satisfies $\mathcal{P}$, then $y$ or the object of $X$ parameterized by $y$ is called *general* or a *general member* with respect to $\mathcal{P}$. In the sequel, whenever we use the phrase an object $G$ is general or a general member, it is understood that it refers to a general property for a pair $X$ and $Y$, and $G$ a general member with respect to this general property.\
We will use this notion of a general property in the following setup. Let ${{\mathbf k}}$ be a field of characteristic zero, $R = {{\mathbf k}}[x_1, \dots, x_n]$ a polynomial ring, and $A$ another polynomial ring over ${{\mathbf k}}$. Further, let $B = A \otimes_{{{\mathbf k}}} R \cong A[x_1,\dots,x_n]$ and $J$ a homogeneous ideal of $B$ which does not contain any element of $A$ other than zero, i.e., $J \cap A = 0$. Here, we set $X = \operatorname{Proj}_Y B/J$, where $Y = \operatorname{Spec}A$. Our parameter space $Y$ will be always affine space over a field of characteristic zero, and we will consider only closed points of $Y$. By abuse of terminology, we say an $R$-ideal $I$ is *general* or *a general member* with respect to some general property $\mathcal{P}$ if $\operatorname{Proj}_{{\mathbf k}}R/I$ is a general member with respect to $\mathcal{P}$ for the pair $X$ and $Y$. We also say a sequence of elements $g_1, \dots, g_s$ of $R$ (or a matrix $M$ whose entries are in $R$) is *general* or *a general member* with respect to some general property $\mathcal{P}$ if the ideal $(g_1,\dots, g_s)$ (or $I_s(M)$ for some fixed $s$) is a general member with respect to $\mathcal{P}$. For the sake of completeness, we list the setups of two cases in detail. We will use Case 1 in and Case 2 for and in its proof.\
**Case 1:**\[defGenEle\] Let $R ={{\mathbf k}}[x_1,\dots, x_n]$ be a polynomial ring over a field of characteristic zero ${{\mathbf k}}$ and $f_1,\dots, f_l$ homogeneous polynomials of degree $d \ge 1$ in $R$. Fix a positive integer $s$. Consider the parameter space $A := {{\mathbf k}}[ u_{ij} \mid 1 \le i \le l, 1 \le j \le s]$, and for $1 \le j \le s$, set $$F_j := u_{1j} f_1 + u_{2j} f_2 + \cdots + u_{lj} f_l,$$ in $B := A \otimes_{{{\mathbf k}}} R \cong A[x_1,\dots,x_n]$. Write $J = (F_1,\dots, F_s) \subset B$. Here $Y = \operatorname{Spec}A$ and $X = \operatorname{Proj}_Y B/J$. A general member satisfying some property with respect to the pair $X$ and $Y$ is also often called as *$s$-general elements* or *$s$-general ${{\mathbf k}}$-linear combination* of $f_1,\dots, f_l$.\
**Case 2:**\[defGenMat\] $R = {{\mathbf k}}[x_1, \dots, x_n]$ be a polynomial ring over a field of characteristic zero ${{\mathbf k}}$. Fix positive integers $m,l$ and $d_1, \dots, d_l$. For $1 \le j \le l$, let $h_j$ denote the number of monomials of $R$ of degree $d_j$, and define $h := h_1 + \dots + h_l$. Our parameter space is $A = \otimes_{{{\mathbf k}}} \, A_{ij}$, where $A_{ij} = {{\mathbf k}}[ u_{ij,k} \mid 1 \le k \le h_1 ]$ for $1 \le i \le m, 1 \le j \le l$. In $B := A \otimes_{{{\mathbf k}}} R = A[x_1,\dots,x_n]$, consider the $m$ by $l$ matrix ${M}^\sim$ whose $(i,j)$-entry is $$u_{ij,1} m_{d_j,1} + u_{ij,2} m_{d_j,2} + \cdots + u_{ij,h_j} m_{d_j,h_j},$$ where $\{m_{d_j,1}, \dots, m_{d_j,h_j} \}$ is a fixed monomial basis of $R$ of degree $d_j$. Let $J = I_s (M^\sim)$, where $s$ is an integer. Here, $X = \operatorname{Proj}_Y B/J$ with $Y = \operatorname{Spec}A$. We say that an $m \times l$ matrix $M$ with entries in $R$ is *general of type $(d_1, \dots, d_l)$* if for some $s$, $I_s(M)$ is a general member for some general property with respect to this pair $X$ and $Y$.\
We apply the general property in Case 2 to the height of $I_s(M)$ for some integer $s$. The open subset defining such a general property corresponds to the complement of the closed subset defined by the ideal $I_e$ in (b) (in our notation, $S = B/J$), provided that such complement is non-empty. We will use this correspondence in the proof of the main theorem ().
\[corGeneralHeight\] Suppose that $A$ is a Noetherian ring and $S = S_0 \oplus S_1 \oplus \cdots$ is a positively graded ring which is finitely generated over $A = S_0$. Then we have the following statements.
1. [[@Eis Theorem 14.8(b)]]{} For any integer $e$, there exists an $A$-ideal $I_e$ (depending on $e$) such that for any maximal ideal ${\mathfrak{m}}$ of $A$, $$\label{labUSC}
\dim A/{\mathfrak{m}}\otimes_A S \ge e~\quad {\text{if and only if}}~\quad {\mathfrak{m}}\supset I_e.$$
2. Write $S = A[x_1,\dots, x_n]/J$. For any integer $e$, there exists an $A$-ideal $I_e$ (depending on $e$) such that for any maximal ideal ${\mathfrak{m}}$ of $A$, $$\label{labUSC}
\operatorname{ht}J (A/{\mathfrak{m}}) [x_1,\dots, x_n] \le n-e~\quad {\text{if and only if}}~\quad {\mathfrak{m}}\supset I_e.$$
We show part (b). Fix a maximal ideal ${\mathfrak{m}}$ of $A$ and write $T = A/{\mathfrak{m}}\otimes_A A[x_1,\dots,x_n] = (A/ {\mathfrak{m}})[x_1,\dots, x_n]$. Notice that $T/JT = A/ {\mathfrak{m}}\otimes_A S$. Now, part (b) follows from the first part and the following identity ([@Har Theorem 1.8A]) $$\operatorname{ht}J T + \dim T/ J T = \dim T (= n).$$
We note that [[@Eis Theorem 14.8(b)]]{} is stated for prime ideals of $A$. In this paper, we only consider closed points of a parameter space, so we stated part (a) for maximal ideals.
1. We list examples of general properties.
- For an ideal $I = (f_1,\dots,f_l)$ of grade $\ge s$, the property of the grade of the ideal generated by $s$-general linear combinations of $f_1,\dots,f_l$ being at least $s$ is a general property.
- For an ideal $I = (f_1,\dots,f_l)$ minimally generated by the $f_i$’s and $s \le l$, the condition that $s$-general linear combinations of $f_1,\dots,f_l$ are part of minimal generating set is a general property.
2. An advantage of having a general property is that one may ask finitely many general properties simultaneously since a finite intersection of non-empty dense open subsets remains non-empty dense open.
3. For those who are familiar with algebraic geometry, generic freeness and generic smoothness of $\mathbb{C}$-varieties are examples of general properties [@Eis Theorem 14.4].
Defining ideals of Rees algebras and special fiber rings {#setupSec}
========================================================
Let $R = {{\mathbf k}}[x_1,\dots, x_n]$ be a polynomial ring in $n$ variables over a field ${{\mathbf k}}$, ${\mathfrak{m}}= ( x_1,\dots, x_n)R$ the homogeneous maximal ideal, $f_0,\dots, f_n$ homogeneous polynomials of degree $d$ in $R$, and $I = (f_0,\dots, f_n)$. Consider a homogeneous surjective $R$-linear map $\pi$ $$\label{piMapDefEqReesAlg}
\pi: R[T_0,\dots,T_n] \to R[It],$$ where $T_i \mapsto f_i t$ for $0 \le i \le n$. Since $\mathcal{F}(I) = R[It] / {\mathfrak{m}}R[It]$, the map $\pi$ induces a surjective ${{\mathbf k}}$-linear map $\overline{\pi}$ for $\mathcal{F}(I)$ $$\overline{\pi}: {{\mathbf k}}[T_0,\dots,T_n] \to \mathcal{F}(I).$$
The kernel of $\pi$ is called the *defining ideal* of the Rees algebra of $I$, and the kernel of $\overline{\pi}$ is called the *defining ideal* of the special fiber ring $\mathcal{F}(I)$. Recall that $f_0, \dots, f_n$ define a rational map $$\label{parametrizationmap}
\phi: \operatorname{Proj}(R) = {{\mathbb P}_{{\mathbf k}}}^{n-1} {{\ext@arrow 0359\rightarrowfill@@{}{[f_0:\cdots: f_n]}}} \mathbb{P}^{n}_{{\mathbf k}},$$ $V(\ker \pi)$ defines the closure of the graph of $\phi$ in ${{\mathbb P}_{{\mathbf k}}}^{n-1} \times {{\mathbb P}_{{\mathbf k}}}^{n}$, and $V(\ker \overline{\pi})$ defines the closure of the image of $\phi$ in $ {{\mathbb P}_{{\mathbf k}}}^n$.\
The study of $\ker \pi$ can be simplified via the symmetric algebra of $I$. For an ideal $I$, the *symmetric algebra* of $I$ is the graded ring $$\operatorname{Sym}_R(I) := R \oplus I \oplus \operatorname{Sym}_R^2(I) \oplus \cdots,$$ where $\operatorname{Sym}_R^k(I)$ denotes the $k$th symmetric power of $I$. By the universal property of $\operatorname{Sym}_R(I)$, the homogeneous presentation $\pi$ factors through $\operatorname{Sym}_R(I)$ $$\label{presSym}
\xymatrix{
R[T_0,\dots, T_n] \ar[rr]^{\pi} \ar[rd]_{\pi'} && R[It]\\
& \operatorname{Sym}_R(I) \ar[ru]_{\pi''}.
}$$ The kernel of $\pi'$ is easy to describe from the graded presentation matrix of $I$: Consider a homogeneous presentation $\varphi$ of the ideal $I$ $$\label{presI}
\oplus^m R(-d-d_i) \stackrel{\varphi}{\to} R(-d)^{n+1} \to I \to 0,$$ where $d_i$ are positive integers. Notice that $\varphi$ is a $n+1 \times m$ matrix. Then $\ker \pi' = I_1 ( [T_0 \, \dots \, T_n] \, \varphi)$. Hence by abuse of notation, we call $\mathcal{A} = (\ker \pi) \operatorname{Sym}_R(I)$ the defining ideal of the Rees algebra.\
Since $R$ is a graded ring, $R[T_0,\dots, T_n]$ has a natural bi-graded structure; we set $\deg x_i = (1,0)$ and $\deg T_i = (0,1)$. Hence $\ker \pi$ is a bi-graded ideal of $R[T_0,\dots, T_n]$, and $\ker \overline{\pi}$ is a homogeneous ideal of ${{\mathbf k}}[T_0,\dots, T_n]$. Write $B = R[T_0,\dots, T_n]$ and $S = {{\mathbf k}}[T_0,\dots, T_n]$. It is often convenient to view $B$ as $B \cong R \otimes_{{\mathbf k}}S = {{\mathbf k}}[x_1, \dots, x_n ] \otimes_{{\mathbf k}}{{\mathbf k}}[T_0,\dots, T_n]$. Hence $B$ is free as an $S$-module. For a graded module $M$, $M(a)$ denotes the grade shift by $a$. That is $[M(a)]_{i} = M_{a+i}$ for any $i \in \mathbb{Z}$. Similarly, for a bigraded module $M$, $M(a,b)$ means $[M(a,b)]_{(i,j)} = M_{(a+i, b+j)}$ for any $i,j \in \mathbb{Z}$.\
Our main theorems are stated in the same hypothesis of the following proposition. The statement (as well as its proof) is well-known. We present this proposition to fix the notation, and we will use it as a quick reference for our setup.
\[setupProp\] Let $R = {{\mathbf k}}[x_1,\dots, x_n]$ be a polynomial ring in $n$ variables over a field ${{\mathbf k}}$, and $I$ an $R$-ideal, and ${\mathfrak{m}}= (x_1,\dots, x_n)R$. Assume that $I$ is codimension $2$ perfect, that the degrees of the entries of the columns of a presentation matrix $\varphi$ of $I$ in \[presI\] are $d_1 \le d_2 \le \cdots \le d_n$, and that $\mu(I_p) \le \dim
R_p$ for all $p \in \operatorname{Spec}(R) \setminus \{{\mathfrak{m}}\}$. Then we have the following:
1. The complex $0 \to \oplus^n R(-d_i) \stackrel{\varphi}{\to} R^{n+1} \to I (d)\to 0$ is exact.
2. $I$ is generated in degree $d = \Sigma^n d_i$.
3. $\operatorname{Sym}_I (R)$ is a complete intersection.
4. $ \mathcal{A} = H^0_{\mathfrak{m}}(\operatorname{Sym}_I (R)),$ where $H^0_{\mathfrak{m}}(\operatorname{Sym}_I (R))$ denotes the 0-th local cohomology module of $\operatorname{Sym}_I(R)$ with support in ${\mathfrak{m}}$.
$(1)$ and $(2)$ follow from the Hilbert-Burch theorem [@Eis Theorem 20.15] since $I$ is a codimension two perfect graded ideal, $(3)$ follows from the fact that $\dim \operatorname{Sym}_R(I) = n + 1$ and $\ker \pi'$ in \[presSym\] is generated by $n$ elements, and $(4)$ follows from [@HSV1 Theorem 2.6].
\[Gs\] For a ring and an ideal $I$, we say that $I$ satisfies the condition $(G_s)$ if $\mu(I_p) \le \dim
R_p$ for all $p \in \operatorname{Spec}(R)$ such that $\dim R_p \lneq s$. Hence the ideal $I$ in satisfies $(G_{\dim R})$ condition. There are plenty of ideals satisfying this condition, e.g., complete intersection ideals, (homogeneous) ideals primary to the homogeneous maximal ideal, and this condition can be checked with the Fitting ideals of $I$ (that is $I$ satisfies $(G_s) \iff$ $\operatorname{ht}\operatorname{F}_i(I) \ge i + 1$ for $1 \le i \le s-1$).
\[pieceRes\] With the notation and hypothesis of \[setupProp\], let $l_1, \dots, l_n$ be in $B = R[T_0,\dots, T_n]$ such that $[ \, l_1 \, \dots \, l_n] = [\, T_0 \, \dots \, T_n] \, \varphi$; hence $\operatorname{Sym}_R(I) = B / (l_1, \dots, l_n)$, and $l_1, \dots, l_n$ form a bi-homogeneous regular sequence in $B$. Then the Koszul complex $\mathcal{K}:= \mathcal{K} ( \underline{l_i} ; B)$ is a bi-graded $B$-resolution for $\operatorname{Sym}_R(I)$; $$\mathcal{K}: 0 \to B( -\Sigma^n d_i, -n) \to \cdots \to \oplus^n B(-d_i, -1) \to B.$$
Since $B$ is a free graded $S = k[T_0,\dots, T_n]$-module, $\mathcal{K}$ is a graded free $S$-resolution of $\operatorname{Sym}_R(I)$. Each component of $\mathcal{K}$ is not of finite rank as an $S$-module. However, for each $k \in \mathbb{Z}$, $\mathcal{K}_k := \mathcal{K}_{(k,*)}$ is a finite free graded $S$-resolution for $[\operatorname{Sym}_R(I)]_{(k,*)}$, and each component of $\mathcal{K}_k$ is of finite rank as $S$-module; $$\label{koszulSk}
\mathcal{K}_k : 0 \to F_n \stackrel{}{\to} \cdots \to F_1 \stackrel{}{\to} F_0,$$ where $F_i=\bigoplus_{1\leq j_1\leq j_2\leq\cdots\leq j_i\leq n}S(-i)^{k-(d_{j_1}+\cdots+d_{j_i})+n-1\choose n-1}$. Furthermore, these $S$-resolutions are linear resolutions, i.e., the non-zero entries of the differential maps are of degree 1.\
Recall the notation that for a bigraded module $M = \oplus_{i,j \in \mathbb{Z}} M_{i,j}$ in $B$, $M_i := \oplus_{j \in \mathbb{Z}} M_{i,j}$. Hence $\operatorname{Sym}(I)_k=[\operatorname{Sym}_R(I)]_{(k,*)}$ and $\mathcal{A}_k=[\mathcal{A}]_{(k,*)}$. In the sequel, when we use a single grading for $B = R \otimes S$, we will always follow this convention.
\[perfPair\] For each $i \in \{ 0,\dots, \delta \}$, there exists an isomorphism of finitely generated graded $S$-modules $$\mathcal{A}_i\cong \operatorname{Hom}_S (\operatorname{Sym}(I)_{\delta-i},S(-n)),$$ where $\delta= d-n = d_1 + \dots + d_n - n$.
In [@Jou97 Section 3.6], Jouanolou describes the isomorphism in terms of Morley forms (for instance, see [@KPU17 Chapter 4]). Therefore, an explicit computation of $\operatorname{Hom}_S (\operatorname{Sym}(I)_{\delta-i},S(-n))$ leads to a generating set, not only their bidegrees, of $\mathcal{A}_i$. In the following two sections, we study $\mathcal{A}_i$ and give a generalized method to compute a generating set for $\mathcal{A}_i$.
The dual generators of the defining ideal {#sectiononfunctionsinA1}
=========================================
We will adapt the notation in and the notation and hypothesis of . The graded free $S$-resolution of $\operatorname{Sym}(I)_{\delta-i}$ in is $$\label{delta-ires}
\mathcal{K}_{\delta-i} : 0 \to F_m \stackrel{\alpha_m}{\to} \cdots \to F_1 \stackrel{\alpha_1}{\to} F_0,$$ where $F_t=\bigoplus_{1\leq j_1\leq j_2\leq\cdots\leq j_i\leq n}S(-t)^{\delta-i-(d_{j_1}+\cdots+d_{j_i})+n-1\choose n-1}$ for $t = 0,\dots, m$. Let $r_t=\operatorname{rk}\alpha_t$ and $f_t=\operatorname{rk}F_t$. By (1), $f_t = r_t + r_{t+1}$ for $t = 1,\dots, m$, and by , $\mathcal{A}_i \cong \operatorname{Hom}_S(\operatorname{Sym}(I)_{\delta-i},S(-n))= \ker\alpha_i^*(-n)$. For the following construction, it is worth mentioning that the above minimal graded resolution is linear.\
\[reviewBE74\] In order to apply the theorems in [@BE74] to our set up, we need to specify shifts. It is not a hard task to do, but for the convenience of the reader and to set up the notation, we review their construction below. We also note that their constructions are for projective modules, but in our paper we only need their theorems for (graded) free modules. Hence our review is written for *graded* free modules of their work.\
1. First for $0\leq t\leq m$, fix a basis for $F_t$. This enables us to have a matrix representation of $\alpha_t$ and $\wedge^{r_t}\alpha_t$ for $1\leq t\leq m$, respectively. Henceforth whenever we talk about these maps, we use their matrix representations. We use on \[delta-ires\] to construct maps $a_t$ for $t=1,\dots,m$. Since $\alpha_m:F_m\rightarrow F_{m-1}$ is an injective map, the entries of $a_m=\wedge^{r_m}\alpha_m=\wedge^{f_m}\alpha_m$ are of degree $r_m$. Thus $a_m:S(-r_m m))\rightarrow \wedge^{r_m}F_{m-1}$. (Recall that since \[delta-ires\] is a linear resolution, the shift of $F_t$ is $-t$ for all $t = 0,\dots,m$.) Using , we have have the commutative diagram $$\xymatrix{
\wedge^{r_{m-1}}F_{m-1}\ar[rr]^{\wedge^{r_{m-1}}\alpha_{m-1}}\ar[rd]_{a_m^*} & & \wedge^{r_{m-1}}F_{m-2}\\
& S(-r_{m-1}(m-1) + r_m)\ar[ru]_{a_{m-1}}&
}$$ Notice that $a_m^*$ is a row matrix whereas $a_{m-1}$ is a column matrix; $$\begin{aligned}
a_{m}^*=[a^*_{m,1}\cdots a^*_{m,{f_{m-1}\choose r_{m}}}] \qquad a_{m-1}=\begin{bmatrix}
a_{m-1,1}\\
\vdots\\
a_{m-1, {f_{m-2}\choose r_{m-1}}}
\end{bmatrix}.
\end{aligned}$$
Then one has the factorization $$\wedge^{r_{m-1}}\alpha_{m-1}=a_{m-1}\circ a_m^*.$$ Not all of the entries of $a^*_{m}$ is zero as $a_m^*=(\wedge^{r_m}\alpha_m)^*$ and $\operatorname{rk}\alpha_m=r_m$. Thus there exists an entry, say $a^*_{m,u}\neq 0$. Now to compute $a_{m-1}$, consider the $u$-th column of $\wedge^{r_{m-1}}\alpha_{m-1}$ and divide its entries by $a^*_{m,u}$. From this we deduce that the entries of $a_{m-1}$ are of degree $r_{m-1}-r_m$ and that $a_{m-1}:S(-r_{m-1}(m-1)+r_m)\rightarrow\wedge^{r_{m-1}}F_{m-2}$. Iteratively, one can see that $a_1:S(-\sum_{p=1}^m(-1)^{p-1}r_p )\to \wedge^{r_1}F_0$.
2. Let $s_t=r_t \cdot t - \sum_{p=t+1}^m (-1)^{p-t-1}r_p$. Then $a_t:S(-s_t )\rightarrow \wedge^{r_{t}} F_{t-1}$ for $t=1,\dots,m$. In particular, $s_1 = r_1 - r_2 + \cdots +(-1)^{m-1}r_m = \sum_{p=1}^m (-1)^{p-1} r_p$.
3. The map $\alpha_1:F_1\rightarrow F_0$ induces the map $F_0^*\otimes F_1\rightarrow S$. By dualizing it, we obtain the map $\widetilde{\alpha_1}:S\rightarrow F_0\otimes F_1^*$. Consider the map $$\label{mapDelta}
\partial:\wedge^{f_0-1}F_0\rightarrow\wedge^{f_0}F_0\otimes F_1^*,$$ which is the composition of the following maps $$\wedge^{f_0-1}F_0 \cong \wedge^{f_0-1}F_0\otimes S\xrightarrow{\operatorname{id}\otimes \widetilde{\alpha_1}}\wedge^{f_0-1}F_0\otimes F_0\otimes F_1^*\xrightarrow{m_\wedge \otimes \operatorname{id}}\wedge^{f_0}F_0\otimes F_1^*,$$ where $m_\wedge:\wedge^{f_0-1}F_0\otimes F_0\rightarrow \wedge^{f_0}F_0$ is the usual multiplication in the exterior algebra $\wedge F_0$. Fix an orientation $\eta$ for $F_0$ and let $a_1:S(-s_1) \rightarrow \wedge^{r_1}F_0$ be the map in (2). This map $\partial$ is used in the proof of .
\[thm1\] With the hypothesis of , for any $i \in \{ 0, \dots, \delta \}$ and $\alpha_1$ in the graded free $S$-resolution of $\operatorname{Sym}_R (I)_{\delta-i}$ in \[delta-ires\], we have the following statements.
1. The following is a complex of graded free $S$-modules $$\wedge^{f_0-r_1-1}F_0\otimes S(-s_1)\xrightarrow{\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*,$$ where $a_1$ is the map in . In particular, if $f_0 = r_1 + 1$, then the image of $a_1$ is in $\ker \alpha_1^*$.
2. In addition, if $\operatorname{rk}\operatorname{Sym}(I)_{\delta-i}>1$, then $$\wedge^{f_0-r_1-1}F_1\otimes S(-s_1)\xrightarrow{\wedge^{f_0-r_1-1}\alpha_1\otimes \operatorname{id}}\wedge^{f_0-r_1-1}F_0\otimes S(-s_1)\xrightarrow{\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*$$ is a complex of graded free $S$-modules.
(1): [@BE74 Lemma 3.2(a)] (the map $\partial$ in \[mapDelta\] is equal to $d_{f_0-r_1-1}^{\, \alpha_1}$) implies that the composition $$\label{lem3.2a}
\wedge^{f_0-r_1-1}F_0\otimes \wedge^{r_1}F_1\xrightarrow{m_\wedge\circ(\operatorname{id}\otimes\wedge^{r_1}\alpha_1)}\wedge^{f_0-1}F_0\xrightarrow{\partial}\wedge^{f_0}F_0\otimes F_1^*$$ is zero.\
We identify $\wedge^{f_0-1}F_0 \cong F_0^*$ and $\wedge^{f_0}F_0\cong S$ following the fixed orientation $\eta$ of $F_0$. Lemma 3.2(c) in [@BE74] implies that the following diagram commutes up to sign: $$\label{lem3.2c}
\xymatrix{
\wedge^{f_0-1}F_0 \ar[rr]^{\partial} \ar[d]_{\eta} && \wedge^{f_0}F_0\otimes F_1^* \ar[d]^{\eta \otimes \operatorname{id}}\\
\wedge^1 F_0^* \ar[rr]^{(m_\wedge\circ(\operatorname{id}\otimes \alpha_1))^*} && \wedge^0 F_0^* \otimes F_1^*.
}$$ Notice that in the above diagram, the vertical maps are isomorphisms and the bottom map, i.e., $(m_\wedge\circ (1\otimes \alpha_1))^*$ is a composition of the following maps $$\begin{aligned}
\wedge^1F_0^*\xrightarrow{m_\wedge^*}\wedge^0F_0^*\otimes \wedge^1F_0^* \xrightarrow{(\operatorname{id}\otimes \alpha_1)^*}\wedge^0F_0^*\otimes \wedge^1F_1^*.\end{aligned}$$ Since $\wedge^0F_0^* \cong S$, the map $(m_\wedge\circ (1\otimes \alpha_1))^*$ can be identified with the map $$\label{lemAlpha}
F_0^* \stackrel{\alpha_1^*}{\longrightarrow} F_1^*.$$
From , , and , we conclude that $$\label{seqTop}
\wedge^{f_0-r_1-1}F_0\otimes \wedge^{r_1}F_1\xrightarrow{\eta\circ m_\wedge\circ(\operatorname{id}\otimes\wedge^{r_1}\alpha_1)}F_0^*\xrightarrow{\alpha_1^*} F_1^*$$ is a complex. Thus $\operatorname{Im}(\eta\circ m_\wedge\circ(\operatorname{id}\otimes \wedge^{r_1}\alpha_1))\subseteq\ker\alpha_1^*$.\
With the commutative diagram in Theorem \[BEprelim\](a) $$\xymatrix{
\wedge^{r_1}F_1\ar[rd]_{a_2^*}\ar[rr]^{\wedge^{r_1}\alpha_1} & &\wedge^{r_1}F_0\\
& S(-s_1),\ar[ur]_{a_1} &
}$$ the first morphism in \[seqTop\] factors as follows $$\label{thm1commdiag2}
\xymatrix{
\wedge^{f_0-r_1-1} F_0 \otimes \wedge^{r_1} F_1\ar[rr]^{\eta\circ m_\wedge\circ (\operatorname{id}\otimes \wedge^{r_1}\alpha_1)} \ar[d]_{\operatorname{id}\otimes a_{2}^*} && F_0^*\ar[r]^{\alpha_1^*} & F_1^*
\\
\wedge^{f_0-r_1-1}F_0\otimes S(-s_1) \ar[r]^{\operatorname{id}\otimes a_1}& \wedge^{f_0-r_1-1}F_0 \otimes \wedge^{r_1} F_0 \ar[r]^{\qquad m_\wedge} & \wedge^{f_0-1}F_0 \ar[u]^{\eta}.
}$$ Now let $J=\operatorname{Im}a_2^*$ be an $S$-ideal (with a shift). Then $\operatorname{Im}\operatorname{id}\otimes a_2^*=\wedge^{f_0-r_1-1}F_0\otimes J$. By (c), $\sqrt{I(\alpha_2)}=\sqrt{I(a_2)}$. So, the ideal $J$ has a positive grade. Since the composition of maps in the top row is zero and the above diagram commutes, we have $$\begin{aligned}
\label{thm1ProofRepeat}
0&=(\alpha_1^*\circ \eta\circ m_\wedge\circ(\operatorname{id}\otimes \wedge^{r_1}\alpha_1))(\wedge^{f_0-r_1-1}F_0\otimes \wedge^{r_1}F_1)\\
&=(\alpha_1^*\circ \eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)\circ(\operatorname{id}\otimes a_2^*))(\wedge^{f_0-r_1-1}F_0\otimes \wedge^{r_1}F_1)\\
&=(\alpha_1^*\circ \eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1))\circ(\operatorname{id}\otimes a_2^*)(\wedge^{f_0-r_1-1}F_0\otimes \wedge^{r_1}F_1)\\
&=(\alpha_1^*\circ \eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1))(\wedge^{f_0-r_1-1}F_0\otimes J(-s_1))\\
&=J(\alpha_1^*\circ \eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1))(\wedge^{f_0-r_1-1}F_0\otimes S(-s_1)).
\end{aligned}$$ The last equality holds as the maps are $S$-module homomorphisms and tensor products are over $S$. Since the last equality holds in the free module $F_1^*$, and $J$ has positive grade, $\alpha_1^*\circ \eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)=0$. Thus, $$\wedge^{f_0-r_1-1}F_0\otimes S(-s_1)\xrightarrow{\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*$$ is a complex of graded free $S$-modules.\
(2): By item (1), it suffices to show that $[\eta\circ m_\wedge\circ (\operatorname{id}\otimes a_1)]\circ [\wedge^{f_0-r_1-1}\alpha_1\otimes \operatorname{id}_S ]=0$. We first extend the complex in the top row of \[thm1commdiag2\] to $$\label{cor1complex1}
\wedge^{f_0-r_1-1}F_1\otimes \wedge^{r_1}F_1\xrightarrow{\wedge^{f_0-r_1-1}\alpha_1\otimes \operatorname{id}_1}\wedge^{f_0-r_1-1}F_0\otimes \wedge^{r_1}F_1\xrightarrow{\eta\circ m_\wedge\circ(\operatorname{id}\otimes\wedge^{r_1}\alpha_1)}F_0^*\xrightarrow{\alpha_1^*} F_1^*.$$ Since \[delta-ires\] is a free $S$-resolution for $\operatorname{Sym}(I)_{\delta-i}$, we have $\operatorname{rk}\operatorname{Sym}(I)_{\delta-i}= \operatorname{rk}F_0 - \operatorname{rk}\alpha_1 = f_0-r_1 > 1$ by the hypothesis. Equivalently, we have $f_0 -1 > r_1$. Hence from $$\eta\circ m_\wedge\circ(\operatorname{id}\otimes\wedge^{r_1}\alpha_1)\circ(\wedge^{f_0-r_1-1}\alpha_1\otimes \operatorname{id}_1)=\eta\circ m_\wedge\circ (\wedge^{f_0-r_1-1}\alpha_1\otimes \wedge^{r_1}\alpha_1)=\eta\circ \wedge^{f_0-1}\alpha_1 = 0,$$ we conclude that \[cor1complex1\] is a complex.\
With the factorization $\wedge^{r_1} \alpha_1 = a_1 \circ a_2^*$, we obtain the following a commutative diagram $$\xymatrix{
\wedge^{f_0-r_1-1}F_1\otimes \wedge^{r_1}F_1\ar[rr]^{\hspace{1cm}\wedge^{f_0-r_1-1}\alpha_1\otimes \operatorname{id}_1\hspace{1cm}}\ar[d]^{\operatorname{id}_2\otimes a_2^*}&&\wedge^{f_0-r_1-1}F_0\otimes \wedge^{r_1}F_1\ar[rr]^{\hspace{1cm}\eta\circ m\circ(\operatorname{id}\otimes\wedge^{r_1}\alpha_1)}& \hspace{1cm}&F_0^*\ar[r]^{\alpha_1^*} &F_1^*.\\
\wedge^{f_0-r_1-1}F_1\otimes S(-s_1)\ar[rr]^{\wedge^{f_0-r_1-1}\alpha_1\otimes \operatorname{id}_S}&&\wedge^{f_0-r_1-1}F_0\otimes S(-s_1)\ar[rru]_{\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)}& \hspace{1cm}& &
}$$ Let $J=\operatorname{Im}a_2^*$ be an $S$-ideal. Then the argument used to prove Item (1) (from \[thm1ProofRepeat\]) shows that $\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)\circ \wedge^{f_0-r_1-1}\alpha_1\otimes \operatorname{id}_S=0$. This complete the proof of the theorem.
We present two examples demonstrating (b). It is worth mentioning that in the first example, the corresponding complex is exact whereas in the second example, it is not.
\[thm1Example\]
1. (cf. ) Let $R={{\mathbf k}}[x,y,z]$ and $I=I_3(\varphi)$, where $$\varphi=\begin{bmatrix}
x & y^2 & z^3\\
y & z^2 & yz^2\\
z & x^2 & y^3\\
0 & xy+yz+xz & 0
\end{bmatrix}.$$ It is easy to verify that this example satisfies the hypothesis of . Consider $B=R[T_0,T_1,T_2,T_3]$ and let $\mathcal{K}$ denote the Koszul complex on $l_1,l_2,l_3$ where $[l_1~l_2~l_3]=[T_0~T_1~T_2~T_3]\cdot\varphi$. Since $\mathcal{K}$ is a graded free resolution of $\operatorname{Sym}(I)$, we can extract the $S$-resolution of $\operatorname{Sym}(I)_1$ $$0\rightarrow S(-1)\xrightarrow{\alpha_1}S^3\rightarrow\operatorname{Sym}(I)_1\rightarrow 0,$$ where $\alpha_1 = \begin{bmatrix}T_0&T_1&T_2\end{bmatrix}^t$. Here, we have $f_0 = 3, f_1=1, s_1 = r_1=1$. (b) shows that $$\wedge^1F_1\otimes S(-s_1)\xrightarrow{\wedge^1\alpha_1\otimes \operatorname{id}} \wedge^1 F_0\otimes S(-s_1)\xrightarrow{m_\wedge\circ(\operatorname{id}\otimes a_1)}\wedge^1F_0\cong^\eta F_0^*\xrightarrow{\alpha_1^*}F_1^*$$ is a complex. For this example, we will verify that $\alpha_1^*\circ \eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)=0$ and $\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)=0\circ \wedge^1\alpha_1\circ\operatorname{id}=0$.\
Fixing a basis for $F_0,F_1$ as $\{e_1,e_2,e_3\},\{g_1\}$ respectively, we can fix a basis $\{e_1\wedge e_2,e_1\wedge e_3,e_2\wedge e_3 \}$, $\{g_1\}$ for $\wedge^1F_0$ and $\wedge^1F_1$, respectively. A basis $\{e_1\otimes g_1,e_2\otimes g_1,e_3\otimes g_1 \}$ of $\wedge^1F_0\otimes \wedge^1F_1$ gives us a following matrix representation of $\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)$ $$\label{rk2Example}
\begin{bmatrix}
0 & T_2 & -T_1\\
-T_2 & 0 & T_0\\
T_1 & -T_0 & 0
\end{bmatrix}.$$ As $\wedge^1\alpha_1\otimes \operatorname{id}\cong \alpha_1$ and $\alpha_1^*$ is the transpose $\alpha_1$, we have $$\begin{aligned}
\wedge^1F_1\otimes S(-s_1)
\xrightarrow{\begin{bmatrix}T_0\\T_1\\T_2\end{bmatrix}} \wedge^1F_0\otimes S(-s_1)
\xrightarrow{\begin{bmatrix} 0 & T_2 & -T_1\\ -T_2 & 0 & T_0\\ T_1 & -T_0 & 0 \end{bmatrix}} F_0^*\xrightarrow{\begin{bmatrix}T_0 & T_1 & T_2\end{bmatrix}} F_1^*.
\end{aligned}$$ Then it is easy to see that this is a complex. In fact, the above complex is a graded minimal free $S$-resolution of $\operatorname{coker}\alpha_1^*$ $$0\rightarrow \wedge^1F_1\otimes S(-s_1)\to
\wedge^1F_0\otimes S(-s_1)\to
F_0^* \to F_1^*\rightarrow \operatorname{coker}\alpha_1^*\rightarrow 0.$$
2. Let $R={{\mathbf k}}[x,y,z]$ and $I=I_3(\varphi)$, where $$\varphi=\begin{bmatrix}
x^2 & 0 & x^4\\
y^2 & x^2 & y^4\\
z^2 & y^2 & z^4\\
0 & z^2 & x^3z
\end{bmatrix}.$$ It is easy to verify that this example satisfies the hypothesis of . We will use the same notation as in part (a). The number $\delta$ is $d - n= 8-3 = 5$. The graded free $S$-resolution of $\operatorname{Sym}(I)_\delta$ is $$0\rightarrow S(-2)^3 \xrightarrow{\alpha_2} S(-1)^{23} \xrightarrow{\alpha_1} S^{21 }\rightarrow\operatorname{Sym}(I)_5 \rightarrow 0.$$ Here, we have $f_0 = 21, f_1=23, f_2 = 3, \operatorname{rk}\operatorname{Sym}(I)_5 = 1$ and $r_1=20, r_2 = 3$. Therefore, $f_0-r_1-1 = 21 - 20 - 1 = 0$, $s_1 = r_1 - r_2 = 17$. (a) implies that $$\wedge^0 F_0\otimes S(-s_1)\xrightarrow{m_\wedge\circ(\operatorname{id}\otimes a_1)}\wedge^1F_0\cong^\eta F_0^*\xrightarrow{\alpha_1^*}F_1^*$$ is a complex. However, this complex does *not* extend to a free resolution of $\operatorname{coker}\alpha_1^*$ (cf. ).
As the examples above demonstrate the complexes in are not exact in general. It is natural to ask which conditions guarantee the exactness of these complexes. A positive answer to this question provides part of minimal generating equations for the defining ideal of the corresponding Rees algebra by Jouanolou. In the following theorems, we provide sufficient conditions for the exactness in the case where $\operatorname{Sym}(I)_k$ is of rank 1 or 2, respectively.
\[thmRk1\] For a fixed integer $i$, consider the complex $$\wedge^{f_0-r_1-1}F_0\otimes S(-s_1)\xrightarrow{\eta\circ m_\wedge \circ(\operatorname{id}\otimes a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*,$$ in . Assume that $\operatorname{rk}\operatorname{Sym}(I)_{\delta-i}=1$. Then the complex is exact if and only if $\operatorname{grade}I(a_1)\geq 2$.
Since $\operatorname{rk}\operatorname{Sym}(I)_{\delta-i}=1$, we have $f_0-r_1-1=0$. Therefore, the complex is isomorphic to the complex $$\label{rk1complex}
S(-s_1)\xrightarrow{\eta \circ a_1}F_0^*\xrightarrow{\alpha_1^*}F_1^*.$$ Notice that $\operatorname{rk}\alpha_1^* \ge 1$ and $\operatorname{grade}I(\eta(a_1)) = \operatorname{grade}I(a_1)$. Hence by , the complex is acyclic if and only if $\operatorname{grade}I( a_1)\geq 2$.
\[thmRk2\] For a fixed integer $i$, consider the complex $$\wedge^{f_0-r_1-1}F_0\otimes S(-s_1)\xrightarrow{\eta\circ m_\wedge \circ(\operatorname{id}\otimes a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*,$$ in . Assume that $\operatorname{rk}\operatorname{Sym}(I)_{\delta-i}=2$. Consider the following statements:
1. $\operatorname{Sym}(I)_{\delta-i}$ satisfies Serre’s condition $(S_2)$.
2. $\operatorname{grade}I(\alpha_t) \ge t+2$ for $t = 1,\dots, m$.
3. The complex above is exact.
Then we have $(a) \Rightarrow (b) \Rightarrow (c)$.
Before proving the theorem, we present a lemma which explains the relationship between the conditions $(a)$ and $(b)$.
\[lemRk2\] With the setup of , we have a complex of graded free $S$-modules $$\label{resolution of coker}
0\rightarrow F_m\otimes S(-s_1)\xrightarrow{\alpha_m\otimes\operatorname{id}} \cdots \to F_1\otimes S(-s_1)\xrightarrow{\alpha_1\otimes \operatorname{id}} F_0\otimes S(-s_1)\xrightarrow{\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*.$$ Furthermore, this complex is acyclic if $\operatorname{Sym}(I)_{\delta-i}$ satisfies Serre’s condition $(S_2)$.
Recall that $$0\rightarrow F_m \xrightarrow{\alpha_m} F_{m-1} \rightarrow\cdots\rightarrow F_2 \xrightarrow{\alpha_2} F_1\xrightarrow{\alpha_1} F_0$$ is a graded minimal free $S$-resolution of $\operatorname{Sym}(I)_{\delta-i}$, and so is its shift. Hence for the first part of the statement, it suffices to show that $$F_1\otimes S(-s_1)\xrightarrow{\alpha_1\otimes \operatorname{id}} F_0\otimes S(-s_1)\xrightarrow{\eta\circ m_\wedge \circ(\operatorname{id}\otimes a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*$$ is a complex. But this follows from (b) with $f_0 - r_1 - 1 = \operatorname{rk}Sym(I)_{\delta-i} - 1 = 1$.\
We first describe the map $\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1): F_0 \otimes S(-s_1) \to F_0^*$. By choosing a basis $\{ e_1,\dots, e_{f_0} \}$ of $F_0$, this can be described as a square matrix. As $a_1 : S(-s_1) \to \wedge^{r_1} F_0 \cong \wedge^{f_0-2} F_0^*$, by abuse of notation, we list the basis for $\wedge^{r_1} F_0$ as $e_i^* \wedge e_j^*$ and $a_1 = \oplus c_{i,j} e_i^* \wedge e_j^*$, where $c_{i,j} \in S$. With this notation, one sees that $\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)$ a skew symmetric matrix whose $i,j$th entry for $i<j$ is $c_{i,j}$ up to sign. Hence we have $\sqrt{I_2 (\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1))} \supset I_1 (a_1)$ and $\operatorname{grade}I(\eta\circ m_\wedge\circ(\operatorname{id}\otimes a_1)) \ge \operatorname{grade}I_1 (a_1) \ge \operatorname{grade}I(\alpha_1)$. Furthermore, as $\operatorname{grade}\alpha_1^* = \operatorname{grade}\alpha_1$, by , the complex in \[resolution of coker\] is exact if and only if $\operatorname{grade}\alpha_m \ge m + 2$ for $i = 1,\dots, m$.\
It is a well-known theorem of Auslander and Bridger [@AB Theorem 4.25] that a module $M$ is $k$th syzygy if and only if $M$ satisfies Serre’s condition $(S_k)$. Hence $\operatorname{Sym}(I)_{\delta-i}$ satisfies $(S_2)$ if and only if for some free $S$-modules $F_{-1}, F_{-2}$, the following complex is acyclic $$0\rightarrow F_m \xrightarrow{\alpha_m} F_{m-1} \rightarrow\cdots\rightarrow F_2 \xrightarrow{\alpha_2} F_1\xrightarrow{\alpha_1} F_0 \to F_{-1} \to F_{-2}.$$ By , this condition implies that $\operatorname{grade}\alpha_m \ge m + 2$ for $i = 1,\dots, m$.
Now the proof follows from and its proof.
A similar argument as in the proof of , one can show that the exactness of the complex in follows from the condition that $\operatorname{rk}\operatorname{Sym}(I)_{\delta-i}=1$ and $\operatorname{Sym}(I)_{\delta-i}$ satisfies Serre’s condition $(S_1)$. Hence it is natural to ask the following question.
\[rge3Question\] Consider the complex $$\wedge^{f_0-r_1-1}F_0\otimes S(-s_1)\xrightarrow{\eta\circ m_\wedge \circ(\operatorname{id}\otimes a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*$$ in . Assume that $\operatorname{rk}\operatorname{Sym}(I)_{\delta-i}= r > 0$. Then is this complex exact if $\operatorname{Sym}(I)_{\delta-i}$ satisfies Serre’s condition $(S_r)$? Furthermore, in the case where this complex is exact, is there a natural way to construct the minimal free resolution of $\operatorname{coker}\alpha_1^*$?
This question has an affirmative answer when $r \le 2$ by .
Degrees of Implicit equation {#sectionOnDegreesofImplicit}
============================
In this section, we provide a concrete description of the bi-degree of the equation defining the special fiber ring $\mathcal{F}(I)$. With the help of Morley forms, the results in this section can be used to find the equation defining $\mathcal{F}(I)$. We assume the notation and set up of , but for the convenience of the reader we recall the key notation and setup. By $d_1, \dots, d_n$ we denote the column degrees of the graded presentation matrix $\varphi$ of $I$, $n$ is the dimension of $R$, $I = I_n(\varphi)$, and $\delta=d_1+\cdots+ d_n-n = d-n$. Consider the diagram, where the rows are exact $$\xymatrix{
0 \ar[r] & \mathcal{A}\ar[r]\ar[d] & \operatorname{Sym}(I)\ar[r]\ar[d] & R[It]\ar[r]\ar[d] & 0\\
0\ar[r] & \ker\overline{\pi}\ar[r] & S={{\mathbf k}}[T_0,\dots,T_n] \ar[r]^{\hspace{1cm}\overline{\pi}}& \mathcal{F}(I)\ar[r] & 0.
}$$ The defining ideal of $\mathcal{F}(I)$ is $\ker\overline{\pi}\cong\mathcal{A}\otimes{{\mathbf k}}\cong \mathcal{A}_0 S$. By , to find the defining ideal of $\mathcal{F}(I)$, it is enough to compute $\operatorname{Hom}_S(\operatorname{Sym}_R(I)_\delta,S(-n))$.\
We begin with the resolution of $\operatorname{Sym}(I)_{\delta}$, which is induced from the resolution in . $$\begin{aligned}
\label{symdeltares}
0\rightarrow F_m\xrightarrow{\alpha_m}F_{m-1}\rightarrow\cdots\rightarrow F_2\xrightarrow{\alpha_2}F_1\xrightarrow{\alpha_1}F_0\rightarrow\operatorname{Sym}(I)_\delta\rightarrow 0,\end{aligned}$$ where $m\leq n-1$ and since $\delta + n = d = d_1 + \cdots + d_n$, we may write $$\begin{aligned}
F_t=\bigoplus_{1\leq j_1\leq j_2\leq\cdots\leq j_t\leq n}S(-t)^{\delta-(d_{j_1}+\cdots+d_{j_t})+n-1\choose n-1}=\bigoplus_{1\leq k_1\leq k_2\leq\cdots\leq k_{n-t}\leq n}S(-t)^{d_{k_1}+\cdots+d_{k_{n-t}}-1\choose n-1},\end{aligned}$$ where $\{ j_1, ,\dots, j_t, k_1,\dots,k_{n-t} \} = \{1,\dots,n\}$. As in the previous section, we let $r_t=\operatorname{rk}~\alpha_t, f_t=\operatorname{rk}~F_t$ and $\mathcal{A}_0\cong \operatorname{Hom}(\operatorname{Sym}(I)_\delta,S(-n))=\ker\alpha_1^*(-n)$.
\[remarkonextraranks\] The length of the Koszul complex in is $n$ whereas the length of the resolution of $\operatorname{Sym}(I)_\delta$ in \[symdeltares\] can be strictly less than $n-1$. In particular, $F_t = 0$ for $m+1 \le t \le n-1$. In other words, we have $${d_{k_1}+\cdots+d_{k_{n-t}}-1\choose n-1} = 0~{\text{if and only if}}~ m+1\le t.$$
By [@UV Proposition 2.4], $\ker\overline{\pi}$ is a principal prime ideal and hence, so is $\ker\alpha_1^*$. By Theorem \[thm1\], we have a complex $$\wedge^{f_0-r_1-1}F_1\otimes S(-s_1)\xrightarrow{\eta\circ m_\wedge \circ (\operatorname{id}\otimes a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*,$$ where $\eta$ is a fixed orientation of $F_0$ and $a_1:S(-s_1)\rightarrow\wedge^{r_1}F_0$ is the map obtained by applying to the resolution \[symdeltares\]. We refer to (2) for the description of the shifts $s_1$ in the preceding complex. Since $\operatorname{rk}\operatorname{Sym}(I)_\delta=1$, we have $f_0-r_1-1=0$, and the above complex simplifies to $S(-s_1)\xrightarrow{\eta(a_1)}F_0^*\xrightarrow{\alpha_1^*}F_1^*$. Thus $\eta(a_1(1_{S(-s_1)}))\in \ker\alpha_1^*$. Henceforth, we will often denote the basis element $1_{S(-s_1)}$ of $S(-s_1)$ simply by $1$ if no confusion arises.\
Our main result of this section is the following theorem.
\[thm5\]Assume ${{\mathbf k}}$ is a field of characteristic zero and the notation and hypothesis of . Let $R={{\mathbf k}}[x_1,\dots,x_n]$ with $n \ge 3$, and let $\phi : \operatorname{Proj}R={{\mathbb P}_{{\mathbf k}}}^{n-1} {{\ext@arrow 0359\rightarrowfill@@{}{[f_0:\cdots: f_n]}}} \mathbb{P}^{n}_{{\mathbf k}}$ be the rational map defined in \[parametrizationmap\]. Write $I = (f_0,\dots, f_n)$. Then the following statements are equivalent:
1. The rational map $\phi$ is birational onto its image.
2. The $\ker\alpha_1^*$ is generated by $(\eta \circ a_1)(1)$.
3. $\operatorname{grade}I(\eta(a_1))\geq 2$.
4. The greatest common divisor of the entries of $\eta \circ a_1$ (this is a column matrix) is 1.
5. $e(\mathcal{F}(I))= e ({R}/{(g_0,\dots,g_{n-2}):I})$, where $g_0, \dots, g_{n-2}$ are general ${{\mathbf k}}$-linear combinations of the $f_i'$s. Here, $e(-)$ denotes, the Hilbert-Samuel multiplicity.
The Hilbert-Samuel multiplicity $e(\mathcal{F}(I))$ is also called the *degree* of the projective variety $\operatorname{Proj}\mathcal{F}{(I)} = \overline{\Phi (\mathbb{P}^{n-1})}$. We use the next two lemmas to prove .
\[KPUmultiplicity\] Let ${{\mathbf k}}$ be a field of characteristic zero, $R={{\mathbf k}}[x_1,\dots,x_n]$ with $n \ge 3$. Further, let $I$ be a homogeneous ideal generated by forms of the same degree $d$ and ${{\mathbf k}}[R_d]$ the $d$-th Veronese subring of $R$. If $g_0,\dots,g_{n-2}$ are general ${{\mathbf k}}$-linear combinations of the generators $f_0,\dots,f_n$ of $I$, then the following equality holds. $$\label{KPUmultiplicityeq}
e(\mathcal{F}(I))=\frac{1}{[{{\mathbf k}}[R_d]:{{\mathbf k}}[f_0,\dots,f_n]]}\cdot e\left(\frac{R}{(g_0,\dots,g_{n-2}):I^\infty} \right).$$
In the above theorem $[{{\mathbf k}}[R_d]:{{\mathbf k}}[f_0,\dots,f_n]]$ denotes the field extension degree $[\operatorname{Quot}({{\mathbf k}}[R_d]):\operatorname{Quot}({{\mathbf k}}[f_0,\dots,f_n])]$. Using , we have that the rational map $\phi$ (as defined in \[parametrizationmap\]) is birational onto its image if and only if $[{{\mathbf k}}[R_d]:{{\mathbf k}}[f_0,\dots,f_n]]=1$.
For any matrix $A$ with entries of the same degree, we define $\deg~ A$ to be the degree of the entries of $A$.
\[thm4\] Assume the notation and hypothesis of . For general ${{\mathbf k}}$-linear combinations $g_0,\dots,g_{n-2}$ of the generators $f_0,\dots,f_n$ of $I$, we have $$\label{deg of entries and multiplicity}
\deg \operatorname{Im} (\eta \circ a_1)+\dim R=e\left(\frac{R}{(g_0,\dots,g_{n-2}):I} \right) = e\left(\frac{R}{(g_0,\dots,g_{n-2}):I^\infty} \right).$$
First we show that for general linear combinations $g_0,\dots,g_{n-2}$, $$e({R}/{(g_0,\dots,g_{n-2}):I^\infty} ) = e({R}/{(g_0,\dots,g_{n-2}):I} ).$$
For each $\lambda=(\lambda_{0},\cdots,\lambda_{n}) \in (\mathbb{A}_{{\mathbf k}}^{n+1})^{n+1}$, where $\lambda_i=(\lambda_{i0}:\cdots:\lambda_{in}) \in\mathbb{A}^{n+1}_{{\mathbf k}}$ we set $g_{i}=\sum_{j=0}^n\lambda_{ij}f_j\in R$ for $0\leq i\leq n$. Let $U'$ be an open subset of $(\mathbb{A}_{{\mathbf k}}^{n+1})^{n+1}$ such that $\det (a_{ij})\in{{\mathbf k}}\backslash\{ 0\}$. This shows that $\{g_0,\dots,g_{n}\}$ is a minimal generating set of $I$ if $\lambda \in U'$. Further, let $U_1'=U'\cap (\mathbb{A}_{{\mathbf k}}^{n+1})^{n-1}$ (the first $n-1$ components of $(\mathbb{A}_{{\mathbf k}}^{n+1})^{n+1}$).\
Notice that the ideal $I$ satisfies $\mu(I_p)\leq \dim R_p$ for every $p\in \operatorname{Spec}(R)\backslash\{{\mathfrak{m}}\}$. By [@Ul1 Corollary 1.6, Proposition 1.7] (see also [@PoliniXie Lemma 3.1]), there exists a dense open set $U''\subseteq (\mathbb{A}_{{\mathbf k}}^{n+1})^{n-1}$ such that for every $\lambda \in U''$, $(g_0,\dots,g_{n-2}):I$ is a geometric $(n-1)$-residual intersection. (We note that in [@PoliniXie], the non-empty open subset $U''$ corresponds to general $R$-linear combinations. Thus, the same $U''$ can be used as the desired non-empty open subset for ${{\mathbf k}}$-linear combinations.)\
It follows that $(g_0,\dots,g_{n-2})_P=I_P$ for $P\in V(I)$ such that $\operatorname{ht}P\leq n-1$. Since $I$ is a codimension two perfect ideal, $I$ is a strongly Cohen-Macaulay ideal. Now we apply [@Hun1 Theorem 3.1] to conclude that $(g_0,\dots,g_{n-2}):I^\infty=(g_0,\dots,g_{n-2}):I$. The open set $U = U'_1 \cap U''$ provides the desired non-empty open subset.\
As $g_0,\dots, g_{n}$ form a minimal homogeneous generators of $I$, we may choose a presentation matrix $\varphi'$ for $(g_0,\dots, g_n)$ with the same column degrees $d_1, \dots, d_n$ of $\varphi$. This follows from the uniqueness of the graded Betti numbers of $I$. Furthermore, we assume that $g_i$ are signed minors of the $(n+1) \times n$ matrix $\varphi'$.\
Let $M=I/(g_0,\dots,g_{n-2})$. Notice that the last two rows of $\varphi'$ is a presentation matrix for $M$, and we call this presentation matrix by $\varphi''$. We note that $\varphi''$ is a $2 \times n$ matrix. Furthermore, $I_2 (\varphi'') = {\operatorname{Fitt}_0(M)}={\operatorname{ann}M}={(g_0,\dots,g_{n-2}):I}$ (cf. [@BE77 Theorem 3.1(2)]) has the maximum possible height $n-1$. Thus we reduce to the case of computing the multiplicity $e({R}/{I_2(\varphi'')} )$. Since the $\operatorname{grade}I_2(\varphi'')$ is the maximum possible number, $n-1$, the Eagon-Northcott complex of $\varphi''$, $\operatorname{EN}(\varphi'')$, forms a graded (minimal) free $R$-resolution for $R/I_2(\varphi'')$. Let $N=R/I_2(\varphi'')$. Clearly, $e(N)=e(N(1))$, and the shifted Eagon-Northcott complex of $\varphi''$ is the resolution of $N(1)$. $$\begin{gathered}
\operatorname{EN}(\varphi'')(1): \\
0 \to (\operatorname{Sym}(R^2)_{n-2})^* \otimes R^{n-1}(-(d_1+\cdots d_n-1)) \to (\operatorname{Sym}(R^2)_{n-3})^* \otimes \bigoplus_{1\leq j_1\leq\cdots\leq j_{n-1}\leq n}R^{n-2}(-(d_{j_1}+\cdots+d_{j_{n-1}}-1)) \\
\to \cdots \to (\operatorname{Sym}(R^2)_0)^* \otimes \bigoplus_{1\leq j_1\leq j_2\leq n}R(-(d_{j_1}+d_{j_2}-1)) \to R(1)\rightarrow 0.\end{gathered}$$ Furthermore, the Hilbert Series of $N(1)$ is $$\frac{\sum\limits_{t=1}^{n-1}\left[(-1)^t\cdot t\sum\limits_{1\leq j_1\leq\cdots\leq j_{t+1}\leq n} z^{(d_{j_1}+\cdots+d_{j_{t+1}}-1)}\right]+z^{-1}}{(1-z)^n}.$$ Let $p(z)={\sum\limits_{t=1}^{n-1}\left[(-1)^t\cdot t\sum\limits_{1\leq j_1\leq\cdots\leq j_{t+1}\leq n} z^{(d_{j_1}+\cdots+d_{j_{t+1}}-1)}\right]+z^{-1}}$. Then $$\begin{aligned}
e(N(1)) & = (-1)^{n-1}\frac{p^{(n-1)}(1)}{(n-1)!}\\
& = (-1)^{n-1}\left[\sum_{t=1} ^{n-1}(-1)^t\cdot t\cdot \sum_{1\leq j_1\leq\cdots\leq j_{t+1}\leq n}{d_{j_1}+\cdots + d_{j_{t+1}}-1\choose n-1} \right] + (-1)^{n-1}\\
& = \sum_{t=1} ^{n-1}\left[(-1)^{n-1+t}\cdot t\cdot \sum_{1\leq j_1\leq\cdots\leq j_{t+1}\leq n}{d_{j_1}+\cdots + d_{j_{t+1}}-1\choose n-1}\right] + 1, \end{aligned}$$ where $p^{(n-1)}(z)$ denotes the $(n-1)^{\text{st}}$ derivative of $p(z)$. We compare the terms with the ranks in \[symdeltares\]. By , we have the following equality. $$\sum_{1\leq j_1\leq\cdots\leq j_{t+1}\leq n}{d_{j_1}+\cdots + d_{j_{t+1}}-1\choose n-1}=
\begin{cases}
\operatorname{rk} F_{n-t-1} = f_{n-t-1} & t\geq n-m-1, \\
0 & t<n-m-1.
\end{cases}$$
Under this identification, we may rewrite $e(N(1))$ as follows $$\begin{aligned}
e(N(1)) & =\sum_{t=n-m-1} ^{n-1}\left[(-1)^{n-1-t}\cdot t\cdot f_{n-t-1}\right] + 1\\
& = (-1)^m(n-m-1)f_m+(-1)^{m-1}(n-m)f_{m-1}+(-1)^{m-2}(n-m+1)f_{m-2}+\cdots +(n-1)f_0 + 1.\end{aligned}$$ We note that since $\operatorname{rk}\operatorname{Sym}(I)_\delta=1$, $\sum_{j=0}^m(-1)^jf_j=1$. Then one has $$\begin{aligned}
n-m - 1 &= (n-m - 1) \sum_{j=0}^m(-1)^jf_j \\
&= (-1)^m (n-m-1) f_m + (-1)^{m-1}(n-m-1) f_{m-1} + (-1)^{m-2}(n-m-1) f_{m-2} + \cdots + (n-m-1) f_0 \\
&= (-1)^m (n-m-1) f_m + (-1)^{m-1}(n-m) f_{m-1} + (-1)^{m-2}(n-m+1) f_{m-2} + \cdots + (n-1) f_0 \\
&\phantom{(-1)^m (n-m-1) f_m +} - [(-1)^{m-1}(1) f_{m-1} + (-1)^{m-2}(2) f_{m-2}+\cdots + (m) f_0] \\
&= [e(N(1)) - 1] - [(-1)^{m-1}(1) f_{m-1} + (-1)^{m-2}(2) f_{m-2}+\cdots + (m) f_0] .\end{aligned}$$ Hence $$\begin{aligned}
e(N(1)) &= (-1)^{m-1}(1) f_{m-1} + (-1)^{m-2}(2) f_{m-2}+\cdots + (m) f_0 + n - m \\
&= \sum_{k=1}^{m}(-1)^{m-k}\cdot k\cdot f_{m-k} + n - m . \end{aligned}$$ Thus, we have shown that $$\label{multplicity of saturation}
e({R}/{(g_0,\dots,g_{n-2}):I^\infty} ) = \sum_{k=1}^{m}(-1)^{m-k}\cdot k\cdot f_{m-k}+(n-m).$$ We will complete the proof by showing that the right hand side of the equation is equal to $\deg \operatorname{Im} (\eta \circ a_1)+n$. First observe that $\deg \operatorname{Im} (\eta \circ a_1)+n =\deg (a_1)+n$ since $\eta$ is an isomorphism of degree 0. In , we showed that the degree of $\operatorname{Im} {a_1}$ is $s_1 = \sum_{t=1}^{m}(-1)^{t-1}r_t$, and $r_t=\sum_{j=t}^{m}(-1)^{j-t}f_j$. Note that since $\sum_{j=0}^m(-1)^jf_j=1$, we have $$1 = f_0 - f_1 + \cdots + (-1)^{t-1} f_{t-1} + \underbrace{(-1)^t f_t + (-1)^{t+1} f_{t+1} + \cdots + (-1)^m f_m}_{(-1)^t r_t}.$$ Therefore, $$r_t = (-1)^{t+1} (f_0 - f_1 + \cdots (-1)^{t-1} f_{t-1}) + (-1)^t = \sum_{j=0}^{t-1}(-1)^{t+j+1}f_j+(-1)^t.$$ Finally, we have $$\begin{aligned}
\deg \operatorname{Im} (\eta \circ a_1)+n &=\deg (a_1)+n\nonumber \\
&= \sum_{t=1}^{m}(-1)^{t-1}r_t +n\nonumber\\
& = \sum_{t=1}^m (-1)^{t-1}\left(\sum_{j=0}^{t-1}(-1)^{t+j+1}f_j+(-1)^t \right)+n\nonumber\\
& = \sum_{t=1}^m \left(\sum_{j=0}^{t-1} (-1)^j f_j-1\right) +n\nonumber\\
& = \sum_{t=1}^m \sum_{j=0}^{t-1} (-1)^j f_j+\sum_{t=1}^m (-1) +n \\
&= \sum_{t=1}^m \sum_{j=0}^{t-1} (-1)^j f_j+m(-1) +n\nonumber\\
& = \left(f_0+(f_0-f_1)+(f_0-f_1+f_2)+\cdots +\left(\sum_{j=0}^{m-1}(-1)^jf_j\right)\right)+(n-m)\nonumber\\
&= \sum_{k=1}^{m}(-1)^{m-k}\cdot k\cdot f_{m-k}+(n-m).\end{aligned}$$ This completes the proof.
We note that the rational map $\phi$ is birational to the image if and only if $[{{\mathbf k}}[R_d]:{{\mathbf k}}[f_0,\dots,f_n]] = 1$.\
$(1)\Leftrightarrow (5)$: This follows from .
$(2)\Leftrightarrow (5)$: As $\mathcal{F}(I)$ is a hypersurface, $e(\mathcal{F}(I))$ is the degree of a generating element of $\ker \overline{\pi}$. The equivalence follows from which say $$e(\mathcal{F}(I))=\frac{\deg \operatorname{Im} (\eta \circ a_1)+n}{[{{\mathbf k}}[R_d]:{{\mathbf k}}[f_0,\dots,f_n]]}.$$
$(2)\Leftrightarrow(3)$: This is a consequence of .
$(3)\Leftrightarrow (4)$: This equivalence is clear as the ambient ring is a UFD.
\(b) is an example where $I$ satisfies the conditions in , but the rational map is not birational to its image. In this example $\deg a_1 + n = 20$, but $e(\mathcal{F}(I) ) = 10$, so that the extension degree is $2$.
The Main result and application {#mainResultsec}
===============================
Let $R = {{\mathbf k}}[x,y,z]$ be a polynomial ring over a characteristic zero field ${{\mathbf k}}$ and $M$ a $4$ by $3$ matrix with entries in $R$. Our main result states that if $M$ is general of type $(1, 2, {{\mathbf q}})$ for some ${{\mathbf q}}> 2$ (see Case 2 in for the notion of general), then $I = I_3(M)$ satisfies the equivalence conditions in . Therefore, we are able to state the bi-degrees and the minimal number of equations in each bi-degree of the defining ideal of the Rees algebra $R[It]$, explicitly (hence for $\mathcal{F}(I)$ as well). Furthermore, by the duality of Jouanolou (), one may recover the explicit equations in $\operatorname{Sym}_R(I)$.
\[rmkMatBlock\] Let $B = S[x_1,\dots, x_d]$, where $S$ is a ring, and $f(x_1,\dots,x_d)$ a non-zero homogeneous polynomial of degree ${{\mathbf q}}$ in $B$. The graded homomorphism $\Phi: B(-{{\mathbf q}}) \stackrel{\cdot f}{\to} B$ induces a homomorphism in each degree $i$. We call the induced map $\Phi^f_i : [B(-{{\mathbf q}})]_i \to B_i$. This map can be described explicitly once we fix a basis of $B_i$. In this note, we will always use the Lex (monomial) order on $x_1,\dots, x_d$. For instance, when $d = 3$, ${{\mathbf q}}= 2$, $f = A_0 x_1^2 + A_1 x_1x_2 + A_2x_3^2$, where $A_i \in S$, then $\Phi^f_3$ has a matrix representation $$\begin{bmatrix}
A_0 &0 &0 \\
A_1 & A_0 &0 \\
0 &0 & A_0 \\
0 & A_1 &0 \\
0 &0 & A_1 \\
A_2 &0 &0 \\
0 &0 &0\\
0 &0 &0\\
0 & A_2 &0\\
0 &0 & A_2
\end{bmatrix}
\,
\overbrace{
\begin{array}{:l}
{x_1}^3 \\
{x_1}^2{x_2} \\
{x_1}^2{x_3} \\
{x_1}{x_2}^2 \\
{x_1}{x_2}{x_3} \\
{x_1}{x_3}^2 \\
{x_2}^3 \\
{x_2}^2{x_3} \\
{x_2}{x_3}^2 \\
{x_3}^3
\end{array}
}^{basis} .$$ Here each column is the presentation of $x_i f$ by the monomials of degree $3$.
To prove our main theorem, we need to compute the height of the maximal minors of $\Phi^f_i$. The following lemma generalizes [@Eis05 Theorem A.2.60] in which there is no gap between nonzero entries.
\[lemStaircase\] Let $R$ be a ring and $g_1, \dots, g_t$ are elements of $R$. Let $M=(m_{ij})$ be a $r\times s$ matrix ($r\geq s$) satsifying
1. each entry of $M$ is either zero or $g_1,\dots,g_t$,
2. the $i$-th non zero entry of each column is $g_i$,
3. the last non-zero entry of each column is $g_t$, and
4. if $m_{ij}=g_q$, then $m_{k j+1}=g_q$ for some $k>i$.
Then $I_s(M)=(g_1,\dots,g_t)^s$.
It suffices to show the statement that if $R' = \mathbb{Z}[y_1\cdots y_t]$ and $M=(m_{ij})$ a $r\times s$ matrix ($r\geq s$) satisfying
1. each entry of $M$ is either zero or $y_1,\dots,y_t$,
2. the $i$-th non zero entry of each column is $y_i$,
3. the last non-zero entry of each column is $y_t$, and
4. if $m_{ij}=y_q$, then $m_{k j+1}=y_q$ for some $k>i$,
then $I_s(M)=(y_1,\dots,y_t)^s$.
Clearly, $ I_s(M)\subseteq (y_1,\dots,y_t)^s$. We now show the other inclusion. We prove by induction on the number of variables $t$ and the number of columns $s$ of the matrix. Suppose $t=1$. The case of $s=1$ is clear, by the pattern of the matrix, that $I_t(M)=(y_1)$. In fact, if $s>1$, then it is easy to see from item (4) that $I_s(M)=(y_1)^s$ settling the base case of induction. Now suppose by induction hypothesis, the result is true for all matrices $M$ following the pattern in the hypothesis, with entries in the ring $T=\mathbb{Z}[y_1,\dots,y_q]$ where $q\leq t-1$ and for all $s$.\
If $s=1$, in which case $M$ has only one column, then clearly $I_s(M)=(y_1,\dots,y_t)$. By induction hypothesis on $s$ assume that the result is true for all matrices $M$, satisfying the pattern in the hypothesis, with the number of columns being strictly less than $s$. Now let $M$ be a matrix with $s$ columns and $j_1,\dots,j_s$ be indices such that $m_{1j_1}=m_{2j_2}=\cdots=m_{sj_s}=y_1$. If $M'$ denotes the matrix obtained by deleting the $j_1$-th row and the first column of $M$, then by induction hypothesis on $s$, it is clear that $I_{s-1}(M')=(y_1,\dots,y_t)^{s-1}$. Since the $j_1$-th row of $M$ has $y_1$ appearing in the first entry and zero otherwise, we have $f=y_1\cdot g$ where $g\in I_{s-1}(M')=(y_1,\cdots,y_t)^{s-1}$. Thus one has $y_1(y_1,\dots,y_t)^{s-1}\subseteq I_s(M)$. We show similarly that $y_1^u(y_1,\dots,y_t)^{s-u}\subseteq I_s(M)$. Let $M_u$ be the matrix obtained by removing the rows $j_1,\dots,j_u$ and the columns $1,\dots,u$ from $M$. Then by induction hypothesis $I_{s-u}(M_u)=(y_1,\dots,y_t)^{s-u}$. Now consider $M$ and an $s$-minor $f'$ obtained by choosing the rows $j_1,\dots,j_u,i_1,\dots,i_{s-u}$ where $1\leq i_q\leq r, i_q\not\in\{j_1,\dots,j_{u}\}$ for $1\leq q\leq s-u$, of $M$. Then we have that $f'=y_1^ug'$ where $g'\in I_{s-u}(M_u)=(y_1,\dots,y_t)^{s-u}$. Thus we have $y_1^u(y_1,\dots,y_t)^{s-u}\subseteq I_s(M)$ for $1\leq u\leq s$.\
Now, consider the ring $\overline{T}=T/(y_1)$ and the matrix $\overline{M}$ obtained by extending the matrix $M$ to $\overline{T}$. $\overline{M}$ follows the pattern in the hypothesis and hence by induction hypothesis on $t$, we have $I_{s}(\overline{M})=(\overline{y_2},\cdots,\overline{y_t})^{s}$. Thus we have $I_s(M) + (y_1) \supseteq (y_2,\dots,y_t)^s$. It suffices to show that for each monominal generator $m$ for $(y_2,\dots,y_t)^s$ is in $I_s(M)$. This follows immediately as $y_i$ are variables. Thus, we have $(y_1,\dots,y_t)^s\subseteq I_s(M)$.
\[mainThm\] Let $R = {{\mathbf k}}[x,y,z]$, where ${{\mathbf k}}$ is a field of characteristic $0$. Let $M$ be a $4 \times 3$ matrix whose entries are in $R$ and let $I = I_3(M)$. If $M$ is general of type $(1,2, {{\mathbf q}})$, where ${{\mathbf q}}> 2$ (see Case 2 in ), then the defining ideal of the Rees algebra $R[It]$ is minimally generated by $$\def\arraystretch{1.2}
\begin{array}{c;{2pt/2pt}c|c}
\text{~} & \text{bidegree} & \text{number of elements} \\
\hline
l_1 & (1,1) & 1 \\ \hline
l_2 & (2,1) & 1 \\ \hline
l_3 & ({{\mathbf q}},1) & 1 \\ \hline
\mathcal{A}_0 & (0, 3{{\mathbf q}}+2 ) & 1 \\ \hline
\mathcal{A}_{{{\mathbf q}}- i} & ({{\mathbf q}}-i, 3i+1 )& {2 + i \choose 2} \\ \hline
\mathcal{A}_{{\mathbf q}}& ({{\mathbf q}}, 3) & 1
\end{array}$$ for $1 \le i \le {{\mathbf q}}-1$. In particular, the defining ideal is minimally generated by ${{{\mathbf q}}+2 \choose 3} + 4$ elements.
The general property $\mathcal{P}$ needs to satisfy the following conditions.
1. $I_3(M)$ is of height $2$, and $I$ satisfies $(G_3)$, i.e, $\operatorname{ht}I_3(M) \ge 2$ and $\operatorname{ht}I_2(M) \ge 3$.
2. For all $1 \le i \le {{\mathbf q}}$, $\operatorname{rk}\operatorname{Sym}_i = 1$ or $2$, and the complex in or is exact. (This exactness can be verified by checking the height of $I (\alpha_j)$, where $\alpha_j$ are the differentials in the complex by .)
Notice that we have finitely many conditions. If each condition is general corresponding to a non-empty open subset $U_k$, then the desired non-empty subset for $\mathcal{P}$ is the intersection of the open subsets $U_k$. To show that each condition is general, first we apply (b) to show the existence of such open subsets. To finish the proof, we need to show that each open subset is not empty. This can be done by demonstrating an example satisfying the conditions in (1) and (2). We do this in . The generating degrees follow from . This completes the proof.
\[bigExample\] Let $R = {{\mathbf k}}[x,y,z]$, where ${{\mathbf k}}$ is a field of characteristic $0$, ${\mathfrak{m}}= (x,y,z)R$, and $M$ the matrix $$M = \begin{bmatrix}
x & y^2 & \gamma z^{{{\mathbf q}}} \\
y & z^2 & yz^{{{\mathbf q}}-1}\\
z & x^2 & y^{{{\mathbf q}}} \\
0 &xy + yz +xz & 0
\end{bmatrix},$$ where $\gamma \in {{\mathbf k}}\setminus \{0\}$. We will show that there exists a nonzero $\gamma$ such that the ideal $I = I_3(M)$ satisfies the desired properties required in the proof of .\
**Claim 1:** $I$ is a height two perfect ideal satisfying the $G_3$ condition.
By Krull’s principal ideal theorem, it suffices to show that $(I,x)$ is of height three. (We will use this technique multiple times in the proof.) Equivalently, we show that height $I (R/xR)$ is two. By abuse of notation, we use the same symbols for $y,z$ for their images in $\overline{R} := R/xR$. Let $$\overline{M} = \begin{bmatrix}
0 & y^2 & \gamma z^{{\mathbf q}}\\
y & z^2 & yz^{{{\mathbf q}}-1}\\
z & 0 & y^{{\mathbf q}}\\
0 & yz & 0
\end{bmatrix}.$$ Notice that from the determinant after deleting the second row of $\overline{M}$ is $\gamma y z^{{{\mathbf q}}+2}$, and the determinant after deleting the last row of $\overline{M}$ is $y^{{{\mathbf q}}+3} - y^3 z^{{{\mathbf q}}} + \gamma z^{{{\mathbf q}}+3}$. Hence any prime containing $I_3(\overline{M})$ contains $(y,z)$. Since $I_3(M) \overline{R}= I_3 (\overline{M})$, this shows that $\operatorname{ht}I \ge 2$. Then by the Hilbert-Burch theorem [@Eis Theorem 20.15], $\operatorname{ht}I = 2$.\
For the $G_3$ condition, it suffices to show that $\operatorname{ht}F_2 (M) = I_2 (M) \ge 3$ (). We will show that any prime ideal containing $I_2(M)$ is of height $3$ (indeed, it will turn out that $(x,y,z)$ is the only prime containing $I_2(M)$). Let ${\mathfrak{p}}$ be a prime ideal containing $I_2(M)$. From the minor (of rows 2,3 and columns 1,3 of $M$) $\begin{vmatrix} y & yz^{{{\mathbf q}}-1} \\ z & y^{{{\mathbf q}}} \end{vmatrix} = y^{{{\mathbf q}}+1} - yz^{{{\mathbf q}}} = y^{{{\mathbf q}}} (y-z)$, we see that $y$ or $y-z$ is in ${\mathfrak{p}}$. If $y \in {\mathfrak{p}}$, then from the minors $yx^2-z^3, x^3-zy^2$ (from the first two columns of $M$), we see that $x,z$ are also in ${\mathfrak{p}}$. So, ${\mathfrak{p}}= (x,y,z)$. If $y-z \in {\mathfrak{p}}$, then from the minor $y^3-xz^2$, we have $y^3 - xy^2 = y^2(x-y)$ is in ${\mathfrak{p}}$. If $y$ is in ${\mathfrak{p}}$, then we are done by the previous argument. So, we assume that $x-y \in {\mathfrak{p}}$. Then from the minor $x ( xy+yz+xz)$, we see that $x(3x^2) = 3x^3$ is in ${\mathfrak{p}}$ as $x-y,y-z \in {\mathfrak{p}}$. Hence $x, y-x, z-x$ is in ${\mathfrak{p}}$, and this implies ${\mathfrak{p}}= (x,y,z)$. Thus, we have $\operatorname{ht}I_2(M) = 3$, and this completes the proof of claim 1.
**Claim 2:** Let $S = {{\mathbf k}}[T_0,T_1,T_2,T_3], B = R \otimes_{{{\mathbf k}}} S = S[x,y,z]$, and $\delta = 1 + 2 + {{\mathbf q}}-3 = {{\mathbf q}}$. Then we have the following.
1. $[\operatorname{Sym}_R(I)]_0$ is $S$-free of rank $1$, and for $1 \le i \le 2$, the graded $S$-resolutions of $[\operatorname{Sym}_R(I)]_i$ are $$0 \to S(-1) \stackrel{\alpha_1}{\to} S^{3} \to [\operatorname{Sym}_R(I)]_1 \to 0$$ and $$\label{eqSym1}
0 \to S^{3}(-1) \oplus S(-1) \stackrel{\alpha_1}{\to} S^{6} \to [\operatorname{Sym}_R(I)]_2 \to 0,$$ respectively, and $\operatorname{ht}I(\alpha_1) = 3$.
2. For $3 \le i \le \delta -1$, the graded $S$-resolution of $[\operatorname{Sym}_R(I)]_i$ is $$\label{midRes}
0 \to S^{2+i-3 \choose 2}(-2) \stackrel{\alpha_2}\to S^{2+ i-1 \choose 2}(-1) \oplus S^{2+i-2 \choose 2}(-1) \stackrel{\alpha_1}{\to} S^{2 + i \choose 2} \to [\operatorname{Sym}_R(I)]_i \to 0,$$
and $\operatorname{ht}I(\alpha_1) = 3$ and $\operatorname{ht}I(\alpha_2) = 4$.
3. The graded $S$-resolution of $[\operatorname{Sym}_R(I)]_\delta=[\operatorname{Sym}_R(I)]_{{{\mathbf q}}}$ is $$0 \to S^{2+{{\mathbf q}}-3 \choose 2}(-2) \stackrel{\alpha_2}\to S^{2+ {{\mathbf q}}-1 \choose 2}(-1) \oplus S^{2+{{\mathbf q}}-2 \choose 2}(-1) \oplus S(-1) \stackrel{\alpha_1}\to S^{2 + {{\mathbf q}}\choose 2} \to[\operatorname{Sym}_R(I)]_{{\mathbf q}}\to 0,$$ and $\operatorname{ht}I(\alpha_1) = 2$.
The shape of the complexes follows from . The rank of each free module in these complexes can be calculated easily as they are the number of monomials in $(x,y,z)$ of degree $i$, and the rank of $[\operatorname{Sym}_R(I)]_i = 1$ if $i = 0,1,\delta$ and $2$ otherwise. By , to check the exactness of these complexes, it suffices to verify the claimed heights of $I(\alpha_j)$ ().\
We will use the notation $[ \, l_0 \, l_1 \, l_2] = [\, T_0 \, \dots \, T_3] \, M$. That is $$\begin{aligned}
l_0 &= xT_0+y T_1+z T_2,\\
l_1 &= y^2T_0 +z^2T_1 + x^2T_2+(xy+yz+zx) T_3,\\
l_2 &= \gamma z^{{{\mathbf q}}} T_0 + yz^{{{\mathbf q}}-1} T_1 + y^{{{\mathbf q}}}T_2.\end{aligned}$$ First, we show that $I(\alpha_2) \ge 4$ for the cases (b) and (c). By , the matrix representation of $\alpha_2$ is $$\def\arraystretch{1.8}
\left[
\begin{array}{c}
\Phi^{-l_1}_{i} \\ \hdashline[2pt/2pt]
\Phi^{l_0}_{i}
\end{array}
\right].$$ Each column consists of exactly 7 non-zero entries including $T_0,T_1,T_2,T_3$, and it satisfies the conditions of . Thus, we conclude that $\operatorname{ht}I(\alpha_2) = 4$.\
When $i = 1$, $\alpha_1$ involves only $l_0$, and the map $\alpha_1 = \Phi^{l_0}_1$ is the column matrix consists of $T_0,T_1,T_2$. Hence $\operatorname{ht}I(\alpha_1) = 3$. Notice that for $2 \le i \le \delta -1$ the free $S$-resolution in \[eqSym1,midRes\] involves only the elements $l_0,l_1$. To show that $\operatorname{ht}I(\alpha_1) \ge 3$, by Krull’s principal ideal theorem, it suffices to show that $\operatorname{ht}I(\alpha_1) + (T_3) \ge 4$. Equivalently, we show that $\operatorname{ht}I(\alpha_1)(S/(T_3)) \ge 3$. In fact, $[\operatorname{Sym}_R(I)]_i = [\operatorname{Sym}_R(J)]_i$ for $J = I_2 (M')$, where $$M' = \begin{bmatrix}
x & y^2 \\
y & z^2 \\
z & x^2 \\
\end{bmatrix}.$$
So, we may assume that $S = {{\mathbf k}}[T_0,T_1,T_2], M = M'$, and $I = J$. We show that $\operatorname{ht}\sqrt{I(\alpha_1)} = 3$. Since $[l_0~ l_1] = [T_0~ T_1~ T_2] M$, and the shape of $M$, the matrices $\Phi^{l_0}_i, \Phi^{l_1}_i$ is stable under the action by the cycle $(T_0~T_1~T_2)$. In other words, if $f(T_0,T_1,T_2) \in I(\alpha_1)$, then $f(T_1,T_2,T_0), f(T_2,T_0,T_1)$ are also in $I(\alpha_1)$. Thus, to show that $\operatorname{ht}\sqrt{I(\alpha_1)} = 3$, if suffices to show that $T_0 \in \sqrt{I(\alpha_1)}$.\
Assume to the contrary $T_0$ is not in $\sqrt{I(\alpha_1)}$. Then $\sqrt{I(\alpha_1)} S_{T_0}$ is a proper ideal of $S_{T_0}$. Consider $$g_0 = l_0 \quad g_1 = A_0 y^2+ A_1 yz + A_2 z^2,$$ where $A_0 = T_0^3 + T_1^2 T_2, A_1 = 2T_1T_2^2, A_2 = T_0^2T_1 + T_2^3$, and $\mathcal{K}' := \mathcal{K}(g_0,g_1;B)$. Even though $(g_0,g_1) \subsetneq (l_0,l_1)$ in $B$, these two ideals agree in $B_{T_0}$. (In the localization at $T_0$, $l_0 = xT_0 + yT_1 + zT_2 \Longleftrightarrow x = l_0 - yT_1/T_0 - zT_2/T_0$, so we may replace $x$ by a combination of $l_0,y,z,T_0,T_1,T_2$.) In other words, the localization of $[\mathcal{K}']_i$ at $T_0$ is also a free $S_{T_0}$-resolution for $([\operatorname{Sym}_R(I)]_i)_{T_0}$. By below, we conclude that $(A_0,A_1, A_2)S_{T_0} \in \sqrt{ I(\alpha_1') S_{T_0} }$. Notice that $(A_0,A_1, A_2)S_{T_0}$ is a unit ideal in $S_{T_0}$. Hence the Fitting lemma [@Eis Cor-Def 20.4, Cor. 20.5] imples that $I(\alpha_1)B_{T_0} = S_{T_0}$ as well. This is a contradiction and completes the proof of $\operatorname{ht}(\alpha_1) = 3$.\
It remains to show that $\operatorname{ht}I(\alpha_1) = 2$ in the case (c). Write $r = \operatorname{rank}\alpha_1$. If $T_0^r \in I(\alpha_1)$, then modding out by $(T_1,T_3)$, we see that the image of $T_2^r$ is in the image of $I(\alpha_1)$ in $S/(T_1,T_3)$. Hence we are done by Krull’s principal ideal theorem[^2]. Now, suppose that $T_0 \not\in \sqrt{I(\alpha_1)}$. We will show that $\operatorname{ht}\sqrt{I(\alpha_1)}S_{T_0} = 2$. By the assumption, $ \sqrt{I(\alpha_1)}S_{T_0} \neq S_{T_0}$. As in the previous case, we will use $$g_0 = l_0, g_1 = A_0 y^2+ A_1 yz + A_2 z^2, g_2 = l_2,$$ where $A_0 = T_0^3 + T_1^2 T_2, A_1 = 2T_1T_2^2, A_2 = T_0^2T_1 + T_2^3$, and $\mathcal{K}' := \mathcal{K}(g_0,g_1,g_2;B)$. Then the Koszul complex $\mathcal{K}'$ in degree ${{\mathbf q}}$ is a free $S_{T_0}$-resolution of $[\operatorname{Sym}_R(I)]_{{{\mathbf q}}}$. If we name $\alpha_1'$ the first differential of $[\mathcal{K}']_{{{\mathbf q}}}$, then $I(\alpha_1)S_{T_0} = I(\alpha_1')S_{T_0}$ since they are both the Fitting ideal of the same module. Therefore, we will show that $\operatorname{ht}I(\alpha_1')S_{T_0} = 2$. Notice that $\alpha'$ is $$\def\arraystretch{1.75}
\left[
\begin{array}{c;{2pt/2pt}c;{2pt/2pt}c}
\Phi^{g_0}_{{\mathbf q}}& \Phi^{g_1}_{{\mathbf q}}& \Phi^{g_2}_{{\mathbf q}}\end{array} \right]
=
\left. {\vphantom{{\begin{matrix}1\\2\\3\\4\\5\\6\\7\\8\\9\\10 \end{matrix}}}}\right[
\overbrace{
\underbrace{
\begin{array}{cccc}
T_0 & 0 & \cdots & 0 \\
T_1 & T_0 & \cdots & 0 \\
\vdots & \vdots & \ddots & 0\\
* & * & \cdots & T_0 \\
\hdashline[1pt/1pt]~
* & * & \cdots & * \\
\vdots & \vdots & \ddots & \vdots \\
\vdots & \vdots & \ddots & \vdots \\
\vdots & \vdots & \ddots & \vdots \\
\vdots & \vdots & \ddots & \vdots \\
* & * & \cdots & *
\end{array}
}_{ {2 + {{\mathbf q}}- 1\choose {{\mathbf q}}- 1}~\text{columns}} }^{\Phi^{g_0}_{{{\mathbf q}}}} \hspace{-1em}\begin{array}{c;{2pt/2pt}c}
~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\~&~\\~&~\\
\end{array}\hspace{-1em}
\overbrace{
\begin{array}{ccc}
* & \cdots & *\\
* & \cdots & *\\
\vdots & \vdots & \vdots \\
\vdots & \vdots & \vdots \\
\hdashline[1pt/1pt]~
\vdots & \vdots & \vdots \\
\vdots & \vdots & \vdots \\
\vdots & \vdots & \vdots \\
\vdots & \vdots & \vdots \\
\vdots & \vdots & \vdots \\
* & \cdots & *
\end{array}
\hspace{-1em}\begin{array}{c;{2pt/2pt}c}
~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\~&~\\~&~\\
\end{array}\hspace{-1em}
\underbrace{
\begin{array}{cccc}
0 & 0 & \cdots & 0 \\
\vdots & \vdots & \cdots & \vdots \\
\vdots & \vdots & \cdots & \vdots \\
\vdots & \vdots & \cdots & \vdots \\
\hdashline[1pt/1pt]~
A_0 & 0 & \cdots & 0 \\
A_1 & A_0 & \cdots & 0 \\
A_2 & A_1 & \ddots & 0 \\
\vdots & \vdots & \vdots & \vdots \\
0 & 0 & \cdots & A_1 \\
0 & 0 & \cdots & A_2
\end{array}
}_{ {1 + {{\mathbf q}}-2 \choose {{\mathbf q}}-2 }~\text{columns}}
}^{\Phi^{g_1}_{{{\mathbf q}}}}
\hspace{-1em}\begin{array}{c;{2pt/2pt}c}
~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\~&~\\~&~\\
\end{array}\hspace{-1em}
\overbrace{
\begin{array}{c}
0 \\ \vdots \\ \vdots \\ \vdots \\ \hdashline[1pt/1pt]~ T_2 \\ 0 \\ 0 \\ \vdots \\ T_1 \\ \gamma T_0
\end{array}
}^{\Phi^{g_2}_{{\mathbf q}}}
\left. {\vphantom{{\begin{matrix}1\\2\\3\\4\\5\\6\\7\\8\\9\\10 \end{matrix}}}}\right]
\hspace{-1em}\begin{array}{c;{2pt/2pt}c}
~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\~&~\\~&~\\
\end{array}\hspace{-1em}
\overbrace{
\begin{array}{c}
x^{{\mathbf q}}\\ x^{{{\mathbf q}}-1}y \\ \vdots \\ x z^{{{\mathbf q}}-1} \\ \hdashline[1pt/1pt] y^{{\mathbf q}}\\ y^{{{\mathbf q}}-1}z \\ \vdots \\ \vdots \\ y z^{{{\mathbf q}}-1} \\ z^{{\mathbf q}}.
\end{array}
}^{basis}$$ Since $T_0$ is invertible in $S_{T_0}$, $I(N) \subset I(\alpha') S_{T_0}$, where $N$ is the matrix obtained from the lower right corner of $\alpha'$. $$N =
\left.
\hspace{-1em}\begin{array}{cc}
~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\
\end{array}\hspace{-1em}
\right[
\begin{array}{ccccc}
A_0 & 0 & \cdots & 0 & 0 \\
A_1 & A_0 & \cdots & 0 & 0 \\
A_2 & A_1 & \ddots & 0 & 0 \\
\vdots & \vdots & \vdots & \vdots & \vdots \\
0 & 0 & \cdots & A_2 & A_1 \\
0 & 0 & \cdots & 0 & A_2
\end{array}
\hspace{-1em}\begin{array}{c;{2pt/2pt}c}
~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\
\end{array}\hspace{-1em}
\left. \begin{array}{c}
T_2 \\ 0 \\ 0 \\ \vdots \\ T_1 \\ \gamma T_0
\end{array} \right].$$ We will show that after modifying $N$, $\operatorname{ht}I(N) \ge 2$, so $\operatorname{ht}I(\alpha_1')S_{T_0} = 2$ by the containment above. (We note that $N$ and its variations $N', N'', N_T$ below will be matrices of rank $j$ and size $(j + 1) \times j$ (with different $j$).) Since $T_0$ is invertible in $S_{T_0}$, we may multiply the last row by $-T_2/\gamma T_0$ and add it to the first row. Also, we multiply the last row by $-T_1/\gamma T_0$ and add it to the second last row. We call this new matrix $N'$ $$N' =
\left[
\begin{array}{ccccc;{2pt/2pt}c}
A_0 & 0 & \cdots & 0 & -A_2 \, T_2/\gamma T_0 & 0\\
A_1 & A_0 & \cdots & 0 & 0 & 0 \\
A_2 & A_1 & \ddots & 0 & 0 &0 \\
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots \\
0 & 0 & \cdots & A_2 & A_1 - A_2T_1/\gamma T_0& 0 \\
\hdashline[1pt/1pt]
0 & 0 & \cdots & 0 & A_2 & \gamma T_0
\end{array}
\right]
:=
\left[
\begin{array}{c;{2pt/2pt}c}
N'' & \mathbf{0}\\
\hdashline[1pt/1pt]
*& \gamma T_0
\end{array}
\right].$$ Here $N''$ is the matrix obtained by deleting the last row and column of $N'$. It is straightforward to check that $I(N') = I(N'')$ as $T_0$ is a unit in $S_{T_0}$. Consider the following matrix $N_{T}$ in $S_{T_0}[T]$ (here, T is a new variable) $$N_{T} =
\left[
\begin{array}{ccccc}
A_0 T & 0 & \cdots & 0 & -A_2 \, T_2/ (\gamma T_0) \\
A_1 & A_0 & \cdots & 0 & 0 \\
A_2 & A_1 & \ddots & 0 & 0 \\
\vdots & \vdots & \vdots & \vdots & \vdots \\
0 & 0 & \cdots & A_1 & A_0 \\
0 & 0 & \cdots & A_2 & A_1 - A_2T_1/\gamma T_0
\end{array}
\right]$$ and let $J_T := I(N_T)$. Since $J_1 \subseteq I(\alpha_1')S_{T_0}[T]$, $\operatorname{ht}(J_T) < \infty$. We claim that $\operatorname{ht}J_T \ge 2$. We will prove this by showing that $\sqrt{(J_T, T)} = (T_1,T_2,T)$ is of height $3$. Once we have proven the claim, then by , there exists a non-empty open subset $U$ of $\operatorname{Spec}k[T]$ such that for any (closed) point $\gamma'$ in $U$, $\operatorname{ht}J_{\gamma'} \ge 2$.\
First of all, we have $(J_T, T) \subset (A_1, A_2, T)$ since the ideal generated by $A_1,A_2,T$ contains all the entries of the first column. Furthermore, if a prime ${\mathfrak{p}}$ contains $(A_1,A_2,T) = (2T_1T_2^2, T_0^2T_1 + T_2^3, T)$, then $T_1,T_2,T$ are in ${\mathfrak{p}}$. (Recall that we are in $S_{T_0}$, so $T_0$ is a unit.) Thus, we conclude that $\sqrt{ (A_1, A_2, T) }= (T_1,T_2,T)$. Consider the minor obtained by deleting the second row of $N_T$ and then putting the first row last $$\label{labA2}
\left|
\begin{array}{ccccccc}
A_2 & A_1 & 0 & 0 &\cdots & 0 \\
0 & A_2 & A_1 & A_0 & \cdots & 0 \\
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots \\
0 & 0 & \cdots & A_2 & A_1 & A_0 \\
0 & 0 & \cdots & 0 & A_2 & A_1 - A_2T_1/\gamma T_0 \\
A_0 T & 0 & \cdots & 0 & 0 & -A_2 \, T_2/ (\gamma T_0) \\
\end{array}
\right|.$$ From this minor, we conclude that $A_2^{\operatorname{rank}N_T}T_2/ (\gamma T_0)$ is in $(J_T,T)$. Therefore, if ${\mathfrak{p}}$ is a prime ideal containing $(J_T,T)$, then it contains $A_2$ or $T_2$. If $A_2 \in {\mathfrak{p}}$, then from the minor obtained by deleting the first row of $N_T$, we have $A_1 \in {\mathfrak{p}}$. In this case, we have $\sqrt{(J_T,T)} = \sqrt{(A_1,A_2,T)} = (T_1,T_2,T) \subset {\mathfrak{p}}$, so $(T_1,T_2,T) = {\mathfrak{p}}$ since $(T_1,T_2,T)$ is a prime of height $3$. We show that we have the same conclusion in the case where $T_2 \in {\mathfrak{p}}$. Assume $T_2 \in {\mathfrak{p}}$, then since $A_1 = 2T_1T_2^2$, $A_1$ is in ${\mathfrak{p}}$. Hence from the minor obtained by deleting the first row of $N_T$ and from the fact that $T, T_2, A_1 \in {\mathfrak{p}}$, the determinant $$\left| \begin{array}{cccccc}
0 & A_0 & 0 & 0&\cdots & 0 \\
A_2 & 0 & A_0 &0 &\cdots & 0 \\
0 & A_2 & 0 & 0&\cdots & 0 \\
\vdots & \vdots & \vdots&\vdots & \vdots & \vdots\\
0 & 0 & \cdots &A_2& 0& A_0\\
0 & 0 & \cdots &0& A_2 & - A_2T_1/\gamma T_0
\end{array} \right|$$ is in ${\mathfrak{p}}$. By , $A_2^{{{\mathbf q}}/2}A_0^{{{\mathbf q}}/2}$ is in ${\mathfrak{p}}$ if ${{\mathbf q}}$ is even, and $A_2^{({{\mathbf q}}-1)/2}A_0^{({{\mathbf q}}-1)/2}(- A_2T_1/\gamma T_0)$ is in ${\mathfrak{p}}$ if ${{\mathbf q}}$ is odd. Since $T_0^3 \in (A_0, T_2) \subset (A_0, {\mathfrak{p}})$ is a unit ideal, $A_0 \not\in {\mathfrak{p}}$. Therefore, in both cases we have $A_2$ or $T_1$ is in ${\mathfrak{p}}$. Since $T_2$ is in ${\mathfrak{p}}$ and $T_0$ is a unit, $A_2 = T_0^2T_1 +T_2^3$ is in ${\mathfrak{p}}$ if and only if $T_1$ is in ${\mathfrak{p}}$. Hence in both cases we have $A_2 \in {\mathfrak{p}}$. We already showed that if $A_2 \in {\mathfrak{p}}$, then ${\mathfrak{p}}= (T_1,T_2,T)$ (in the paragraph before \[labA2\]). Hence we have $\sqrt{(J_T,T)} = \sqrt{(A_2,A_2,T)} = (T_1,T_2,T)$, and this proves the claim.\
By , there exists a non zero $\gamma'$ in ${{\mathbf k}}$ such that $\operatorname{ht}N_{\gamma'} = 2$. But the maximal minors of the following matrices $$N_{\gamma'} =
\left[
\begin{array}{ccccc}
A_0 \gamma' & 0 & \cdots & 0 & -A_2 \, T_2/ (\gamma T_0) \\
A_1 & A_0 & \cdots & 0 & 0 \\
A_2 & A_1 & \ddots & 0 & 0 \\
\vdots & \vdots & \vdots & \vdots & \vdots \\
0 & 0 & \cdots & A_2 & A_1
\end{array}
\right],
\quad
N'' =
\left[
\begin{array}{ccccc}
A_0 & 0 & \cdots & 0 & -A_2 \, T_2/ (\gamma \gamma' T_0) \\
A_1 & A_0 & \cdots & 0 & 0 \\
A_2 & A_1 & \ddots & 0 & 0 \\
\vdots & \vdots & \vdots & \vdots & \vdots \\
0 & 0 & \cdots & A_2 & A_1
\end{array}
\right]$$ generate the same ideal. We replace $\gamma$ by $\gamma \gamma'$. This does not change our argument since the previous calculation and proof depended only on the fact that $\gamma$ was not zero. Therefore, with new $\gamma$, we obtain the desired height for $I (N)$, and this completes the proof.
\[23case\] Let $S = {{\mathbf k}}[T_0,T_1,T_2]$, where ${{\mathbf k}}$ is a field, $B = S[x,y,z], {\mathfrak{m}}= (x,y,z)B, g_0 = T_0x + T_1 y+ T_2 z, g_1= A_0 y^2+ A_1 yz + A_2 z^2$ with $A_i \in S$. If $g_0,g_1$ form an $B$-regular sequence, then for any $i \ge 2$, the ith graded component (with respect to $x,y,z$ degree) of $B/(g_0,g_1)$ has a representation $\alpha$ of the form $$\def\arraystretch{1.8}
\left[
\begin{array}{c;{2pt/2pt}c}
\Phi^{g_0}_{i} & \Phi^{g_1}_{i}
\end{array} \right]
=
\left. {\vphantom{{\begin{matrix}1\\2\\3\\4\\5\\6\\7\\8\\9 \end{matrix}}}}\right[
\overbrace{
\underbrace{
\begin{array}{cccc}
T_0 & 0 & \cdots & 0 \\
T_1 & T_0 & \cdots & 0 \\
\vdots & \vdots & \ddots & 0\\
* & * & \cdots & T_0 \\
\hdashline[1pt/1pt]~
* & * & \cdots & * \\
\vdots & \vdots & \ddots & \vdots \\
\vdots & \vdots & \ddots & \vdots \\
\vdots & \vdots & \ddots & \vdots \\
* & * & \cdots & *
\end{array}
}_{ {2 + i - 1\choose i - 1}~\text{columns}} }^{\Phi^{g_0}_{i}} \hspace{-1em}\begin{array}{c;{2pt/2pt}c}
~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\~&~\\
\end{array}\hspace{-1em}
\overbrace{
\begin{array}{ccc}
* & \cdots & *\\
* & \cdots & *\\
\vdots & \vdots & \vdots \\
\vdots & \vdots & \vdots \\
\hdashline[1pt/1pt]~
\vdots & \vdots & \vdots \\
\vdots & \vdots & \vdots \\
\vdots & \vdots & \vdots \\
\vdots & \vdots & \vdots \\
* & \cdots & *
\end{array}
\hspace{-1em}\begin{array}{c;{2pt/2pt}c}
~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\~&~\\
\end{array}\hspace{-1em}
\underbrace{
\begin{array}{cccc}
0 & 0 & \cdots & 0 \\
\vdots & \vdots & \cdots & \vdots \\
\vdots & \vdots & \cdots & \vdots \\
\vdots & \vdots & \cdots & \vdots \\
\hdashline[1pt/1pt]~
A_0 & 0 & \cdots & 0 \\
A_1 & A_0 & \cdots & 0 \\
A_2 & A_1 & \ddots & 0 \\
\vdots & \vdots & \vdots & \vdots \\
0 & 0 & \cdots & A_2
\end{array}
}_{ {1 + i -2 \choose i-2 }~\text{columns}}
}^{\Phi^{g_1}_{i}}
\left. {\vphantom{{\begin{matrix}1\\2\\3\\4\\5\\6\\7\\8\\9 \end{matrix}}}}\right]
\hspace{-1em}\begin{array}{c;{2pt/2pt}c}
~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\ ~&~\\~&~\\
\end{array}\hspace{-1em}
\overbrace{
\begin{array}{c}
x^i \\ x^{i-1}y \\ \vdots \\ x z^{i-1} \\ \hdashline[1pt/1pt] y^i \\ y^{i-1}z \\ \vdots \\ \vdots \\ z^i
\end{array}
}^{basis}
$$ Furthermore, we have $\operatorname{rk}\alpha = {2+i-1 \choose 2} + {i-1}$ and $T_0^{2+i-1 \choose 2}(A_0,A_1,A_2)^{i-1} \subset I(\alpha)$.
We use the polynomial grading on $B$ and write $M := B/(g_0,g_1)$. Since $g_0,g_1$ form a regular sequence, the Koszul complex $\mathcal{K}:=\mathcal{K}(g_0,g_1;B)$ is a grade $B$-resolution of $M$. By , $\mathcal{K}_i$ is a free $S$-resolution of $M_i$. That is $$0 \to S^{2+i-3 \choose 2} \stackrel{\alpha_2}\to S^{2+ i-1 \choose 2} \oplus S^{2+i-2 \choose2} \stackrel{\alpha}{\to} S^{2 + i \choose 2} \to M_i \to 0$$ is exact. Thus, $$\begin{aligned}
\operatorname{rk}M_i &= {2 + i \choose 2} + {2+i-3 \choose 2} - {2+ i-1 \choose 2} - {2+i-2 \choose 2} \\
&= \left[{2 + i \choose 2} - {2+ i-1 \choose 2}\right] - \left[{2+i-2 \choose2} - {2+i-3 \choose 2}\right] \\
&= {2 + i - 1 \choose 1} - {2 + i -3 \choose 1} = 2,\end{aligned}$$
and $\operatorname{rk}\alpha = {2 + i \choose 2} - 2$. The presentation follows by using the Lex order on the monomials (in terms of $x,y,z$) of $B$. Since $$\begin{aligned}
{2+i-1 \choose 2} + \dim_{{{\mathbf k}}} (y,z)^{i-2} &=(i+1)i/2 + (i -1) = (i^2 + 3i)/2 - 1 = (i^2+3i+2)/2 - 2 \\
&= (i+2)(i+1)/2 - 2={2+i \choose 2}-2 = \operatorname{rank}\alpha,\end{aligned}$$ we see that the $T_0^{2+i-1 \choose 2}I(N) \subset I(\alpha)$, where $$N = \left[ \begin{array}{cccc}
A_0 & 0 & \cdots & 0 \\
A_1 & A_0 & \cdots & 0 \\
A_2 & A_1 & \cdots & 0 \\
\vdots & \vdots & \vdots & \vdots \\
0 & 0 & \cdots & A_2
\end{array} \right].$$ By , $I(N) = (A_0,A_1,A_2)^{i-1}$, and this proves the statement.
\[lemOffDiag\] Let $x,y,z$ be variables over $\mathbb{Z}$. Then we have $$\left| \begin{array}{cccccc}
0 & x & 0 & 0&\cdots & 0 \\
y & 0 & x &0 &\cdots & 0 \\
0 & y & 0 & x&\cdots & 0 \\
\vdots & \vdots & \vdots&\vdots & \vdots & \vdots\\
0 & 0 & \cdots &y& 0& x\\
0 & 0 & \cdots &0& y& 0
\end{array} \right|
= \begin{cases}
(-1)^{\frac n2}x^{\frac n2} y^{\frac n2}, & if~n~is~even \\
0, & if~n~is~odd
\end{cases}
~~~~~and~
\left| \begin{array}{cccccc}
0 & x & 0 & 0&\cdots & 0 \\
y & 0 & x &0 &\cdots & 0 \\
0 & y & 0 & x&\cdots & 0 \\
\vdots & \vdots & \vdots&\vdots & \vdots & \vdots\\
0 & 0 & \cdots &y& 0& x\\
0 & 0 & \cdots &0& y& z
\end{array} \right|
= \begin{cases}
(-1)^{\frac n2} x^{\frac n2} y^{\frac n2}, & if~n~is~even \\
(-1)^{\frac{n-1}2}x^{\frac{n-1}2} y^{\frac{n-1}2} z, & if~n~is~odd.
\end{cases}$$
For the first determinant, one can use induction on the size of the matrix, and for the second determinant, one can use the Laplace expansion along the last row (or column) and use the result of the first determinant.
The non-empty open subset for the conditions in is a proper subset of the parameter space. The following matrix provides an example where $I:= I_3(M)$ satisfies condition (1) in the proof of , but not condition (2). $$M =
\begin{bmatrix}
x &0 &x^3\\
y &x^2 &0\\
0 &y^2 &z^3\\
0 &z^2 &0
\end{bmatrix}.$$ It is interesting that the rational map induced by $I$ is birational to its image.
We will end the paper with a couple of questions.
\[lastQues\]
1. Let $U$ be the non-empty open subset corresponds to the general condition in . Can we determine the closed set which is the complement of $U$? In the case where $R = {{\mathbf k}}[x_1,x_2]$, Kustin, Polini, Ulrich were able to connect this condition to a geometric condition (See [@KPU17 Lemma 2.10]).\
2. In the general case where $R = {{\mathbf k}}[x_1,\dots, x_n]$, does the following presentation matrix (or its variations) of $I = I_n(\varphi)$ $$\varphi = \begin{bmatrix}
x_1^{d_1} & x_1^{d_2} & \cdots & x_1^{d_n} \\
x_2^{d_1} & x_2^{d_2} & \cdots & x_2^{d_n} \\
\vdots & \vdots & \ddots & \vdots \\
x_n^{d_1} & x_n^{d_n} & \cdots & x_n^{d_n} \\
g_1 & \cdots & \cdots & g_n
\end{bmatrix},$$ where $g_i$ are symmetric polynomials of degree $d_i$, provide an example of defining a similar general property described in ?\
Note that the corresponding presentation matrices for the graded pieces of $\operatorname{Sym}_R(I)$ are invariant under the action of the cyclic group generated by the cycle $(T_0, \dots, T_{n-1})$. We hope an expert can utilize this fact and provide an alternative way to verify . Recall that the more general $\varphi$ becomes, the more challenging to verify conditions.
[CBtD14]{}
Maurice Auslander and Mark Bridger. . Memoirs of the American Mathematical Society, No. 94. American Mathematical Society, Providence, R.I., 1969.
Laurent Busé, Marc Chardin, and Jean-Pierre Jouanolou. Torsion of the symmetric algebra and implicitization. , 137(6):1855–1865, 2009.
David A. Buchsbaum and David Eisenbud. What makes a complex exact? , 25:259–268, 1973.
David A. Buchsbaum and David Eisenbud. Some structure theorems for finite free resolutions. , 12:84–139, 1974.
David A. Buchsbaum and David Eisenbud. What annihilates a module? , 47(2):231–243, 1977.
Winfried Bruns and J[ü]{}rgen Herzog. , volume 39 of [*Cambridge Studies in Advanced Mathematics*]{}. Cambridge University Press, Cambridge, 1993.
Laurent Busé and Jean-Pierre Jouanolou. On the closed image of a rational map and the implicitization problem. , 265(1):312–357, 2003.
Jacob A. Boswell and Vivek Mukundan. Rees algebras and almost linearly presented ideals. , 460:102–127, 2016.
Laurent Busé. On the equations of the moving curve ideal of a rational algebraic plane curve. , 321(8):2317–2344, 2009.
Teresa Cortadellas Bení tez and Carlos D’Andrea. Rational plane curves parameterizable by conics. , 373:453–480, 2013.
Teresa Cortadellas Bení tez and Carlos D’Andrea. Minimal generators of the defining ideal of the [R]{}ees algebra associated with a rational plane parametrization with [$\mu=2$]{}. , 66(6):1225–1249, 2014.
David Cox, J. William Hoffman, and Haohao Wang. Syzygies and the [R]{}ees algebra. , 212(7):1787–1796, 2008.
A.V. Doria, S.H. Hassanzadeh, and A. Simis. A characteristic-free criterion of birationality. , 230(1):390 – 413, 2012.
David Eisenbud. , volume 150 of [*Graduate Texts in Mathematics*]{}. Springer-Verlag, New York, 1995. With a view toward algebraic geometry.
David Eisenbud. , volume 229 of [*Graduate Texts in Mathematics*]{}. Springer-Verlag, New York, 2005. A second course in commutative algebra and algebraic geometry.
Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in algebraic geometry. Available at <https://faculty.math.illinois.edu/Macaulay2/>.
Robin Hartshorne. . Springer-Verlag, New York-Heidelberg, 1977. Graduate Texts in Mathematics, No. 52.
Joe Harris. , volume 133 of [*Graduate Texts in Mathematics*]{}. Springer-Verlag, New York, 1992. A first course.
Craig Huneke and Irena Swanson. , volume 336 of [*London Mathematical Society Lecture Note Series*]{}. Cambridge University Press, Cambridge, 2006.
J. Herzog, A. Simis, and W. V. Vasconcelos. Approximation complexes of blowing-up rings. , 74(2):466–493, 1982.
J. Herzog, A. Simis, and W. V. Vasconcelos. Koszul homology and blowing-up rings. In [*Commutative algebra ([T]{}rento, 1981)*]{}, volume 84 of [ *Lecture Notes in Pure and Appl. Math.*]{}, pages 79–169. Dekker, New York, 1983.
Jooyoun Hong, Aron Simis, and Wolmer V. Vasconcelos. On the homology of two-dimensional elimination. , 43(4):275–292, 2008.
Craig Huneke. Strongly [C]{}ohen-[M]{}acaulay schemes and residual intersections. , 277(2):739–763, 1983.
Jean-Pierre Jouanolou. Résultant anisotrope, compléments et applications. , 3(2):Research Paper 2, approx. 91 pp. (electronic), 1996. The Foata Festschrift.
Andrew R. Kustin, Claudia Polini, and Bernd Ulrich. Rational normal scrolls and the defining equations of [R]{}ees algebras. , 650:23–65, 2011.
Andrew R. Kustin, Claudia Polini, and Bernd Ulrich. Blowups and fibers of morphisms. , 224(1):168–201, 2016.
Andrew Kustin, Claudia Polini, and Bernd Ulrich. The bi-graded structure of symmetric algebras with applications to [R]{}ees rings. , 469:188–250, 2017.
Nguyen P. H. Lan. On [R]{}ees algebras of linearly presented ideals. , 420:186–200, 2014.
Jeff Madsen. Equations of rees algebras of ideals in two variables, 2015.
Susan Morey. Equations of blowups of ideals of codimension two and three. , 109(2):197–211, 1996.
Susan Morey and Bernd Ulrich. Rees algebras of ideals with low codimension. , 124(12):3653–3661, 1996.
Claudia Polini and Yu Xie. -multiplicity and depth of associated graded modules. , 379:31–49, 2013.
Hà Huy Tài. On the [R]{}ees algebra of certain codimension two perfect ideals. , 107(4):479–501, 2002.
Bernd Ulrich. Artin-[N]{}agata properties and reductions of ideals. In [*Commutative algebra: syzygies, multiplicities, and birational algebra ([S]{}outh [H]{}adley, [MA]{}, 1992)*]{}, volume 159 of [*Contemp. Math.*]{}, pages 373–400. Amer. Math. Soc., Providence, RI, 1994.
Bernd Ulrich and Wolmer V. Vasconcelos. The equations of [R]{}ees algebras of ideals with linear presentation. , 214(1):79–92, 1993.
[^1]: Such construction holds for arbitrary set of homogeneous ideals of the same degree in $S$, but this is the set up we will work in this paper.
[^2]: Computations using Macaulay2 [@M2] for small ${{\mathbf q}}$ shows that this case does not happen, but this is necessary for our proof.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The mathematical modelling of species transport in the turbulent boundary layer of fluids that precipitate on the wall, is an important topic at the heart of one of the biggest challenges in efficient energy utilization in all process industries; namely the fouling of heat exchangers. A major contributor to the complexity of the problem is the multi-length-scale nature of the governing phenomena. That is, transport mechanisms dominating at the nano-scale may be responsible for the macroscopic performance of the industrial process. This paper addresses some of the challenges that need to be met in modelling the boundary conditions, i.e. the atomic/molecular-scale conditions, for the species-specific mass conservation equations, at the wall, for single-phase, multi-component fluids that precipitate at the wall.'
author:
- 'Sverre G. Johnsen , Tiina M. Pääkkönen, Stefan Andersson, Stein Tore Johansen,'
- Bernd Wittgens
bibliography:
- 'References.bib'
title: 'On the wall boundary conditions for species-specific mass conservation equations in mathematical modelling of direct precipitation fouling from supersaturated, multi-component fluid mixtures'
---
Authors should not enter keywords on the manuscript, as these must be chosen by the author during the online submission process and will then be added during the typesetting process (see http://journals.cambridge.org/data/relatedlink/jfm-keywords.pdf for the full list)
Introduction {#sec:introduction}
============
Fouling of solid surfaces and heat exchanger surfaces in particular, is a common and much studied problem in most process industries [@MullerSteinhagen11]. Fouling is defined as the unwanted accumulation of solid (or semi-solid) material on solid surfaces. Similar is the desired accumulation of solids e.g. in chemical vapour deposition (CVD) [@Kleijn89; @Krishnan94]. A common and costly problem in many industrial applications is the direct precipitation of super saturated fluids on heat exchanger surfaces. Typical examples are found in e.g. waste incineration, metal production, or in power plants, where efficient heat recovery is key to sustainable production, and a combination of direct precipitation and deposition of e.g. solid metal oxides is a major show-stopper. By precipitation, we understand all types of phase transitions from a fluid to a relatively denser phase, e.g., gas $\to$ liquid (condensation), gas $\to$ solid (sublimation), liquid $\to$ solid (solidification). Our work is relevant for both physisorption and chemisorption.
In previous papers, we have presented the mathematical modelling of solid deposition processes [@Johansen91; @Johnsen09; @Johnsen15]. The most recent work [@Johnsen15] has been on developing a generic modelling framework for the dissolved species mass transport through the turbulent, reactive boundary layer of multi-component fluid mixtures that precipitate on the wall. The modelling is based on Maxwell-Stefan diffusion in multi-component mixtures, and the turbulent, single-phase Navier-Stokes equations. The governing equations are simplified in accordance with common assumptions of computational fluid dynamics (CFD), and the developed framework can be employed as a sub-grid model for direct precipitation/ crystallization/solidification fouling in coarse grid CFD models.
The deposition rate is determined by transportation processes taking place at three main length scales (see figure \[fig:Fig1\]): 1) The nano-scale determines if deposits will stick to the wall as well as the rate of phase-change at the wall. 2) The micro-scale determines the efficiency of turbulent/diffusive transport through the laminar sub layer close to the wall. 3) The meso-scale determines the bulk conditions at the outskirts of the turbulent boundary layer and transport in the log-layer. Our work has so far focused on the micro- and meso-scales. This paper discusses some of the aspects of the processes taking place at the nano-scale, and argues that these are crucial for the accurate/predictive modelling of precipitation from industry-scale process streams.
The main motivation for this paper is to discuss the importance of understanding the detailed precipitation kinetics taking place at the wall, in order to utilize correct boundary conditions for the species-specific mass conservation equations. In order to do this, we need to understand how the boundary conditions for the non-depositing species arise, and what the main differences between the non-depositing and depositing species are, with respect to boundary conditions. Finally, we suggest a way forward in how to obtain useful boundary conditions for the depositing species.
Throughout the paper we will employ the notation wall, when discussing the properties at the wall-fluid interface. That is, even if a thick deposit layer is present between the actual wall and the flowing fluid, and a more accurate term would be e.g. the *moving* or *time-evolving solid-fluid interface*, we will, for consistency, use the index $w$ and refer to the wall when discussing the boundary condition seen by the fluid, at the solid-fluid interface. Thus, it should be kept in mind that the wall properties, as we are referring to them, are, in reality, not constant values, but rather time-dependent and depending on the deposit layer thickness, porosity, density, etc. It is out of scope for the current paper to go into all these details, however, and wall properties will be treated as constants without further consideration of the exact physical or chemical nature of the solid-fluid interface.
![ Illustration to demonstrating the various length scales applying in fouling, in the context of the “law of the wall”, for the velocity parallel to the wall. The over-all mass flux, $j_{BL,\delta}$, represents the transport of a depositing species, $\delta$, through the turbulent boundary layer, and the interface reaction rate, $j_{IR,\delta}$, represents the mass integration flux of the species, $\delta$, into the deposit layer at the wall. []{data-label="fig:Fig1"}](Fig1){width="\figwidth"}
Theory {#sec:theory}
======
We are considering a single-phase fluid mixture consisting of a set, $\Omega$, of $N$ unique, distinguishable, chemical species. The set of species furthermore consists of two non-intersecting subsets, namely the set of depositing species, $\Delta$, and the set of non-depositing species, $H$, such that $\Omega=\Delta\cup H$, but $\Delta\cap H=\varnothing$. We will reserve the indexes $\delta \in \Delta $ and $\eta \in H $ to denote, exclusively, the species in the sets of depositing and non-depositing species, respectively. It is assumed that each species field, hence the fluid itself, can be modelled as a continuum. This implies that species properties can be considered as continuously varying physical fields that are well defined throughout the fluid domain. Furthermore, we assume homogeneous mixing in the sense that local species properties are taken as volume averages over infinitesimal volumes. These assumptions allow us to utilize differential calculus in deriving governing equations for the species transport.
Definitions {#sec:definitions}
-----------
An arbitrary species, contained in $\Omega$, and denoted by $i\in \left\{ 1,\ldots ,N \right\}$ moves with the local convective velocity ${\mathbi{u}}_i$, and its local convective mass flux is defined as ${\mathbi{j}}_i=\rho_i{\mathbi{u}}_i$, where $\rho_i$ is the species’ mass concentration. The advective mass flux is given by ${\mathbi{j}}=\sum\nolimits_{i=1}^{N}{{\mathbi{j}}_i}=\rho_f{\mathbi{u}}_f$, where $\rho_f\equiv\sum\nolimits_{i=1}^{N}\rho_i$is the fluid mass density. The mass-averaged advective velocity is defined as ${\mathbi{u}}_f\equiv\sum\nolimits_{i=1}^{N}{X_i{\mathbi{u}}_i}$, where the species mass fractions are expressed as $X_i=\nicefrac{\rho_i}{\rho_f}$ and can be related to the species mole fractions, $z_i$, via $X_i=\left(\nicefrac{M_{w,i}}{\overline{M}_{w}}\right)z_i$, where $M_{w,i}$ are the species molar masses, and the mean molar mass is given by $\overline{M}_w=\sum\nolimits_{i}{M_{w,i}z_i}$. The species’ diffusive mass flux is given by the difference between the species’ convective and advective mass fluxes; $$\label{eq:difffluxdef}
{\mathbi{j}}_{i,d}=\rho_i\left({\mathbi{u}}_i-{\mathbi{u}}_{f}\right)~.$$ It follows that the diffusive mass fluxes sum to zero.
Mass conservation equation {#sec:mass_conservation_equation}
--------------------------
The steady-state, mass conservation equation for the species $i$ is the Advection-Diffusion Equation, $$\label{eq:ADE}
{\mathbi{\nabla}}\boldsymbol\cdot\left(\rho_i{\mathbi{u}}_{i}\right) = {\mathbi{\nabla}}\boldsymbol\cdot\left(\rho_fX_i{\mathbi{u}}_f\right) + {\mathbi{\nabla}}\boldsymbol\cdot{\mathbi{j}}_{i,d} = {{R}_{i}}~,$$ which, by summation over the species index, $i$, becomes the continuity equation for the fluid mixture,${\mathbi{\nabla}}\left(\rho_f{\mathbi{u}}_f\right)=0$. $R_i$ is a source-term due to chemical reactions.
Diffusion {#sec:diffusion}
---------
The diffusive mass flux of the species $i$ can be modelled by Maxwell-Stefan theory, $$\label{eq:MSdiff}
{\mathbi{j}}_{i,d}=-\rho_f\sum\nolimits_{j=1}^{N-1}{D_{ij}{\mathbi{d}}_j}~,$$ where the $D_{ij}$ are the binary diffusivities of components $i$ in components $j$, and the driving force vector is proportional to the gradient of a scalar potential energy field, namely the species’ total chemical potential, ${\mathbi{d}}_j=\left(\nicefrac{1}{\mathcal{R}T}\right){\mathbi{\nabla}}\mu_{tot,j}$. It follows from the definitions above, that the diffusive mass flux of the $N$th, dependent species is given by ${\mathbi{j}}_{N,d}=-\sum\nolimits_{i=1}^{N-1}{{\mathbi{j}}_{i,d}}$. The dependent species is conveniently referred to as the solvent whereas the remaining species are referred to as the solutes.
We consider, now, a total chemical potential that can be split into a pressure-, temperature- and composition-dependent chemical term, and terms containing contributions from supplementary fields, $\mu_{tot,j} = \mu_{chem,j}\left(P,T,\left\{z_j\right\}\right) + \sum{\mu_{\psi ,j}\left(\psi \right)}$. The supplementary potential fields, may be due to externally imposed fields (e.g. gravitational or accelerating motion, or electrical/magnetic fields), but may also be due to internally induced fields (e.g. electric fields in inhomogeneous ionic mixtures). Considering, for simplicity, just one supplementary field, we can expand [Eq. (\[eq:MSdiff\])]{} in terms of its partial derivatives, and express the diffusive mass flux in terms of field gradients; $$\label{genfluxvec}
{\mathbi{j}}_{i,d} = -\rho_fD_{ij}\left[d_{T,j}{\mathbi{\nabla}}\left(\ln T\right) + d_{P,j}{\mathbi{\nabla}}P+d_{\psi ,j}{\mathbi{\nabla}}\psi +\Gamma_{jk}{\mathbi{\nabla}}z_k\right]~,$$ where $d_{T,j}\equiv\nicefrac{\partial_{T}\mu_{chem,j}}{\mathcal{R}}$, $d_{P,j}\equiv\nicefrac{\partial_{P}\mu_{chem,j}}{\mathcal{R}T}$, $d_{\psi ,j}\equiv\nicefrac{\partial_{\psi }\mu_{\psi ,j}}{\mathcal{R}T}$, and $\Gamma_{jk}\equiv\nicefrac{\partial_{z_k}\mu_{chem,j}}{\mathcal{R}T}$, and Einstein’s summation convention is employed. We realize that the net diffusive driving force is a combination of a temperature gradient force (thermophoresis), a pressure gradient force (barophoresis), supplementary potential gradient forces (e.g. gravitophoresis, electrophoresis, etc.), and the compositional gradient force (diffusiophoresis). Thermophoresis and barophoresis are related to the mixture molar entropy and volume, respectively. The diffusiophoretic contribution is related to the species’ activities and involves the compositional gradients of all the species in the mixture, but reduces to the Fickian diffusion term under certain conditions (e.g. binary or dilute mixtures, or mixtures where the constituents are similar in size and nature [@Taylor93]). Other phoretic behaviours are related to the species charges in relation to the supplementary fields (e.g. an ion species of molar electric charge $Q_i$ and mass (“gravitational charge”) $M_{w,i}$, will experience molar electrophoretic and gravitophoretic driving forces, $\left(\nicefrac{Q_i}{RT}\right){\mathbi{\nabla}}\psi_E$ and $\left(\nicefrac{M_{w,i}}{RT}\right){\mathbi{\nabla}}\psi_G$, in the electrical and gravitational potential fields, $\psi_E$ and $\psi_G$, respectively). Converting the mole fractions to mass fractions, $z_k=\Lambda_{kl}X_l$, where $\Lambda_{kl}=\overline{M}_w\left[\nicefrac{z_k}{M_{w,N}} + \nicefrac{\left(\delta_{kl}-z_k\right)}{M_{w,l}}\right]$, and $\delta_{kl}$ is the Kronecker delta, the diffusive mass flux of species $i$, in the $\alpha$ direction, is expressed as $$\label{eq:genflux}
j_{i,d,\alpha} = -\rho_f\left[D_{T,i}\partial_\alpha\left(\ln T\right) + D_{P,i}\partial_\alpha P + D_{\psi ,i}\partial_\alpha\psi + D_{X,il}\partial_\alpha X_l\right]~,$$ where $D_{T,i}\equiv D_{ij}d_{T,j}$ , $D_{P,i}\equiv D_{ij}d_{P,j}$, $D_{\psi ,i}\equiv D_{ij}d_{\psi ,j}$, and $D_{X,il}\equiv D_{ij}\Gamma_{jk}\Lambda_{kl}$.
Advection {#sec:advection}
---------
Due to the no-slip condition, effects of turbulence vanish close to the wall, and molecular diffusion is the only remaining transport mechanism. To conserve the total mass of the mixture, however, a non-zero, advective velocity component normal to the wall is required, as will be shown below. In the following, we will consider mass fluxes in the direction normal to the wall ($\bot$), only. The total deposition flux is found by summing [Eq. (\[eq:difffluxdef\])]{}, evaluated at the wall, over all species, $$\label{eq:jdep}
j_{dep}\equiv \rho_{f,w}u_{f,\bot ,w}=\sum\nolimits_\delta{\rho_{\delta ,w}u_{\delta ,\bot ,w}}~,$$ and using that 1) the convective mass fluxes of the non-depositing species are zero at the wall; and 2) that the diffusive mass fluxes sum to zero. It follows directly that the advective fluid velocity at the wall is given by the sum of the mass fraction weighted convective velocities of the depositing species, $$\label{eq:advvel}
u_{f,\bot ,w}=\sum\nolimits_\delta{X_{\delta ,w}u_{\delta ,\bot ,w}}~,$$ and employing the definition in [Eq. (\[eq:difffluxdef\])]{}, it follows that [Eq. (\[eq:advvel\])]{} can be expressed as $$\label{eq:advvelfrac}
u_{f,\bot ,w}=\frac{\sum\nolimits_\delta{j_{\delta ,d,\bot}}}{\rho_{f,w}\left(1 - \sum\nolimits_\delta{X_{\delta ,w}}\right)}~.$$ Thus, the advective fluid velocity at the wall is uniquely defined by the diffusive mass fluxes and mass fractions of the depositing species, at the wall.
Species mass fraction boundary conditions at the wall {#sec:species_mass_frac_bc}
-----------------------------------------------------
Provided the convective mass fluxes of the non-depositing species vanish at the wall, [Eq. (\[eq:difffluxdef\])]{} dictates that the diffusive mass fluxes of the non-depositing species be exactly cancelled by the advective mass flux, at the wall; $$\label{eq:fluxnondep}
j_{\eta ,d,\bot ,w}=-\rho_{\eta ,w}u_{f,\bot ,w}~.$$ By inserting for equations \[eq:genflux\] and \[eq:advvel\] in [Eq. (\[eq:fluxnondep\])]{}, relationships between the mass fraction gradient of the non-depositing species, the mass fractions of the depositing species, and the temperature, pressure, and supplementary field gradients, evaluated at the wall, are obtained; $$\label{eq:nondepBC}
\left[-D_{T,\eta }\partial_\bot\left(\ln T\right) - D_{P,\eta}\partial_\bot P - D_{\psi ,\eta }\partial_\bot\psi - D_{X,\eta l}\partial_\bot X_l + X_\eta\sum\nolimits_\delta{X_\delta u_{\delta ,\bot}}\right]_w=0~.$$ [Eq. (\[eq:nondepBC\])]{} acts as the wall boundary condition for [Eq. (\[eq:ADE\])]{}, for the arbitrary, non-depositing species $\eta$. It provides one equation for the mass fraction gradients for each of the non-depositing solute species. The solution of this linear system of equations provides the Neumann boundary conditions, for the mass conservation of the non-depositing species, at the wall. We realize, however, that [Eq. (\[eq:nondepBC\])]{} needs estimates of the temperature, pressure and supplementary potential gradients, the various diffusivities of all the species, the mass fractions of all the depositing species, and finally the convective velocities of the depositing species, at the wall. Thus, these will have to be updated for each iteration when solving the governing equations numerically. Their profiles are, in fact, part of the final solution to the problem. The possibility to derive [Eq. (\[eq:nondepBC\])]{} for the non-depositing species arises from the fact that we know their mass fluxes at the wall. The mass fluxes of the depositing species, however, are a priori unknown. Thus, there is no such simple approach to find their boundary conditions. The natural choice is to impose Dirichlet boundary conditions on the depositing species. In the Discussion section, below, we will make some comments on how to obtain these numbers.
One single depositing species {#sec:one_single_dep_sp}
-----------------------------
In the special case that only one of the species deposits, the summation signs in equations \[eq:jdep\] and \[eq:advvel\] are superfluous, and we can express the total deposition flux as $$j_{dep} \equiv \rho_{f,w}u_{f,\bot ,w}=\rho_{\delta ,w}u_{\delta ,\bot ,w}~.$$ [Eq. (\[eq:advvelfrac\])]{} can be expressed in terms of the diffusive mass flux of the depositing species and the difference between the fluid mass density and the depositing species’ mass concentration, at the wall; $$\label{eq:advveldeprate}
u_{f,\bot ,w}=\frac{j_{\delta ,d,\bot ,w}}{\rho_{f,w}-\rho_{\delta ,w}}~.$$ Inserting for [Eq. (\[eq:advveldeprate\])]{} in [Eq. (\[eq:fluxnondep\])]{}, we get a relationship between the non-depositing and depositing species’ diffusive mass fluxes at the wall; $$\label{eq:jnondepjdep}
j_{\eta ,d,\bot ,w}=-\frac{X_{\eta ,w}}{1-X_{\delta ,w}}j_{\delta ,d,\bot ,w}~.$$ Now, in the case of one single depositing species, the combination of equations \[eq:genflux\] and \[eq:jnondepjdep\] reveals that we must require that the temperature, pressure, supplementary potential, and mass fraction gradients at the wall, balance in accordance with the equation $$\begin{gathered}
\label{eq:singledepspbc}
\Bigg[\left(1 + \frac{X_{\eta }}{1-X_{\delta }}\frac{D_{T,\delta }}{D_{T,\eta }}\right)D_{T,\eta }\partial_{\bot }\left(\ln T\right) + \left(1 + \frac{X_{\eta }}{1-X_{\delta }}\frac{D_{P,\delta }}{D_{P,\eta }}\right)D_{P,\eta }\partial_{\bot }P \\
+ \left(1 + \frac{X_{\eta }}{1-X_{\delta }}\frac{D_{\psi ,\delta }}{D_{\psi ,\eta }}\right)D_{\psi ,\eta }\partial_{\bot }\psi + \sum\limits_{\begin{smallmatrix}
l=1, \\
l\ne \delta
\end{smallmatrix}}^{N-1}{\left(1 + \frac{X_{\eta }}{1-X_{\delta }}\frac{D_{X,\delta l}}{D_{X,\eta l}}\right)D_{X,\eta l}\partial_{\bot }X_{l}}\Bigg]_{w}=0~,\end{gathered}$$ for each non-depositing species, $\eta$. In the case of one single depositing species, [Eq. (\[eq:singledepspbc\])]{} replaces [Eq. (\[eq:nondepBC\])]{} in determining the Neumann boundary conditions for the non-depositing species. Furthermore, in the special case that the fluid consists of two species, only, where one is depositing, and the other not, we can express $X_{\eta }=1-X_{\delta }$, and [Eq. (\[eq:singledepspbc\])]{} reduces to $$\left[\partial_T\left(\mu_{chem,\eta } +\mu_{chem,\delta }\right)\partial_{\bot }T + \partial_P\left(\mu_{chem,\eta} + \mu_{chem,\delta }\right)\partial_{\bot}P + \partial_\psi\left(\mu_{\psi,\eta} + \mu_{\psi,\delta}\right)\partial_{\bot}\psi\right]_w=0~.$$
Simplified boundary conditions for the non-depositing species {#sec:simplified_bc}
-------------------------------------------------------------
In the case that the advective fluid velocity at the wall is negligible, the boundary conditions simplify significantly. In this case, it suffices to require that the wall-normal diffusive mass fluxes of the non-depositing species vanish, at the wall. Thus, we require that the wall-normal right-hand-side terms in [Eq. (\[eq:genflux\])]{} cancel out, and we get the simplified Neumann wall boundary condition for the non-depositing species; $$\left.\partial_\bot z_\eta\right|_w = -\left[\left(\Gamma_{i\eta}\right)^{-1}\left(d_{T,i}\partial_\bot\ln T + d_{P,i}\partial_\bot P + d_{\psi ,i}\partial_\bot\psi\right)\right]_w~.$$
Discussion {#sec:discussion}
==========
From the theoretical considerations above, two questions naturally arise:
1. How can we estimate the mole fractions of the depositing species, at the wall?
2. How do we know which species are depositing and which are non-depositing, in the first place?
These questions will only partly be answered in this paper, but a discussion of what is required to handle these questions is provided below. To be able to answer these questions, a profound understanding of the fundamentals of the precipitation kinetics at a molecular/atomic level, at the wall and at the deposit layer surface, is required. This understanding can only be obtained through the combination of experimental studies and atomic-scale modelling of the solid-fluid interface reactions (e.g. atomistic/molecular dynamics simulations). Then, it is required to translate modelling results from molecular dynamics calculations into the language of thermodynamics, to be able to utilize the chemical potential gradient in the calculation of the deposition fluxes. In particular, the partial derivative of the chemical potential with respect to temperature, or rather the Soret coefficient responsible for the thermophoretic behaviour, is currently poorly understood.
The direct precipitation fouling process consists of two major steps, namely the mass transport to the wall and the phase change (a.k.a. surface integration) at the wall. The former is governed by diffusive, advective, and turbulent transport mechanisms, in the fluid phase. The latter depends on the thermodynamic/chemical integration of fluid phase species into the foulant layer. In addition, other processes may take place simultaneously; diffusion and counter diffusion in the porous foulant layer; adhesion probability less than one due to short residence times at the wall; and re-entrainment of deposited material [@Bott95]. If the time-scale of mass transport to the wall is much longer than that of the phase-change, the fouling process is characterized as diffusion controlled. The regime is characterized by e.g. increasing fouling rate with increasing flow velocity. Opposite, if the phase-change at the wall is limiting the fouling rate, the fouling process is denoted interface controlled. The regime is characterized by constant (or even decreasing) fouling rate with increasing flow velocity. In the literature there are reports of both diffusion controlled [@Hasson68] and interface controlled [@Augustin95; @Bansal93; @Mwaba06a; @Mwaba06b] fouling regimes.
Direct precipitation fouling has traditionally been modelled in accordance with the Kern-Seaton approach, where the two main items for deposition, are the boundary layer mass transfer and the surface integration. These are typically modelled by two rate equations, for the depositing species, $\delta$, where the mass transfer rate is proportional to some power of a concentration difference [@MullerSteinhagen11; @Bott95; @Konak74]. For the boundary layer: $$\label{eq:jBL}
j_{BL,\delta}=k_{BL,\delta}\left(X_{\delta ,Bulk}-X_{\delta ,w}\right)~,$$ where $k_{BL}$is an effective mass transfer coefficient due to all transport phenomena in the boundary layer, and $X_{\delta ,Bulk}$ and $X_{\delta ,w}$ are the bulk and solid-fluid interface (wall) mass fractions of species $\delta$, respectively. For the surface integration: $$\label{eq:jIR}
j_{IR,\delta}=k_{IR,\delta}\left(X_{\delta ,w}-X_{\delta ,Sat,w}\right)^{n_{IR,\delta}}~,$$ where $k_{IR}$ is an effective mass transfer coefficient at the solid-fluid interface (interface reaction rate) and $X_{\delta ,Sat,w}$ is the saturation mass fraction of species $\delta$, at the solid-fluid interface. Mass conservation requires that $j_{BL}=j_{IR}$, which determines the interface mass fraction. E.g., by linearising [Eq. (\[eq:jIR\])]{}, it can be shown that the wall mass fractions of species $\delta$ can be approximated by $$X_{\delta ,w}\approx\frac{k_{BL,\delta }X_{\delta ,bulk}-A_{IR,\delta}}{k_{BL,\delta}+B_{IR,\delta}}~,$$ resulting in a mass flux of $$j_{\delta}=j_{BL,\delta }=j_{IR,\delta }\approx\frac{k_{BL,\delta}\left(B_{IR,\delta}X_{\delta,bulk}+A_{IR,\delta}\right)}{k_{BL,\delta}+B_{IR,\delta}}~,$$ where $A_{IR,\delta}\equiv\frac{B_{IR,\delta}}{n_{IR,\delta}}\left[\left(1-n_{IR,\delta}\right)X_{\delta,0}-X_{\delta,Sat,w}\right]$, $B_{IR,\delta}\equiv n_{IR,\delta}k_{IR,\delta}\left(X_{\delta,0}-X_{\delta,Sat,w}\right)^{n_{IR,\delta}-1}$, and $X_{\delta,0}$ is the approximate wall mass fraction. Ultimately, it is the magnitude of the interface reaction rate, $k_{IR,\delta}$, that determines the fouling regime classification. That is, in the extremes of fully diffusion- or interface-controlled fouling we have $k_{IR}\to\infty$ and $k_{IR}\to0$, respectively. In the fully diffusion-controlled regime, it is evident that [Eq. (\[eq:jIR\])]{} dictates $X_{\delta,w}\xrightarrow[k_{IR}\to\infty]{}X_{\delta ,Sat,w}$ , to ensure finite interface mass-transfer. In the fully interface-controlled regime, [Eq. (\[eq:jBL\])]{} requires that $X_{\delta,w}\xrightarrow[k_{IR}\to0]{}X_{\delta ,Bulk}$ [@Mwaba06a]. This modelling approach is, however, only able to capture deposition due to downhill mass diffusion, i.e. where diffusion is in the direction of decreasing concentration, and it relies on supersaturated bulk conditions, i.e. $X_{\delta,Sat}\le X_{\delta,w}\le X_{\delta,Bulk}$. It fails to account for the complex behaviour that can arise from the thorough thermodynamic considerations introduced in the Theory section, above. That is, the traditional approach is well suited for simple scenarios (e.g. isothermal/isobaric conditions without supplementary fields), but it cannot be expected to produce accurate results in the presence of e.g. strong temperature gradients.
Even for simple mixtures, however, it is possible to achieve deposition due to uphill diffusion, i.e. where the net diffusive mass flux is in the direction of increasing concentrations of the depositing species. E.g., consider a two-component, ideal mixture, where one species can deposit on the wall, under isobaric conditions and the absence of supplementary fields. The wall-normal component of [Eq. (\[eq:genflux\])]{}, for the depositing species, can be written $$\label{eq:jSoret}
j_{\delta,d,\bot,w}=-\left.\rho_fD_\delta\left[\partial_\bot z_\delta + z_\delta\left(1 - z_\delta\right)S_{T,\delta}\partial_\bot T\right]\right|_w<0~,$$ where the overall diffusivity is expressed as $D_\delta\equiv 2\frac{D_{\delta \delta }\left(1 - z_\delta\right) + D_{\delta\eta}z_\delta}{z_\delta\left(1 - z_\delta\right)}$, and the Soret coefficient is given by $S_{T,\delta}=\frac{1}{2RT}\frac{D_{\delta\delta}\partial_T\mu_\delta + D_{\delta\eta}\partial_T\mu_\eta}{D_{\delta\delta}\left(1 - z_\delta\right) + D_{\delta\eta}z_\delta}$ , and the positive mass flux points in the direction away from the wall, into the fluid, per definition. In the limit of dilute mixture ($z_\delta\ll1$), and similar diffusivities ($D_{\delta\delta}\approx D_{\delta\eta}$), [Eq. (\[eq:jSoret\])]{} yields the requirement $$\left.S_{T,\delta}\partial_\bot T\right|_w>-\left.\partial_\bot\ln{z_\delta}\right|_w~.$$ Now, for a negative composition gradient (mole fraction decreasing in the direction pointing away from the wall), we see that $$\label{eq:StdT}
\left.S_{T,\delta}\partial_\bot T\right|_w>-\left.\partial_\bot\ln{z_\delta}\right|_w>0~.$$ That is, for a steep enough temperature gradient (sign depending on the sign of the Soret coefficient), a positive deposition rate (negative mass flux at the wall) can occur even for negative mole fraction gradients (uphill diffusion). This is illustrated in figure \[fig:Fig2\], which shows a monotonously decreasing mole fraction profile, $\ln{z_\delta}$, along with two different monotonously increasing temperature profiles, $T_1$ and $T_2$. As indicated in the figure, $T_1$is steep enough to give a negative mass flux, i.e. deposition at the wall, for a certain Soret coefficient, $S_{T,\delta}>0$. $T_2$, however, results in a positive mass flux, hence no deposition. It can be deduced that there exists a critical temperature gradient, $\left.\partial_\bot T_2\right|_w<\left.\partial_\bot T_c\right|_w<\left.\partial_\bot T_1\right|_w$, that results in zero mass flux, $j_{\delta,d,w}=0$. In the presence of additional fluid components or supplementary influencing fields, there are virtually endless opportunities for uphill diffusion-based deposition. In conclusion, the true upper bound for the solid-fluid interface mass fraction, in a single-phase system, is the critical supersaturation, $X_{\delta,CSat}$, not the bulk mass fraction. Even saturations above the critical supersaturation may be permitted if transport equations for additional (solid) phases, dispersed in the fluid phase, are taken into account.
![ Sketch of two scenarios with a given mole fraction profile decreasing away from the wall (negative gradient), positive Soret coefficient, and two different temperature profiles resulting in negative (deposition) and positive (no deposition) mass fluxes, respectively (See [Eq. (\[eq:StdT\])]{}). []{data-label="fig:Fig2"}](Fig2){width="\figwidth"}
Since the $X_{i,Sat}$ generally are functions of the overall composition, temperature, pressure, and the effect of supplementary, influencing fields, we deduce that $X_{\delta,w}=X_{\delta,Sat,w}$ implies that $X_{i,w}=X_{i,Sat,w}\forall i\in\Omega $. This situation is commonly referred to as a situation of local thermodynamic equilibrium at the solid-fluid interface. It further implies that the chemical potentials of the solid and fluid are identical at the interface. Assuming that the chemical potentials of the depositing species increase with their mass fractions, it will be energetically favourable to undergo a fluid-solid phase-change, and thus deposit on the wall, in the situation where $X_{\delta ,w}>X_{\delta,Sat,w}$. In the opposite situation, where $X_{\delta,w}<X_{\delta,Sat,w}$, however, it will be energetically unfavourable. In this sense, we accept $X_{\delta,Sat,w}$ as the lower bound for the interface mass fraction. Thus, we have established that $X_{\delta,Sat,w}\le X_{\delta,w}\le X_{\delta,CSat}~\forall~\delta\in\Delta$ in single-phase flow, but $X_{\delta,Sat,w}\le X_{\delta,w}\le 1~\forall~\delta\in\Delta$ in a multi-phase framework.
We are now ready to address the two questions posed at the beginning of the Discussion section. First, having established the lower and upper bounds for the depositing species’ mass fractions at the solid-fluid interface, we still need to estimate the actual interface mass fractions to be able to solve the governing equations for mass-transport through the boundary layer. The interface mass fractions are, however, part of the solution, and thus the estimation of the interface mass fractions must be part of the iterative, numerical scheme to solve the governing equations. Employing [Eq. (\[eq:jIR\])]{}, we can calculate the interface mass fractions from the equation $$X_{\delta,w}=X_{\delta,Sat,w} + \left(\nicefrac{j_{\delta,\bot,w}}{k_{IR}}\right)^{\nicefrac{1}{n_{IR}}}~,$$ where $X_{\delta,Sat,w}$ must be determined from thermodynamic/chemical equilibrium calculations at wall conditions or from interpolation in tables of experimental data, $j_{\delta,\bot,w}$ is the deposition rate of species $\delta$, from the previous iteration, and the interface reaction coefficient can be determined from e.g. the Arrhenius equation, $k_{IR}=k_{IR,0}\exp\left(\nicefrac{-E_a}{RT_w}\right)$, where the kinetic parameters $k_{IR,0}$ and $E_a$ must be determined experimentally or from molecular dynamics simulations.
To answer which species will deposit, it is necessary to perform a full thermodynamic/chemical equilibrium analysis of the fluid-substrate system, to provide information on e.g. the supersaturation degrees with respect to different solid stages and fluid compositions. This is beyond the scope of this paper, but we point the reader’s attention in the direction of recent developments in atomic/molecular-scale modelling and simulation. There has been done a substantial amount of work on studying different types of atomic and molecular deposition at solid surfaces using molecular dynamics and other atomic-scale modelling techniques Results that can be derived from these simulations are, in a rough order of required computational complexity, e.g., nano-scale surface structures, growth and reaction mechanisms, thermodynamics data (such as reaction enthalpies, free energy differences and equilibrium constants), and growth kinetics parameters (rate coefficients and corresponding Arrhenius parameters). That is, molecular/atomic scale simulations may provide essential information on how surface structures may affect the formation of solid phases at the surface.
An additional topic that may be solved by employing advanced molecular dynamics simulations is the accurate, predictive modelling of the thermophoretic driving force coefficients (e.g. the Soret coefficients). There has in recent years been much progress in applying molecular dynamics simulations for the study of thermophoretic effects [@Artola07; @Galliero08; @Reith00], including the calculation of Soret coefficients, so the practical application of such methods for quantitative predictions is certainly plausible. This is a crucial achievement in the predictive modelling of molecular transport in the turbulent boundary layer, yet it is currently poorly understood, especially for particles and large molecules, but also for smaller molecules. E.g., for colloidal particles suspended in a liquid, there are still no models that can even predict the sign of the Soret coefficient [@Geelhoed14]. The sign of the Soret coefficient is generally a function of temperature, thus, the fluid mixture’s deposition behaviour can be very sensitive to the temperature. That is, there may exist a critical wall temperature, where the Soret coefficient changes sign, with the implication that deposition can be turned on/off by selecting the wall temperature carefully.
Conclusion {#sec:conclusion}
==========
The direct precipitation fouling rate of a supersaturated, multi-component fluid mixture is depending on the flow conditions in the fluid and a host of fluid properties and parameters. What ultimately determines the fouling rate, however, is the surface reaction/phase change that takes place at the nano-scale, at the wall, initially, and at the deposit layer surface, eventually. To obtain accurate/predictive mathematical models for the direct precipitation fouling rate, it is crucial to establish fundamental models for the solid-fluid interface conditions that act as wall boundary conditions in the species mass conservation equations. In this way, we can combine the three very different length-scales at which the transportation of depositing species takes place. Thus, even for industrial processes, taking place at very large length scales, it is crucial to understand the atomistic/molecular scale-phenomena that occur at the wall, to fully appreciate predictive fouling models coupled with CFD models.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was funded by the Research Council of Norway and The Norwegian Ferroalloy Producers Research Association, through the SCORe project[@SCORE].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- János Kollár
title: Effective Nullstellensatz for Arbitrary Ideals
---
=c
=ø
=
2328 1310 1310
\[section\] \[thm\][Question]{} \[thm\][Lemma]{} \[thm\][Corollary]{} \[thm\][Corollaries]{} \[thm\][Proposition]{} \[thm\][Criterion]{} \[thm\][Conjecture]{} \[thm\][Principle]{} \[thm\][Complement]{} \[thm\][Definition]{} \[thm\][Condition]{} \[thm\] \[thm\][Example]{} \[thm\][Hint]{} \[thm\][Exercise]{} \[thm\][Problem]{} \[thm\][Construction]{} \[thm\][Question]{} \[thm\][Algorithm]{} \[thm\][Remark]{} \[thm\][Note]{} \[thm\][Summary]{}
\[thm\][Notation]{}
\[thm\][Claim]{}
1374 1265
Introduction
============
Let $X,Y\subset {{\mathbb P}}^n$ be closed irreducible subvarieties and $Z_i$ the irreducible components of $X\cap Y$. One variant of the theorem of Bézout (cf. [@fulton 8.4.6]) says that $$\sum_i\deg Z_i\leq \deg X\cdot \deg Y.$$ This result holds without any restriction on the dimensions of $X,Y, Z_i$ and it can be easily generalized to the case when $X_1,\dots,X_s$ are arbitrary subschemes of ${{\mathbb P}}^n$ and the $Z_i$ are the reduced irreducible components of $X_1\cap\dots\cap X_s$.
It is frequently of interest to study finer algebraic or metric properties of intersections of varieties. In recent years considerable attention was paid to the case when the $X_i$ are all hypersurfaces, in connection with the effective versions of Hilbert’s [*Nullstellensatz*]{}. Assume that we have polynomials $f_1,\dots,f_s\in
{{\mathbb C}}[x_1,\dots,x_n]$ of degrees $d_i=\deg f_i$. There are three related questions one can ask about the intersection of the hypersurfaces $(f_i=0)$, in each case attempting to minimize a bound $B(d_1,\dots,d_s)$.
Algebraic Bézout version
: [@brownaw-ppp] Find prime ideals $P_j\supset (f_1,\dots,f_s)$ and natural numbers $a_j$ such that $$\prod_jP_j^{a_j}\subset (f_1,\dots,f_s)
{\quad\mbox{and}\quad} \sum_j a_j\deg P_j\leq B(d_1,\dots,d_s).$$
Effective Nullstellensatz version
: [@null] If the $f_i$ have no common zeros in ${{\mathbb C}}^n$, find polynomials $g_i$ such that $$\sum_i f_ig_i=1
{\quad\mbox{and}\quad} \deg (f_ig_i)\leq B(d_1,\dots,d_s).$$
Łojasiewicz inequality version
: [@jks] Fix a metric on ${{\mathbb C}}^n$ and let $Z$ be the intersection of the hypersurfaces $(f_i=0)$. Prove that if $x$ varies in a bounded subset of ${{\mathbb C}}^n$ then $${\operatorname{dist}}(Z,x)^{B(d_1,\dots,d_s)}\leq C\cdot \max_i|f_i(x)|{\quad\mbox{for some
$C>0$.}\quad}$$
The optimal value of the bound $B(d_1,\dots,d_s)$ is known in almost all cases. If we assume that $d_i\geq 3$ for every $i$, then $$B(d_1,\dots,d_s)=d_1\cdots d_s,$$ is best possible for $s\leq n$. (See [@null 1.5] for the case $s>n$.)
The algebraic Bézout version is also called the [*prime power product*]{} variant of the Nullstellensatz.
The aim of this paper is to consider these problems in case the $f_j$ are replaced by arbitrary ideals. The first step in this direction was taken in [@sombra]. His methods can deal with special cases of the above problems if the ideals are Cohen–Macaulay. Łojasiewicz–type inequalities for arbitrary analytic sets were studied in the works of Cygan, Krasiński and Tworzewski, see especially [@twor; @cygan; @ckt]. Although they consider the related problem of separation exponents, their proof can easily be modified to give a general Łojasiewicz inequality for reduced subschemes.
My proofs grew out of an attempt to understand their work in algebraic terms. This leads to a general Łojasiewicz inequality in the optimal form and to an effective Nullstellensatz with a slightly worse bound. In the algebraic Bézout version my results are considerably weaker. It should be noted, however, that the straightforward generalization of the algebraic Bézout version fails to hold (\[algbez.false.exmp\]).
All three of these results can be formulated for arbitrary ideals, but for simplicity here I state them for unmixed ideals. ($I$ is called [*unmixed*]{} if all primary components of $I$ have the same dimension.) These are the ideals that correspond to the usual setting of intersection theory. For such ideals the degree of $I$ gives a good generalization of the degree of a hypersurface. The precise versions for arbitrary ideals are stated in (\[bez.thm\]), (\[eff.null.thm\]) and (\[lojas.thm\]).
\[main.intr.thm\] Let $K$ be any field and $I_1,\dots,I_m$ unmixed ideals in $K[x_1,\dots,x_n]$. Then there are prime ideals $P_j\supset (I_1,\dots,I_m)$ and natural numbers $a_j$ such that
1. $\prod_j P_j^{a_j}\subset (I_1,\dots,I_s)$, and
2. $\sum_j a_j\leq n\cdot \prod_i\deg I_i$.
\[eff.null.intr.thm\] Let $K$ be any field and $I_1,\dots,I_m$ unmixed ideals in $K[x_1,\dots,x_n]$. The following are equivalent:
1. $I_1,\dots,I_m$ have no common zero in $\bar K^n$.
2. There are polynomials $f_j\in I_j$ such that $$\sum_jf_j=1{\quad\mbox{and}\quad} \deg f_j
\leq (n+1)\cdot
\prod_i\deg I_i.$$
\[lojas.intr.thm\] (cf. [@cygan; @ckt]) Let $I_1,\dots,I_m$ be unmixed ideals in ${{\mathbb C}}[x_1,\dots,x_n]$ and $X_1,\dots,X_m\subset {{\mathbb C}}^n$ the corresponding subschemes. Let $f_{ij}$ be generators of $I_i$. Then for every bounded set $B\subset {{\mathbb C}}^n$ there is a $C>0$ such that for every $x\in B$, $${\operatorname{dist}}(X_1\cap \cdots \cap X_m,x)^{\prod_i\deg I_i}\leq C\cdot
\max_{ij}|f_{ij}(x)|.$$
The difference between the geometric and algebraic versions of the Bézout theorem can be seen already in the case when an irreducible variety is intersected with a hyperplane.
\[algbez.false.exmp\] Pick coordinates $u,v$ in ${{\mathbb C}}^2$ and $x,y,z,s$ in ${{\mathbb C}}^4$. For every odd $n$ consider the morphism $$F_n:{{\mathbb C}}^2\to {{\mathbb C}}^4{\quad\mbox{given by}\quad} F_n(u,v)=(u^n,u^2,uv,v).$$ Let $S_n\subset {{\mathbb C}}^4$ be the image of $F_n$. It is easy to see that $\deg S_n=n$ and the ideal of $S_n$ in ${{\mathbb C}}[x,y,z,s]$ is $$I_n=(x^2-y^n,z^2-ys^2, z^n-xs^n, xz-ys^{\frac{n+1}{2}},
xs-y^{\frac{n-1}{2}}z).$$ Let us intersect $S_n$ with the hyperplane $(s=0)$ to get a curve $C_n$. Set theoretically, the intersection is the image of $f_n:{{\mathbb C}}\to {{\mathbb C}}^4$ given by $f_n(u)=(u^n,u^2,0,0)$ and its ideal is $$J_n=(x^2-y^n,z,s).$$ On the other hand, $$(I_n,s)=(x^2-y^n,z^2, xz, y^{\frac{n-1}{2}}z,s),$$ and we see that, as a vectorspace, $$J_n/(I_n,s) \cong \langle z,yz,\dots, y^{\frac{n-3}{2}}z\rangle.$$ Let $m=(x,y,z,s)$ be the ideal of the origin. There are two minimal ways of writing an algebraic Bézout form of this example: $$J_n^2\subset (I_n,s){\quad\mbox{and}\quad} m^{\frac{n-1}{2}}\cdot J_n\subset
(I_n,s).$$ Taking degrees we get $2\deg J_n=2n>n=\deg I_n$ and $\frac{n-1}{2}\deg m+\deg J_n=
\frac{n-1}{2}+n>n=\deg I_n$.
This example illustrates the nature of the difficulties, but it does not seem to give pointers as to the general shape of the theory. Unfortunately, I do not have any plausible conjectures about what happens in general. As in [@null], the effect of embedded primes seems small, but the correct way of estimating it is still elusive.
Instead, I approach the question as follows. There are many different varieties $S^{\lambda}_n\subset{{\mathbb C}}^4$ whose intersection with the hyperplane $(s=0)$ is $C_n$. (For instance, pick polynomials $f(u,v),g(u,v)$ with no common zero and let $S^{f,g}_n$ be the image of $(u,v)\mapsto
(u^n,u^2,vf(u,v),vg(u,v))$.) Each $S^{\lambda}_n$ gives an ideal $I^{\lambda}_n$ and one can ask about all the quotients $$J_n/(I^{\lambda}_n,s).$$ It turns out that their length is bounded independent of $\lambda$ and it is not too big. The main lemma of [@null 3.4] is a formalization of this observation using local cohomology groups in some special cases.
This paper develops another approach to this problem, going back to [@cayley]. For any space curve $C\subset {{\mathbb P}}^3$ Cayley considered all cones defined by $C$ with a variable point $p\in {{\mathbb P}}^3$ as vertex. These cones can be encoded as one equation on the Grassmannian of lines in ${{\mathbb P}}^3$. More generally, for any pure dimensional subscheme $Y^d\subset {{\mathbb C}}^n$ (or for any pure dimensional algebraic cycle on ${{\mathbb C}}^n$) consider the ideal $I^{ch}(Y)$ generated by all cones defined by $Y$ with a variable $(n-d-1)$-dimensional linear space as its vertex. Following [@ds95], it is called the [*ideal of Chow equations*]{} (\[con-id.def\]). It turns out that this ideal controls the length of the embedded components of any intersection. With this observation at hand, the rest of the arguments turn out to be not very complicated.
Section 2 reviews some basic facts about algebraic cycles and their intersection theory on ${{\mathbb C}}^n$. Section 3 collects known results about integral closures of ideals.
The ideal of Chow equations is defined and studied in section 4. The connection between the ideal of Chow equations and intersection theory is established in section 5.
Finally the main results are proved in sections 6 and 7.
Another approach to such theorems is to reduce them to the hypersurface case. If $X\subset {{\mathbb C}}^n$ is a subscheme of degree $m$ then, set theoretically, $X$ can be defined by degree $m$ equations. This gives reasonable bounds for each problem, roughly like $(\max_i\{\deg I_i\})^n$. For many ideals of about the same degree this is close to the optimal bound for the Nullstellensatz, but it is considerably worse in general. For the algebraic Bézout version this method and [@brownaw-ppp] gives a bound in the original form taking into account the degrees of the $P_j$.
A modified version of this idea is to reduce everything to intersecting with the diagonal and then using the methods of [@null] directly. This gives $3^n$-times the optimal bounds. If, however, the quotients $K[x_1,\dots,x_n]/I_j$ (or more precisely, their homogenizations) are Cohen-Macaulay, then the methods of [@sombra] give better bounds. The factor $(n+1)$ in (\[eff.null.intr.thm\]) can be replaced by 2.
I thank J. Johnson and P. Roberts for their help with many of the computations. I am greatful for the comments of A. Beauville, P. Philippon and P. Tworzewski and for several e–mails of B. Sturmfels which helped to eliminate many mistakes.
Partial financial support was provided by the NSF under grant number DMS-9622394.
Intersection of cycles on ${{\mathbb A}}^n$
===========================================
\[cyc.def\] Let $Y$ be a scheme. An [*algebraic cycle*]{} on $Y$ is a formal linear combination of reduced and irreducible subschemes $Z=\sum a_i[Z_i]$, $a_i\in {{\mathbb Z}}$. (I do not assume that the $Z_i$ have the same dimension.) The cycles form a free Abelian group $Z_*(Y)$. The subgroup generated by all reduced and irreducible subschemes of dimension $d$ is denoted by $Z_d(Y)$.
If $Y$ is proper and $L$ is a line bundle on $Y$ then one can define the $L$-degree of a cycle $$\deg_LZ:=\sum a_i(Z_i\cdot L^{\dim Z_i}),$$ where $(Z_i\cdot L^{\dim Z_i})$ denotes the top selfintersection number of the first Chern class of $L|_{Z_i}$. The function $Z\mapsto \deg_LZ$ is linear.
Let $Y$ be a scheme with a compactification $Y\subset\bar Y$ and assume that $L$ is the restriction of a line bundle $\bar L$ from $\bar Y$ to $Y$. For a cycle $Z=\sum a_iZ_i$ on $Y$ set $\bar Z=\sum a_i\bar Z_i$ where $\bar Z_i$ is the closure of $Z_i$ in $\bar Y$. Then one can define the degree of a cycle $Z=\sum a_iZ_i$ on $Y$ by $$\deg_LZ:=\deg_{\bar L}\bar Z.$$ It is important to note that this depends on the choice of $\bar
Y$ and $\bar L$. I use this version of the degree only for the pair $Y={{\mathbb A}}^n$ and $\bar Y={{\mathbb P}}^n$.
\[ncap.def\] Let $X$ be a scheme and $D$ an effective Cartier divisior on $X$. Let $[Z]\in Z_d(X)$ be an irreducible $d$-cycle on $X$. Define $Z\ncap D \in Z_*(X)$ as follows.
1. If $Z\subset {\operatorname{Supp}}D$ then set $[Z]\ncap D:=[Z]\in Z_d(X)$.
2. If $Z\not\subset {\operatorname{Supp}}D$ then $D|_{Z}$ makes sense as a Cartier divisor. Set $[Z]\ncap D:=[D|_{Z}]\in Z_{d-1}(X)$.
This definition can be extended to $Z_*(X)$ by linearity.
If $f$ is a defining equation of $D$ then I also use $Z\ncap f$ to denote $Z\ncap D$.
It should be emphasized that this definition is not at all well behaved functorially. While it is well defined on cycles, it is not well defined on the Chow group. Furthermore, if $D_1,D_2$ are two Cartier divisors then in general $$(Z\ncap D_1)\ncap D_2\neq (Z\ncap D_2)\ncap D_1.$$ (For instance let $X={{\mathbb A}}^2, Z=(y-x^2=0), D_1=(x=0)$ and $D_2=(y=0)$.)
\[ncap-deg.lem\] Let $L$ be an ample line bundle on $X$, $D$ a section of $L^{\otimes d}$ and $Z$ an effective cycle on $X$. Then
1. $\deg_L(Z\ncap D)\leq d\cdot \deg_L(Z)$.
2. If $X$ is proper, $d=1$ and all the components of $Z$ have positive dimension then $\deg_L(Z\ncap
D)= \deg_L(Z)$.
Proof. By linearity it is sufficient to check this when $Z$ is an irreducible and reduced subvariety. If $Z\subset {\operatorname{Supp}}D$ then $\deg_L(Z\ncap D)= \deg_L(Z)$, and otherwise $\deg_L(Z\ncap D)\leq d\cdot \deg_L(Z)$ with equality holding if $X$ is proper and $\dim Z\geq 1$ by the usual Bézout theorem.
One would like to define $Z_1\ncap Z_2$ for any two cycles $Z_i$ on a scheme $X$. As usual, this is reduced to intersecting $Z_1\times Z_2$ with the diagonal $\Delta\subset X\times X$. Traditional intersection theory works if $X$ is smooth since in this case $\Delta\subset X\times X$ is a local complete intersection (cf. [@fulton Chap.8]). The usual intersection product $Z_1\cdot Z_2$ is then a cycle of the expected dimension $d=\dim Z_1+\dim Z_2-\dim X$. If $\dim (Z_1\cap Z_2)=d$ then $Z_1\cdot Z_2$ is well defined as a cycle, but if $\dim (Z_1\cap Z_2)>d$ then $Z_1\cdot Z_2$ is defined only as a rational equivalence class inside ${\operatorname{Supp}}(Z_1\cap Z_2)$.
Here I follow the path of [@stu-vo; @vogel] and try to define $Z_1\ncap Z_2$ as a well defined cycle which may have components of different dimension. If $X={{\mathbb P}}^n$, the $Z_i$ are pure dimensional and $d\geq 0$ then $Z_1\ncap Z_2$ is a cycle such that $$\deg (Z_1\ncap Z_2)=\deg Z_1\cdot\deg Z_2.$$ The cases when $d<0$ were not considered to have much meaning traditionally. [@twor] realized that the definition is meaningful and gives an interesting invariant.
The construction of (\[ncap.def\]) needs $\Delta$ to be a global complete intersection. Unfortunately this happens very rarely. The only such example that comes to mind is $X={{\mathbb A}}^n$, or more generally, any scheme $X$ which admits an étale map to ${{\mathbb A}}^n$. For simplicity of exposition, I work with $X={{\mathbb A}}^n$. Homogenity considerations can then be used to define $\ncap$ for a few other interesting cases, most importantly for $X={{\mathbb P}}^n$.
\[vt-cyc.def\] Let $X_i=\sum_j a_{ij}X_{ij}$ be effective cycles on ${{\mathbb A}}^n$ for $i=1,\dots,s$. We would like to define a cycle which can reasonably be called the intersection of these cycles. This is done as follows.
Choose an identification ${{\mathbb A}}^{ns}={{\mathbb A}}^n\times\cdots\times{{\mathbb A}}^n$. Using this identification define $$\prod_{i=1}^s X_i:=\sum_{j_1,\dots,j_s}
\left(\prod_{i=1}^sa_{ij_i}\right)
\left(\prod_{i=1}^sX_{ij_i}\right)$$ as a cycle in $Z_*({{\mathbb A}}^{ns})$.
Let $\Delta\subset {{\mathbb A}}^n\times\cdots\times{{\mathbb A}}^n$ denote the diagonal. Each coordinate projection $$\Pi_r:{{\mathbb A}}^n\times\cdots\times{{\mathbb A}}^n\to {{\mathbb A}}^n{\quad\mbox{(onto the $r$th
factor)}\quad}$$ gives an isomorphism $\Pi_r:\Delta\cong {{\mathbb A}}^n$ which is independent of $r$.
Let ${\cal L}:=(L_1,\dots,L_{n(s-1)})$ be an ordered set of hyperplanes in ${{\mathbb A}}^{ns}$ such that their intersection is $\Delta$. Set $$(X_1\ncap\cdots\ncap X_s, {\cal L}):=
\left(\prod_{i=1}^s X_i\right)\ncap L_1\ncap \cdots\ncap L_{n(s-1)},$$ where the right hand side means that we first intersect with $L_1$, then with $L_2$ and so on. To be precise, the right hand side is in $Z_*({{\mathbb A}}^{ns})$, but every irreducible component of it is contained in $\Delta$. Thus it can be viewed as a cycle in $Z_*(\Delta)$ and so it can be identified with a cycle in $Z_*({{\mathbb A}}^n)$ using any of the projections $\Pi_r$.
$(X_1\ncap\cdots\ncap X_s, {\cal L})$ is called an [*intersection cycle*]{} of $X_1,\dots,X_n$. Any of these cycles is denoted by $X_1\ncap\cdots\ncap X_s$.
It should be emphasized that $X_1\ncap\cdots\ncap X_s$ is not a well defined cycle since it depends on the choice of ${\cal L}$. In the papers [@vogel; @vangast] the $L_i$ are chosen generic and then $(X_1\ncap\cdots\ncap X_s, {\cal L})$ is well defined as an element of a suitable Chow group. We would like to get a cycle which is defined over our field $K$. As long as $K$ is infinite, a general choice of the $L_i$ would work but there are some problems when $K$ is finite. (It is for such reasons that [@brownaw-ppp] does not work for all finite fields.) Furthermore, in our applications it is sometimes advantageous to make a special choice of the $L_i$. For these reasons I allow any choice of the $L_i$. The price we pay is that even the degree of $(X_1\ncap\cdots\ncap X_s, {\cal L})$ depends on the $L_i$. This, however, does not seem to cause problems in the applications.
We obtain the following Bézout type inequality.
\[bezout.thm\] Let $X_1,\dots,X_s$ be effective cycles on ${{\mathbb A}}^n$. Then $$\deg (X_1\ncap\cdots\ncap X_s)\leq \prod_j\deg X_j.$$
Proof. $\deg \prod_{i=1}^s X_i=\prod_{i=1}^s \deg X_i$ and cutting with a hyperplane does not increase the degree by (\[ncap-deg.lem\]).
\[vog.cyc.def\]
Let $K$ be an infinite field. For a scheme $Y$ let $B(Y)$ denote all subvarieties of $Y$ which can be obtained by repeatedly taking irreducible components and their intersections. For general $\cal L$ we can write $$(X_1\ncap\cdots\ncap X_s, {\cal L})=\sum a_i[Z_i({\cal L})],$$ where the $Z_i({\cal L})$ depend algebraically on $\cal L$. For each $Z_i({\cal L})$ there is a smallest $W\in B(X_1\cap\cdots\cap X_s)$ such that $Z_i({\cal L})\subset W$ for every general choice of ${\cal L}$. For each $W\in B(X_1\cap\cdots\cap X_s)$, the sum of these cycles gives a well defined element of the Chow group $A_*(W)$. This cycle is denoted by $(X_1\dcap\cdots\dcap X_s,W)$. Thus we obtain a refined intersection cycle $$X_1\dcap\cdots\dcap X_s := \sum_{W\in B(X_1\cap\cdots\cap
X_s)}(X_1\dcap\cdots\dcap X_s,W)$$ If $Z\subset X_1\cap\cdots\cap X_s$ is a connected component then $$\deg (X_1\dcap\cdots\dcap X_s,Z):=\sum_{W\subset Z}
\deg (X_1\dcap\cdots\dcap X_s,W)$$ is well defined. It is called the [*equivalence*]{} of $Z$ in $X_1\dcap\cdots\dcap X_s$ (cf. [@fulton 9.1]).
In analogy with [@twor], one can define a local variant of this number as follows. For every $p$, $$\sum a_i{\operatorname{mult}}_pZ_i({\cal L})$$ is constant on a Zariski open subset of the ${\cal L}$-s. I denote it by $${\operatorname{mult}}_p (X_1\dcap\cdots\dcap X_s).$$ There is an inequality $${\operatorname{mult}}_p(X_1\dcap\cdots\dcap X_s)\leq
\sum_{p\in W\in B(X_1\cap\cdots\cap X_s)}\deg (X_1\dcap\cdots\dcap
X_s,W).$$
We need to set up a correspondence between ideal sheaves and algebraic cycles. This does not work as well as the usual correspondence between subschemes and ideal sheaves, but it is better suited for our purposes. Another way of going from cycles to ideal sheaves is studied in section 4.
\[ass-id.def\] Let $X$ be a scheme and $Z=\sum a_i[Z_i]$ an effective cycle. Let $I(Z_i)\subset {{\cal O}}_X$ denote the ideal sheaf of $Z_i$. Define the [*ideal sheaf of $Z$*]{} by $$I(Z):=\prod_iI(Z_i)^{a_i}\subset {{\cal O}}_X.$$ It is clear that $I(Z_1+Z_2)=I(Z_1)I(Z_2)$.
\[ass-cyc.def\] Let $F$ be any coherent sheaf on $X$ and $F_i\subset F$ the subsheaf of sections whose support has codimension at most $i$. Let $x_{ij}$ be the generic points of the irreducible components $X_{ij}\subset {\operatorname{Supp}}(F_i/F_{i-1})$. Set $$Z(F):=\sum_{ij} ({\operatorname{length}}_{x_{ij}}F_i)\cdot [X_{ij}].$$ $Z(F)$ is called the [*cycle associated to $F$*]{}.
Let $Q_{ij}\subset {{\cal O}}_X$ be the ideal sheaf of $X_{ij}$ and $b_{ij}:={\operatorname{length}}_{x_{ij}}F_i$. Then $\prod_jQ_{ij}^{b_{ij}}$ maps $F_i$ to $F_{i-1}$, thus $I(Z(F))\subset {\operatorname{Ann}}(F)$.
In particular, if $J\subset {{\cal O}}_X$ is an ideal sheaf then $$I(Z({{\cal O}}_X/J))\subset J.$$
If $X$ is proper and $L$ is a line bundle on $X$ then one can define the [*$L$-arithmetic degree*]{} of a sheaf $F$ by $${\operatorname{arith-deg}}_LF:=\deg_L Z(F).$$ If $I\subset {{\cal O}}_X$ is an ideal sheaf then the arithmetic degree of ${{\cal O}}_X/I$ is also called the arithmetic degree of $I$ and denoted by ${\operatorname{arith-deg}}_LI$. Note that there is a possibility of confusion since $I$ is also a sheaf.
This definition is very natural and it appeared in several different places (see, for instance, [@harts; @null; @bay-mumf]). The concept was used extensively in many papers (cf. [@stv]).
\[I-prod.lem\] Let $X_1,\dots,X_m$ be schemes and $Z_i$ a cycle on $X_i$ for every $i$. Let $\pi_i:\prod_j X_j\to
X_i$ be the $i$-th coordinate projection. Then $$I(\prod_j Z_j)\subset (\pi_1^*I(Z_1),\dots,\pi_m^*I(Z_m)).$$
Proof. Using induction, it is sufficient to prove the case $m=2$. Let $Z_k=\sum_j a_{kj}Z_{kj}$, then $$Z_1\times Z_2=\sum_{ij}a_{1i}a_{2j}(Z_{1i}\times Z_{2j}).$$ If $I,J$ are arbitrary ideals and $a,b\geq 1$, then $$(I,J)^{ab}\subset (I,J)^{a+b-1}\subset (I^a,J^b).$$ Using this on $X_1\times X_2$, we obtain that $$\begin{aligned}
I(Z_1\times Z_2)&=&
\prod_{ij}I(Z_{1i}\times Z_{2j})^{a_{1i}a_{2j}}\\ &=&
\prod_{ij}(\pi_1^*I(Z_{1i}), \pi_2^*I(Z_{2j}))^{a_{1i}a_{2j}}\\
&\subset &
\prod_{ij}(\pi_1^*I(Z_{1i})^{a_{1i}},
\pi_2^*I(Z_{2j})^{a_{2j}})\\ &\subset &
\prod_{j}(\pi_1^*\prod_{i}I(Z_{1i})^{a_{1i}},
\pi_2^*I(Z_{2j})^{a_{2j}})\\ &= &
\prod_{j}(\pi_1^*I(Z_1),
\pi_2^*I(Z_{2j})^{a_{2j}})\\ &\subset & (\pi_1^*I(Z_1),
\pi_2^*I(Z_2)).\qed\end{aligned}$$
Integral closure of ideals
==========================
In this section we recall some relevant facts concerning integral closure of ideals. [@teiss Chap.I] serves as a good general reference.
\[intclos.def\] Let $R$ be a ring and $I\subset R$ an ideal. $r\in R$ is called [*integral over $I$*]{} if $r$ satisfies an equation $$r^k+\sum_{j=1}^k i_jr^{k-j}=0{\quad\mbox{where $i_j\in I^j$.}\quad}$$ All elements integral over $I$ form an ideal $\overline{I}$, called the [*integral closure*]{} of $I$.
We use the following easy properties of the integral closure.
1. $(\overline{I}, \overline{J})\subset \overline{(I,J)}$,
2. $\overline{I_1}\cdot\overline{I_2}\subset \overline{I_1I_2}$, and so $(\overline{I})^m\subset \overline{I^m}$.
We also need the following special case of the Briancon–Skoda theorem. A short proof of it can be found in [@li-te p.101].
[@bri-sko]\[bri-sko\] If $R=K[x_1,\dots,x_n]$ (or more generally, if $R$ is regular of dimension $n$) then $\overline{I^n}\subset I$.
The following result gives the best way to compare integral closures (cf. [@teiss I.1.3.4]).
\[val-crit.thm\] Let $R$ be a ring and $I, J\subset R$ two ideal. The following are equivalent.
1. $J\subset \bar I$.
2. If $p:R\to S$ is any homomorphism of $R$ to a DVR $S$ then $p(J)\subset p(I)$.
If $K$ is an algebraically closed field and $R$ a finitely generated $K$-algebra then in (2) it is sufficient to use homomorphisms to the power series ring $K[[t]]$.
Integral closures usually do not commute with taking quotients, but this holds in some special cases.
\[intcl.rest.lem\] Let $I\subset K[x_1,\dots,x_n]$ be an ideal. Then $$\overline{(I,x_n)}/(x_n)= \overline{(I,x_n)/(x_n)}.$$
Proof. If $J_1\subset J_2\subset R$ are ideals then $\overline{J_2}/J_1\subset \overline{J_2/J_1}$ always holds using (\[intclos.def\]). If $R\to R/J_1$ splits (as a ring homomorphism) then any equation over $R/J_1$ can be lifted to an equation over $R$, showing the other containment.
We need two lemmas about ideals given by algebraic families of generators.
\[zdense.gens.lem\] Let $K$ be an infinite field, $R$ a $K$-algebra and $L\subset R$ a finite dimensional $K$-vectorspace. Let $U$ be a $K$-variety and $$F:U\to L{\quad\mbox{given by}\quad} u\mapsto r_u$$ a $K$-morphism. Let $V\subset U$ be Zariski dense. Then there is an equality of ideals $$(r_u:u\in V)=(r_u:u\in U).$$
Proof. If $J$ is any ideal in $R$ then $L\cap J$ is a sub vector space in $L$. Thus $\{u\in U:r_u\in J\}$ is Zariski closed in $U$. Set $J=(r_u:u\in V)$. Since $V$ is dense in $U$, we obtain that $r_u\in J$ for every $u\in U$.
\[zdense.prod.lem\] Notation as in (\[zdense.gens.lem\]). Assume in addition that $U$ is irreducible. Let $u\mapsto r_u$ and $u\mapsto s_u$ be $K$-morphisms from $U$ to $L$. Let $V\subset U$ be Zariski dense. Then we have an equality of ideals $$\overline{(r_us_u:u\in V)}=
\overline{(r_u:u\in U)\cdot (s_u:u\in U)}.$$
Proof. Let $p:R\to S$ be any homomorphism to a DVR. An ideal in $S$ is characterized by the minimum order of vanishing of its elements. We need to prove that both ideals above give the same number.
The order of vanishing of each $p(r_u)$ in $S$ is a lower semi continuous function of $U$, thus it achieves the minimum value on a dense open subset of $U$. Similarly for $p(s_u)$. Thus we can choose a $u\in V$ where both $p(r_u)$ and $p(s_u)$ achieve their minimum.
Let $R=K[x,y], L=\{ax+by\}, U=K, r_u=x-uy,
s_u=x+uy$. Then $(r_u:u\in U)\cdot (s_u:u\in U)=(x,y)^2$ is different from $(r_us_u:u\in U)=(x^2,y^2)$. This shows that (\[zdense.prod.lem\]) fails without integral closure.
Another such example is given in (\[axes.exmp\]).
It is easy to check that the ideals $(I_n,s)$ in (\[algbez.false.exmp\]) are integrally closed, hence integral closure alone cannot remove the embedded primes, even in a geometrically very simple situation.
The ideal of Chow equations
===========================
Let $K$ be a field and $Z$ any effective cycle in ${{\mathbb A}}^n$. In this section we define an ideal in $K[x_1,\dots,x_n]$, called the ideal of Chow equations of $Z$. The main advantage of this notion is that it behaves well with respect to arbitrary hyperplane sections. This is the crucial property that one needs for the applications. On the other hand, the ideal of Chow equations is quite difficult to analyze and I leave several basic questions unresolved. (The explanation of the name and other variants are discussed in (\[why-chow?\]).)
\[con-id.def\] Let $Z=\sum a_iZ_i$ be a purely $d$-dimensional cycle in ${{\mathbb A}}^n$. Let $\pi:{{\mathbb A}}^n\to {{\mathbb A}}^{d+1}$ be a projection such that $\pi:Z_i\to
{{\mathbb A}}^{d+1}$ is finite for every $i$. We call such a projection [*allowable*]{}.
The center of the projection $\pi$ is a linear space $L\subset {{\mathbb P}}^n\setminus {{\mathbb A}}^n$ of dimension $n-d-2$ and $\pi$ is allowable iff $L$ is disjoint from $\cup_i\bar Z_i$. This shows that allowable projections can be parametrized by an irreducible quasiprojective variety.
If $\pi$ is allowable then $\pi_*(Z)$ is a well defined codimension 1 cycle in ${{\mathbb A}}^{d+1}$, and so it corresponds to a hypersurface. Choose an equation of this hypersurface and pull it back by $\pi$ to obtain a polynomial $f(\pi,Z)$.
Assume first that $K$ is infinite. Define the [*ideal of Chow equations*]{} of $Z$ in the polynomial ring $K[{{\mathbb A}}^n]\cong
K[x_1,\dots,x_n]$ as $$I^{ch}(Z):=(f(\pi,Z): \mbox{ $\pi$ is allowable})\subset K[{{\mathbb A}}^n].$$ For technical reasons we frequently work with the integral closure of this ideal, denoted by $\overline{I^{ch}}(Z)$.
We see in (\[coneq.indep.cor\]) that these are independent of the base field. Thus if $K$ is finite, one can define $I^{ch}(Z)$ by taking any infinite field extension of $K$ first.
Finally, if $Z=\sum a_iZ_i$ is any effective cycle then write $Z$ as a sum $Z=\sum Z^d$ where $Z^d$ has pure dimension $d$ and set $$I^{ch}(Z)=\prod_d I^{ch}(Z^d).$$ Its integral closure is denoted by $\overline{I^{ch}}(Z)$. A product formula in terms of the $Z_i$ is given in (\[prod-lem\]), but this only works for the integral closures.
\[why-chow?\]
The ideals $I^{ch}(Z)$ were first considered by [@cayley] and $I^{ch}(Z)$ is essentially equivalent to the Chow form of $Z$, as explained in [@catan; @ds95]. This equivalence clarifies the definition of $I^{ch}(Z)$, but it obscures other versions of this concept.
In (\[con-id.def\]) we consider [*linear*]{} projections $\pi:{{\mathbb A}}^n\to {{\mathbb A}}^{d+1}$. It is, however, possible to use larger classes of morphisms. For instance we can allow $\pi$ to be any algebraic automorphism of ${{\mathbb A}}^n$ followed by a projection or we can even allow $\pi$ to be any smooth morphism. The latter case can be localized in various topologies.
More generally, if $R$ is any smooth $K$-algebra and $Z$ a $d$-cycle on ${\operatorname{Spec}}R$ then one can define the ideal of locally Chow equations (using étale or analytic topology or even working formally) and these ideals behave well with respect to intersections with smooth divisors. Here I concentrate on the simpler case of linear projections. I was unable to decide if the various definitions give the same ideals for a cycle in ${{\mathbb A}}^n$.
I do not know if it is essential to consider the integral closure or not in the definition above. The examples (\[noncl.id.exmp\], \[charp.exmp\]) show that $I^{ch}(Z)$ is not integrally closed in general. More importantly, the crucial property (\[prod-lem\]) fails without integral closure as shown by (\[axes.exmp\]).
The main question is whether (\[main.tech.thm\]) holds without integral closure on the right hand side. This would eliminate the extra factor $(n+1)$ in (\[eff.null.intr.thm\]). I do not know the answer. This question is related to the degree bounds considered in [@sturmf Sec. 4].
As a special case of (\[zdense.gens.lem\]) we obtain:
\[zdense-lem\] Let $\{\pi_{\lambda}:\lambda \in \Lambda\}$ be a Zariski dense set of allowable projections as in (\[con-id.def\]). Then $$I^{ch}(Z)=(f(\pi_{\lambda},Z)\vert \lambda \in
\Lambda).\qed$$
\[coneq.indep.cor\] $I^{ch}(Z)$ is independent of the base field $K$. That is, if $L\supset K$ is a field extension, then $$I^{ch}(Z)\otimes_KL=I^{ch}(Z_L).$$
Proof. If $K$ is infinite, then the projections defined over $K$ form a Zariski dense set of the projections defined over $L$. Thus by (\[zdense-lem\]) we obtain the same ideals.
For finite $K$ we defined $I^{ch}(Z)$ by forcing the above formula to hold.
Let $X\subset {{\mathbb A}}^n$ be a smooth subvariety with ideal sheaf $I(X)$. Then $I^{ch}(X)=I(X)$ and $I^{ch}(a\cdot X)=I(X)^a$. More generally, let $Z=\sum a_iZ_i$ be any cycle. Then the above relationship holds near any smooth point of ${\operatorname{Supp}}Z$.
Thus $I^{ch}(Z)$ is interesting only near the singular points of ${\operatorname{Supp}}Z$.
Let $p\in {\operatorname{Supp}}Z$ be a point of multiplicity $d$ and $m_p$ the ideal of $p$. A general projection $\pi(Z)$ has multiplicity $d$ at $\pi(p)$, thus each $f(\pi,Z)$ has multiplicity $\geq d$ at $p$. This shows that $I^{ch}(Z)\subset m_p^d\cap I(Z)$.
By [@catan 1.14], if $Z$ has codimension at least 2 then $I(Z)\neq I^{ch}(Z)$ along the singular locus of $Z$.
\[low.embdim.exmp\] Let ${{\mathbb A}}^{n-k}\subset {{\mathbb A}}^n$ be the subspace $(x_n=\cdots=x_{n-k+1}=0)$. Let $Z$ be a $d$-cycle on ${{\mathbb A}}^{n-k}$ and $j_*Z$ the corresponding cycle on ${{\mathbb A}}^n$. We would like to compare $I^{ch}(Z)$ and $I^{ch}(j_*Z)$.
A general projection of $j_*Z$ can be obtained as a projection $\rho:{{\mathbb A}}^n\to {{\mathbb A}}^{n-k}$ followed by a general projection $\pi:{{\mathbb A}}^{n-k}\to {{\mathbb A}}^{d+1}$. This shows that $$f(\pi\circ \rho, j_*Z)=f(\pi, Z)(x_1+L_1,\dots,x_{n-k}+L_{n-k}),$$ where the $L_i$ are linear forms in $x_{n-k+1},\dots, x_n$ defining $\rho$.
This shows that the restriction map $$I^{ch}(j_*Z)\onto I^{ch}(Z){\quad\mbox{is surjective.}\quad}$$
\[noncl.id.exmp\] Let $X\subset {{\mathbb A}}^n$ be defined by equations $g(x_1,\dots,x_{n-1})=x_n=0$. A general projection of $X$ is isomorphic to $X$ and, at least in characteristic zero, $$\begin{aligned}
f(\pi,X)&=&g(x_1+a_1x_n,\dots,x_{n-1}+a_{n-1}x_n)\\ &=& \sum_I
c_Ia^Ix_n^{|I|}\frac{\partial^Ig}{\partial x^I},\end{aligned}$$ where the $c_I$ are nonzero constants and $a_i\in
K$. Since the $a_i$ can vary independently, we see that the $f(\pi,X)$ generate the ideal $$\left(x_n^{|I|}\frac{\partial^Ig}{\partial x^I}:
I=(i_1,\dots,i_{n-1})\right).$$
Consider for instance the case $n=3$ and $g=x_1^3+x_2^5$. Then $$I^{ch}(X)=(x_1^3+x_2^5,x_1^2x_3,x_1x_3^2,x_3^3,
x_2^4x_3,x_2^3x_3^2,x_2^2x_3^3,x_2x_3^4,x_3^5).$$ $x_2^2x_3^2$ is integral over $I^{ch}(X)$ (since $(x_2^2x_3^2)^2-x_3^3\cdot x_2^4x_3=0$) but it is not in $I^{ch}(X)$. Hence $\overline{I^{ch}}(X)\neq I^{ch}(X)$.
\[charp.exmp\] Assume that $K$ has characteristic $p$ and let $0\in {{\mathbb A}}^2$ be the origin with ideal $(x,y)$. Set $Z=p[0]$. Then $f(\pi,Z)=(ax+by)^p$ for some $a,b$, thus $I^{ch}(Z)=(x^p,y^p)$. Its integral closure is the much bigger ideal $(x,y)^p$.
The next lemma gives a product formula for $\overline{I^{ch}}(Z)$. This result is crucial for the applications and it fails if we do not take the integral closure, as the example after the lemma shows.
\[prod-lem\] Let $Z=\sum a_iZ_i$ be an effective cycle. Then $$\overline{I^{ch}}(Z)=\overline{\prod_i I^{ch}(Z_i)^{a_i}}.$$
Proof. It is enough to check this for pure dimensional cycles.
Let $\pi$ be any allowable projection for $Z$. Then $\pi$ is allowable for every $Z_i$ and $f(\pi,Z)=\prod_if(\pi,Z_i)^{a_i}$ which proves the containment $\subset$. The converse follows by a repeated application of (\[zdense.prod.lem\]).
\[axes.exmp\] Choose $n\geq 3$ odd and in ${{\mathbb A}}^n$ consider the 1–cycle of the $n$ coordinate axes $Z=\sum_{i=1}^n Z_i$. Then $$\prod_i I^{ch}(Z_i)=\prod_i(x_1,\dots,\widehat{x_i},\dots,x_n).$$ As a vectorspace this has a basis consisting of all monomials of degree at least $n$ which involve at least 2 variables.
On the other hand, I claim that $I^{ch}(Z)$ does not contain the monomial $x_1\cdots x_n$.
Let $\pi:(x_1,\dots,x_n)\mapsto (\sum a_ix_i,\sum b_ix_i)$ be a projection. This gives the equation $$f(\pi,Z)=\prod_i\left(\sum_j m_{ji}x_j\right)
{\quad\mbox{where $m_{ji}=a_jb_i-a_ib_j$.}\quad}$$ Thus twice the coefficient of the $x_1\cdots x_n$ term is
$$\begin{aligned}
2\cdot \sum_{\sigma\in S_n}\prod_j m_{j\sigma(j)}
&=&\sum_{\sigma\in S_n}\left(\prod_j m_{j\sigma(j)}+\prod_j
m_{j\sigma^{-1}(j)}\right)\\ &=&\sum_{\sigma\in S_n}\left(\prod_j
m_{j\sigma(j)}+\prod_j m_{\sigma(j)j}\right)\\ &=&\sum_{\sigma\in
S_n}\left(\prod_j m_{j\sigma(j)}+(-1)^n\prod_j
m_{j\sigma(j)}\right)=0.\end{aligned}$$
If $n=3$ then it is easy to compute that $\prod_i I^{ch}(Z_i)=(I^{ch}(Z),x_1x_2x_3)$. I have not checked what happens for $n\geq 5$ or for even values of $n$.
The following result of [@amor Thm.B] shows that $\overline{I^{ch}}(Z)$ contains a fairly small power of $I(Z)$. (The statement in [@amor] is slightly different since he is working with $I^{ch}(Z)$, but his proof actually gives this version.)
\[jks.lem\][@amor] Let $Z=\sum a_iZ_i$ be a cycle in ${{\mathbb A}}^n$. Let $I(Z_i)$ denote the ideal of $Z_i$. Then $$\prod_iI(Z_i)^{a_i\deg Z_i}\subset \overline{I^{ch}}(Z).$$ More precisely, if $x\in {{\mathbb A}}^n$ is a point then $$\prod_iI(Z_i)^{a_i{\operatorname{mult}}_xZ_i}\subset \overline{I^{ch}}(Z)
{\quad\mbox{in a neighborhood of $x$.}\quad}\qed$$
\[con<ord.lem\] Let $Z$ be a cycle in ${{\mathbb A}}^n$. Then $$I^{ch}(Z)\subset I(Z).$$ Let $J\subset K[x_1,\dots,x_n]$ be an ideal. Then $$I^{ch}(Z(K[x_1,\dots,x_n]/J))\subset J.$$
Proof. Because of the multiplicative definitions of $I(Z)$ (\[ass-id.def\]) and of $I^{ch}(Z)$ (\[con-id.def\]), it is sufficient to prove the first claim in case $Z$ is pure dimensional.
Write $Z=\sum a_iZ_i$ and let $\pi:{{\mathbb A}}^n\to {{\mathbb A}}^{d+1}$ be an allowable projection. Then $\pi_*[Z]=\sum a_i\pi_*[Z_i]$, so $f(\pi,Z)=\prod f(\pi,Z_i)^{a_i}$. $f(\pi,Z_i)\in I(Z_i)$, so $f(\pi,Z)\in \prod I(Z_i)^{a_i}=I(Z)$.
The second part follows from the first and from (\[ass-cyc.def\]).
The ideal of Chow equations and intersection theory
===================================================
The next result is the key property of the ideal of Chow equations.
\[coneq.main.lem\] Let $X\subset {{\mathbb A}}^n$ be an irreducible and reduced subvariety and $H=(x_n=0)$ a hyperplane not containing $X$. Then $$I^{ch}(X\ncap H)\subset (I^{ch}(X),x_n).$$
Proof. By (\[coneq.indep.cor\]) we may assume that the base field is infinite. Choose a general linear subspace $L\subset
\bar H \setminus H \setminus\bar X$ of dimension $n-d-2$. $\dim (\bar H \setminus H)=n-2$ and $\dim (\bar X\cap \bar H)\leq
d-1$. Since $(n-d-2)+(d-1)<n-2$, $L$ is disjoint from $\bar X$. Let $\pi':H\to H'$ and $\pi:{{\mathbb A}}^n\to {{\mathbb A}}^{d+1}$ be the projections with center $L$. Let $\rho:{{\mathbb A}}^n\to H$ be a projection and set $\pi'':=\pi\circ \rho: {{\mathbb A}}^n\to H'$. The 3 projections appear in the following diagram: $$\begin{array}{cll} H & \subset & {{\mathbb A}}^n\\
\pi'\downarrow{\ \ } & \swarrow \pi'' & \downarrow\pi\\ H'& \subset
& {{\mathbb A}}^{d+1}
\end{array}$$ $ X\ncap H$ can be viewed as a cycle on $H$; in such a case I denote it by $Z$.
$\pi_*(X)$ is a hypersurface in ${{\mathbb A}}^{d+1}$ and $\pi'_*(Z)$ is a hypersurface in $H'$ such that $\pi'_*(Z)=\pi_*(X)\cap H'$. Thus $$f(\pi',Z)(x_1,\dots,x_{n-1})=f(\pi,X)(x_1,\dots,x_{n-1},0).$$
As in (\[low.embdim.exmp\]), the generators of $I^{ch}(X\ncap
H)$ are of the form $$\begin{aligned}
f(\pi'', X\ncap H)&=&
f(\pi',Z)(x_1+a_1x_n,\dots,x_{n-1}+a_{n-1}x_n)\\ &\equiv
&f(\pi',Z)(x_1,\dots,x_{n-1})\mod (x_n)\\ &\equiv
&f(\pi,X)(x_1,\dots,x_{n-1}, x_n)\mod (x_n). \qed\end{aligned}$$
\[coneq.main.lem-rem\] More generally, (\[coneq.main.lem\]) also holds if $X$ is a pure dimensional cycle and $H$ does not contain any of its irreducible components.
The generalization to intersecting with several linear equations is formal, but the induction seems to require the use of integral closure, as shown by the following example. The final result itself, however, may not need integral closure.
The 3 coordinate axes $C_x,C_y,C_z$ in ${{\mathbb A}}^3$ can be defined by the determinantal equations $${\operatorname{rank}}\left(
\begin{array}{ccc} x&y&0\\ z&y&z
\end{array}
\right)
\leq 1.$$ $C_x+C_y+C_z$ is a hyperplane section of the surface $$Z_1 {\quad\mbox{given by equations}\quad}
{\operatorname{rank}}\left(
\begin{array}{ccc} x&y+as&bs\\ z&y+cs&z+ds
\end{array}
\right)
\leq 1.$$ For general $a,b,c,d$, $Z_1$ is the cone over a rational normal curve. By explicit computation, $xyz\in I^{ch}(Z_1)$. As we remarked in (\[axes.exmp\]), this implies that $$I^{ch}(C_x)\cdot I^{ch}(C_y)\cdot I^{ch}(C_z)
\subset (I^{ch}(Z_1),s).$$ Next consider the surface $$Z_2 {\quad\mbox{with equations}\quad}
{\operatorname{rank}}\left(
\begin{array}{ccc} x&y+as^2&bs^2\\ z&y+cs^2&z+ds^2
\end{array}
\right)
\leq 1.$$ For general $a,b,c,d$ this defines a rational triple point. By explicit computation, $xyz\not\in (I^{ch}(Z_2),s)$, which implies that $$I^{ch}(C_x)\cdot I^{ch}(C_y)\cdot I^{ch}(C_z)
\not\subset (I^{ch}(Z_2),s).$$
\[surfcut.ind\] Let $Z$ be a cycle on ${{\mathbb A}}^n$ and $H_i=(\ell_i=0)$ hyperplanes. Then $$I^{ch}(Z\ncap H_1\ncap \cdots\ncap H_m)\subset
\overline{(I^{ch}(Z),\ell_1,\dots,\ell_m)}.$$
Proof. Consider first the case when $Z$ is irreducible and $m=1$. The claim is trivial if $Z\subset H_1$ and the $Z\not\subset H_1$ case is treated in (\[coneq.main.lem\]).
Next we prove the $m=1$ case by induction on the number of irreducible components of $Z$. $$\begin{aligned}
I^{ch}((Z_1+Z_2)\ncap \ell_1)&\subset &
\overline{I^{ch}(Z_1\ncap \ell_1)\cdot I^{ch}(Z_2\ncap \ell_1)}
{\quad\mbox{(by (\ref{prod-lem}))}\quad}\\ &\subset &
\overline{(I^{ch}(Z_1),\ell_1)\cdot (I^{ch}(Z_2),\ell_1)}
{\quad\mbox{(by induction)}\quad}\\ &\subset &
\overline{(I^{ch}(Z_1)I^{ch}(Z_2),\ell_1)}\\
&\subset &
\overline{(I^{ch}(Z_1+Z_2),\ell_1)}{\quad\mbox{(by (\ref{prod-lem}))}\quad}\end{aligned}$$
Finally the case $m>1$ is established by induction using the chain of inclusions $$\begin{aligned}
I^{ch}(Z\ncap \ell_1\ncap \ell_2)&\subset &
\overline{(I^{ch}(Z\ncap\ell_1),\ell_2)}\\ &\subset&
\overline{(\overline{(I^{ch}(Z), \ell_1)},\ell_2)}\\ &\subset&
\overline{(I^{ch}(Z),\ell_1,\ell_2)}.\qed\end{aligned}$$
We are ready to formulate our main technical theorem.
\[main.tech.thm\] Let $K$ be a field and $Z_1,\dots,Z_m$ cycles in ${{\mathbb A}}^n$. Let $Z_1\ncap\cdots\ncap Z_m$ be any of the intersection cycles. Then $$I^{ch}(Z_1\ncap\cdots\ncap Z_m) \subset
\overline{(I(Z_1),\dots,I(Z_m))}.$$
Proof. Choose an identification ${{\mathbb A}}^{nm}={{\mathbb A}}^n\times\cdots\times{{\mathbb A}}^n$ ($m$-times) and let $\Pi_r:{{\mathbb A}}^{nm}\to {{\mathbb A}}^n$ be the projection onto the $r$-th factor. Let $\Delta\subset {{\mathbb A}}^n\times\cdots\times{{\mathbb A}}^n$ denote the diagonal.
Choose an ordered set of hyperplanes ${\cal L}:=(L_i=(\ell_i=0):
i=1,\dots,n(m-1))$ in ${{\mathbb A}}^{nm}$ whose intersection is $\Delta$. This gives us a cycle $(Z_1\ncap\cdots\ncap Z_m,{\cal L})$ which we view as a cycle in ${{\mathbb A}}^{nm}$.
Applying (\[surfcut.ind\]) we obtain that $$I^{ch}(Z_1\ncap\cdots\ncap Z_m,{\cal L})
\subset
\overline{(I^{ch}(Z_1\times\cdots\times Z_m),
\ell_1,\dots,\ell_{n(m-1)})}.$$ $I^{ch}(Z_1\times\cdots\times Z_m)
\subset I(Z_1\times\cdots\times Z_m)$ by (\[con<ord.lem\]) and using (\[I-prod.lem\]) this gives the inclusion $$I^{ch}(Z_1\ncap\cdots\ncap Z_m,{\cal L})
\subset
\overline{(\Pi_1^*I(Z_1),\dots, \Pi_m^*I(Z_m),
\ell_1,\dots,\ell_{n(m-1)})}.$$ Let us restrict to $\Delta$. The left hand side becomes $I^{ch}(Z_1\ncap\cdots\ncap Z_m)$ by (\[low.embdim.exmp\]), and the right hand side becomes $\overline{(I(Z_1),\dots,I(Z_m))}$ by (\[intcl.rest.lem\]).
It is possible that (\[main.tech.thm\]) can be considerably sharpened. The strongest and most natural statement would be $$I^{ch}(Z_1\ncap\cdots\ncap Z_m) \subset
(I^{ch}(Z_1),\dots,I^{ch}(Z_m)).$$ For the applications the main point would be to get rid of the integral closure since this would eliminate the extra factor $(n+1)$ in (\[eff.null.intr.thm\]).
Effective Nullstellensatz
=========================
We are ready to formulate and prove the precise technical versions of our main theorems, using the notion of arithmetic degree as defined in (\[ass-cyc.def\]).
\[bez.thm\] Let $K$ be any field and $I_1,\dots,I_m$ ideals in $K[x_1,\dots,x_n]$. Then there are prime ideals $P_j\supset (I_1,\dots,I_m)$ and natural numbers $a_j$ such that
1. $\prod_j P_j^{a_j}\subset (I_1,\dots,I_m)$, and
2. $\sum_j a_j\leq n\cdot \prod_i{\operatorname{arith-deg}}I_i$.
Proof. As in (\[ass-cyc.def\]), set $Z_i=Z(I_i)$ and let $\sum b_jX_i=(Z_1\ncap\cdots\ncap Z_m,{\cal L})$ be any of the intersection cycles defined in (\[vt-cyc.def\]). Set $d_j:=\deg
X_j$, then $\sum_j b_jd_j\leq \prod_i \deg Z_i$ by (\[bezout.thm\]). By (\[main.tech.thm\]), $$\prod_jI^{ch}(X_j)^{b_j} \subset
\overline{(I(Z_1),\dots,I(Z_m))}.$$ $I(X_j)^{d_j}\subset \overline{I^{ch}}(X_j)$ by (\[jks.lem\]), and so we obtain that $$\prod_jI(X_j)^{b_jd_j} \subset
\overline{(I(Z_1),\dots,I(Z_m))}.$$ $I(Z_s)\subset I_s$ by (\[con<ord.lem\]), hence $$\overline{(I(Z_1),\dots,I(Z_m))}\subset \overline{(I_1,\dots,I_m)}.$$ $(I_1,\dots,I_m)^n\subset \overline{(I_1,\dots,I_m)}$ by (\[bri-sko\]). Putting these together we get that $$\prod_jI(X_j)^{nb_jd_j} \subset (I_1,\dots,I_m).$$ Setting $P_j:=I(X_j)$ and $a_j:=nb_jd_j$ gives (\[bez.thm\]).
\[eff.null.thm\] Let $K$ be any field and $I_1,\dots,I_m$ ideals in $K[x_1,\dots,x_n]$. The following are equivalent:
1. $I_1,\dots,I_m$ have no common zero in $\bar K^n$.
2. There are polynomials $f_j\in I_j$ such that $$\sum_j f_j=1{\quad\mbox{and}\quad} \deg f_j
\leq (n+1)\cdot
\prod_i{\operatorname{arith-deg}}I_i.$$
Proof. It is clear that $(2)\Rightarrow (1)$. To see the converse, introduce a new variable $x_0$ and let $\tilde I_s\subset K[x_0,\dots,x_n]$ denote the homogenization of $I_s \subset K[x_1,\dots,x_n]$. Then ${\operatorname{arith-deg}}\tilde I_s={\operatorname{arith-deg}}I_s$ and $x_0$ is contained in the radical of $(\tilde I_1,\dots, \tilde I_m)$, hence it is contained in any prime ideal containing $(\tilde I_1,\dots, \tilde I_m)$. By (\[bez.thm\]) there are prime ideals $P_j$ and natural numbers $a_j$ such that
1. $\prod_j P_j^{a_j}\subset (\tilde I_1,\dots,\tilde I_m)$, and
2. $\sum_j a_j\leq (n+1)\cdot \prod_i{\operatorname{arith-deg}}I_i$.
Since $x_0\in P_j$ for every $j$, we see that $$x_0^{\sum a_j}\in (\tilde I_1,\dots,\tilde I_m).$$ Thus there are $f_i\in I_i$ with homogenizations $\tilde f_i$ such that $$x_0^{\sum a_j}=\sum_i \tilde f_i
{\quad\mbox{and}\quad} \deg \tilde f_i= \sum_j a_j.$$ Setting $x_0=1$ we obtain (\[eff.null.thm\]).
Łojasiewicz inequalities
========================
Next we turn to applications of these results to the study of Łojasiewicz inequalities and separation exponents. These results are essentially reformulations of [@cygan; @ckt].
\[loj.ineq\] Let $f$ be a real analytic function on ${{\mathbb R}}^n$ and $Z:=(f=0)$. Fix a norm on ${{\mathbb R}}^n$ and set ${\operatorname{dist}}(Z,x):=\inf_{z\in Z}{||x-z||}$. [@loj-stud p.124] proved that for every compact set $K$ there are $m, C>0$ such that $${\operatorname{dist}}(Z,x)^m\leq C\cdot |f(x)|{\quad\mbox{for $x\in K$.}\quad}$$ Any inequality of this type is called a [*Łojasiewicz inequality*]{}.
In general it is rather difficult to obtain an upper bound for $m$ in terms of other invariants of $f$. The problem becomes easier if ${{\mathbb R}}^n$ is replaced by ${{\mathbb C}}^n$, but even in this case it is not straightforward to obtain sharp upper bounds for $m$. The question was investigated in [@brownaw-loc] and [@jks]. Instead of ${{\mathbb C}}$, one can work over any algebraically closed field with an absolute value.
Let $K$ be a field with an absolute value $|\ |$. (The case when $K={{\mathbb C}}$ and $|\ |$ is the usual absolute value is the most interesting, but the cases when $K$ is of positive characteristic or $|\ |$ is nonarchimedian are also of interest.) $|\ |$ induces a norm on $K^n$ by ${||{\bold
x}||}:=(|x_1|+\cdots+|x_n|)^{1/2}$. This defines a distance on $K^n$ as in (\[loj.ineq\]).
\[gen.int.def\] Let $X$ be any topological space and $F,G$ two sets of $K$-valued functions on $X$. We say that $F$ is [*integral over $G$*]{}, denoted by $F\ll G$, if the following condition holds:
1. For every $f\in F$ and $x\in X$ there are $g_1,\dots,g_m\in G$ and a constant $C$ such that $|f(x')|\leq
C\max_i|g_i(x')|$ for every $x'$ in a neighborhood of $x$.
If $F$ and $G$ are continuous (which will always be the case for us) then $(*)$ is automatic if $g(x)\neq 0$ for some $g\in G$. Thus $(*)$ is a local growth condition near the common zeros of $G$.
The two notions of integral dependence are closely related:
\[int.same.lem\] (cf. [@teiss 1.3.1]) Let $K$ be an algebraically closed field with an absolute value $|\ |$. Let $X$ be an affine variety over $K$ (with the metric topology) and $I\subset
{{\cal O}}_X$ an ideal sheaf. A polynomial function is integral over $I$ in the sense of (\[intclos.def\]) iff it is integral over $I$ in the sense of (\[gen.int.def\]).
The relationship between the distance function and the ideal of Chow equations was established in earlier papers.
\[dist-id.lem\] (cf. [@jks 8], [@cygan 3.7]) Let $K$ be an algebraically closed field with an absolute value $|\
|$. Let $Z\subset K^n$ be an irreducible subvariety, $z\in Z$ a point and $m={\operatorname{mult}}_zZ$. Then, in a neighborhood of $z$, $${\operatorname{dist}}(Z,x)^m\ll I^{ch}_Z\ll I_Z\ll {\operatorname{dist}}(Z,x).\qed$$
The main result about Łojasiewicz inequalities is the following.
\[lojas.thm\] (cf.[@ckt]) Let $K$ be an algebraically closed field with an absolute value $|\ |$. Let $I_1,\dots,I_m $ be ideals in $K[x_1,\dots,x_n]$ and $X_1,\dots,X_m\subset K^n$ the corresponding subschemes. Set $D:={\operatorname{arith-deg}}I_1\cdots{\operatorname{arith-deg}}I_m$. Then $${\operatorname{dist}}(X_1\cap \cdots \cap X_m,z)^D\ll (I_1,\dots,I_m)\ll
\max_i\{{\operatorname{dist}}(X_i,z)\}.$$
Proof. Let $(X_1\ncap\dots\ncap X_m)=\sum a_iZ_i$ be one of the intersection cycles. $\sum a_i\deg Z_i\leq D$ by (\[bezout.thm\]) and $Z_i\subset
X_1\cap \cdots \cap X_m$ by construction. Thus $$\begin{aligned}
{\operatorname{dist}}(X_1\cap \cdots \cap X_m,z)^D&\ll&
{\operatorname{dist}}(X_1\cap \cdots \cap X_m,z)^{\sum a_i\deg Z_i}\\ &\leq &
\prod_i {\operatorname{dist}}(Z_i,z)^{a_i\deg Z_i}\\ & \ll &
\prod_i I^{ch}(Z_i)^{a_i}{\quad\mbox{(by (\ref{dist-id.lem}))}\quad}\\ & \subset
&\overline{(I_1,\dots,I_m)}{\quad\mbox{(by (\ref{bez.thm}))}\quad}\\ & \ll &
(I_1,\dots,I_m) {\quad\mbox{(by (\ref{int.same.lem}))}\quad}\\ &\ll &
\max_i\{{\operatorname{dist}}(X_i,z)\}{\quad\mbox{(by (\ref{dist-id.lem})).\qed}\quad}\end{aligned}$$
With a similar proof we obtain the following local version.
\[local.sep.cor\](cf. [@cygan 4.5]) Let $K$ be an algebraically closed field with an absolute value $|\ |$. Let $I_1,\dots,I_m $ be ideals in $K[x_1,\dots,x_n]$ and set $D_p:={\operatorname{mult}}_p(Z(I_1)\dcap\cdots\dcap Z(I_m))$. Then, in a neighborhood of $p$, $${\operatorname{dist}}(X_1\cap \cdots \cap X_m,z)^{D_p}\ll (I_1,\dots,I_m)\ll
\max_i\{{\operatorname{dist}}(X_i,z)\}. \qed$$
Application to deformation theory
=================================
In usual deformation theory we are given a scheme $X_0$ and we would like to understand all flat families $\{X_t:t\in \Delta\}$ where $\Delta$ is the unit disc. There are, however, some deformation problems where we are interested in flat families $\{Y_t:t\in
\Delta\}$ where $X_0=Y_0/(\mbox{embedded points})$, or, more generally, when $X_0$ and $Y_0$ have the same fundamental cycles. This question arises for instance in studying the Chow varieties. (See [@hp] or [@koll96 Chap.I] for definitions and properties of the Chow varieties.) A point in the Chow variety of ${{\mathbb P}}^n$ is not a subscheme but a pure dimensional cycle $W\in Z_d({{\mathbb P}}^n)$. Thus if we want to study the Chow variety near $W$ then we need to understand the deformations of all subschemes $X\subset {{\mathbb P}}^n$ whose fundamental cycle is $W$. If $d\geq 1$ then there are infinitely many such subschemes $X$ since adding embedded points does not change the fundamental cycle.
Assume that we find a subscheme $X_0\subset {{\mathbb P}}^n$ whose fundamental cycle is $W$ and a deformation $\{X_t:t\in \Delta\}$. From the point of view of the Chow variety we are interested only in the fundamental cycle of $X_t$ and not in $X_t$ itself. Hence the only case we need to study is when $X_t$ has no embedded points for general $t$. The affine version of this problem can be stated as follows.
Let $W\in Z_d({{\mathbb A}}^n)$ be a $d$-cycle. Let $Y\subset
{{\mathbb A}}^n\times {{\mathbb A}}$ be a subscheme of pure dimension $(d+1)$ without embedded points such that the second projection $\pi:Y\to {{\mathbb A}}^1$ is flat. Let $Y_0=\pi^{-1}(0)$ be the central fiber and assume that $Z(Y_0)=W$.
What can we say about $Y_0$ in terms of $W$?
This question is related to the problems considered in [@flat-crit].
As an application of (\[coneq.main.lem-rem\]) we obtain the following partial answer. This is a place where it would be more natural to use the ideal of locally Chow equations (\[why-chow?\]).
\[def.prop\] With the above notation, $I^{ch}(W)\subset I(Y_0)$.
Consider the case when $W=[z=x^2-y^n=0]\in Z_1({{\mathbb A}}^3)$. As in (\[noncl.id.exmp\]) we obtain that $$I^{ch}(W)=(x^2-y^n,z^2,xz, y^{n-1}z).$$ On the other hand, in (\[algbez.false.exmp\]) we found an example of a deformation $S$ such that $$I(S_0)=(x^2-y^n,z^2,xz, y^{\frac{n-1}{2}}z).$$ Using [@teiss-sn] we obtain that the length of $I(W)/I(Y_0)$ is at most the arithmetic genus of $W$ which is $\frac{n-1}{2}$. Comparing these two results we conclude that $$(x^2-y^n,z)\supset I(Y_0)\supset (x^2-y^n,z^2,xz,
y^{\frac{n-1}{2}}z)$$ for every deformation $Y$. It is not hard to see that for every ideal satisfying the above condition there is a corresponding deformation.
Using (\[noncl.id.exmp\]) one can compute $I^{ch}$ for all monomial plane curves in ${{\mathbb A}}^3$. The results give strong restrictions on $Y_0$ but I do not see how to get a complete answer as in the above example. P. Roberts computed several examples of monomial space curves and in each case $I^{ch}$ turned out to be quite close to the ideal of the curve.
In higher dimensions $I^{ch}$ gives a very unsatisfactory answer when $W=[X_0]$ where $X_0$ is normal. By [@harts-book III.9.12] in this case $Y_0=X_0$ but $I^{ch}(X_0)\neq I(X_0)$ if $X_0$ is singular.
On the other hand, (\[def.prop\]) gives information about $Y_0$ even if $W$ has multiple components. I do not see how to get any information about $Y_0$ by other methods in this case.
[Chislenko88]{}
F. Amoroso, Multiplicité et formes éliminantes, Bull. Soc. Math. France 122 (1994) 149-162
D. Bayer - D. Mumford, What can be computed in algebraic geometry? in: Computational algebraic geometry and commutative algebra, Cortona, Cambridge Univ. Press, 1993, 1-41
J. Briancon - H. Skoda, Sur la clôture intégrale d’un idéale de germes de fonctions holomorphes en un point de $C^n$, C.R. Acad. Sci. Paris, 278 (1974) 949-951
W. D. Brownawell, Local diophantine Nullstellen inequalities, Journal AMS 1 (1988) 311-322
W. D. Brownawell, A prime power product version of the Nullstellensatz (preprint)
F. Catanese, Chow Varieties, Hilbert Scemes and Moduli Spaces of Surfaces of General Type, J. Alg. Geom. 1 (1992) 561-596
A. Cayley, A new analytic representation of curves in space, Quarterly J. Math. 3(1860) 225-236 = Collected Works, vol. 4, 446-455
E. Cygan, Intersection theory and separation exponent in complex analytic geomtry, (preprint)
E. Cygan, T. Krasiński and P. Tworzewski, Separation at infinity and the Łojasiewicz exponent of polynomial mappings, (preprint)
J. Dalbec and B. Sturmfels, Introduction to Chow forms, in Invariant Methods in Discrete and Computational Geometry, N. White, ed., Proceedings Curacao, Kluwer, 1995, pp. 37–58.
W. Fulton, Intersection Theory, Springer (1984)
L. J. van Gastell, Excess intersection and a correspondence principle, Inv. Math. 103 (1991) 197-222
R. Hartshorne, Connectedness of the Hilbert scheme, Publ. Math. IHES 29 (1966) 5-48
R. Hartshorne, Algebraic Geometry, Springer, 1977
W. Hodge - D. Pedoe, Methods of Algebraic Geometry, Cambridge Univ. Press, 1952
S. Ji - J. Kollár - B. Shiffman, A Global Łojasiewicz Inequality for Algebraic Varieties, Trans. AMS, 329 (1992) 813-818
J. Kollár, Sharp effective Nullstellensatz, Jour. AMS , 1 (1988) 963-975
J. Kollár, Flatness criteria, J. Algebra, 175 (1995) 715-727
J. Kollár, Rational Curves on Algebraic Varieties, Springer Verlag, Ergebnisse der Math. vol. 32, 1996
J. Lipman - B. Teissier, Pseudo-rational local rings and a theorem of Briancon–Skoda, Mich. Math. J 28 (1981) 97-116
S. Łojasiewicz, Sur le problème de la division, Studia Math 18 (1959) 87-136
B. Malgrange, Ideals of Differentiable Functions, Oxford Univ. Press, 1966
M. Sombra, A sparse effective Nullstellensatz, (preprint)
J. Stückrad - W. Vogel, An algebraic approach to intersection theory, Queen’s papers in math. 2 (1982) 1-32
B. Sturmfels, Equations defining toric varieties, in Algebraic Geometry, Santa Cruz 1995, Proc.Symp. Pure Math, vol. 62, AMS 1997
B. Sturmfels - N. V. Trung - W. Vogel, Bounds on degrees of projective schemes, Math. Ann. 302 (1995) 417-432
B. Teissier, Résolution simultanée, in: Séminaire sur les singularités des surfaces, Springer Lecture Notes 777 (1980) 71-146
B. Teissier, Variétés polaires II, Algebraic Geometry, La Rabida, Springer LN 961 (1982) 314-491
P. Tworzewski, Intersection theory in complex analytic geometry, Ann. Polon. Math. 62 (1995) 177-191
W. Vogel, Results on Bézout’s theorem, Tata Inst. Lecture Notes, 1984
University of Utah, Salt Lake City UT 84112
[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a complete noncompact Riemannian manifold $M$ and give conditions on a compact submanifold $K \subset M$ so that the outward normal exponential map off of the boundary of $K$ is a diffeomorphism onto ${M \backslash K}$. We use this to compactify $M$ and show that pinched negative sectional curvature outside $K$ implies $M$ has a compactification with a well defined Hölder structure independent of $K$. The Hölder constant depends on the ratio of the curvature pinching. This extends and generalizes a 1985 result of Anderson and Schoen.'
address: 'University of Washington, Seattle, Washington'
author:
- Eric Bahuaud
- Tracey Marsh
title: |
Hölder Compactification for some\
manifolds with pinched negative\
curvature near infinity
---
The Poincaré model of hyperbolic space has a natural geometric compactification – one can compactify by adding the sphere at infinity. Taking this to be a model case for other simply connected manifolds of negative curvature leads to a classical construction made precise in [@Eberlein-ONeill]: Let $M$ be a Cartan-Hadamard manifold, that is, a complete, simply connected Riemannian manifold with nonpositive sectional curvature. Define an equivalence relation on the set of geodesic rays parametrized by arc length by saying that geodesic rays $\sigma$ and $\tau$ are *asymptotic* if $d_M( \sigma(t), \tau(t))$ remains bounded as $t {\rightarrow}+\infty$. Here $d_M$ is the distance function on $M$ induced by the metric $g$. We define ${M(\infty)}$ to be the set of all equivalence classes of this relation; this is the *geometric boundary at infinity*. \[page:asymp\]
In 1985, Michael Anderson and Richard Schoen proved that given a Cartan-Hadamard manifold $M$ with pinched sectional curvatures like $$\begin{aligned}
-\infty < -b^2 \leq \sec( M ) \leq -a^2 < 0, \end{aligned}$$ where $a$ and $b$ are positive constants, then geometric boundary at infinity has a $C^{a/b}$ structure [@Anderson-Schoen]. Motivated by this result we investigate to what extent the simply connected hypothesis may be relaxed when compactifying the manifold and what resulting regularity may be obtained for the compactified manifold with boundary. In particular, let $M$ be a complete, noncompact Riemannian $(n+1)$-manifold. Define an *essential subset* \[page:es\] $K$ of $M$ to be a compact $(n+1)$-dimensional Riemannian submanifold with boundary, such that $Y := \partial K$ is a smooth hypersurface that is convex with respect to the outward pointing unit normal and such that $\exp\colon N^{+}Y {\rightarrow}\overline{{M \backslash K}}$ is a diffeomorphism, where $N^{+}Y \approx Y \times [0, \infty)$ is the outward normal ray bundle of $Y$. The main result of this paper is:
Let $(M,g)$ be a complete, noncompact Riemannian $(n+1)$-manifold. Suppose that there exists $K \subset M$, a compact $(n+1)$-dimensional Riemannian submanifold with boundary that satisfies:
1. $K$ is totally convex in $M$, i.e. if $p, q \in K$ and $\gamma\colon[0,1] {\rightarrow}M$ is any geodesic with $\gamma(0) = p$ and $\gamma(1) = q$, then $\gamma( [0,1]) \subset K$.
2. ${M \backslash K}$ satisfies the following curvature assumption: $$\begin{aligned}
\label{curvasmp}
-\infty < -b^2 \leq \sec( {M \backslash K}) \leq -a^2 < 0.\end{aligned}$$
Then $K$ is an essential subset of $M$ and ${M^{*}}:= M \cup {M(\infty)}$ is a geometric compactification of $M$ as a topological manifold with boundary. The boundary is homeomorphic to $\partial K$. Further, ${M^{*}}$ is endowed with the structure of a $C^{a/b}$ manifold with boundary, independent of the choice of $K$. \[thm:main\]
Since any point $x$ in a Cartan-Hadamard manifold $M$ may be regarded as a pole, any small closed ball about $x$ is easily seen to be an essential subset for $M$. Therefore Theorem \[thm:main\] generalizes and strengthens the Anderson-Schoen result, for it allows for much greater variety in the topology of $M$; essential subsets relax the stringent hypothesis of simple connectedness used in the Anderson-Schoen paper. In addition, the result here proves the regularity of the *entire* compactification ${M^{*}}= M \cup {M(\infty)}$. The Anderson-Schoen theorem only proves the regularity for the boundary at infinity ${M(\infty)}$.
The outline of this paper is as follows. In Section \[sec:notation\] we outline our notation and explain our comparison theorems. In Section \[sec:tconvex\] we provide a condition for an essential subset. In Section \[sec:tracey\] we describe the compactification of $M$ as a topological manifold, and then in Section \[sec:compreg\] we set up the necessary estimates to show the compactified manifold has a well defined $C^{a/b}$ structure.
We would like to thank our advisor Jack Lee for suggesting this problem to us and for his constant guidance. We would also like to thank Robin Graham for suggestions on an early draft of this paper.
Notation and basic estimates {#sec:notation}
============================
In this section we outline our notation and provide the estimates that will be used in the subsequent comparison geometry. Throughout this paper $M$ denotes a complete, connected, noncompact Riemannian $(n+1)$-manifold with metric $g$. The letter $K$ will always denote an essential subset and $Y:=\partial K$ will denote the smooth hypersurface boundary. Throughout this paper we assume the curvature assumption and write $\alpha = a/b$. There is no loss in generality in assuming that the pinching constants in satisfy $a \leq 1 \leq b$, for the ratio of maximum to minimum sectional curvature, $a/b$, is invariant under a homothety of the metric. Further, we follow the curvature sign conventions given in [@Lee] or [@Petersen]: for orthonormal vectors $X,Z$, the sectional curvature of the plane they span is $ \sec(X,Z) = R(X,Z,Z,X)$, where $R$ is the Riemannian curvature 4-tensor.
We trivialize the normal ray bundle with respect to the outward unit normal for $Y$ as $N^{+}Y \approx Y \times [0, \infty)$. The exponential map restricted to $N^{+}Y$ is written ${E}$. For $p \in Y$, the notation $\gamma_p$ denotes the geodesic normal to $Y$ emanating outwards from $p$. We call a geodesic ray *untrapped* if it eventually escapes any compact set.
Following Petersen [@Petersen] we reserve the term *geodesic segment* for a distance minimizing geodesic curve between two points.
It is easy to verify that $E\colon Y \times [0, \infty) {\rightarrow}M$ is a local diffeomorphism at every point of $Y \times \{0\}$. Therefore by compactness of $Y$ we may obtain an $\epsilon > 0$ and a one-sided collar neighbourhood $T$ of $Y$ so that $E\colon Y\times [0, \epsilon) {\rightarrow}T$ is a diffeomorphism. Let $r\colon T {\rightarrow}{\mathbb{R}}$ denote the distance to $Y$. Adapting the proof of the classical Gauss lemma (see [@Lee] or [@Klingenberg] for a precise statement of the classical form) we obtain a decomposition of the metric as $g = dr^2 + g_Y(y,r)$. Further, if we choose any coordinates $\{y^{\beta}\}$ on an open set $U \subset Y$ we may get coordinates on $T$ by extending $y^{\beta}$ to be constant along the integral curves of $\operatorname{grad}r$, and then $(y^{\beta}, r)$ form coordinates on $E(U \times [0, \epsilon)) \subset T$. We will refer to such coordinates as *Fermi coordinates for $Y$*, and in Section \[sec:tracey\] we will see that total convexity implies Fermi coordinates for $Y$ exist on neighbourhoods of the form $U \times [0, \infty)$.
We use Latin indices to index directions in $M$ and consequently these indices range from $0$ to $n$. We use Greek indices to index the directions along $Y$ which range from $1$ to $n$, and a zero or ‘$r$’ to index the direction normal to $Y$.
We now consider the second fundamental form of $Y$ in $M$ and we fix signs by taking our definition as $h(X,Z) = g( \nabla_X Z, -\partial_r)$, where $\nabla$ is the connection in $M$, and $X, Z$ are vector fields on $Y$ extended arbitrarily to vector fields on $M$. Note that this definition uses the *inward* pointing unit normal. Given this convention, we say $Y$ is *convex* (respectively *strictly convex*) with respect to the outward unit normal if the scalar second fundamental form is positive semidefinite (respectively positive definite).
In Fermi coordinates the second fundamental form of $r$-level sets may be written as $(h_r)_{\beta \gamma} = \frac{1}{2} \partial_r g_{\beta \gamma}$. We raise an index to obtain a family of shape operators $S(r)$, where $S(r)^{\beta}_{\gamma} = g^{\beta \nu} (h_r)_{\nu \gamma}$. A computation shows that $S$ satisfies a Riccati equation involving curvature, namely $$\begin{aligned}
\label{eqn:ricatti}
( \partial_r S(r) + S(r)^2 )_{\beta}^{\nu} = -R_{0 \beta \; 0}^{\; \; \; \; \nu}.\end{aligned}$$
We will make use of Jacobi fields suitable to our coordinates. Fix $p \in Y$ and consider the outward normal geodesic $\gamma_p$. Choose any curve $\sigma$ in $Y$ such that $\sigma(0) = p$ and define a variation through geodesics by $\Gamma(s,t) = {E}(\sigma(s), t )$. This gives rise to a Jacobi field $J(t) = \partial_s \Gamma(s,t) |_{s=0}$ along $\gamma_p$. Explicitly, $$J(t) = \dot\sigma^{\beta}(0) \partial_{\beta} |_{(\sigma(0),t)}.$$ So these special Jacobi fields have constant components in Fermi coordinates. Convexity of $Y$ easily implies the following estimates:
Let $J(t) = \dot\sigma^{\beta}(0) \partial_{\beta} |_{(\sigma(0),t)}$ be the Jacobi field along $\gamma$ described above. Then
1. ${\left\langle J(0) \;, \; D_t J(0) \right\rangle} \geq 0$,
2. $|J(t)| \geq |J(0)| \; \cosh(at)$.
\[lemma:jacobigrowth\]
The comparison theorems we use are based on the treatment given in [@Petersen]. These are obtained by analysis of the Riccati differential equation . In what follows an inequality involving the shape operator of the form $S \geq c$ means that every eigenvalue of $S$ is greater than or equal to $c$. Inequalities involving a metric are to be interpreted as inequalities between quadratic forms.
For the metric comparisons that follow we require a covering of the compact hypersurface $Y$. Fix $\epsilon = \frac{1}{2} \min\{ \operatorname{inj}(Y), \pi \}$, where $\operatorname{inj}(Y)$ is the injectivity radius of $g_Y(y,0)$. For any $y \in Y$, the ball $B^Y_{\epsilon}(y)$ is the domain of a convex normal coordinate chart with image $B_{\epsilon}(0) \subset {\mathbb{R}}^n$. On the ball $B_{\epsilon}(0)$, we will need to consider three metrics, the original $g_Y$ (transfered to $B_{\epsilon}(0)$ by means of normal coordinates), the round metric on the unit sphere ${\mathbb{S}}^{n}$ in normal coordinates, and the flat metric in coordinates. We will denote the round metric hereafter by ${\mathring{g}}$ and the flat metric by ${\overline{g}}$. On compact subsets of $B_{\epsilon}(0)$ all three of these metrics are comparable. Since $g_y(0,0)_{\beta \nu} = {\mathring{g}}(0)_{\beta \nu} = {\overline{g}}(0)_{\beta \nu} = \delta_{\beta \nu}$, continuity of the metrics implies we may find an $r = r(y)$ with $0 < r < \epsilon/2$ so that
- $\frac{1}{4} {\mathring{g}}_{\beta \nu} \leq (g_Y)_{\beta \nu} \leq 4 {\mathring{g}}_{\beta \nu}$ on $B_r(0)$,
- $\frac{1}{4} {\overline{g}}_{\beta \nu} \leq (g_Y)_{\beta \nu} \leq 4 {\overline{g}}_{\beta \nu}$ on $B_r(0)$,
- $B_r(0) \subset B_{2r}(0) \subset B_{\epsilon}(0)$.
Compactness of $Y$ yields a finite subcover of the balls $B_{r(y)}(y)$ that cover $Y$. Label these finitely many balls $W_i$. Label the balls with the same centres and radius $2r(y)$ as $V_i$ and observe $W_i \subset \overline{W_i} \subset V_i$. We refer to the covering of $Y$ by $\{W_i\}$ as the *reference covering for Y*.
The choice of this covering ensures that distances between points in $W_i \subset Y$ with respect to the metrics $(g_Y)_i$, ${\mathring{g}}_i$, and ${\overline{g}}_i$ are all comparable. We refer to this property again as the *distance comparison principle*. \[dcp\]
In the theorem that follows and throughout the note we take eigenvalues of the metric $g_Y$ with respect to the euclidean metric ${\overline{g}}_i$ in normal coordinates for the $V_i$. We let $\Omega_i$ denote the maximum eigenvalue of $g_Y(y,0)$ in each $\overline{W_i}$, and then set $\Omega = \max_i \; \Omega_i$. Similarly, let $\omega$ be the minimum eigenvalue of $g_Y(y,0)$ over the cover $\overline{W_i}$. As this covering and constants will be used throughout the paper, we always use normal coordinates along $Y$ in any choice of Fermi coordinates that follows.
An adaptation of the comparison theorems in [@Petersen] yields the following theorem.
Let $(y^{\beta}, r)$ be Fermi coordinates for $Y$ on $W_i \times [0,\infty)$ for an open set $W_i$ in the reference covering described above. Let $\Lambda, \lambda$ denote the maximal and minimal eigenvalues of the shape operator over $Y$. We have: \[thm:comparison\]
**Shape operator estimate:** $$\begin{aligned}
\label{eqn:shapeinequality}
a \tanh( a( r + L_1) ) \; \delta^{\beta}_{\gamma} \; \leq \; S^{\beta}_{\gamma}(y, r) \; \leq \; b \coth( b( r + L_2) ) \; \delta^{\beta}_{\gamma},\end{aligned}$$ where $L_1 = \frac{1}{a} \tanh^{-1}(\frac{\lambda'}{a})$, $L_2 = \frac{1}{b} \coth^{-1}(\frac{\Lambda'}{b})$, and $$\begin{aligned}
\label{eqn:constants}
\Lambda' = \begin{cases} \Lambda & \text{ if } \Lambda > b, \\
2b & \text{ if } \Lambda \leq b . \end{cases} \nonumber \\
\lambda' = \begin{cases} \lambda & \text{ if } \lambda < a, \\
\frac{a}{2} & \text{ if } \lambda \geq a . \end{cases}\end{aligned}$$ **Metric estimate:** $$\begin{aligned}
\label{eqn:metric}
L_3 \; \cosh^2( a( r+ L_1 ) ) \; \delta_{\beta \nu} \leq g_{\beta \nu}(y, r) \leq L_4 \; \sinh^2( b( r+ L_2 ) ) \; \delta_{\beta \nu}, \end{aligned}$$ where $\displaystyle L_3 = \left( \frac{\omega}{\cosh^2{ a L_1 } } \right)$, $\displaystyle L_4 = \left( \frac{\Omega}{\sinh^2{ b L_2 } } \right)$, and $\Omega, \omega$ are described above.
\[lemma:metriccomp\]
Essential Subsets {#sec:tconvex}
=================
In this section we provide a sufficient condition for the submanifold $K \subset M$ to be an essential subset. We assume that $K$ is a compact $(n+1)$-dimensional Riemannian submanifold with boundary, such that $Y := \partial K$ is a smooth hypersurface that is convex with respect to the outward pointing unit normal. We discuss a condition that ensures that $E\colon Y \times [0, \infty) {\rightarrow}\overline{{M \backslash K}}$ is a diffeomorphism. This property allows us to relax the hypothesis that $M$ be simply connected in the Anderson-Schoen result; essential subsets replace the requirement that the map $exp_p\colon T_p M {\rightarrow}M$ be a diffeomorphism which is ensured by the Cartan-Hadamard theorem.
The basic idea of an essential subset $K$ is to capture the topology of the manifold $M$ inside $K$ in a such way that the normal geodesic flow off of the boundary of $K$ is a diffeomorphism of the outward normal bundle $Y \times [0, \infty)$ onto $\overline{{M \backslash K}}$. Some sort of hypothesis on the topology and geometry of $M \backslash K$ is necessary in order for the exponential map to be injective, as can be seen by considering the orbit space obtained by taking the upper half space model of the two-dimensional hyperbolic plane under the action of the discrete group of dilations $G = \{ 2^n: n \in {\mathbb{Z}}\}$. Given a sufficiently small ball centred at any point along the ‘waist’ of this quotient space, i.e. the image of $x=0$, one finds normal geodesics off the ball that intersect.
A subset $K \subset M$ is *totally convex in M* if whenever $p, q \in K$ and $\sigma\colon[0,1] \longrightarrow M$ is a geodesic such that $\sigma(0)=p$, $\sigma(1) = q$ then $\sigma([0,1]) \subset K$. The inclusion of $K$ into $M$ is a homotopy equivalence; see [@Klingenberg] for details. It is interesting that totally convex sets play an important and somewhat analogous role in the of theory of souls of positively curved manifolds. We again refer the interested reader to [@Klingenberg] and the references therein.
We have the following sufficient condition for an essential subset.
Let $K \subset M$ be a compact Riemannian submanifold with hypersurface boundary $Y$. Suppose that \[thm:characterizationES\]
- K is totally convex in M, and
- $\sec({M \backslash K}) \leq 0$.
Then $K$ is an essential subset for $M$.
As $K$ is totally convex, it is also geodesically convex ( i.e. $K$ contains a geodesic segment between any two of its points ). It is well known that $K$ geodesically convex implies that $Y= \partial K$ is convex.
We now check that the image $E(Y\times [0, \infty))$ is a subset of $\overline{{M \backslash K}}$. To see this, notice that any normal geodesic $\gamma_p$ that re-enters $K$ must lie entirely inside $K$ since $K$ is totally convex, but this violates the fact $\gamma_p$ is a geodesic ray with an outward pointing tangent vector at $p$.
Next, Jacobi field estimates and the nonpositive curvature assumption on ${M \backslash K}$ imply that $E$ is a local diffeomorphism on $Y \times [0, \infty)$. We need only argue that $E$ is bijective. Surjectivity of $E$ onto ${M \backslash K}$ is easy to see: for any point $q \in {M \backslash K}$ there is a closest point $p \in K$ to $q$, and it is straightforward to argue that $\gamma_p(t_0) = q$ for some $t_0$. In order to argue injectivity of $E$, let $$\begin{split}
\epsilon_0 = \sup \{ \epsilon > 0 \; | \; & E\colon Y \times [0, \epsilon) {\rightarrow}\overline{{M \backslash K}} \; \text{is a diffeomorphism} \\
& \text{onto its image} \}.
\end{split}$$ If $E$ is not injective then for $n > 0$ we have distinct points $p_n, q_n \in Y$ and $r$-values $t_n, s_n \in [0, \epsilon_0 + 1/n)$ such that $E(p_n, t_n) = E(q_n, s_n)$. By the choice of $\epsilon_0$ it is impossible that both $t_n < \epsilon_0$ and $s_n < \epsilon_0$, so we may assume $ \epsilon_0 < s_n < \epsilon_0 + 1/n$. It is straightforward to argue that a bound of the form $\epsilon_0 - 2/n < t_n < \epsilon_0 + 1/n$ holds as well. By compactness of $Y$ we may pass to convergent subsequences and obtain points $p$ and $q$ such that $E(p, \epsilon_0) = E(q, \epsilon_0)$. The points $p$ and $q$ are distinct as $p_n$ and $q_n$ are distinct and $E$ is a local diffeomorphism in a neighbourhood of $(p, \epsilon_0)$. We set $m = E(p, \epsilon_0) = E(q, \epsilon_0)$.
We now show that $\gamma_p$ and $\gamma_q$ meet ‘head on’, i.e. that $\gamma_p'(\epsilon_0) = -\gamma_q'(\epsilon_0)$. If this is not the case, choose a vector $X \in T_{m} ({M \backslash K})$ such that $$\label{neginnerprod}
g( X, \gamma_p'(\epsilon_0) ) < 0 \; \text{and} \; g( X, \gamma_q'(\epsilon_0) ) < 0.$$
Consider any path $\sigma\colon (-\delta, \delta) \rightarrow {M \backslash K}$ for some $\delta > 0$ with $\sigma(0) = m$, $\sigma'(0) = X$. Because $E$ is a local diffeomorphism, near $p$ we have a smooth curve $\sigma_Y^p\colon (-\delta, \delta) \rightarrow Y$ and a positive smooth function $r^p$ such that $\sigma(t) = E( \sigma_Y^p(t), r^p(t))$. We may obtain a variation through $Y$-normal geodesics by considering the map $\Gamma(s,t) = E( \sigma_Y^p(s), t r^p(s) )$. The first variation formula and equation above imply that $r^p(s)$ is a decreasing function of $s$; in particular for $s$ sufficiently small and positive, $r^p(s) < \epsilon_0$. The same argument may be applied near $q$, and this implies for small enough $s$ that $E( \sigma_Y^p(s), r^p(s) ) = E( \sigma_Y^q(s), r^q(s) )$ which contradicts the choice of $\epsilon_0$.
We have proved that $\gamma_p(\epsilon_0) = \gamma_q(\epsilon_0)$ and $\gamma_p'(\epsilon_0) = -\gamma_q'(\epsilon_0)$. Therefore $\gamma_p$ is a geodesic such that $\gamma_p(0) = p \in K$ and $\gamma_p(2 \epsilon_0) = q \in K$, and so the image of $\gamma_p$ is contained in $K$ by total convexity, a contradiction. Thus $E$ is a bijective local diffeomorphism, and consequently $E\colon [0,\infty) {\rightarrow}\overline{{M \backslash K}}$ is a diffeomorphism. This means $K$ is an essential subset of $M$.
The topological compactification {#sec:tracey}
================================
\[sec:topcomp\]
In this section we explain how to compactify $M$ given a choice of essential subset $K$ by extending the exponential map to take values in ${M^{*}}\backslash K = ({M \backslash K}) \cup {M(\infty)}$. Fix an essential subset $K$ and define an extension ${\overline{E}}\colon Y \times (0,1] {\rightarrow}{M^{*}}\backslash K$ by
$${\overline{E}}(p,s) = \begin{cases} {E}(p, 2 \tanh^{-1}{s}) & \text{ if } s \in (0,1), \\
[ \; {E}( p, t): t \geq 0\; ] & \text{ if } s = 1. \end{cases}$$
\[ebarref\]
Recall that the notation $[\gamma]$ above represents the equivalence class of the geodesic ray $\gamma$ under the asymptotic equivalence relation, see page . We have also collapsed the normal component using a diffeomorphism. We now verify that ${\overline{E}}$ is a bijection. Relative to the diffeomorphism $E\colon Y \times [0, \infty) {\rightarrow}{M \backslash K}$ of the previous section, we write a generic curve $\sigma$ in ${M \backslash K}$ as $\sigma = (\sigma_Y, \sigma_r)$.
${\overline{E}}$ is injective. \[prop:Ebaroneone\]
We must show that given distinct points $p, q \in Y$ that the normal geodesics $\gamma_p, \gamma_q$ have unbounded distance as a function of time. It suffices to show given any $t > 0$ that the length of any curve $\sigma$ from $\gamma_p(t)$ to $\gamma_q(t)$ is bounded below by an unbounded function of time.
Suppose that $\sigma$ leaves the collar $Y \times [t/2, \infty)$. Then the normal contribution of the length integral and the decomposition of the metric imply $len(\sigma) \geq t$. In case that $\sigma$ remains in the collar, $len(\sigma) \geq len(\sigma^P)$, where $\sigma^P$ is the projection of $\sigma$ onto $Y \times \{t/2\}$, i.e. $\sigma^P(s) = (\sigma_Y(s), t/2)$. Then Jacobi field estimates imply that $$len(\sigma) \geq len( \sigma^P ) \geq \cosh(a t/2) d_Y(p,q).$$ Therefore the length of any curve from $\gamma_p(t)$ to $\gamma_q(t)$ is bounded below by an unbounded function of time.
In order to show that ${\overline{E}}$ is surjective we consider an untrapped geodesic ray $\sigma$ parametrized by arc length. Recall that untrapped means that $\sigma$ eventually escapes every compact set. Eventually $\sigma$ remains inside ${M \backslash K}$ and we take $\sigma(0)$ to be any point inside ${M \backslash K}$. In this parametrization, the growth of $\sigma_r$ is bounded below by a linear function.
Let $\sigma = (\sigma_Y, \sigma_r)$ be an untrapped geodesic ray in ${M \backslash K}$ parametrized by arc length. Then there exist constants $C, B, t_0 > 0$ such that $$\sigma_r(t)\geq C t + B, \mbox{ for all } t > t_0.$$ \[lemma:geodesicrgrowth\]
As a geodesic, the normal component of $\sigma$ satisfies $$\ddot\sigma_r + \dot\sigma^{\alpha} \dot\sigma^{\beta} \Gamma^0_{\alpha \beta} = 0,$$ where we have used the index conventions from Section \[sec:notation\] and the decomposition of the metric $g = dr^2 + g_Y(r)$ to obtain the vanishing of the $\Gamma^0_{00}$ and $\Gamma^{0}_{\alpha 0}$-Christoffel symbols. The $\Gamma^0_{\alpha \beta}$-Christoffel symbol is the scalar second fundamental form and so by our comparison results of Theorem \[lemma:metriccomp\] we have: $$\begin{split}
\label{eqn:sigmar}
\ddot\sigma_r & = - \dot\sigma^{\alpha} \dot\sigma^{\beta} \Gamma^0_{\alpha \beta} \\
& \geq a \tanh(a(r + L_1) ) |\dot\sigma_Y|^2_{g} \\
& \geq 0.
\end{split}$$
Thus $\dot\sigma_r$ is nondecreasing. Since $\sigma_r$ is eventually unbounded, there is a $t_0$ where $\dot\sigma_r(t_0) > 0$ and so $\dot\sigma_r(t) \geq \dot\sigma_r(t_0) > 0$ for $t > t_0$. Upon integration we find that $\sigma_r(t) \geq \dot\sigma_r(t_0) (t - t_0) + \sigma_r(t_0)$.
We now find a candidate base point for a normal geodesic asymptotic to $\sigma$.
Let $\sigma_Y \colon [0, \infty) {\rightarrow}Y$ be the projection of $\sigma$ onto $Y$. Then $len( \sigma_Y ) < \infty$.
Since $\sigma$ is parametrized by arc length, $ \dot\sigma^{\alpha} \dot\sigma^{\beta} g_{\alpha \beta}(\sigma(t)) = |\dot\sigma_Y|_g \leq 1$. The metric estimate of Theorem \[thm:comparison\] implies $$L_3 \; \cosh^2( a( \sigma_r(t)+ L_1 ) ) \sum_{\alpha} (\dot\sigma_Y^{\alpha})^2 \leq \dot\sigma^{\alpha} \dot\sigma^{\beta} g_{\alpha \beta}( \sigma(t) ) \leq 1.$$ We may also consider the projected curve $\sigma_Y$. Here the metric estimates imply $$\begin{aligned}
\dot\sigma^{\alpha} \dot\sigma^{\beta} g_{\alpha \beta}( \sigma_Y(t),0 ) \leq 4 L_4 \; \sinh^2( b L_2 ) \sum_{\alpha} (\dot\sigma_Y^{\alpha})^2.\end{aligned}$$ Combining these estimates and the result of Lemma \[lemma:geodesicrgrowth\] we obtain: $$\begin{aligned}
\dot\sigma^{\alpha} \dot\sigma^{\beta} g_{\alpha \beta}( \sigma_Y(t) ) \leq \frac{C(L_2, L_3, L_4)}{ \cosh^2( a(Ct + B) + L_1)}\end{aligned}$$ Integrating the square root of both sides we find that $len( \sigma_Y ) < \infty$.
Since $len( \sigma_Y ) < \infty$, the completeness of $M$ implies that $\sigma_Y$ has a limit, and since $Y$ is closed this limit is a point $p \in Y$. Let $\gamma_p(t)$ denote the outward normal geodesic emanating from $Y$ at $p$. We now show that ${\overline{E}}$ is surjective by showing that $\gamma_p$ is asymptotic to $\sigma$.
${\overline{E}}$ is surjective. \[prop:Ebaronto\]
Given the untrapped geodesic ray $\sigma$ above, the previous lemma establishes the existence of a candidate normal geodesic $\gamma_p$ to represent the equivalence class $[\sigma]$ in ${M(\infty)}$. We prove that $d( \gamma_p(t), \sigma(t) )$ remains bounded as $t {\rightarrow}\infty$. The triangle inequality implies that it is sufficient to show separately that $$d\big( (p, \sigma_r(t) ), (\sigma_Y(t), \sigma_r(t) ) \big) \; \text{and} \; d \big( (p, t), (p, \sigma_r(t)) \big)$$ remain bounded as functions of time.
*Step 1.* *$d( (p, t), (p, \sigma_r(t)) )$ is bounded as $t {\rightarrow}+\infty$:* In this situation, $d( (p, t), (p, \sigma_r(t)) ) = |t - \sigma_r(t)|$, so we must show the quantity $|t - \sigma_r(t)|$ is bounded as $t {\rightarrow}\infty$. By the fundamental theorem of calculus this is equivalent to the statement $1 - \dot\sigma_r(t) \in L^1(t_0, \infty)$ for $t_0$ sufficiently large: $$|t - \sigma_r(t)| \leq \int^{t}_{t_0} 1 - \dot\sigma_r(s) ds + |t_0 - \sigma_r(t_0)|,$$ Since $\dot\sigma_r^2 + |\dot\sigma_Y|^2 = 1$, the integrability of $1 - \dot\sigma_r(t)$ is related to the integrability of $|\dot\sigma_Y|^2$, for $$1 - \dot\sigma_r(t) \leq (1 - \dot\sigma_r(t))(1 + \dot\sigma_r(t)) = |\dot\sigma_Y|^2,$$ for large $t$. Just as in the proof of Lemma \[lemma:geodesicrgrowth\], the special form of the $r$-component of the geodesic equation in Fermi coordinates and the estimates of Theorem \[thm:comparison\] imply that $\ddot\sigma_r \geq 0$ and $$|\dot\sigma_Y|^2 \leq (1/a) \coth(a(r + L_1) \ddot\sigma_r.$$ The fundamental theorem of calculus applied to $\ddot\sigma_r$ implies that $\ddot\sigma_r \in L^1(t_0, \infty)$ and consequently since $\coth(a(r + L_1)$ is bounded, shows that $|\dot\sigma_Y|^2 \leq C \ddot\sigma_r$ for large enough $t$. Thus $|\dot\sigma_Y|^2$ is integrable and the quantity $|t-\sigma_r(t)|$ remains bounded as $t {\rightarrow}+\infty$.
*Step 2.* *$d( (p, \sigma_r(t)), (\sigma_Y(t), \sigma_r(t) ))$ is bounded as $t {\rightarrow}+\infty$:* For each $t_0$ consider the curve $$\tau^{(t_0)}(s) = ( \sigma_Y(s), \sigma_r(t_0) ), \mbox{ on} \; [t_0, \infty).$$ Clearly $ d( ( \sigma_Y(t_0), \sigma_r(t_0) ), (p, \sigma_r(t_0)) ) \leq len( \tau^{(t_0)}) $, and so it suffices to show that $len( \tau^{(t_0)})$ is bounded above by a constant independent of $t_0$. To this end we use Jacobi field estimates based at the $r$-level set of value $\sigma_r(t_0)$.
In particular consider the Jacobi field along $\gamma_{\sigma_Y(s)}(t)$ given in Fermi coordinates as the constant vector field $$J_s(t) = (\dot\sigma_Y(s), t) \in T_{(\sigma_Y(s), t)} \overline{{M \backslash K}}.$$ See Section \[sec:notation\] for a description of these special Jacobi fields. In order to make our application of Lemma \[lemma:jacobigrowth\] transparent, rescale the time parameter by $\lambda = t - \sigma_r(t_0)$. Then $J_s( 0 ) = \dot\tau^{(t_0)}(s)$, and for $s \in (t_0, \infty)$, $t \in (\sigma_r(t_0), \infty)$ and $\lambda \in (0, \infty)$, Lemma \[lemma:jacobigrowth\] implies $$|J_s(\lambda)| \geq |J_s(0)| \cosh( a \lambda ).$$ Therefore we may write: $$1 \geq |\dot\sigma_Y(s)|_{g_Y(\sigma_r(s))} = |J_s(\sigma_r(s))| \geq |\dot\tau^{(t_0)}(s)| \cosh( a ( \sigma_r(s) - \sigma_r(t_0))).$$ Consequently using this estimate and the estimate on $\sigma_r$ from Lemma \[lemma:geodesicrgrowth\], we find that $$|\dot\tau^{(t_0)}(s)| \leq \frac{1}{\cosh( a( C s + B - \sigma_r(t_0)) )},$$ for constants, $C, B$. Upon integration of this expression we find $$\begin{aligned}
len( \tau^{(t_0)} ) &\leq& \int^{\infty}_{t_0} \frac{1}{\cosh( a( C s + B - \sigma_r(t_0)) )}\\
&\leq& C(a) \tanh^{-1} (e^{ a( Cs + B - \sigma_r(t_0) ) } ) \big|^{s=\infty}_{s=t_0}\\
&\leq& C(a) \frac{\pi}{2}.\end{aligned}$$
Thus $len( \tau^{(t_0)})$ is bounded independent of $t_0$. This completes step 2.
The proof of the above proposition can be extended to yield a stronger result that will be useful in proving that the topology on ${M^{*}}$ is well defined. In the lemma below we consider two hypersurfaces $Y_1, Y_2$ where $Y_i$ is the boundary of an essential subset $K_i$. We use the notation $\gamma_{p'}$ to denote the normal geodesic to $Y_1$ emanating from the point $p' \in Y_1$, and $\sigma_{q'}$ to denote the normal geodesic to $Y_2$ emanating from the point $q' \in Y_2$.
Let $K_1$ and $K_2$ be essential subsets of $M$ with $Y_i = \partial K_i$. Given any point $q \in Y_2$ there exists an open neighbourhood $V_q \subset Y_2$ of $q$ and $B > 0$ such that for every $q' \in V_q$: $$d(\sigma_{q'}(t), \gamma_{p'}(t)) \leq B, \mbox{for all } t \geq 0,$$ where $\gamma_{p'}$ is the unique normal geodesic ray emanating from $Y_1$ that is asymptotic to $\sigma_{q'}$. \[lemma:unifbdd\]
Let the exponential map ${E}\colon Y_1 \times [0, \infty) {\rightarrow}{M \backslash K}_1$ be denoted ${E}_1$.
Fix $q \in Y_2$ and $R>0$. Now $\sigma_{q}$ is eventually outside every compact set, so there exists $T \geq 0$ such that $\sigma_{q}(t) \in {E}_1(Y_1 \times [R, \infty))$ for $t\geq T$. Further the $r$-component of this curve, $(\sigma_{q})_r$, is eventually strictly monotone increasing so we may increase $T$ if necessary to ensure that $(\dot\sigma_{q})_r(t) > 0$ for $t \geq T$. By continuity of the exponential map, there is a precompact open ball $V_q$ in $Y_2$ about $q$ such that for any $q' \in V_q$ we have both $\sigma_{q'}(t) \subset {E}_1( Y_1 \times [R, \infty) )$ and $\dot\sigma_{q'}(t) > 0$ for all $t \geq T$. For each such $q'$, let $p'$ be the unique element of $Y_1$ such that $\gamma_{p'}$ is asymptotic to $\sigma_{q'}$. By compactness, for any $q' \in \overline{V_q}$, $d(\sigma_{q'}(t), \gamma_{p'}(t))$ is uniformly bounded for $t \in [0, T]$. We need only check that a uniform bound holds for the tails of the geodesics emanating from $V_q$.
In order to estimate $d(\sigma_{q'}(t), \gamma_{p'}(t))$ for $t\geq T$, proceed as in the proof of Proposition \[prop:Ebaronto\]. By continuity of the exponential map and by shrinking $V_q$ if necessary all constants may be chosen independently of $q'$.
We may now prove the topological part of Theorem \[thm:main\]:
\[thmmarsh\] If $M$ is a complete, noncompact Riemannian manifold with an essential subset $K$ with pinched negative curvature as in (\[curvasmp\]), then $M$ admits a geometric compactification as a topological manifold with boundary.
Throughout the proof we use the notation for $\gamma$ and $\sigma$ as described preceding Lemma \[lemma:unifbdd\].
Suppose $K_1, K_2$ are two essential subsets of $M$. The propositions above imply that we get bijections ${\overline{E}}_i\colon Y_i \times (0,1] {\rightarrow}{M^{*}}\backslash K_i$. Endow each subset ${M^{*}}\backslash K_i$ with the topology $\tau_i$ that makes ${\overline{E}}_i$ a homeomorphism. We now show that these topologies are equivalent. Let $K$ be a compact set such that $K \supset K_1 \cup K_2$. Consider the identity map from $({M^{*}}\backslash K, \tau_1) {\rightarrow}({M^{*}}\backslash K, \tau_2)$, i.e. the composition $\operatorname{\overline{\psi}}= {\overline{E}}_2^{-1} \circ {\overline{E}}_1$. To show that the topology on ${M^{*}}\backslash K$ is independent of $K_i$, it suffices to show that $\operatorname{\overline{\psi}}$ is a homeomorphism. By the symmetric roles of the $K_i$, it suffices to prove that $\operatorname{\overline{\psi}}$ is an open map.
We already have that $\psi = {E}_2^{-1} \circ {E}_1$ is a diffeomorphism. We need only check that open neighbourhoods in $\tau_1$ of points in ${M(\infty)}$ are taken to open neighbourhoods in $\tau_2$. Choose a basis element of the form ${\overline{E}}_1(U \times (c,1])$ where $U$ is open in $Y_1$. For every $[\gamma_p] \in {M(\infty)}\cap {\overline{E}}_1(U \times (c,1])$ where $\gamma_p$ is asymptotic to $\sigma_q$ we must find a neighbourhood $V_q \subset Y_2$ of $q$ and $d > 0$ so that ${\overline{E}}_2( V_q \times (d, 1] ) \subset {\overline{E}}_1(U \times (c,1])$. It is sufficient to show that ${\overline{E}}_2( V_q \times (d, 1) ) \subset {\overline{E}}_1(U \times (c,1))$; equivalently we may show $E_2( V_q \times (d, \infty) ) \subset E_1(U \times (c,\infty))$ for some different constants $c, d$.
Set $W := {\overline{E}}_1(U \times (c,1))$. The tail of $\sigma_q$ is eventually in W; we may choose a $T>0$ so that $\sigma_q(t) \in W$ for $t\geq T$. Since $\psi$ is a diffeomorphism, for each $t > T$ we may get a ball $V_q(t)$ about $q$ and $\epsilon(t) > 0$ such that ${\overline{E}}_2(V_q(t) \times (t-\epsilon(t), t+\epsilon(t))) \subset W$. We may assume that for $t_2 > t_1$ we have $V_q(t_2) \supset V_q(t_1)$, and that the radii of these balls are less than the injectivity radius of $Y_2$. Now $\cap_{t > T} V_q(t)$ is either a ball or is the singleton set $\{q\}$. In case the intersection is a ball, $V_q$, we have that ${\overline{E}}_2( V_q \times (T, 1) ) \subset W$, which completes the argument. Otherwise choose $q_n {\rightarrow}q$ such that $q_n$ enters $W$ and eventually leaves it. Let $p_n$ be the corresponding points on $Y_1$ so that $\sigma_{q_n}$ is asymptotic to $\gamma_{p_n}$. By compactness of $Y_1$, we may pass to a convergent subsequence and assume that $p_n {\rightarrow}p_0$. But now the uniform bound of Lemma \[lemma:unifbdd\] and continuity of the exponential map imply that $$d( \sigma_{q}(t), \sigma_{p_0}(t) ) = \lim_{n{\rightarrow}\infty} d( \sigma_{q_n}(t), \gamma_{p_n}(t) ) \leq B.$$ This means that $\sigma_q$ is asymptotic to $\gamma_{p_0}$, and so by injectivity of ${\overline{E}}_1$, $p = p_0$. This implies $p_n {\rightarrow}p$, and so $p_n$ is eventually inside $U$, a contradiction. Thus $\cap_{t > T} V_q(t)$ contains a ball. Therefore the topology on ${M^{*}}\backslash K$ is well defined.
Regularity of the compactification {#sec:compreg}
==================================
The results of the previous section described how to compactify $M$ as a topological manifold with boundary given a specific choice of essential subset $K$, and that the topology of the compactification is independent of $K$. In this section we lay the groundwork and prove that the compactification ${M^{*}}$ has a $C^{a/b}$ structure. In order to do this we will first have to describe our explicit comparison with hyperbolic space and how this relates to Fermi coordinates. Next, since the manifold ${M \backslash K}$ is not complete and we estimate distances in ${M \backslash K}$ as compared to hyperbolic space we explain how to refine the reference covering for $Y$. Just as in hyperbolic space with a compact set $K$ removed, we have to check that for points $p$ and $q$ far enough from $K$ but with closest points $p'$ to $p$ and $q'$ to $q$ on $K$ sufficiently close, the geodesic segment from $p$ to $q$ remains in ${M \backslash K}$. Such a refined covering will be called a *special covering for Y*.
Given these geometric preliminaries we define a bounded metric $d_K$ on ${M \backslash K}$. Given two essential subsets $K_1$, $K_2$, each endowed with a special covering for $Y_i$, we establish a $C^{a/b}$ comparability estimate for distances in a subset of $M\backslash (K_1 \cup K_2)$. Then in the proof of the main theorem we explain why the distance estimate yields a $C^{a/b}$ structure for ${M^{*}}$.
We now describe our comparison geometry and modification of the metric comparison described at the end of Section \[sec:notation\]. In particular, consider the reference covering of $Y$ by small normal coordinate balls $W_i \subset \overline{W_i} \subset V_i$ as described preceding Theorem \[thm:comparison\]. In each $W_i$ we may use the metric ${\mathring{g}}_i$ to obtain a hyperbolic metric of constant curvature $-\lambda^2$ given by $$(h_{\lambda})_i = dr^2 + \frac{\sinh^2(\lambda r)}{\lambda^2} {\mathring{g}}_i.$$ \[hmetric\] We will call these metrics *hyperbolic comparison metrics*. A little algebra applied to the metric estimates of Theorem \[thm:comparison\] implies:
Consider Fermi coordinates $(y^{\beta}, r)$ for $Y$ on $W_i \times [0, \infty)$. There exists an $R = R( \Lambda, \lambda, \Omega, \omega, a, b)$ independent of $i$ such that for every $r > R$: \[thm:hcomp\] $$\begin{aligned}
\frac{ \sinh^2( a( r - R ) )} {a^2} \; { (\mathring{g}_i) }_{\beta \nu} \; \leq g_{\beta \nu}(y,r) \; \leq \;\frac{ \sinh^2( b( r+ R ) ) }{b^2}\; { (\mathring{g}_i) }_{\beta \nu}.\end{aligned}$$ In particular, for any points $p, q \in W_i \times [R, \infty)$ such that a geodesic segment[^1] from $p$ to $q$ lies entirely inside $W_i \times [R, \infty)$ then: $$d_a(p, q) \leq d(p,q) \leq d_b(p,q),$$ where $d_{\lambda}$ is the distance in the hyperbolic comparison metric on $W_i \times [R, \infty)$ of constant curvature $-\lambda^2$ described above.
We now provide an adaptation of the estimates used in [@Anderson-Schoen]. We first begin with some estimates in the two-dimensional hyperbolic plane of curvature $-\lambda^2$, $H^2(-\lambda^2)$. Let $p, q \in H^2(-\lambda^2)$, and take measurements from a third point $x \in H^2(-\lambda^2)$. Suppose that $s=d_{\lambda}(p,x), t = d_{\lambda}(q, x)$, and let $\theta$ be the angle between the radial geodesic connecting $x$ to $p$ and the radial geodesic connecting $x$ to $q$. The well known law of hyperbolic cosines [@Petersen] yields a formula involving the distance between $p$ and $q$ and these parameters: $$\label{eqn:hypdist}
\cosh(\lambda \; d_{\lambda}(p, q)) = \cosh(\lambda s) \cosh(\lambda t) - \sinh(\lambda s) \sinh(\lambda t) \cos(\theta).$$ We use this formula throughout this section. In the special case that $\theta = \pi/2$ we obtain the hyperbolic Pythagorean theorem, $$\cosh(\lambda \; d_{\lambda}(p, q)) = \cosh(\lambda s) \cosh(\lambda t).$$
Assume $t \geq s > 2R$. We have:
In a two-dimensional hyperbolic plane, there exist positive constants $c_1, c_2, c_3 > 0$ depending on $\lambda$ so that the following estimates hold: \[lemma:hcomp1\]
1. $d_{\lambda}(p,q) \leq \begin{cases} s + t + \frac{2}{\lambda} \log \theta + c_1 & \text{when } e^{\lambda s} \theta \geq 1, \\
t - s + c_2 & \text{when } e^{\lambda s} \theta \leq 4. \end{cases}$
2. $d_{\lambda}(p,q) \geq s + t + \frac{2}{\lambda} \log \theta - c_3$.
The above lemma is proved by straightforward estimation of and is essentially the form that Anderson-Schoen obtained in [@Anderson-Schoen].
For the next estimate, let $p, q, x \in H^2(-\lambda^2)$ be as before except assume that $s=t$. We need to estimate the distance from $x$ to the geodesic segment $\sigma$ from $p$ to $q$. Again straightforward estimation yields:
In a two-dimensional hyperbolic plane, there exists positive constants $c_4, c_5$ such that \[lemma:hcomp2\] $$d_{\lambda}(x, \sigma) \geq \begin{cases} - \frac{1}{\lambda} \log \theta - c_4 & \text{when }\; e^{\lambda s} \theta \geq 1, \\
s - c_5 & \text{when }\; e^{\lambda s} \theta \leq 4. \end{cases}$$
We now convert the above estimates in hyperbolic space into estimates suited to our Fermi coordinates.
Consider Fermi coordinates $(y^{\beta}, r)$ for $Y$ on $W_i \times [R, \infty)$[^2]. Let $p,q \in W_i \times [R, \infty)$. Then there exist positive constants $\{c_j\}_{j=1}^{8}$ depending only on $R$ and the reference covering such that:
1. $d_{b}(p,q) \leq \begin{cases} s_p + s_q + c_1 \log d_Y(p',q') + c_2 & \text{when } e^{b s_p} d_Y(p',q') \geq 2, \\
s_q - s_p + c_3 & \text{when } e^{b s_p} d_Y(p',q') \leq 2. \end{cases}$
2. $d_a(p, q) \geq s_p + s_q + c_4 \log d_Y(p',q') - c_5$,
where $p = (p', s_p)$, $q=(q', s_q)$ in coordinates, and $s_q \geq s_p$.
Also if $s_p = s_q$ and $\sigma$ is a geodesic segment in the hyperbolic comparison metric (cf. page ) from $p$ to $q$, then the minimum $r$-value of $\sigma$ satisfies:
1. $\sigma_{r_{min}} \geq \begin{cases} -c_6 \log d_Y(p',q') - c_7, &\text{when}\; e^{b s_p} d_Y(p',q') \geq 2, \\
s - c_8, & \text{when}\;e^{b s_p} d_Y(p',q') \leq 2. \end{cases} $
\[lemma:as\]
**Note:** In order to avoid a proliferation of constants we reuse the labels $c_1$ through $c_8$ above, hence these constants are not the same as the constants in Lemmas \[lemma:hcomp1\] and \[lemma:hcomp2\].
The points $p$ and $q$ lie in exactly one coordinate 2-plane $\Pi$ perpendicular to $Y$. Distances between $p$ and $q$ in the hyperbolic comparison metrics are realized by geodesics lying entirely in $\Pi$ and so we may use Lemmas \[lemma:hcomp1\] and \[lemma:hcomp2\] specific to two dimensions stated above in our metric comparisons. Further by the choice of reference covering and the distance comparison principle (cf. page ) the distance $\theta$ along the unit sphere is comparable to distance along $Y$. From the choice of reference covering it follows that if $e^{b s_p} d_Y(p',q') \geq 2$, then $e^{bs} \theta \geq 1$; similarly if $e^{b s_p} d_Y(p',q') \leq 2$, then $e^{bs} \theta \leq 4$.
The manifold ${M \backslash K}$ is not complete. Therefore we need to be careful when applying comparison geometry to estimate distances in ${M \backslash K}$, for geodesic segments could potentially leave the manifold ${M \backslash K}$ entirely. Fortunately the curvature assumptions imply that at least for points far enough from $Y$ whose nearest points on $Y$ are close enough, geodesic segments remain in the domain of a Fermi chart. We now explain how to obtain these special charts. For $x \in Y, \mu > 0$ and $t_0 > 0$, we call $$\begin{aligned}
TC(x, \mu, t_0) = \{ (y^{\beta}, t) \in W_i \times [0, \infty): d_{Y}(x, y) \leq \mu \; \mbox{and } t \geq t_0 \}, \end{aligned}$$ a *truncated cylinder* about $x$ in Fermi coordinates $(W_i\times [0,\infty), (y^{\beta},t))$.
Fix $x \in W_i \times \{0\}$ in the domain of a Fermi coordinate chart. Then there exist positive constants $\epsilon$, $\delta$, $T_{OB}$, $T_{IB}$ depending on $x$ and the constant $R$ from Theorem \[thm:hcomp\] such that if we define $$\begin{aligned}
OB & = & TC( x, \epsilon+\delta, T_{OB}),\; \mbox{the ``outer'' buffer}, \\
IB & = & TC( x, \epsilon, T_{IB}),\; \mbox{the ``inner'' buffer}, \end{aligned}$$ then if $p, q \in IB$, the $g$-geodesic segment from $p$ to $q$ remains entirely inside OB. \[lemma:db\]
We will determine the above constants such that if $p, q \in IB$, then
1. There is a curve $\sigma$ from $p$ to $q$ with $\sigma \subset OB$ such that $len(\sigma) \leq d_b(p,q)$.
2. For any curve $\sigma'$ from $p$ to $q$ that escapes $OB$, $len(\sigma') > d_b(p,q)$.
This implies that a geodesic segment between $p$ and $q$ lies in $OB$, and hence in the domain of a Fermi chart. See Figure \[fig:double\].
![Situation of Lemma \[lemma:db\] []{data-label="fig:double"}](double){width="60.00000%"}
*Proof of Step 1:* Consider “extremal” choices of $p$ and $q$. For $\epsilon$ and $T_{IB}$ to be determined, choose $p$ and $q$ to be any points on $B^Y_{\epsilon}(x) \times \{T_{IB}\}$ and write $p = (p', T_{IB})$, $q = (q', T_{IB})$. Now set $\xi = d_Y(p',q')$. Using Lemma \[lemma:as\] we see that the $b$-hyperbolic geodesic segment $\sigma$ from $p$ to $q$ has minimal $r$-value $\sigma_{r_{min}}$ greater than or equal to $$\sigma_{r_{min}} \geq \begin{cases} -c_6 \log \xi - c_7, &\text{when}\; e^{b T_{IB}} \xi \geq 2 \\
T_{IB} - c_8, & \text{when}\;e^{b T_{IB}} \xi \leq 2. \end{cases}$$ We impose the condition that $\sigma \subset OB$, i.e. that $\sigma_{r_{min}} \geq T_{OB}$. Since $\xi \leq 2 \epsilon$, this imposes the conditions: $$\begin{split}
\label{eqn:loccond1}
T_{OB} & \leq c_6 \log\left( \frac{1}{2\epsilon}\right) - c_7, \\
T_{IB} & \geq T_{OB} + c_8.
\end{split}$$
We now argue that there are no other conditions imposed on the constants after considering the extremal case. Pick arbitrary points $u,v \in IB$ and let $\tau$ be the geodesic segment from $u$ to $v$ in the $b$-hyperbolic metric. We must check that $\tau$ remains in $OB$. Now $\tau$ must lie in a two-dimensional plane. In this plane, let $p'$, $q'$ be the normal projections of $u, v$ on $ B^Y_{\epsilon}(x) \times \{0\}$, and consider $p=(p', T_{IB})$ and $q=(q',T_{IB})$ with $\sigma$ as before. The two-dimensional region between the normal geodesics containing $u$ and $v$ and with $r$-values greater than those of $\sigma$ is convex in the hyperbolic metric. This implies that $\tau$ remains in this region and so inside $OB$.
Thus given conditions above, the geodesic segment in the $b$-hyperbolic metric is a curve from $p$ to $q$ that remains in $OB$.
*Proof of Step 2:* Let $p, q \in IB.$ Suppose that $\sigma'$ is a curve between $p$ and $q$ that leaves $OB$. The length of such a curve satisfies: $$len(\sigma')~\geq~d(p,\partial OB)~+~d(q,\partial OB).$$
The boundary of $OB$ is the union $\partial OB_1 \cup \partial OB_2$ where $\partial OB_1$ is the “bottom” disc $B^Y_{\epsilon+\delta}(x) \times \{T_{OB}\}$ and $\partial OB_2$ the “vertical walls”, $\partial B^Y_{\epsilon+\delta}(x) \times [T_{OB}, \infty)$, for some $\delta$ to be determined. Now $d(p, \partial OB_1) = r_p - T_{OB}$, as vertical geodesics are always minimizing.
To estimate $d(p, \partial OB_2)$, suppose $\eta = (\eta', r_{\eta})$ is a point on $\partial OB_2$ closest to $p$. Note that a $g$-geodesic segment from $p$ to $\eta$ must lie in $OB$, and so by Theorem \[thm:hcomp\] we may compare to the $a$-hyperbolic metric, i.e. $$d(p, \partial OB_2) \geq d_a(p, \partial OB_2).$$ Applying the comparison from Lemma \[lemma:as\] we find $$\begin{aligned}
d_a(p, \partial OB_2) & \geq & r_p + r_{\eta} + c_4 \log{ d_Y(p', \eta') } - c_5 \\
& \geq & r_p + r_{\eta} + c_4 \log{\delta} - c_5 \\
& \geq & r_p + c_4 \log{\delta} - c_5,\end{aligned}$$ as $r_{\eta} \geq 0$. Therefore: $$\begin{aligned}
d(p,\partial OB) \geq \min\left\{ r_p - T_{OB}, r_p + c_4 \log{ \delta } - c_5 \right\}.\end{aligned}$$ We impose the condition that $c_4 \log{1/\delta} + c_5 = -c_4 \log{\delta} + c_5 \leq T_{OB}$, and find that $$len(\sigma') \geq d(p,\partial OB) + d(q,\partial OB) \geq r_p + r_q - 2 T_{OB}.$$ We need to apply Lemma \[lemma:as\] once more to estimate $d_b(p,q)$. Set $\xi = d_Y(p',q')$ and Lemma \[lemma:as\] implies $$d_{b}(p,q) \leq \begin{cases} r_p + r_q + c_1 \log \xi + c_2 & \text{when } e^{b r_p} \xi \geq 2, \\
r_q - r_p + c_3 & \text{when } e^{b r_p} \xi \leq 2. \end{cases}$$ Consider each case separately. In order to guarantee that $len(\sigma') > d_b(p,q)$ we may therefore impose $$r_p + r_q - 2 T_{OB} > r_p + r_q + c_1 \log(\xi) + c_2,$$ in case $e^{b r_p} \xi \geq 2$, and $$r_p + r_q - 2 T_{OB} > r_q - r_p + c_3,$$ when $e^{b r_p} \xi \leq 2$. For the first case this is implied by the condition $$T_{OB} < \frac{c_1}{2} \log\left(\frac{1}{\xi}\right) - \frac{c_2}{2},$$ and since this condition must hold for any $p, q \in IB$, it must hold when the $\log$-term is as small as possible, i.e. when $\xi = 2 \epsilon$. So case 1 imposes $$T_{OB} < \frac{c_1}{2} \log\left(\frac{1}{2 \epsilon}\right) - \frac{c_2}{2}.$$ The second case is implied by $$T_{OB} < r_p - \frac{1}{2} c_3 \leq \frac{1}{b}\log\left( \frac{2}{ \xi} \right) - \frac{1}{2} c_3,$$ and once again in order for this to hold for all $p, q$ we must impose: $$T_{OB} < r_p - \frac{1}{2} c_3 \leq \frac{1}{b}\log\left( \frac{2}{ 2 \epsilon} \right) - \frac{1}{2} c_3.$$
This completes the proof of Step 2. To conclude the proof, we have seen that the conditions that must be satisfied in order to satisfy both items 1 and 2 above are:
1. $\delta + \epsilon$ is small enough so that $B^Y_{\delta+\epsilon}(x) \subset W_i$.
2. $T_{OB} \geq c_4 \log(\frac{1}{\delta}) + c_5$
3. $T_{OB} < \min \left\{ \frac{c_1}{2} \log\left(\frac{1}{2 \epsilon}\right) - \frac{c_2}{2},\frac{1}{b}\log\left( \frac{1}{ \epsilon} \right) - \frac{1}{2} c_3 , c_6 \log\left( \frac{1}{2\epsilon}\right) - c_7 \right\}.$
4. $T_{IB} \geq T_{OB} + c_8$.
Recall $\log(\frac{1}{z}) {\rightarrow}+\infty$ as $z {\rightarrow}0^{+}$. First choose $\epsilon$ and $\delta$ to meet condition 1; clearly any smaller $\epsilon$ will also work. This fixes $\delta$, so now choose $T_{OB}$ subject to condition 2. Shrinking $\epsilon$ if necessary, we may also satisfy condition 3. Finally choose $T_{IB}$ large enough to meet condition 4.
Having finished the geometric preliminaries, we are now ready to describe the $C^{a/b}$ structure for ${M^{*}}$ that is independent of essential subset. We begin by describing the basic philosophy of the proof. In order to show that ${M^{*}}$ has a $C^{a/b}$ structure we must construct a $C^{a/b}$ atlas for ${M^{*}}$. Given an essential subset $K_1 \subset M$, we use the double buffer lemma to obtain a collection of truncated cylinders that cover a neighbourhood of infinity in a sense that we make precise below. We then obtain Fermi coordinates on these cylinders, and by collapsing the normal $r$-coordinate by a diffeomorphism, we obtain a coordinate cylinder that covers a deleted neighbourhood of the boundary ${M(\infty)}\subset {M^{*}}$. We will show that transition functions from these cylinders to the collapsed truncated cylinders emanating from a second essential subset $K_2$ are $C^{a/b}$ functions. As will be seen in the proof of Theorem \[thm:maincal\] below, the transition functions will then extend by uniform continuity to $C^{a/b}$ functions on a coordinate cylinder including an open subset of ${M(\infty)}$.
Consider two essential subsets $K_1, K_2$ for $M$. We begin with $K_1$. By Theorem \[thmmarsh\], every point $p \in {M(\infty)}$ is the image under ${\overline{E}}_1$ of exactly one point $p' \in Y_1$. By Lemma \[lemma:db\] we obtain parameters $\epsilon(p)$, $\delta(p)$, $T_{IB}(p)$, $T_{OB}(p)$. Since the collection $\{B_{\epsilon(p)}(p')\}$ covers $Y_1$, we pass to a finite subcover $${{ {\mathcal B} }}_1 := \{B_{\epsilon(p_k)}(p_k')\}_{k=1}^{N_1}.$$ Set $T_1 = \max \{ T_{IB}(p_k): 0 \leq k \leq N_1 \}$. Notice in Fermi coordinates relative to $Y_1$ if we write $p = E_1(p', r_p)$, $q= E_1(q', r_q)$ and assume $p',q' \in B_{\epsilon(p_k)}(p_k')$ for some $k$ and $\min\{r_p, r_q\} \geq T_1$ then a $g$-geodesic segment from $p$ to $q$ remains in some “double buffer” where we have comparison to hyperbolic metrics. In what follows we only use this property and we will not mention the underlying double buffer structure explicitly again.
The same procedure may be repeated to obtain a collar neighbourhood of infinity relative to $Y_2$, and we let ${{ {\mathcal B} }}_2, N_2, T_2$ denote the corresponding data for $Y_2$ as described above for $Y_1$. We set $$\label{Tdefn}
T = \max\left\{T_1 , T_2, \frac{1}{2a} ( 1 - \log(e-2) ) + \operatorname{diam}(K_1 \cup K_2)\right\}.$$ The reason for the last term in the definition of $T$ will become apparent during the proof of Proposition \[prop:calphonm\]. We call ${ {\mathcal B} }_j$ the *special coverings for $Y_j$*, $j=1, 2$, and the region $E_1( Y_1 \times (T, \infty) ) \cap E_2( Y_2 \times (T, \infty) )$ the *special (deleted) neighbourhood of infinity*. Observe also that every $p \in {M(\infty)}$ is in the intersection of the ${M^{*}}$-closure of two truncated cylinders, one emanating from each of $Y_1$ and $Y_2$. As such, the truncated cylinders are deleted neighbourhoods of points in ${M(\infty)}$. We introduce notation for these truncated cylinders and their images. For $j = 1, 2$ let $$\begin{aligned}
TC^k_j & := B^k \times [T, \infty) \subset Y_j \times [0, \infty), \text{where } { {\mathcal B} }_j = \{B^k\}_{k=1}^{N_j}, \\
{ {\mathcal TC} }_j & := \{ TC_j^k: 0 \leq k \leq N_j \}, \\
C_j^k & := E_j( TC_j^k ) \subset M, \\
{ {\mathcal C} }_j & := \{ C_j^k: 0 \leq k \leq N_j \}.\end{aligned}$$ Observe that in this notation, the lower index denotes the essential subset index and the upper index denotes an element of the special cover.
Since each $B \in { {\mathcal B} }_j$ is contained in some $W_i$ from the reference covering for $Y_j$, $B$ is the image of a coordinate parametrization $\phi\colon \widetilde{B} \subset {\mathbb{R}}^{n+1} {\rightarrow}B \subset M$. As $E\colon Y \times [0, \infty) {\rightarrow}M$, we define $E_{coord} := E \circ (\phi \times Id)\colon \widetilde{B} \times [0, \infty) \subset {\mathbb{R}}^{n+1} {\rightarrow}M$.
In what follows, we also consider the above constructions with $r$-coordinate collapsed by the diffeomorphism $\zeta\colon Y \times [0, \infty) {\rightarrow}Y \times [0, 1)$ given by $\zeta(p, r) = (p, \tanh(r/2))$. We use a circumflex to denote the collapsed version of subsets of $Y \times [0, \infty)$. For example if $P\subset Y \times [0, \infty)$ then $\hat{P} = \zeta(P) \subset Y \times [0,1)$. We also write the restriction of the map ${\overline{E}}$ (cf. page ) to $Y \times [0,1)$ as ${\hat{E}}$. Thus $$C_j^k = E_j( TC_j^k) = {\hat{E}}_j( \widehat{TC}_j^k ).$$
To proceed we need to check that distances in the special neighbourhood of infinity measured relative to each essential subset are $C^{a/b}$ comparable. This will be the key ingredient in showing that ${M^{*}}$ has a $C^{a/b}$ structure. To facilitate this, given an essential subset $K_j$, for $p, q \in {M \backslash K}_j$ with $p = E_j(p', r_p)$, $q=E_j(q', r_q)$ in Fermi coordinates, define $$\begin{aligned}
d_{K_j} (p, q) = | e^{-r_p} - e^{-r_q} | + d_{Y_j} ( p', q' ).\end{aligned}$$ This metric is defined on the entire set ${M \backslash K}_j$. It is easy to verify that when restricted to a particular truncated cylinder $C^k_j$ with coordinate parametrization $({\hat{E}}_{coord})_j^k$, this metric is equivalent to the Euclidean metric in collapsed Fermi coordinates, i.e. that $$| e^{-r_p} - e^{-r_q} | + d_Y ( p', q' ), \mbox{and }$$ $$\begin{gathered}
| (({\hat{E}}_{coord})_j^k)^{-1}(p) - (({\hat{E}}_{coord})_j^k)^{-1}(q) | \\
= |( (p')^{\alpha}, \tanh(r_p/2)) - ((q')^{\alpha}, \tanh(r_q/2))|\end{gathered}$$ are equivalent on $(({\hat{E}}_{coord})_j^k)^{-1}(C^k_j)$.
We now prove the main $C^{a/b}$ distance estimate.
There exists a positive constant $C$ depending on $a, b$ and $\operatorname{diam}(K_1 \cup K_2))$ such that whenever $C_1^k \in { {\mathcal C} }_1$ and $C_2^{k'} \in { {\mathcal C} }_2$ and $$C_1^k \cap C_2^{k'} \neq \emptyset,$$ then $$\begin{aligned}
\label{maincheck}
d_{K_2}( p, q ) \leq C \left(d_{K_1} (p, q) \right)^{a/b}, \end{aligned}$$ for all $p, q \in C_1^k \cap C_2^{k'}$. \[prop:calphonm\]
Throughout the proof recall that we assume $a \leq 1 \leq b$. We write $\alpha = a/b$.
In Fermi coordinates relative to $K_1$ we write $p = E_1(p', r_p)$, $q=E_1(q',r_q)$, and with respect to $K_2$ we write $p = E_2({\widetilde{p}'}, {\widetilde{r}_p})$, $q=E_2({\widetilde{q}'},{\widetilde{r}_q})$. By our assumption on $p, q$ and construction of the special covering and neighbourhood of infinity we have $$\label{dcomp}
d_a(p, q) \leq d_M(p,q) \leq d_b(p, q),$$ where $d_{\lambda}$ is the distance in the hyperbolic comparison metric (cf. page ). By the distance comparison principle (cf. page ), $\theta$ is comparable to $d_{Y_1}( p',q')$, and ${\widetilde{\theta}}$ is comparable to $d_{Y_2}( {\widetilde{p}'},{\widetilde{q}'})$. We are thus free to work with the angle $\theta$ and replace angles by a constant times distances along hypersurfaces upon obtaining the final estimate.
The inequality and the hyperbolic cosine law, equation , imply that $$\begin{gathered}
\frac{1}{a} \cosh^{-1} \left(\cosh{(a{\widetilde{r}_p})\cosh(a{\widetilde{r}_q}) } - \sinh{(a{\widetilde{r}_p})} \sinh{(a{\widetilde{r}_q})} \cos{{\widetilde{\theta}}} \right) \\
\leq \frac{1}{b} \cosh^{-1} \left(\cosh{ (b r_p) } \cosh{(b r_q)} - \sinh{(b r_p)} \sinh{(b r_q)} \cos{\theta} \right). \end{gathered}$$
We use the estimate $1-\theta^2/2 \leq \cos(\theta) \leq 1-\theta^2/8$ for $0 \leq \theta \leq \pi$, and then the angle-sum formulas for hyperbolic cosine imply: $$\begin{gathered}
\label{mainineq}
\frac{1}{a} \cosh^{-1} \left(\cosh{ (a({\widetilde{r}_p}- {\widetilde{r}_q})) } + \sinh{(a{\widetilde{r}_p})} \sinh{(a{\widetilde{r}_q})} \frac{{\widetilde{\theta}}^2}{8} \right) \\
\leq \frac{1}{b} \cosh^{-1} \left(\cosh{ (b(r_p-r_q)) } + \sinh{(b r_p)} \sinh{(b r_q)} \frac{\theta^2}{2} \right). \end{gathered}$$ Set $D := \operatorname{diam}_M(K_1 \cup K_2)$. The triangle inequality implies that $$\begin{split}
\label{sstilcomp}
r_p - D & \leq {\widetilde{r}_p}\leq r_p + D,\\
r_q - D & \leq {\widetilde{r}_q}\leq r_q + D.
\end{split}$$
Assume that $r_q \geq r_p$ and the proof now breaks into two cases, first when $r_q-r_p \geq \log{(2)}$ and then when $0 \leq r_q - r_p \leq \log(2)$.
### Case 1: $r_q-r_p \geq \log{(2)}$ {#case-1-r_q-r_p-geq-log2 .unnumbered}
The main idea in this case is that we have $e^{-r_q} \leq \frac{1}{2} e^{-r_p}$, and therefore $$\begin{aligned}
\label{case1main}
e^{-r_p} \leq 2 (e^{-r_p} - e^{-r_q}).\end{aligned}$$ The main inequality above in conjunction with the estimate $$\alpha \cosh^{-1}(z) \leq \cosh^{-1}(z^{\alpha}),$$ valid for $z \geq 1$ and $0 \leq \alpha \leq 1$, imply $$\begin{gathered}
\label{case1red}
\cosh{ (a({\widetilde{r}_p}- {\widetilde{r}_q})) } + \sinh{(a{\widetilde{r}_p})} \sinh{(a{\widetilde{r}_q})} \frac{{\widetilde{\theta}}^2}{8} \\
\leq \left(\cosh{ (b(r_p-r_q)) } + \sinh{(b r_p)} \sinh{(b r_q)} \frac{\theta^2}{2} \right)^{\alpha}. \end{gathered}$$ When $z \geq \frac{1}{2a} ( 1 - \log( e - 2 ) )$ we have $$e^{az - 1} \leq \sinh(az) \leq e^{a z}.
\label{tech:sinh}$$ By our choice of $T$ in , $r_q \geq r_p \geq \frac{1}{2a} ( 1 - \log( e - 2 ) )$. Consequently: $$\begin{aligned}
\label{case1left}
\cosh{ (a({\widetilde{r}_p}- {\widetilde{r}_q})) } & + \sinh{(a{\widetilde{r}_p})} \sinh{(a{\widetilde{r}_q})} \frac{{\widetilde{\theta}}^2}{8} \notag \\
&\geq \frac{e^{a({\widetilde{r}_p}-{\widetilde{r}_q})} + e^{a({\widetilde{r}_q}-{\widetilde{r}_p})}}{2} + e^{-2} e^{a({\widetilde{r}_p}+{\widetilde{r}_q})} \frac{{\widetilde{\theta}}^2}{8} \notag \\
&\geq \frac{e^{a({\widetilde{r}_p}+{\widetilde{r}_q})}}{8e^2} \left( e^{-2a{\widetilde{r}_q}} + e^{-2a{\widetilde{r}_p}} + {\widetilde{\theta}}^2 \right).\end{aligned}$$ Similarly with the right hand side of inequality , we may use the upper bound for hyperbolic sine provided by to obtain $$\begin{aligned}
\label{case1right}
\left(\vphantom{\frac{\theta^2}{2}}\cosh{ (b(r_p-r_q)) }\right. &\left. + \sinh{(b r_p)} \sinh{(b r_q)} \frac{\theta^2}{2}\; \right)^{\alpha} \notag \\
&\leq ( e^{b(r_p-r_q)} + e^{b(r_q-r_p)} + e^{b(r_p+r_q)}\theta^2)^{\alpha} \notag \\
&\leq ( e^{b(r_p+r_q)} ( e^{-2br_q} + e^{-2br_p} + \theta^{2}) )^{\alpha} \notag \\
&= e^{a(r_p+r_q)} ( 2 e^{-2br_p} + \theta^{2} )^{\alpha}.\end{aligned}$$ Combining inequalities , , , dividing by $e^{a({\widetilde{r}_p}+{\widetilde{r}_q})}/(8e^2)$ and using to remove tildes gives: $$\begin{aligned}
\label{ineq2}
e^{-2a{\widetilde{r}_q}} + e^{-2a{\widetilde{r}_p}} + {\widetilde{\theta}}^2 & \leq & 8e^{2aD+2} ( 2 e^{-2br_p} + \theta^{2} )^{\alpha}\end{aligned}$$ An easy computation shows that we always have the estimate: $$\label{tilest}
e^{-2 {\widetilde{r}_q}} + e^{-2{\widetilde{r}_p}} \geq ( e^{-{\widetilde{r}_q}} - e^{-{\widetilde{r}_p}} )^2.$$ Recall that $a \leq 1$, and so $e^{-2az} \geq e^{-2z}$ for $z\geq 0$. Apply this and inequality to the left hand side of to obtain $$\begin{aligned}
\label{ineq3}
e^{-2a{\widetilde{r}_q}} + e^{-2a{\widetilde{r}_p}} + {\widetilde{\theta}}^2 & \geq & e^{-2{\widetilde{r}_q}} + e^{-2{\widetilde{r}_p}} + {\widetilde{\theta}}^2 \notag \\
& \geq & (e^{-{\widetilde{r}_p}} - e^{-{\widetilde{r}_q}})^2 + {\widetilde{\theta}}^2.\end{aligned}$$ For the right hand side of we use $b\geq 1$ and the estimate to see: $$\begin{aligned}
\label{ineq4}
8 e^{2aD+2} ( 2 e^{-2br_p} + \theta^{2} )^{\alpha} & \leq & 8e^{2aD+2} ( 4(e^{-r_p} - e^{-r_q})^2 + \theta^2 )^{\alpha}.\end{aligned}$$ Combining inequalities , and we have: $$\begin{aligned}
(e^{-{\widetilde{r}_p}} - e^{-{\widetilde{r}_q}})^2 + {\widetilde{\theta}}^2 & \leq & 8 \cdot 4^{\alpha} e^{2aD+2} ( (e^{-r_p} - e^{-r_q})^2 + \theta^2 )^{\alpha}.\end{aligned}$$ This now implies , and completes the proof of Case 1.
### Case 2: $0 \leq r_q-r_p \leq \log{(2)}$ {#case-2-0-leq-r_q-r_p-leq-log2 .unnumbered}
The main idea in this case is to use a power series expansion for hyperbolic cosine as $r_q-r_p$ is bounded. Note that if $0 \leq r_q-r_p \leq \log{(2)}$ then $$\label{eqn:rprq}
0 \leq |{\widetilde{r}_q}- {\widetilde{r}_p}| \leq \log(2) + 2D.$$ Simple calculations imply that we may choose constants $k_1, \ldots, k_4$ so that for $0 \leq z \leq \log(2) + 2D$: $$\begin{aligned}
1+k_1 z^2 \leq & \cosh{(z)} & \leq 1+ k_2 z^2, \label{locest1}\\
k_3 z \leq & 1-e^{-z} & \leq k_4 z. \label{locest2}\end{aligned}$$ So these estimates hold when $z = r_q - r_p$ or $z = |{\widetilde{r}_q}-{\widetilde{r}_p}|$.
We begin with inequality . We will first apply estimates and the estimates for hyperbolic sine from . We then apply the estimate $(1+x)^{\alpha} \leq 1 + x^{\alpha}$, valid for $x \geq 0$ and $0 \leq \alpha \leq 1$. This yields $$\begin{aligned}
1 + k_1 (a({\widetilde{r}_p}- {\widetilde{r}_q}))^2 &+ \frac{1}{8e^{2}} e^{a({\widetilde{r}_p}+{\widetilde{r}_q})} {\widetilde{\theta}}^2 \nonumber \\
&\leq 1 + \left( k_2 (b(r_p-r_q))^2 + e^{b(r_p+r_q)}\theta^2 \right)^{\alpha}.\end{aligned}$$ We cancel the ones and divide by $e^{a({\widetilde{r}_p}+{\widetilde{r}_q})}$, absorbing this factor into the right hand side of the inequality, obtaining $$\begin{aligned}
\label{locest4}
k_1 e^{-a({\widetilde{r}_p}+{\widetilde{r}_q})}(a({\widetilde{r}_p}& - {\widetilde{r}_q}))^2 + \frac{1}{8e^{2}} {\widetilde{\theta}}^2 \nonumber \\
& \leq e^{2aD} \left( k_2 e^{-2br_p} (b(r_p-r_q))^2 + \theta^2 \right)^{\alpha}.\end{aligned}$$ We consider the right hand side of this inequality. A little algebraic manipulation, use of estimate , and the fact that $b \geq 1$ give $$\begin{aligned}
\label{locest5}
e^{2aD} ( k_2 e^{-2br_p} ( b (r_p & -r_q) )^2 + \theta^2 )^{\alpha} \notag \\
& \leq e^{2aD} \left( k_2 k_3^{-2} b^2 \left( e^{-r_p} ( 1 - e^{-(r_q - r_p)} ) \right)^2 + \theta^2 \right)^{\alpha} \notag \\
& \leq e^{2aD} \max\{(k_2 k_3^{-2} b^2)^{\alpha}, 1\} \left( (e^{-r_p}-e^{-r_q}) + \theta \right)^{2 \alpha}.\end{aligned}$$ We now consider the left hand side of inequality . When ${\widetilde{r}_q}> {\widetilde{r}_p}$ we find: $$\begin{aligned}
\label{locest6}
k_1 e^{-a({\widetilde{r}_p}+{\widetilde{r}_q})} & (a({\widetilde{r}_p}- {\widetilde{r}_q}))^2 + \frac{1}{8e^{2}} {\widetilde{\theta}}^2 \nonumber \\
&\geq a^2 k_1 k_4^{-2} e^{-2{\widetilde{r}_q}} (1-e^{-({\widetilde{r}_q}-{\widetilde{r}_p})})^2 + \frac{1}{8e^{2}}{\widetilde{\theta}}^2 \nonumber \\
&\geq a^2 k_1 k_4^{-2} (e^{{\widetilde{r}_p}-{\widetilde{r}_q}} (e^{-{\widetilde{r}_p}}-e^{-{\widetilde{r}_q}} ))^2 + \frac{1}{8e^{2}}{\widetilde{\theta}}^2 \nonumber \\
&\geq \min\left\{a^2 k_1 k_4^{-2} e^{-2(\log(2) + 2D)}, \frac{1}{8e^{2}} \right\} \left( (e^{-{\widetilde{r}_p}}-e^{-{\widetilde{r}_q}})^2 + {\widetilde{\theta}}^2 \right),\end{aligned}$$ where the last estimate uses inequality . We may now put estimates , together with to get: $$\begin{gathered}
\label{locest7}
\min\left\{a^2 k_1 k_4^{-2} e^{-2(\log(2) + 2D)}, \frac{1}{8e^{2}} \right\} \left( (e^{-{\widetilde{r}_p}}-e^{-{\widetilde{r}_q}})^2 + {\widetilde{\theta}}^2 \right) \\
\leq e^{2aD} \max\{(k_2 k_3^{-2} b^2)^{\alpha}, 1\} \left( (e^{-r_p}-e^{-r_q}) + \theta \right)^{2 \alpha}.\end{gathered}$$ Observe that for positive $x, y$ and $z$, $x^2 + y^2 \leq z^2$ imply $x + y \leq 2z$. Consequently this estimate and inequality implies that for some constant $C$: $$\begin{aligned}
e^{-{\widetilde{r}_p}}-e^{-{\widetilde{r}_q}} + {\widetilde{\theta}}& \leq & C \left((e^{-r_p}-e^{-r_q} ) + \theta \right)^{\alpha}.\end{aligned}$$ An entirely similar computation holds for ${\widetilde{r}_p}\geq {\widetilde{r}_q}$. This completes the proof of Case 2.
We are now ready to finish the proof of Theorem \[thm:main\], which we obtain immediately from Theorems \[thm:characterizationES\], \[thmmarsh\], \[thm:hcomp\] and the following:
\[thm:maincal\] Let $M$ be a complete Riemannian manifold containing an essential subset. Suppose for every essential subset $K \subset M$ with reference covering $\{W_i\}$ for $Y=\partial K$ that there exists an $R>0$ such that for every $r>R$ $$\begin{aligned}
\frac{ \sinh^2( a( r - R ) )} {a^2} \; { (\mathring{g}_i) }_{\beta \nu} \; \leq g_{\beta \nu}(y,r) \; \leq \;\frac{ \sinh^2( b( r+ R ) ) }{b^2}\; { (\mathring{g}_i) }_{\beta \nu},\end{aligned}$$ for all i, where ${ (\mathring{g}_i) }$ is the round metric in normal coordinates (cf. page ). Then ${M^{*}}$ has a $C^{a/b}$ structure independent of $K$.
We must find an atlas of $C^{a/b}$ compatible charts.
Recall we have earlier defined the special neighbourhood of infinity $S = E_1( Y_1 \times (T, \infty) ) \cap E_2( Y_2 \times (T, \infty) )$, relative to essential subsets $K_1$ and $K_2$. The complement $K_0 = M \backslash S$ is a compact set, and we choose an atlas ${{ {\mathcal C} }}_0$ of normal coordinate balls covering $K_0$ so that the collection of balls of half the radius still cover $K_0$. Preceding Proposition \[prop:calphonm\] we defined a covering ${ {\mathcal C} }_j, j= 1,2$. Every truncated cylinder $C^k_j \in { {\mathcal C} }_j$ is a deleted neighbourhood of points on the boundary of ${M(\infty)}$. Let $\overline{C}^k_j$ be the union of $C^k_j$ and points of ${M(\infty)}$ in the ${M^{*}}$-closure of $C^k_j$; this is an open subset of ${M^{*}}$ containing an open subset of ${M(\infty)}$. Set $\overline{{ {\mathcal C} }}_j = \{ \overline{C}^k_j: 1\leq k \leq N_j \}$. We now show that ${{ {\mathcal C} }} = {{ {\mathcal C} }}_0 \cup \overline{{ {\mathcal C} }}_1 \cup \overline{{ {\mathcal C} }}_2$ is a $C^{a/b}$ compatible atlas for ${M^{*}}$.
Whenever a chart from ${ {\mathcal C} }_0$ overlaps with a chart from any ${ {\mathcal C} }_j, j = 0, 1, 2$ the transition function is smooth, and therefore $C^{a/b}$. Similarly, transition functions from two charts in a single ${ {\mathcal C} }_j$, $j=1, 2$ are $C^{a/b}$ functions.
We now consider the case that a chart $\overline{C}^k_1 \in \overline{{ {\mathcal C} }}_1$ meets a chart $\overline{C}^{k'}_2 \in \overline{{ {\mathcal C} }}_2$. But this is exactly the situation of Proposition \[prop:calphonm\]. We have a $C^{\alpha}$ estimate of the form: $$d_{K_2}( p, q ) \leq C \left(d_{K_1} (p, q) \right)^{\alpha},$$ for points $p, q \in C^k_1 \cap C^{k'}_2$.
Since $d_{K_j}(p, q) = | e^{-r_p} - e^{-r_q} | + d_{Y_j} ( p', q' )$ is equivalent to the Euclidean distance $| ({\hat{E}}_{coord})_j^{-1}(p) - ({\hat{E}}_{coord})_j^{-1}(q)|$ on $({\hat{E}}_{coord})_j^{-1} (C^k_1 \cap C^{k'}_2)$, we have that the transition function $$\psi = (({\hat{E}}_{coord})^{k'}_2)^{-1} \circ ({\hat{E}}_{coord})^k_1$$ is a $C^{\alpha}$ map on $(({\hat{E}}_{coord})_1^k)^{-1} (C^k_1 \cap C^{k'}_2)$.
As in the proof of Theorem \[thmmarsh\], $\psi$ extends to a continuous map $\operatorname{\overline{\psi}}$ on the closure of $(({\hat{E}}_{coord})_1^k)^{-1} (C^k_1 \cap C^{k'}_2)$, and by the result above extends to a $C^{a/b}$ map on the closure as well. Thus $$\operatorname{\overline{\psi}}= (({\overline{E}}_{coord})^{k'}_2)^{-1} \circ ({\overline{E}}_{coord})^k_1$$ is a $C^{a/b}$ map on $(({\overline{E}}_{coord})_1^k)^{-1} (\overline{C}^k_1 \cap \overline{C}^{k'}_2)$.
In summary, we have shown that given any essential subset $K$ we may construct a smooth atlas for ${M^{*}}$, and that any two such atlases are $C^{a/b}$ compatible. These atlases are contained in a maximal atlas, which is a $C^{a/b}$ structure for ${M^{*}}$ independent of essential subset.
[M]{}
Anderson, M. and Schoen, R., *Positive harmonic functions on complete manifolds of negative curvature*, Annals of Math., **121** (1985) 429-461. Eberlein, P. and O’Neill, B., *Visibility Manifolds*, Pacific Journal of Math. **46** (1973) no. 1, 45-109. Klingenberg, W., *Riemannian Geometry*, de Gruyter Studies in Mathematics, New York, 1995. Lee, J. M., *Riemannian Manifolds, An Introduction to Curvature*, Springer-Verlag, New York, 1997. Petersen, P., *Riemannian Geometry*, Springer-Verlag, New York, 1999.
[^1]: Recall from Section \[sec:notation\] that we use *geodesic segment* to mean a distance minimizing geodesic curve between two points.
[^2]: Recall the constant $R$ used here is the constant from Theorem \[thm:hcomp\], which is used throughout this section.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Kirstin Fritz\
Department für Physik und CeNS, Ludwig-Maximilians-Universität,\
München, Germany
- |
Georg Fritz\
Arnold Sommerfeld Center for Theoretical Physics und CeNS,\
Ludwig-Maximilians-Universität,\
München, Germany
- |
Barbara Windschiegl\
Institut für Organische und Biomolekulare Chemie, Georg August Universität\
Göttingen, Germany\
- |
Claudia Steinem\
Institut für Organische und Biomolekulare Chemie, Georg August Universität\
Göttingen, Germany\
- |
Bert Nickel[^1]\
Department für Physik und CeNS, Ludwig-Maximilians-Universität,\
München, Germany
title: Arrangement of Annexin II tetramer and its impact on the structure and diffusivity of supported lipid bilayers
---
Introduction {#introduction .unnumbered}
============
Annexin A2 (Anx A2) is a $\mathrm{Ca}^{2+}$-binding protein which binds to acidic phospholipids and is involved in many cellular regulatory processes, such as the regulation of vesicular trafficking, endosome fusion, insulin signal transduction [@Biener_JBiolChem_96] and RNA binding [@Filipenko_JBiolChem_04]. Like other members of the Annexin protein family, it consists of two domains: the conserved core domain harboring the $\mathrm{Ca}^{2+}$-binding sites and, among different Annexins, a variable N-terminal domain exposing interaction sites for other protein partners. In particular, Anx A2 forms a heterotetrameric complex (Anx A2t) with S100A10 (p11). S100A10 belongs to the S100 protein family, although it is distinct from the other members of this family as it does not undergo $\mathrm{Ca^{2+}}$-dependent conformational changes. Even in the absence of $\mathrm{Ca}^{2+}$, S100A10 is in the active state and, like most other S100 proteins, forms an anti-parallelly packed non-covalent homodimer [@Rety_NatStructBiol_99].
To date, no experimentally resolved high-resolution structure of the full length complex of Anx A2t is available. However, data of the Anx A2 monomer missing the first 19 amino acids [@Burger_JMolBiol_96; @Rosengarth_Annexins_04], as well as of the complex between dimeric p11 and two synthetic N-terminal fragments composed of the first 11 amino acids of Anx A2 exist [@Rety_NatStructBiol_99]. Based on these data, a structural model of the Anx A2t complex has been created by computational modeling [@Sopkova-deOliveiraSantos_BiochimBiophysActa_00]. Yet, there is still a controversial discussion about the organization of Anx A2t when bound to a single membrane and the complex that is formed upon membrane-membrane connection. On the one hand, when binding to a single surface supported membrane, the thickness of an Anx A2t layer obtained from scanning force microscopy experiments suggested that two Anx A2 monomers are connected by a p11 dimer and are bound in a “side-by-side” configuration to the membrane interface [@Menke_ChemBioChem_04], see Fig. \[FIGscheme\] ([*a, top*]{}). It has frequently been proposed that the contact between membranes may then be mediated via the interaction of two opposing Anx A2t complexes by formation of a heterooctameric structure, see Fig. \[FIGscheme\] ([*b, top*]{}) [@Lewit-Bentley_CellBiolInt_00; @Menke_ChemBioChem_04; @Schulz_Proteins_07]. On the other hand, based on cryo-electron microscopy results on Anx A2t-connecting vesicles [@Lambert_JMolBiol_97], it was proposed that the Anx A2t complex bridges membranes in a “vertical” configuration, in which each Anx A2 monomer binds to one of the membranes while the p11 dimer is located in between [@Sopkova-deOliveiraSantos_BiochimBiophysActa_00], see Fig. \[FIGscheme\] ([*b, bottom*]{}). From these measurements the same, albeit hypothetical, structure was proposed for the protein when binding to a single bilayer [@Menke_ChemBioChem_04], see Fig. \[FIGscheme\] ([*a, bottom*]{}). Importantly, the vertical conformation of Annexin A2t, as obtained from membrane-membrane junctions, is predicted to result in a larger thickness of the protein layer than the side-by-side conformation [@Nakata_JCellBiol_90; @Sopkova-deOliveiraSantos_BiochimBiophysActa_00; @Lambert_JBiolChem_04; @Schulz_Proteins_07]. However, during scanning force microscopy experiments as in ref. [@Menke_ChemBioChem_04] a certain force is exerted on the soft protein layer and therefore its thickness might be underestimated. Hence, to date, it cannot be excluded that Anx A2t also binds to a single lipid bilayer in the vertical conformation.
In this paper we used x-ray reflectivity (XR) to investigate the conformation of Anx A2t and its impact on the structure of single supported lipid bilayers (SLBs). In the field of SLBs, XR has been successfully applied as a non-invasive tool to resolve subnano-structural features [@Miller_PhysRevLett_05; @Reich_RevSciInstrum_05; @Nickel_Biointerphases_08]. Differences between gel and fluid lipid bilayers have been characterized [@Novakova_PhysRevE_06; @Reich_BiophysJ_08] in progress toward understanding heterogeneous lipid mixtures. In contrast to neutron scattering techniques, XR measurements using modern synchrotron sources achieve a superior resolution. However, due to the low contrast only few XR measurements have been conducted on protein binding to SLBs [@Horton_Langmuir_07]. Here we demonstrate that a single Anx A2t layer provides enough contrast for a high-resolution XR study of the protein-SLB structure. It enables us to elucidate the thickness of Anx A2t bound to a single bilayer and to discriminate between the side-by-side and vertical configuration. A major advantage of XR is that also the structure of the covered SLB, [*i.e.*]{}, the head-to-head distance and the packing density of the lipids, remains experimentally accessible upon protein binding. Hence, the influence of protein binding and substrate on the bilayer structure can be monitored. In particular, it has been proposed that binding of Anx A2t creates pores in the membrane through densification of the bilayer [@Chasserot2005] and thereby exerts its role in endo- and exocytosis. However, only few measurements investigate the influence of Anx A2t binding on the structure and density of the bilayer. In this paper a careful determination of the electron density profiles is reported and further decomposition into components of the bilayer allow us to analyze properties of both bilayer leaflets independently, so that differences in packing density can be resolved. By using different ratios of anionic POPS to neutral POPC we examine the influence of Anx A2t binding on these structural bilayer parameters. Furthermore, it was hypothesized that the protein collects negatively charged lipids below itself and thus may alter the fluidity and state of the bilayer [@ross2005]. To test this conjecture we complement our x-ray studies by continuous bleaching measurements on fluorescently labeled lipid probes. The combination of both experimental methods allows us to correlate the structural rearrangement induced by protein binding to changes in the diffusion properties of the lipid bilayer. Last, possible implications for endo- and exocytotic processes are discussed.
Materials and Methods {#materials-and-methods .unnumbered}
=====================
#### SLB formation.
Surface supported bilayers of varying molar composition of 1-palmitoyl-2-oleoyl-[*sn*]{}-glycero-3-phosphocholine (POPC), and 1-palmitoyl-2-oleoyl-[*sn*]{}-glycero-3-\[phospho-L-serine\] (sodium salt) (POPS), were prepared on silicon oxide by the following procedure: Appropriate amounts of lipids were dissolved in chloroform and filled in a clean glass vial. 0.5 mol % fluorescent dye (Texas Red 1,2-dihexadecanoyl-[*sn*]{}-glycero-3-phosphoethanolamine, triethylammonium salt, Texas Red DHPE or Oregon Green 488 1,2-dihexadecanoyl-[*sn*]{}-glycero-3-phosphoethanolamine, Oregon Green 488 DHPE, Invitrogen, Germany) were added to the mixture for fluorescence microscopy. The solvent was evaporated by a nitrogen flow. After removal of residual solvent overnight in a dessicator connected to a rotary vacuum pump, the mixture was dispersed in a PBS buffer (Sigma-Aldrich) with 1 M NaCl to a lipid concentration of 1 mg/ml and vortexed to form a lipid suspension. The suspension was kept at $40 \celsius$ for two hours. Large unilamellar vesicles were obtained by extrusion through 100 nm polycarbonate filters using an extruder (Avestin, Ottawa, Canada). 150$\mu$l of the vesicle suspension were injected into the microfluidic chamber, which holds the silicon wafer. For a detailed description of the microfluidic chamber please refer to ref. [@Reich_RevSciInstrum_05]. Silicon wafers with thermal oxide layers were bought from Crystec (Berlin, Germany) and cut to a sample size of 15x20 $\mathrm{mm^2}$. To avoid fluorescence quenching the thickness of the oxide layer on the silicon surfaces ranged from 50 nm to 100 nm. Prior to usage, silicon wafers were cleaned chemically by a three stage protocol [@Reich_RevSciInstrum_05; @Reich_BiophysJ_08] involving $\mathrm{H_2O_2/HCl/H_2O}$ and $\mathrm{H_2O_2/NH_3/H_2O}$. They were stored in deionized water and used within 24h. The bare silicon surface was measured by x-ray reflectivity and the obtained roughness of approximately $4\, \AA$ was fixed during the decomposition of the electron density profile. The sample was incubated for at least three hours at room temperature, then rinsed intensively with buffer to remove excess vesicles, and monitored by fluorescence microscopy for homogeneity. In a following step, the buffer was carefully exchanged by deionized pure water (MilliQ, specific resistivity 18.2 M$\Omega$/cm, Millipore Corp. Billerica, Massachusetts), which was injected into one of the two outlets. This induced an osmotic shock promoting fusion of the adsorbed vesicles to the substrate and resulted in SLB formation. The SLB coverage was controlled by fluorescence microscopy. When homogenous coverage was reached, 20 mM Tris/HCl, 100 mM NaCl, 1 mM $\mathrm{CaCl_2}$, 0.5 mM DTT, pH 7.4 was injected.
#### Protein binding and unbinding.
Annexin A2 tetramer was extracted from bovine intestines as described previously [@Gerke_EMBOJ_84]. The protein was stored at $-20 \celsius$ in 20 mM MES, 100 mM NaCl, 0.5 mM EGTA, pH 6.0. Before usage, the protein solution was carefully warmed up to $4\celsius$ in an ice bath and 1 ml of 20 mM Tris/HCl, 100 mM NaCl, 1 mM $\mathrm{CaCl_2}$, 0.5 mM DTT, pH 7.4 was added to the protein solution. The solution was further diluted to a concentration of 3-6$\mu$M. $150\, \mu l$ of protein solution were then added to the SLB in the microfluidic chamber. Bilayers were incubated for at least three hours at room temperature. After incubation, the microfluidic chamber was rinsed with 20 mM Tris/HCl, 100 mM NaCl, 1 mM $\mathrm{CaCl_2}$, 0.5 mM DTT, pH 7.4 to remove excess protein. To unbind Anx A2t completely from the membrane, the sample was rinsed with 20 mM Tris/HCl, 100 mM NaCl, 10 mM EDTA, pH 7.4.
#### Fluorescence microscopy.
SLBs were controlled by fluorescence microscopy on site at HASYLAB and ESRF using a transportable Zeiss Axiotech vario fluorescence microscope (Oberkochen, Germany) equipped with 10$\times$ (NA 0.3) and long-distance 63$\times$ (NA 0.75) Plan-Neofluar objectives. Images were captured with an ORCA C4742-95 CCD camera and WASABI imaging software from Hamamatsu Photonics (Tutzing, Germany). For continuous bleaching experiments, a 120 W mercury short arc reflector lamp (HXP-R120W) was used. Continuous bleaching data were analyzed using self-written code based on MATLAB, The MathWorks, Inc..
#### X-ray reflectivity.
X-ray reflectivity measurements were performed at the beamline D4 at the Hamburger Synchrotronstrahlungslabor (HASYLAB) at the Deutsches Elektronen Synchrotron (DESY), Hamburg, Germany, and at the beamline ID01 at the European Synchrotron Radiation Facility (ESRF) in Grenoble, France. An x-ray energy of 19.75 keV was chosen to maximize the reflectivity signal while minimizing the beam damage in the microfluidic chambers. [@Reich_RevSciInstrum_05]. Sample chambers were mounted in a horizontal scattering geometry as described previously [@Reich_RevSciInstrum_05; @Horton_Langmuir_07]. Briefly, the incident beam enters the microfluidic chamber through a topas foil and passes through a $200\,\mu$m water-filled channel before hitting the sample. Reflected intensities were collected by tilting the sample between incident angles $\mathrm \theta$ = 0.02 and 2 degrees with a step number 142 at ESRF. For data collected at HASYLAB the sample was tilted between 0.026 and 1.51 degrees with a step number of 243 points. This leads to a momentum transfer $q$ normal to the surface up to $q = 0.53 \,\AA^{-1}$ at HASYLAB and $q= 0.71\,\AA^{-1}$ at ESRF. Here $q$ is defined by $q=2\pi\sin(\theta)/ \lambda$, where $\lambda$ is the wavelength of the incoming beam and $\theta$ the incident angle. The beam cross section was defined by a presample aperture of 80 $\mu$m horizontal and 100 $\mu$m vertical at D4 and 250 $\mu$m horizontal and 1000 $\mu$m vertical at ID01. Evacuated beam guides with Kapton windows positioned close to the sample chamber minimized air scattering. The reflected intensity was collected with a NaI (cyberstar) detector. For each data point the reflected intensity was collected for one second. Automatic attenuators in front of the sample were used in order to reduce exposure to the full beam intensity. Furthermore, the sample was protected by a fast shutter system during motor movement. To control for radiation damage of bilayer and protein the samples were measured at the same sample position for a second time. No change in reflectivity signal of bilayer and protein layer was observed during this procedure. By detuning from the reflection condition, the background was determined and subtracted. The remaining reflection signal was corrected for illumination (footprint correction) and normalized to a reflectivity of one in the total reflection region. For graphical presentation, the data were multiplied by $q^4$ , to compensate for the overall decay of the reflectivity signal.
#### X-ray data analysis.
The reflectivity data were fitted with Parrats algorithm [@Parratt_PhysRev_54; @Nevot_RevPhysAppl_80]. Two different fitting strategies were performed; in a first approach a small number of layers were chosen, each layer representing different regions of the bilayer [@Mattai_Biochemistry_89; @Reich_BiophysJ_08]. Layer thickness and electron density were chosen as fitting parameters. In a second approach, a larger number of layers was created (10-12 layers) and only the electron density of each layer was used as fitting parameter. In this case, the thickness of each layer was set to the minimal layer thickness experimentally accessible (5-7 $\mathrm{\AA}$, see below) [@Daniel_Biointerphases_07]. The layer roughness was set to zero in both approaches. Within the first approach, the electron density of the bilayer was calculated with two slabs for each headgroup [@Mattai_Biochemistry_89; @Reich_BiophysJ_08], and three for the hydrophobic core of the bilayer [@Novakova_PhysRevE_06], for measurements performed at ID01. Due to the lower resolution at D4, the bilayer data from this instrument were fitted with only one slab for each headgroup and three slabs for the hydrophobic core of the bilayer. In both measurements, an additional layer was added to account for the substrateÕs roughness and hydration of the headgroup proximal to the silicon substrate. Within the second approach, the layer thickness was chosen to match the resolution of the experimental data. The resolution in a reflectometry experiment can be estimated by ($\pi/q_{max}$) [@Daniel_Biointerphases_07]. Here $q_{max}$ represents the maximum $q$ value achieved before the reflectivity signal decreases below the background signal. This leads to a slab thickness of 7 $\mathrm{\AA}$ and 5 $\mathrm{\AA}$ for data obtained at D4 and ID01, respectively. Thirteen and ten layers representing the bilayer were chosen for data from ID01 and D4, respectively. Further layers did not improve the quality of the fit. An average electron density was calculated from the fifteen best fits of both models. To further smoothen the profile a running average with a window size corresponding to the resolution of the experiments was applied. A similar approach was chosen to extract the electron density profile of the SLB covered by the protein layer.
#### Decomposition of electron density profiles.
Electron density profiles were decomposed into several component groups as described in the main text. The parameters of the decomposition were estimated by using a trust-region reflective Newton method (MATLAB, The MathWorks, Inc.) to minimize the total $\chi^2$. To quantify the uncertainty in the estimated parameters, we performed 10,000 independent fits with randomly chosen initial parameter sets (within their physiological ranges). This implied the constraint that the spatial order of the components had to be maintained, i.e., we did not allow permutation of the individual groups. In Figs. S1 and S2 of the [*Supplementary Material*]{}, the final $\chi^2$ values are plotted against the final parameters. We followed Ref. [@Fritz_JMB_09] to compute the errors with respect to the optimal parameter values listed in Table \[TABbilayerParameters\]. The squared error for parameter $\theta_k$ was calculated using the following equation: $$\begin{aligned}
\label{EQNerrors}
\sigma_{k}^2 & = & \frac{\displaystyle \sum_{\theta_{k,i}} (\theta_{k,i} - \theta_k^{opt})^2 e^{-\chi_i^2/2}}{ \displaystyle\sum_{\theta_{k,i}} e^{-\chi_i^2/2}}\,,\end{aligned}$$ where $\theta_{k,i}$ is the value of parameter $\theta_k$ in the $i_{th}$ fit, $\theta_k^{opt}$ is the value of $\theta_k$ in the fit with the lowest value of $\chi^2$, and $\chi_i^2$ is the value of $\chi^2$ for the $i_{th}$ fit. In using the likelihood function $e^{-\chi^2/2}$, we assume that the errors in the measurements are independent and normally distributed with widths equal to the standard error of the mean.
#### Calculation of Anx A2t coverage.
The coverage of Anx A2t was derived from the electron density adjacent to the bilayer as follows: The electron density of the total protein layer $\rho$ can be written as a weighted sum the electron densities of pure protein $\rho_{\mathrm prot}$ and pure water $\rho_{\mathrm{H_2O}}$, i.e., $\mathrm{\rho=x \rho_{\mathrm prot}+(1-x) \rho_{\mathrm H_2O}}$, where $x$ and $(1-x)$ are the volume fractions of the protein and water, respectively. As a proxy for the total protein layer we used the maximal electron density in the Anx A2 signature and correspondingly determined the electron density of the pure protein from the chemical sum formula and the mass density of Anx A2. The mass density of Anx A2 is given by $\rho_{m}=\left[1.41+0.145 \exp{(-M/13)}\right]\, g/cm^3$, as described in [@fischer2004]. Here $M$ is the molecular mass of Anx A2 in $kDa$, leading to a mass density of $\rho_{m}=1.418\,g/cm^3$. With this, the electron density of Anx A2 is obtained as $\rho_{prot} = 0.458\,e^-/\AA^{3}$ while the electron density of bulk water is given by $\rho_{H_2O} = 0.336\,e^-/\AA^{3}$. Taken together, we find a protein coverage of 91% in the case of the 50%POPS containing bilayer and 27% in the case of the 25%POPS containing bilayer.
Results {#results .unnumbered}
=======
X-Ray Reflectivity {#x-ray-reflectivity .unnumbered}
------------------
Anx A2t binds to anionic phospholipids in a $\mathrm{Ca^{2+}}$-dependent manner. In order to scrutinize the structural and dynamical properties of Anx A2t binding, SLBs composed of different ratios of anionic (POPS) to neutral (POPC) lipids were prepared on silicon substrates in Tris 1mM $\mathrm{CaCl_2}$ buffer (calcium buffer hereafter). For x-ray reflectivity measurements lipid bilayers were mixed in the molar ratios (POPS:POPC) of (1:3), referred to as 25 mol % POPS and (1:1), referred to as 50 mol % POPS. Figure \[FIGreflectivity\] shows the x-ray reflectivity as a function of the momentum transfer $q$ for both lipid mixtures before and after Anx A2t incubation. The increase in intensity for $q < 0.02\, \mathrm{\AA}^{-1}$ is due to total reflection at the silicon surface and a $q^4$ correction, as described in [*Materials & Methods*]{}. The rapid intensity oscillations (Kiessig fringes) show a periodicity of $\Delta q \approx 0.016\, \mathrm{\AA}^{-1}$ for the 25 mol % POPS containing bilayer and $\Delta q \approx 0.0064\, \mathrm{\AA}^{-1}$ for the 50 mol % POPS containing bilayer. These narrow fringes stem from the interference between reflections at the silicon oxide layer and the silicon substrate. In contrast, the broad Kiessig fringes with $\Delta q \approx 0.15\, \mathrm{\AA}^{-1}$ have been shown to result from the SLB [@Reich_RevSciInstrum_05]. After incubation with Anx A2t, a clear change in signal is observed for both POPS concentrations (compare [*ellipsoids*]{} in Fig. \[FIGreflectivity\]). Importantly, the signal change after Anx A2t incubation is reversible, [*i.e.*]{}, after rinsing with a calcium gelating buffer (cf. [*Materials & Methods*]{}) the A2t-specific signature disappears (data not shown).
To extract structural information of the SLBs before and after protein binding, the x-ray reflectivity data in Fig. \[FIGreflectivity\] were analyzed with the help of Parrats algorithm [@Parratt_PhysRev_54; @Nevot_RevPhysAppl_80]. Briefly, the algorithm estimates the electron density profile that most likely explains the reflectivity data, [*i.e.*]{}, it yields a fit for the electron density of the probe as a function of the distance from the silicon surface. For each data set in Fig. \[FIGreflectivity\] the algorithm was initialized with varying starting parameters and the 15 best fits (lowest $\chi^2$) were selected for arithmetic averaging of the obtained electron density profiles, see exemplarily the inset in Fig. \[FIGelectronDensity\] [*a*]{} for the SLB containing 25 mol % POPS. For all details of the fitting procedure please refer to [*Materials & Methods*]{}. The solid lines in Fig. \[FIGreflectivity\] show the simulated reflected intensities based on the best fit for each data set and indicate good agreement between experimental data and the quantitative fit. The electron density profiles of the membranes with 25 mol % POPS (Fig. \[FIGelectronDensity\] [*a*]{}) and 50 mol % POPS (Fig. \[FIGelectronDensity\] [*b*]{}) show both the typical shape of a supported bilayer [@Reich_RevSciInstrum_05; @Klauda_BiophysJ_06; @Kucerka_CurrOpinColloidInterfaceSci_07]: The head distal to the silicon surface as well as the hydrophobic part are clearly visible as an increase and decrease of electron density compared to the buffer’s electron density. The head proximal to the silicon substrate is not clearly visible in the total profile of the 25 mol % POPS containing bilayer (Fig. \[FIGelectronDensity\] [*a*]{}), however, it is more pronounced for the 50 mol % POPS containing bilayer (cf. Fig. \[FIGelectronDensity\] [*b*]{}). The electron density profiles after Anx A2t incubation are shown in Fig. \[FIGelectronDensity\_w\_prot\]. An increase in electron density adjacent to the distal headgroup is observed in both the 25 mol % POPS and the 50 mol % POPS containing bilayer. We interpret this increase as an Anx A2t layer. In the case of the 50 mol % POPS containing SLB the Anx A2t signature is pronounced more strongly compared to the 25 mol % POPS bilayer, indicating an enhanced Anx A2t coverage for increasing POPS content. Indeed, as detailed in [*Materials and Methods*]{}, the 50 mol % POPS containing bilayer shows 91% coverage of Anx A2t as compared to 27% in the case of the 25 mol % POPS bilayer. Interestingly, both data sets show approximately the same protein layer thickness of about $60\,\AA$, indicating that the higher protein coverage does not influence the Anx A2t configuration. Moreover, already from visual inspection of the electron density profiles the vertical configuration of Anx A2t seems very unlikely, as it would result in a significantly larger thickness of the protein layer [@Nakata_JCellBiol_90; @Sopkova-deOliveiraSantos_BiochimBiophysActa_00; @Lambert_JBiolChem_04; @Schulz_Proteins_07]. A more quantitative analysis of the electron density profiles follows below.
Decomposition of Electron Density Profiles {#decomposition-of-electron-density-profiles .unnumbered}
------------------------------------------
To extract quantitative information about the spatial dimensions of the protein layer and to study possible conformational changes in the SLB, the electron density profiles were analyzed in more detail. To this end, the electron density profiles were divided into several component groups of the lipid bilayer and the Anx A2t complex (Table \[TABinputParameters\]). These groups contribute additively to the total electron density: two Gaussian functions were used for the headgroup (phosphate + $\mathrm{Ca}^{2+}$ + serine/choline) and the backbone (glycerol + carbonyl group), two error functions were used for the alkyl chains and one Gaussian function for the chain termini (methyl groups) of both leaflets [@Kucerka_BiophysJ_08]. The areas of all groups were fixed to the stoichiometric ratios of their electron numbers, cf. Table \[TABinputParameters\], such that only the total area of each leaflet, the width and the position of each group were fitting parameters[^2]. In addition, the area per lipid of each leaflet was constrained to values above $40\, \mathrm{\AA}^2$/lipid, since film balance measurements of PS lipids in calcium buffer revealed this value as an empirical minimum for their area fraction [@Mattai_Biochemistry_89]. Silicon substrate and water are each represented by an error function. The silicon substrate’s roughness was adjusted to the value of $4\,\AA$ obtained by independent measurements of the bare wafers and was kept constant during the fitting process. Anx A2t is represented by 6 error functions, where 2 error functions represent the p11 dimer and 4 error functions account for the hydrophilic and hydrophobic domains of the Annexin monomer, respectively. The thickness of the protein layer was calculated by the half-maximal width of the two enveloping error functions. The separation of the total electron density into component groups is not obvious per se, however, the physical constraint that the total contribution of each lipid group was fixed to its electron number (see Table 1, [@Feigenson_Biochemistry_86; @Mattai_Biochemistry_89; @Klauda_BiophysJ_06]) results in four unambiguous and stable fitting parameters, namely the (inverse) packing density of the lipids, [*i.e.*]{}, the area per lipid in each leaflet, the head-to-head distance and the width of the protein layer. These parameters before and after Anx A2t binding are summarized in Table 2 and will be discussed further below.
#### Conformation of the Anx A2t Complex. {#conformation-of-the-anx-a2t-complex. .unnumbered}
The spatial decomposition of the overall electron density profile into the different groups is shown in Figs. \[FIGelectronDensity\] and \[FIGelectronDensity\_w\_prot\] (see figure caption for color code). The thickness of the resulting Anx A2t layer is $59 \pm 6\, \mathrm{\AA}$ for the 25 mol % and $67\pm5\, \mathrm{\AA}$ for the 50 mol % POPS containing bilayer (see [*red areas*]{} in Figs. \[FIGelectronDensity\_w\_prot\] [*a*]{} and [*b*]{}). This width is remarkably close to the dimensions of the side-by-side configuration of Anx A2t ($56\, \mathrm{\AA}\, \mathrm{to}\, 61\, \mathrm{\AA}$ [@Schulz_Proteins_07]), cf. Fig. \[FIGscheme\] ([*a, top*]{}). In contrast, the vertical arrangement depicted in Fig. \[FIGscheme\] ([*a, bottom*]{}) is expected to result in an Anx A2t thickness between $90\, \mathrm{\AA}\, \mathrm{and}\, 107 \, \mathrm{\AA}$ [@Lambert_JMolBiol_97; @Nakata_JCellBiol_90] and can thus be ruled out by our data. Hence, our reflectometry experiments favor the side-by-side configuration as the most plausible mechanism for Anx A2t binding to single SLBs [*in vitro*]{}.
#### Structural Changes in the SLB. {#structural-changes-in-the-slb. .unnumbered}
Next, we analyzed the structural reorganization of the lipid bilayer associated with changes in the lipid ratio and with binding of Anx A2t. To this end, we used the head-to-head distance of the lipids as a proxy for the thickness and the area per lipid as a measure for the (inverse) packing density of the SLB, see Table \[TABbilayerParameters\]. In the presence of calcium-containing buffer, the head-to-head distance is $44\pm 1\, \mathrm{\AA}$ for the 25 mol % POPS bilayer and $42\pm1\, \mathrm{\AA}$ for the 50 mol % POPS bilayer. After Anx A2t incubation the SLB thickness increases slightly to $48\pm3\, \mathrm{\AA}$ for the 25 mol % POPS bilayer and to $45\pm 3\, \mathrm{\AA}$ for the 50% POPS bilayer, which is in both cases within the resolution tolerance of the experiment. That is, the head to head distance does not vary significantly between the samples. In contrast, the packing density displays a much stronger response to both the fraction of anionic lipid and to binding of Anx A2t: First of all, we observed that all anionic bilayers exhibit an asymmetric packing density of the two leaflets, [*i.e.*]{}, the distal (Annexin-facing) leaflet contains more lipids per unit area than the proximal (surface-facing) leaflet. In detail, for the 25 mol % POPS sample in calcium-containing buffer the area per lipid is $60\pm2\,\mathrm{\AA}^2$/lipid for the distal leaflet compared to $69\pm 3\,\mathrm{\AA}^2$/lipid for the proximal leaflet. An increase of the POPS fraction in the lipid mixture leads to a densification of the bilayer and increases its asymmetry; in the 50 mol % POPS sample the area fraction is $51\pm3 \,\mathrm{\AA}^2$/lipid for the distal as compared to $67\pm 2 \,\mathrm{\AA}^2$/lipid for the proximal leaflet. The binding of Anx A2t to the bilayer potentiates the asymmetric densification of the two leaflets: For the 25 mol % POPS sample the area fraction decreases to $46\pm 3 \,\mathrm{\AA}^2$/lipid in the distal compared to $65\pm 5\,\mathrm{\AA}^2$/lipid in the proximal leaflet. For the 50 mol % POPS sample we obtain $41\pm 4 \,\mathrm{\AA}^2$/lipid in the distal as compared to $59\pm 4 \,\mathrm{\AA}^2$/lipid for the proximal leaflet.
Note that in our decomposition procedure we explicitly allowed for the hypothetical scenario in which Anx A2t penetrates the headgroup of the bilayer to a certain degree. On the first sight it may appear as if this could be an alternative explanation for the increased electron density in the buffer-facing leaflet. However, it turned out that this hypothetical scenario did not reproduce the electron density profiles appropriately. This is because the stoichiometric ratio between the lipid components is fixed during our fitting process. Specifically, if Anx A2t would penetrate the headgroup of the bilayer more deeply, not only the electron density of the headgroups, but also the density of the lipid chains would be reduced. Apparently, this reduction of the density in the central region of the bilayer cannot be compensated by an even deeper penetration of the Anx A2t in our fitting procedure. Therefore we believe that the increased electron density in the buffer-facing leaflet is indeed due to a densification of the lipids.
Continuous Photobleaching {#continuous-photobleaching .unnumbered}
-------------------------
One might expect that the structural changes of the SLB in response to the anionic fraction of lipids and the binding of Anx A2t should also induce changes in the dynamical properties of the bilayer. In particular, the diffusion properties of the lipids should vary strongly with bilayer density. Therefore, we used continuous photobleaching to test whether the densification of the SLBs is correlated with a reduction in lipid mobility. To this end, a small fraction of fluorescently labeled lipids was added to the lipid mixture and a circular spot was continuously illuminated by the lamp of the microscope, see [*Materials & Methods*]{}. The interplay between continuous photobleaching of the dyes inside the spot and diffusion of unbleached molecules into the spot leads to a characteristic, time dependent intensity profile; the higher the diffusion constant, the broader the profile at the rim of the spot. Similar to our XR measurements, we studied SLBs with 50 mol % POPS and 25 mol % POPS, as well as a pure POPC bilayer as a control. A time series of a 25 mol % POPS containing bilayer during continuous bleaching before and after protein incubation is shown in Figs. \[FIGbleaching\] [*a*]{} and [*b*]{}, respectively. The white curves are the radially averaged intensity profiles of each exposure. Both time series show the same bleaching rate in the central region. At the rim of the field-of-view, a diffusion-induced increase of the fluorescence is observed. After incubation with Anx A2t, the fluorescence profile at the rim narrows and weakens (cf. [*arrows*]{} in Fig. \[FIGbleaching\]), indicating a lower diffusion constant of the lipids after Anx A2t binding.
Quantitative information of the diffusion constant of the labeled lipids can be obtained from the evaluation of such profiles [@Dietrich_BiophysJ_97; @Hochrein_Langmuir_06]. Briefly, a timeseries of the radially averaged intensity profiles is extracted from the fluorescence images and the bleaching rate is estimated in the center of the bleached area. The intensity profiles were corrected for uneven illumination and the radial reaction-diffusion equation, describing bleaching and lipid mobility, is solved numerically. Subsequently, the diffusion constant is used as the sole fit-parameter to minimize the $\chi^2$ between experimental and theoretical timeseries of bleaching profiles. In Fig. \[FIGdiffusionConstants\] we show the resulting diffusion constants of the three different bilayer compositions in calcium buffer before ([*dark grey bars*]{}) and after Anx A2t incubation ([*light grey bars*]{}). For the pure POPC bilayer we obtain a diffusion constant of $D = 4.2 \pm 1.4\, \mu m^2/s$ before and $D = 5.0 \pm 1.2\, \mu m^2/s$ after Anx A2t incubation, respectively. This result indicates that the supported lipid bilayer is fluid and displays a diffusion constant within the range observed in previous studies [@Ladha_BiophysJ_96; @Liangfang_JChemPhys_05; @Scomparin_EurPhysJ_09]. It also reveals that the mobility of the lipids in the pure POPC bilayer is not influenced by incubation with Anx A2t, as expected from the fact that the negatively charged POPS is essential for $\mathrm{Ca^{2+}}$-dependent Annexin binding. In calcium-containing buffer the 25 mol % and 50 mol % POPS containing bilayer show similar diffusion constants of $D = 2.3 \pm 1.0 \, \mu m^2/s$ and $D = 2.0 \pm 0.8\, \mu m^2/s$, respectively. After Anx A2t incubation, the diffusion constant of the two POPS containing SLBs is significantly reduced to $D = 0.9 \pm 0.5\, \mu m^2/s$ and $0.7 \pm 0.2\, \mu m^2/s$. Hence, both anionic SLBs show the same reduction in diffusion constant within the error bar. Note that we did not measure the reference value without calcium buffer, since the 50 mol% POPS bilayers tend to delaminate from the silicon substrate without the stabilizing effect of calcium buffer.
Discussion {#discussion .unnumbered}
==========
In this paper we investigated the conformation of the Anx A2t complex upon binding to single surface supported bilayers and the accompanied structural and dynamical changes in the lipid membranes. The thicknesses of the Anx A2t layer obtained from our x-ray measurements indicate that Anx A2t unlikely binds in the vertical conformation to a single membrane, as it would result in a significantly larger protein thickness [@Lambert_JMolBiol_97; @Nakata_JCellBiol_90]. Instead, our results favor the side-by-side configuration of the Anx A2 tetramer [@Schulz_Proteins_07] and thus, provide a non-invasive and independent verification of previous AFM studies [@Menke_ChemBioChem_04], see Fig. \[FIGsummary\] for a summarizing illustration.
Binding of Anx A2t to single membranes in a side-by-side configuration has potential implications for the route of Anx A2t-induced membrane bridging. On the one hand it seems plausible that tetramers in the side-by-side configuration perform some kind of “breathing modes”, in which one of the two Annexin monomers temporarily detaches from its membrane interface and is free to bind to an approaching bilayer. As a result, the membrane-membrane contact would be established by Anx A2t in a vertical configuration, cf. Fig. \[FIGscheme\] ([*b, bottom*]{}), in line with cryo-electron microscopy results on Anx A2t-connecting vesicles [@Lambert_JMolBiol_97]. However, this mode of membrane-bridging demands a high flexibility of the Anx A2t complex. Although previous studies indicated that the p11 dimer displays a certain flexibility [@Rety_NatStructBiol_99], it is currently unknown whether the Anx A2t tetramer is indeed able to fluctuate between vertical and side-by-side configuration. Alternatively, it was suggested that the membrane-membrane contact is mediated by the formation of a heterooctameric structure composed of two opposing Anx A2t complexes, cf. Fig. \[FIGscheme\] ([*b, top*]{}). This molecular arrangement was favored by Waisman in the case of Anx A2t-chromaffin granules interactions [@WaismanM_MolCellBiochem_95] and has also been discussed to occur following disulfide-bridge formation between cysteines within the C-terminal region of p11 [@Lewit-Bentley_CellBiolInt_00]. In this model the dynamics of octamer-formation could take two alternative routes: Either the octamers pre-assemble on a single membrane interface and thus directly allow bridging to a second membrane, or the Anx A2t tetramers distribute among both membrane interfaces and form octamers only upon membrane-membrane contact. From our x-ray experiments we found that the thickness of the protein layer on a single SLB is only compatible with a monolayer of Anx A2t, suggesting that a pre-assembly of octamers on a single membrane interface is not significant under the experimental conditions used here.
#### Structural changes in the lipid bilayer.
Typical densities of uncharged bilayers are around $70\,\mathrm{\AA}^2$/lipid [@Vacklin_Langmuir_05; @Klauda_BiophysJ_06]. Indeed, for the silicon-facing leaflet we found packing densities of $69\pm 3 \,\mathrm{\AA}^2$/lipid for the 25 mol% POPS and $67\pm 2\,\mathrm{\AA}^2$/lipid for the 50 mol% POPS containing bilayer. Hence, the density of the substrate-facing leaflet turns out to be independent on POPS amount and is close to values known for uncharged membranes. In contrast, in the presence of $\mathrm{Ca^{2+}}$ the distal leaflet density is increased to $60\pm 2 \,\mathrm{\AA}^2$/lipid for the 25 mol % POPS and to $51\pm 3\,\mathrm{\AA}^2$/lipid for the 50 mol % POPS containing bilayer. Thus, in all measurements, the distal, buffer-facing leaflet shows a pronounced and POPS dependent response to the presence of calcium-containing buffer. Indeed, it is known that for mixtures of anionic and zwitterionic lipids, $\mathrm{Ca^{2+}}$ induces molecular segregation and clustering, and the formation of domains with a higher density [@Trauble_PNAS_74]. Based on these findings, we propose that the response of the buffer-facing leaflet to $\mathrm{Ca^{2+}}$ results from a chelating effect of calcium ions. Indeed, a chelating effect of calcium ions bridging anionic lipids has been reported before [@Casal_Biochemistry_87; @Mattai_Biochemistry_89]: In the distal leaflet $\mathrm{Ca^{2+}}$ bridges at least two POPS molecules and thus results in a closer packing density, while in the proximal leaflet, if POPS is not depleted by electrostatics (see below), $\mathrm{Ca^{2+}}$ rather bridges each individual POPS molecule to the negatively charged silicon substrate, leaving the density unchanged.
Our data was obtained at a salt concentration of 20 mM Tris/HCl, 100 mM NaCl, 1 mM $\mathrm{CaCl_2}$. Here, the Debye screening length is about $9\,\mathrm{\AA}$ only, and the question emerges whether electrostatic repulsion may also favor an enrichment of the anionic lipid in the distal leaflet, thus contributing to the stronger response of this leaflet to $\mathrm{Ca^{2+}}$. On silicon supports, anionic Texas Red DHPE lipids in a POPC matrix have been found to be enriched in the distal leaflet at moderate (75 mM) monovalent salt concentrations [@Shreve_Langmuir_08]. This observation was accounted to repulsion from anionic hydroxyl groups at the silicon surface screened by the associated Debye length of approximately $11\,\mathrm{\AA}$. Since the water gap between silicon support and surface supported bilayers has been shown to be below detection limit for reflectivity studies [@Reich_RevSciInstrum_05; @Horton_Langmuir_07; @Reich_BiophysJ_08], it is well below the Debye screening length. In this context one should also note that screening within the hydrophobic core of the SLB is expected to be low and thus, if the proximal leaflet is not screened sufficiently, also the distal leaflet will be under the influence of the surface electric field [@Shreve_Langmuir_08]. Therefore a charge-induced POPS enrichment in the distal leaflet may also contribute to the investigated asymmetric densification. It has been pointed out that the enrichment mechanism itself may occur at the early stage of vesicle spreading by rapid diffusion of lipids between the two leaflets via edge effects rather than via flip-flop [@Richter_Langmuir_04]. Here we used the method of vesicle spreading with the help of osmotic pressure. Vesicle spreading was performed in pure water and thus a redistribution of negatively charged lipid during vesicle spreading is very likely. Note that for $\mathrm{SiO_2}$ the density of hydroxyl groups and thus the negative charge at the surface depends crucially on the cleaning procedure. Here we tried to maximize the amount of hydroxyl groups by aggressive wet-chemical cleaning, see [*Materials & Methods*]{}. Less efficient cleaning procedures or longer storage times after cleaning may result in a reduced number of hydroxyl groups, which may be the origin for some controversy in the literature [@Richter_Langmuir_04; @Shreve_Langmuir_08].
Upon binding of Anx A2t, for the 25% POPS containing bilayer the density of the distal leaflet increased significantly, whereas the density of the proximal leaflet was not notably affected. In the 50 % POPS containing bilayer both leaflets show a further, albeit still asymmetric, densification. Interestingly, Langmuir compression experiments with DMPA suggested a phase transition between liquid expanded to liquid condensed phase at a packing density of $40 \,\mathrm{\AA}^2$/lipid [@Schalke_BiochimBiophysActa_00], similar to the densities we found for the buffer-facing leaflet. Likewise, Watkins et al. have reported a comparable packing density for DPPC SLBs [@Watkins_PhysRevLett_09], in agreement with gel phase data. Thus, it appears as if the responses of both leaflets are decoupled unless the density in the distal leaflet reaches gel phase values. Only then also the proximal leaflet “senses” the binding Anx A2t and gets compactified. Hence, it is tempting to speculate that the Anx A2t-mediated densification of the bilayer is accompanied with a transition to gel phase in the distal leaflet. Indeed, asymmetric phase transitions in only a single leaflet of the bilayer have been observed before, [*e.g.*]{}, in dilauroylphosphatidylcholine/distearoylphosphatidylcholine (DLPC/DSPC) mixtures [@Lin_BiophysJ_06].
#### Dynamical changes in the bilayer.
Continuous bleaching measurements show a reduced diffusion constant of $D = 2.3 \pm 1.0\,\mu m^2/s$ for POPS containing SLBs in calcium containing buffer compared to $D = 4.2 \pm 1.4\,\mu m^2/s$ for pure POPC SLBs. This decrease in diffusion constant could be accounted to several mechanisms. The first mechanism is based on obstructed diffusion, in which calcium ions induce lipid domains of higher packing density and thereby form obstacles for the diffusion of fluorescently labeled lipids [@Ratto_BiophysJ_02; @Fenz_Langmuir_09]. Alternatively, since $\mathrm{Ca^{2+}}$ promotes the spreading of negatively charged small unilamellar vesicles on $\mathrm{SiO_2}$ [@Cremer_JPhysChemB_99; @Richter_Langmuir_06], one may attribute the reduced diffusion constant to partial sticking [@Dertinger_Langmuir_06] of POPS to some ion-bridged OH groups. Our x-ray results indicate that most of the structural changes due to $\mathrm{Ca^{2+}}$ are confined to the buffer-facing leaflet, suggesting that the formation of domains could indeed be the origin of the reduced lipid mobility. Yet, with our experiments we cannot exclude that sticking of POPS to the surface also plays a role. However, since only small amounts of lipids are absorbed to the surface at any time [@Dertinger_Langmuir_06], one may speculate that obstructed diffusion is the main mechanism for the reduction of lipid mobility. Obstructed diffusion in phase-separated SLBs has been studied before, and a reduction of the diffusivity by 50 % was observed for an area fraction of the gel-phase of 0.4 [@Ratto_BiophysJ_02]. Hence, due to a higher lipid density in the 50% POPS containing bilayer we expected that a higher POPS content leads to a lower diffusion constant. However, we could not resolve this reduction within experimental error. Interestingly, Gilmanshin [*et. al*]{} have observed a similar behavior for mixtures of POPC and anionic POPG [@gilmanshin1994]: The diffusion constant was not dependent on POPC concentration between values of 0 and 80 mol% in calcium containing buffer, whereas it slightly rose with higher POPC amount. Thus, our measurements show similar behavior for mixtures of POPS and POPC.
Upon Anx A2t incubation, the diffusion constant of the two POPS containing SLBs is significantly reduced to $0.9 \pm 0.5 \,\mu m^2/s$ (25 mol % POPS) and $0.7 \pm 0.2 \,\mu m^2/s$ (50 mol% POPS). Hence, both anionic bilayers show the same reduction within the error bars, while the POPC control sample displays no decrease in mobility upon protein binding. FRAP measurements with surface supported POPC/POPG membranes also revealed a decrease in diffusion constant upon binding of Annexin IV [@gilmanshin1994] and the diffusion constant decreased with increasing POPG fraction. However, above a threshold of 50 mol % POPG no further reduction of the mobility was observed. From our measurements it seems as if in POPC/POPS mixtures after Anx A2t binding this threshold appears at slightly lower amount of anionic POPS. Yet, from our data it cannot be excluded that there is still a modest dependency of the diffusion constant within the experimental errors.
Lipid demixing upon protein binding to multicomponent membranes has been discussed before in the context of oppositely charged lipid-protein pairs [@Heimburg_BiophysJ_99; @May_BiophysJ_00]. Menke et al. [@Menke_ChemBioChem_04] have shown by AFM measurements that an area of POPS enrichment develops around Anx A2t. This suggests that obstructed diffusion by phase separation could also be the origin for the reduction of diffusion in presence of Anx A2t, presumably due to POPS assembly below Anx A2t [@Dermine_JBiolChem_01]. It was speculated that Anx A2t may act directly to trap and cluster PS, thereby creating microdomains in the plasma membrane [@Chasserot2005]. Our x-ray data supports this picture: In all measurements the bilayers show a higher density after Anx A2t binding.
Conclusions {#conclusions .unnumbered}
===========
Our observations might have important consequences for the understanding of the physiological processes induced by membrane-associated Anx A2t. It was proposed that domain formation may act as nucleation site for lipid rafts and promote their clustering. Once domains are formed, raft structures and the associated cholesterol may further stabilize the lipid-Annexin 2 interaction [*in vivo*]{}, resulting in Annexin 2-membrane scaffolds that are required to assemble components of the exocytotic machinery [@Chasserot2005]. However, the role of Anx A2t in this mechanism remained unclear. In addition, it was proposed that Anx A2t creates pores through a densification of the bilayer and thereby facilitates membrane fusion [@faure2002]. Here we could for the first time demonstrate that the protein is indeed able to induce a densification of PS-containing bilayers [*in vitro*]{} and, in addition, resolve that primarily the protein-facing leaflet is compactified. It is tempting to speculate that this asymmetry might be involved in the mechanism of Anx A2t-mediated endo- and exocytosis: The asymmetric insertion of lipids into the outer monolayer of lipid vesicles is often accompanied by positive-curvature strain [@Esteban-Martin2009]. In fact, Monte Carlo simulations showed that phase separation in asymmetric bilayers leads to spontaneous budding of the membrane [@Wallace2005]. In the future it will be interesting to use off-specular neutron scattering techniques for the study of Anx A2t-induced membrane reorganization. For instance, deuterated POPS could be leveraged to enhance the contrast with respect to POPC and thus, to monitor protein-induced lipid segregation and clustering in the two leaflets.
Acknowledgements {#acknowledgements .unnumbered}
================
Financial support was obtained from BMBF (05KN7WMA) and from the DFG Nanosystems-Initiative-Munich (NIM). Support from Wolfgang Caliebe, Oliver Seeck (beamline D4) and Oier Bikondoa (beamline ID01) as well as travel support from HASYLAB/DESY (Hamburg) and ESRF (Grenoble) is gratefully acknowledged.
[57]{} \[1\][\#1]{}
Biener, Y., R. Feinstein, M. Mayak, Y. Kaburagi, T. Kadowaki, and Y. Zick. 1996. . *J. Biol. Chem.* 271:29489–29496.
Filipenko, N., T. MacLeod, C.-S. Yoon, and D. Waisman. 2004. . *J. Biol. Chem.* 279:8723–8731.
R[é]{}ty, S., J. Sopkova, M. Renourd, S. Tabares, D. Osterloh, V. Gerke, F. Russo-Marie, and A. Lewit-Bentley. 1999. . *Nat. Struct. Biol.* 6:89–95.
Burger, A., R. Berendes, S. Liemann, J. Benz, A. Hofmann, P. Göttig, R. Huber, V. Gerke, C. Thiel, J. Römisch, and K. Weber. 1996. . *J. Mol. Biol.* 257:839–847.
Rosengarth, A., and H. Luecke. 2004. *Annexins*. 1:129–136.
Sopkova-de Oliveira Santos, J., F. Oling, S. R[é]{}ty, A. Brisson, J. Smith, and A. Lewit-Bentley. 2000. . *Biochim. Biophys. Acta*. 1498:181–191.
Menke, M., M. Ross, V. Gerke, and C. Steinem. 2004. . *ChemBioChem*. 5:1003–1006.
Lewit-Bentley, A., S. R[é]{}ty, J. Sopkova-de Oliveira Santos, and V. Gerke. 2000. S100 protein complexes: some insights from structural studies. *Cell. Biol. Int.* 24:799–802.
Schulz, D., S. Kalkhof, A. Schmidt, C. Ihling, C. Stingl, K. Mechtler, O. Zschörnig, and A. Sinz. 2007. Annexin a2/p11 interaction: New insights into annexin a2 tetramer structure by chemical crosslinking, high-resolution mass spectrometry, and computational modeling. *Proteins*. 69:254–269.
Lambert, O., V. Gerke, M. F. Bader, F. Porte, and A. Brisson. 1997. Structural analysis of junctions formed between lipid membranes and several annexins by cryo-electron microscopy. *J. Mol. Biol.* 272:42–55.
Nakata, T., K. Sobue, and N. Hirokawa. 1990. . *J. Cell. Biol.* 110:13–25.
Lambert, O., N. Cavusoglu, J. Gallay, M. Vincent, J. L. Rigaud, J.-P. Henry, and J. Alaya-Sanmartin. 2004. Novel organisation and properties of annexin 2-membrane complexes. *J. Biol. Chem.* 279:10872–10882.
Miller, C. E., J. Majewski, T. Gog, and T. L. Kuhl. 2005. . *Phys. Rev. Lett.* 94:238104.
Reich, C., M. B. Hochrein, B. Krause, and B. Nickel. 2005. A microfluidic setup for studies of solid-liquid interfaces using x-ray reflectivity and fluorescence microscopy. *Rev. Sci. Instrum.* 76:095103.
Nickel, B. 2008. Nanostructure of supported lipid bilayers in water. *Biointerphases*. 3:FC40.
Novakova, E., K. Giewekemeyer, and T. Salditt. 2006. . *Phys. Rev. E*. 74:051911.
Reich, C., M. R. Horton, B. Krause, A. P. Gast, J. O. Rädler, and B. Nickel. 2008. Asymmetric structural features in single supported lipid bilayers containing cholesterol and g(m1) resolved with synchrotron x-ray reflectivity. *Biophys. J.* 95:657–668.
Horton, M. R., C. Reich, A. P. Gast, J. O. Rädler, and B. Nickel. 2007. Structure and dynamics of crystalline protein layers bound to supported lipid bilayers. *Langmuir*. 23:6263–6269.
Chasserot-Golaz, S., N. Vitale, E. Umbrecht-Jenck, D. Knight, V. Gerke, and M.-F. Bader. 2005. . *Mol. Biol. Cell*. 16:1108–1119.
Ross, M., V. Gerke, and C. Steinem. [2003]{}. . *Biochemistry*. [42]{}:[3131–3141]{}.
Gerke, V., and K. Weber. 1984. Identity of [p36K]{} phosphorylated upon rous-sarcoma virus transformation with a protein purified from brush-borders - calcium-dependent binding to non-erythroid spectrin and [F-actin]{}. *EMBO J.* 3:227–233.
Parratt, L. G. 1954. Surface studies of solids by total reflection of x-rays. *Phys. Rev.* 95:359.
Nevot, L., and P. Croce. 1980. Characterization of surfaces by grazing x-ray reflection - application to study of polishing of some silicate-glasses. *Rev. Phys. Appl.* 15:761–779.
Mattai, J., H. Hauser, R. A. Demel, and G. G. Shipley. 1989. Interactions of metal ions with phosphatidylserine bilayer membranes: effect of hydrocarbon chain unsaturation. *Biochemistry*. 28:2322.
Daniel, C., K. E. Sohn, T. E. Mates, E. J. Kramer, J. O. Rädler, E. Sackmann, B. Nickel, and L. Andruzzi. 2007. Structural characterization of an elevated lipid bilayer obtained by stepwise functionalization of a self-assembled alkenyl silane film. *Biointerphases*. 2:109–118.
Fritz, G., C. Koller, K. Burdack, L. Tetsch, I. Haneburger, K. Jung, and U. Gerland. 2009. Induction kinetics of a conditional [pH]{} stress response system in [[*Escherichia coli*]{}]{}. *J. Mol. Biol.* 393:272–286.
Fischer, H., I. Polikarpov, and A. F. Craievich. [2004]{}. . *Prot. Sci.* [13]{}:[2825–2828]{}.
Klauda, J. B., N. Kucerka, B. R. Brooks, R. W. Pastor, and J. F. Nagle. 2006. Simulation-based methods for interpreting x-ray data from lipid bilayers. *Biophys. J.* 90:2796–2807.
Kucerka, N., M. P. Nieh, J. Pencer, T. Harroun, and J. Katsaras. 2007. . *Curr. Opin. Colloid Interface Sci.* 12:17–22.
Kucerka, N., J. F. Nagle, J. N. Sachs, S. E. Feller, J. Pencer, A. Jackson, and J. Katsaras. 2008. Lipid bilayer structure determined by the simultaneous analysis of neutron and x-ray scattering data. *Biophys. J.* 95:2356–2367.
Feigenson, G. W. 1986. On the nature of calcium-ion binding between phosphatidylserine lamellae. *Biochemistry*. 25:5819–5825.
Dietrich, C., R. Merkel, and R. Tampe. 1997. Diffusion measurement of fluorescence-labeled amphiphilic molecules with a standard fluorescence microscope. *Biophys. J.* 172:1701–1710.
Hochrein, M. B., C. Reich, B. Krause, J. O. Rädler, and B. Nickel. 2006. Structure and mobility of lipid membranes on a thermoplastic substrate. *Langmuir*. 22:538–545.
Ladha, S., A. R. Mackie, L. J. Harvey, D. C. Clark, E. J. A. Lea, M. Brullemans, and H. Duclohier. 1996. . *Biophys. J.* 71:1364–1373.
Liangfang, Z., and G. Steve. 2005. Lipid diffusion compared in outer and inner leaflets of planar supported bilayers. *J. Chem. Phys.* 123:211104.
Scomparin, C., S. Lecuyer, M. Ferreira, T. Charitat, and B. Tinland. 2009. . *Eur. Phys. J. E*. 28:211–220.
Waisman, D. M. 1995. Annexin [II]{} tetramer: structure and function. *Mol. Cell Biochem.* 149-150:301–22.
Vacklin, H. P., F. Tiberg, G. Fragneto, and R. K. Thomas. 2005. Composition of supported model membranes determined by neutron reflection. *Langmuir*. 21:2827–2837.
Trauble, H., and H. Eibl. 1974. Electrostatic effects on lipid phase-transitions - membrane structure and ionic environment. *Proc. Natl. Acad. Sci. USA*. 71:214–219.
Casal, H. L., A. Martin, H. H. Mantsch, F. Paltauf, and H. Hauser. 1987. . *Biochemistry*. 26:7395–7401.
Shreve, A. P., M. C. Howland, A. R. Sapuri-Butti, T. W. Allen, and A. N. Parikh. 2008. Evidence for leaflet-dependent redistribution of charged molecules in fluid supported phospholipid bilayers. *Langmuir*. 24:13250–13253.
Richter, R., N. Maury, and A. Brisson. 2004. On the effect of the solid support on the interleaflet distribution of lipids in supported lipid bilayers. *Langmuir*. 21:299–304.
Schalke, M., P. Krüger, M. Weygand, and M. Lösche. 2000. . *Biochim. Biophys. Acta*. 1464:113–26.
Watkins, E. B., C. E. Miller, D. J. Mulder, T. L. Kuhl, and J. Majewski. 2009. Structure and orientational texture of self-organizing lipid bilayers. *Phys. Rev. Lett.* 102:238101.
Lin, W.-C., C. D. Blanchette, T. V. Ratto, and M. L. Longo. 2006. Lipid asymmetry in [DLPC/DSPC]{}-supported lipid bilayers: a combined [AFM]{} and fluorescence microscopy study. *Biophys. J.* 90:228–37.
Ratto, T. V., and M. L. Longo. 2002. Obstructed diffusion in phase-separated supported lipid bilayers, a combined [AFM]{} and [FRAP]{} approach. *Biophys. J.* 83:3380–3392.
Fenz, S. F., R. Merkel, and K. Sengupta. 2009. . *Langmuir*. 25:1074–85.
Cremer, P. S., and S. G. Boxer. 1999. Formation and preading of lipid bilayers on planar glass support. *J. Phys. Chem. B*. 103:2554–2559.
Richter, R. P., R. Berat, and A. R. Brisson. 2006. . *Langmuir*. 22:3497–3505.
Dertinger, T., I. von der Hocht, A. Benda, M. Hof, and J. Enderlein. 2006. Surface sticking and lateral diffusion of lipids in supported bilayers. *Langmuir*. 22:9339–9344.
Gilmanshin, R., C. E. Creutz, and L. K. Tamm. [1994]{}. . *Biochemistry*. [33]{}:[8225–8232]{}.
Heimburg, T., B. Angerstein, and D. Marsh. 1999. Binding of peripheral proteins to mixed lipid membranes: The effect of local demixing upon binding. *Biophys. J.* 76:2575–2586.
May, S., D. Harries, and A. Ben-Shaul. 2000. Lipid demixing and protein-protein interactions in the adsorption of charged proteins on mixed membranes. *Biophys. J.* 79:1747–1760.
Dermine, J. F., S. Duclos, J. Garin, F. St-Louis, S. Rea, R. G. Parton, and M. Desjardins. 2001. Flotillin-1-enriched lipid raft domains accumulate on maturing phagosomes. *J. Biol. Chem.* 276:18507–12.
Faure, A. V., C. Migne, G. Devilliers, and J. Ayala-Sanmartin. [2002]{}. . *Exp. Cell Res.* [276]{}:[79–89]{}.
Esteban-Mart[í]{}n, S., H. J. Risselada, J. Salgado, and S. J. Marrink. 2009. Stability of asymmetric lipid bilayers assessed by molecular dynamics simulations. *J. Am. Chem. Soc.* 131:15194–202.
Wallace, E. J., N. M. Hooper, and P. D. Olmsted. 2005. The kinetics of phase separation in asymmetric membranes. *Biophys. J.* 88:4072–83.
Tables {#tables .unnumbered}
======
[ ]{}
[ ]{}
Figure Legends {#figure-legends .unnumbered}
==============
### Figure \[FIGscheme\]. {#figurefigscheme. .unnumbered}
Cartoon of the proposed conformations of Anx A2t when bound to a single bilayer ([*a*]{}) and the resulting conformations when binding to two bilayers ([*b*]{}). The Anx A2 monomers are represented by half spheres while the p11 (S100A10) dimers are represented by ellipsoids. The side-by-side configuration ([*a, top*]{}) is formed by binding of both Anx A2 monomers to one bilayer with the S100A10 dimer on top. From this conformation it was proposed that an octameric complex might be formed in the presence of a second bilayer ([*b, top*]{}). The vertical conformation ([*b, bottom*]{}) is formed by one Anx A2 monomer binding to one bilayer and the second A2 monomer binding to a second membrane. From this model the same, albeit hypothetical, model was proposed for Anx A2t binding to a single bilayer Fig. \[FIGscheme\] ([*a, bottom*]{}).
### Figure \[FIGreflectivity\]. {#figurefigreflectivity. .unnumbered}
X-ray reflectivity data of a 25 mol % POPS containing phospholipid membrane ([*upper data points*]{}) and a 50 mol % POPS containing phospholipid bilayer ([*lower data points*]{}) in calcium buffer ([*squares*]{}) and after incubation with Anx A2t ([*triangles*]{}). The ellipsoids highlight the change in reflectivity signal due to Anx A2t binding. The solid lines represents the 15 best fit obtained by Parrats algorithm, see [*Materials and Methods*]{} for all details. For better comparison data was multiplied by $q^{4}$ and shifted on the y-axis.
### Figure \[FIGelectronDensity\]. {#figurefigelectrondensity. .unnumbered}
Electron density profiles of a 25 mol % POPS ([*a*]{}) and a 50 mol % POPS ([*b*]{}) containing SLB in calcium buffer, as obtained by quantitative fits of the data in Fig. \[FIGreflectivity\]. The inset shows the individual profiles obtained by the 15 best fits ([*grey lines*]{}) as well as their arithmetic mean ([*black lines*]{}). Blue, grey, and light grey areas indicate contributions from buffer, bilayer, and silicon support, respectively. These contributions add up to the total profile ([*red line*]{}), which is superimposed on the arithmetic mean ([*black line*]{}). The contribution from the bilayer can be further separated into contributions from both leaflets ([*dark grey areas*]{}). Each leaflet is separated into its head, its alkyl chain part, and its chain termini ([*green*]{}, [*dark green*]{}, and [*yellow area*]{}, respectively). Individual contributions of the head are shown by turquoise lines (phosphate plus choline/serine group) and green lines (glycerol plus carbonyl group).
### Figure \[FIGelectronDensity\_w\_prot\]. {#figurefigelectrondensity_w_prot. .unnumbered}
Electron density profiles of POPS containing SLBs after Anx A2t incubation. The electron density profile of a 25 mol % POPS ([*a*]{}) and a 50 mol % POPS containing SLB ([*b*]{}) is shown after incubation with Anx A2t. The color code is identical to Fig. \[FIGelectronDensity\]. Additionally, the contribution of Anx A2t to the electron density profile is represented by the red area. Individual contributions to the Anx A2t signal can be assigned to the Anx A2 monomer ([*dark red areas*]{}) and S100A10 dimer ([*thin red line*]{}), for details see text.
### Figure \[FIGbleaching\]. {#figurefigbleaching. .unnumbered}
Continuous bleaching series of a 25 mol % POPS containing bilayer before ([*a*]{}) and after ([*b*]{}) Anx A2t incubation. Images were taken before bleaching, after 60s and 120s of bleaching time. The light intensities were chosen similar in ([*a*]{}) and ([*b*]{}), such that the bleaching rates were comparable in both cases. The white rim at the border of the aperture indicates fresh non-bleached fluorescent lipids diffusing into the bleached area. The white lines show the radially averaged fluorescence profiles of the images. The arrows mark the decay-width of the fluorescence signal at the rim of the aperture: the reduced width after Anx A2t incubation in ([*b*]{}) is indicative of a reduced lipid mobility.
### Figure \[FIGdiffusionConstants\]. {#figurefigdiffusionconstants. .unnumbered}
Diffusion constant of a pure POPC, a 25 mol % POPS and a 50 mol % POPS bilayer in calcium buffer ([*dark grey bars*]{}) and after Anx A2t incubation ([*light grey bars*]{}). Values are mean values from at least 3 independent samples with at least 5 independent measurement points on each sample. Error bars represent standard deviations from the mean.
### Figure \[FIGsummary\]. {#figurefigsummary. .unnumbered}
Proposed quarternary structure of Anx A2t upon binding to a supported lipid bilayer. Two Anx A2 monomers ([*half spheres*]{}) bind to the distal leaflet of the membrane while a p11 dimer ([*ellipsoids*]{}) sits on top of this structure. The presence of calcium and Anx A2t lead to a densification of the protein-facing leaflet of the bilayer, presumably due to an enrichment of POPS as compared to POPC.
![\[FIGscheme\] Fritz et al.](Fritz_etal_Fig1){width="6in"}
![\[FIGreflectivity\] Fritz et al.](Fritz_etal_Fig2){width="7in"}
![\[FIGelectronDensity\] Fritz et al.](Fritz_etal_Fig3){width="6in"}
![\[FIGelectronDensity\_w\_prot\] Fritz et al.](Fritz_etal_Fig4){width="6in"}
![\[FIGbleaching\] Fritz et al. ](Fritz_etal_Fig5){width="6in"}
![\[FIGdiffusionConstants\] Fritz et al.](Fritz_etal_Fig6){width="6in"}
![\[FIGsummary\] Fritz et al. ](Fritz_etal_Fig7){width="6in"}
[^1]: Corresponding author. Full address: Department für Physik and CeNS, Ludwig-Maximilians-Universität, Geschwister-Scholl-Platz 1, D-80539 München, Germany, Tel.: (+49) 89 2180-1460, Fax: (+49) 89 2180-1452, E-mail: [email protected]
[^2]: The area of the gaussian for the methyl groups was adjusted in stoichiometric proportion to the arithmetic mean of the areas of the alkyl chains in both leaflets.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The motivation of this paper is to explore the parameters that affect the performance of Microchannel Plate Photomultiplier Tubes (MCP-PMTs) in magnetic fields with the goal to guide their design to achieve a high magnetic field tolerance. MCP-PMTs based on two different designs were tested. The magnetic field tolerance of MCP-PMT based on a design providing independently biased voltages showed a significant improvement (up to 0.7 T) compared to the one utilizing an internal resistor chain design (up to 0.1 T), indicating the importance of individually adjustable voltages. The effects of the rotation angle of the MCP-PMT relative to the magnetic field direction and of the bias voltage between the photocathode and the top MCP were extensively investigated using the MCP-PMT based on the independently biased voltage design. It was found that the signal amplitude of the MCP-PMT exhibits an enhanced performance at a tilt angle of $\pm$8$^{\circ}$, due to the 8$^{\circ}$ bias angle of the MCP pores. The maximum signal amplitude was observed at different bias voltages depending on the magnetic field strength.'
title: '**Characteristics of fast timing MCP-PMTs in magnetic fields**'
---
Fast timing, Microchannel plate, Photodetector, Electron-Ion Collider, Particle identification detector, Magnetic field.
Introduction {#sec:level1}
============
The Electron-Ion Collider (EIC) [@1], which is recommended in the 2015 Long Range Plan for Nuclear Science [@2] as the highest priority for a new facility construction in the US, aims to revolutionize our understanding both of nucleon and nuclear structure and of nuclear dynamics in the many-body regime, where strongly coupled relativistic quantum fluctuations and non-perturbative effects combine to give a dynamical origin to nuclear mass and spin. The broad physics program of the EIC requires a large multipurpose spectrometer able to measure a plethora of physics processes over a wide range of energies and solid angles. Particular to the EIC is the requirement of particle indentification, i.e., the separation of electrons, pions, kaons, and protons (e/$\pi$/K/p) in the final state in processes such as semi-inclusive deep inelastic scattering and charm production.
To address the broad physics potential of the EIC, several detector concepts are being proposed, including the BeAST [@3] and the sPHENIX [@4] concepts from Brookhaven National Laboratory (BNL), the JLEIC full acceptance detector [@5] from Thomas Jefferson National Accelerator Facility (JLab), and the TOPSiDE 5D particle flow detector [@6] from Argonne National Laboratory (ANL). These detector concepts feature different layouts of sub-systems, which have been worked out to varying detail. Common to all concepts are the use of Time-Of-Flight (TOF) systems and imaging Cherenkov detectors for hadron particle identification. Integration of these sub-systems into the central detector requires to placing their photo-sensors in the non-uniform fringe field of the solenoidal magnet. Thus particle identification at the EIC requires low-cost photon sensors with picosecond timing resolution, millimeter spatial resolution, high rate capability, and last but not least high radiation and magnetic field tolerance.
The microchannel plate photomultiplier tube (MCP-PMT) [@7] is a compact photosensor consisting of a photocathode for photon-electron conversion, two MCPs in a stacked chevron configuration for electron amplification and a readout system for charge collection. The compact design and confined electron amplification by secondary electron emission inside the micron size MCP pores provide the MCP-PMT with picosecond timing resolution and millimeter position resolution, ideal for application in time-of-flight systems and imaging Cherenkov detectors. The LAPPD collaboration [@8] between universities, U.S. national laboratories, and industrial partners developed the technology to manufacture the world’s largest MCP based photosensor, the Large-Area Picosecond Photon Detector (LAPPD$^{TM}$). A critical aspect of the LAPPD$^{TM}$ technology is its use of low-cost, very large area (20 $\times$ 20 cm$^2$) MCPs [@9] within an all glass vacuum envelope. The MCPs used in LAPPDs$^{TM}$ are made from bundled and fused capillaries of borosilicate glass functionalized through atomic-layer deposition [@10; @11; @12] of conductive and secondary-electron emissive material layers. This revolutionary process eliminates the chemical etching and hydrogen firing steps employed in traditional MCP manufacturing, which caused the glass to become brittle and resulted in strong ion feedback. These features and the inherent mechanical stability of borosilicate glass allows the production of exceptionally large area MCPs with long lifetime [@13] and low background noise rates [@14].
As integral part of the LAPPD project, a dedicated fabrication facility [@15] capable of producing 6$\times$6 cm$^2$ MCP-PMTs based on the LAPPD design was built at Argonne National Laboratory. The facility served as intermediate production facility while preparing for mass production with our industrial partner, Incom, Inc [@16]. To date the Argonne facility produced several dozens of 6$\times$6 cm$^2$ MCP-PMTs which were provided to various users for early evaluation. As Incom, Inc. cranks up mass production of LAPPDs$^{TM}$, the Argonne fabrication facility will be converted into an R&D platform for LAPPD$^{TM}$ design optimizations geared to specific applications. Within a fast turn around, small size (6$\times$6 cm$^2$) MCP-PMTs based on different designs can be produced and can be tested either on the test bench or in particle beams. Once availble the optimized design can be transferred directly to Incom, Inc. for LAPPD$^{TM}$ mass production.
In this paper, we report on tests in magentic fields of two 6$\times$6 cm$^2$ MCP-PMTs based on different designs, as produced in the Argonne fabrication facility. We describe the different designs in section 2, the magnetic field tolerance measurement setup in section 3, while the experimental results are presented and discussed in section 4; conclusions are drawn at the end of the paper.
Designs of the MCP photodetector {#sec_design}
================================
Two MCP-PMTs based on different designs were tested in this study: the internal resistor chain design and the independently biased design. The former relies on ALD coated MCPs and spacers inside the MCP-PMT for bias voltage distribution, while the latter relies on an external high voltage divider for bias voltage distribution.
Internal resistor chain MCP-PMT design
--------------------------------------
The internal resistor chain MCP-PMT design is adapted from the original LAPPD$^{TM}$ design [@17]. The left panel of Figure \[fig:design\] shows a schematic of the internal resistor chain MCP-PMT design. The sealed vacuum package consists of a photocathode, two MCPs, three grid spacers and a stripline anode. An air-sensitive alkali antimonide photocathode is deposited on the inside surface of the top glass window, and the electronic connection is provided by a pre-coated nichrome layer at the edges of the top window to apply high voltage. Two MCPs with pores of 8$^{\circ}$ bias angles are placed in chevron geometry to prevent drift of positive ions to the photocathode and to ensure a well-defined first strike of the incoming photoelectrons. The MCPs used here are sliced from the same ALD coated 20 $\times$ 20 cm$^2$ MCPs used for LAPPD$^{TM}$ production, featuring a pore size of 20 $\mu$m, a length to diameter (L:d) ratio of 60:1 and an open to full area ratio of 65%. Glass spacers are used between the photocathode and the top MCP, between the MCPs, and between the bottom MCP and the anode to separate individual components and support the stack configuration. The stripline anode is made through silk-screening of silver strips onto the glass tile base, and each stripline is grounded through a resistor. It is important to note that the MCPs and glass spacers are all coated with resistive materials via the ALD method, making the whole detector stack an internal resistor chain, as indicated by the dashed line circuit in Figure \[fig:design\]. When a single high voltage (HV) is applied to the photocathode, the applied HV is distributed between the internal components, controlled by the resistances of the ALD coated MCPs and glass spacers. Signals generated by incident photons are picked up from the stripline anodes and routed to an oscilloscope or an electronic waveform digitizer.
The internal resistor chain design only requires one HV connection from the outside to the inside of the vacuum as provided by the pre-coated nichrome mask on the top window. This simple design offers the advantage of ease of implementation and potentially low production cost. However, processing and testing of the fabricated MCP-PMTs reveal several drawbacks: (a) the HV distribution relies on the resistance ratios between the spacers and MCPs, where it is challenging to identify precisely matched resistances for the MCPs and spacers; (b) the fabrication of MCP-PMT requires thorough baking and scrubbing of the MCPs under vacuum for outgassing, while it has been shown that the resistances of ALD coated MCPs and spacers are reduced unevenly during this process, possibly resulting in mismatched resistances of MCPs and spacers; (c) once the detector is sealed, there is no way to individually optimize the MCP’s performance as the bias voltage on each MCP cannot be adjusted individually; (d) the absolute quantum efficiency (QE) of the photocathode cannot be measured using the traditional method as the photocurrent (nA level) generated from incident photons is dwarfed by the continuous bias current ($\mu$A level) of the resistor chain.
Independently biased MCP-PMT design {#sec_design}
-----------------------------------
The independently biased MCP-PMT design (IBD) offers the option to optimize the performance of each MCP individually. A schematic of the new IBD configuration is shown in the right panel of Figure \[fig:design\]. The major differences compared to the previously described internal resistor design include: (1) the spacers are bare glass grids with no ALD coating on the surface, so the spacers can be treated as insulators; (2) ultra-thin stainless steel shims with the same pattern as grid spacers are attached between the spacers and the MCP surfaces to provide HV connections; (3) finger tabs are implemented on each shim, leading to the nearest silkscreen printed silver strip contact at one corner, which in turn provides the HV connection to the outside. Four shims are inserted between the upper and lower surfaces of the two MCPs. The new IBD design is based on a minimal modification of the internal resistor chain design, using shims and corner strip lines for the HV connection, while no pins are required to provide high voltage on the MCPs and in the gaps. Figure \[fig:pics\] shows a photograph of a sealed MCP-PMT based on the independently biased design (right). Simple readout circuit boards were designed and fabricated to hold the MCP-PMTs. An external resistor chain HV divider is integrated into the readout board of the independently biased MCP-PMT design so that only one HV source is necessary. The bias voltage of individual MCPs can be independently adjusted by altering the values of the corresponding resistors.
Magnetic field tolerance test facility {#sc}
======================================
Argonne National Laboratory acquired a decommissioned superconducting magnet from a magnetic resonance imaging (MRI) scanner. The primary goal of this magnet is to perform the precise calibration of the various magnetic probes for the g-2 muon experiment [@18]. The MRI magnet provides a large bore with a diameter of 68 cm and a very homogeneous field (7 ppb/cm), with a tunable magnetic field strength of up to 4 Tesla. We assembled a characterization system compatible with the solenoid magnet to test the performance of the 6$\times$6 cm$^2$ MCP-PMTs in strong magnetic fields of this magnet. A non-magnetic, light-tight dark box was built to contain the MCP-PMTs during their tests. The dark box was held on a platform with the detector surface normal to the direction of the magnetic field. The position of the dark box was adjusted so that the center of the MCP photodetector was aligned with the center of the solenoid magnet. A rotation mechanism was integrated into the system, able to rotate the MCP-PMTs by an angle $\theta$ (-90$^{\circ}~\leq~\theta~\leq~$ 90$^{\circ}$), as illustrated in Figure \[fig:theta\].
![Photograph of the magnetic field tolerance testing system.[]{data-label="fig:4"}](Figure4.jpg)
Figure \[fig:4\] shows a picture of the entire magnetic field tolerance testing system. A 405 nm light-emitting diode (LED) driven by a pulse generator provided the light source. The light was guided into the dark box via an optical fiber. High voltage was applied to the MCP-PMTs from a power supply with continuous voltage control. Signals collected at the striplines were read out through a DT5742 desktop digitizer [@19] with a sampling rate of 5 GS/s, as produced by CAEN (Costruzioni Apparecchiature Elettroniche Nucleari S.p.A.). The digitizer is based on a switched capacitor array of DRS4 (Domino Ring Sampler) chips [@20] and features 16 analog input channels, and one additional analog input for a fast trigger.
A similar MRI magnet with tunable magnetic field up to 3 Tesla and a similar platform but without the rotation mechanism were available for MCP-PMT testing at the University of Virginia. The following measurements of the MCP-PMT based on the internal resistor chain design were performed at University of Virginia, while measurements of the MCP-PMT based on the independently biased design were performed at Argonne.
Results and discussion
======================
The operational principle of MCP-PMTs relies on the electron multiplication process where the MCP pore walls are bombarded with secondary electrons. Each pore of the MCP has an internal diameter of 20 $\mu$m with the inner wall processed with resistive and secondary emissive coating layers, which act as an independent electron multiplier. When the MCP-PMT is operated in a magnetic field, the trajectories of electrons during the electron multiplication process are affected by the Lorentz force due to the presence of both electric and magnetic fields. We studied the MCP-PMT performance as a function of magnetic field strength, rotation angle, and photocathode to MCP electric field strength.
Dependence on the magnetic field strength
-----------------------------------------
The performance of MCP-PMTs based on the above two designs was tested in the magnetic field at a zero rotation angle $\theta$, i.e., where the direction of the magnetic field is normal to the surface of the MCP photodetector. A 405 nm pulsed LED with a fixed light intensity corresponding to 10 photoelectrons was used as light source. The signal amplitude versus magnetic field strength is shown in Fig. \[fig:5\].
![The response of the MCP-PMT as a function of the tilt angle $\theta$ between the normal to the MCP-PMT window and the direction of the magnetic field. The two peaks around -8$^{\circ}$ and 8$^{\circ}$ are related to the bias angle of the MCP pores. Note that the intensities of the two peaks are not the same due to the different effect of the top and bottom MCPs. []{data-label="fig:6"}](Figure6.jpg)
The MCP-PMT based on the internal resistor chain design shows a poor magnetic field tolerance, the signal amplitude drops by a factor of 6 when the magnetic field increases from 0 to 0.1 Tesla, and another factor of 6 when the field increases to 0.2 Tesla. This rapid decrease is mainly due to the resistances of the MCPs and spacers having been significantly changed during the baking and scrubbing process, resulting in a bias voltage mismatch of the two MCPs. These results indicate that the MCP-PMT based on the resistor chain design is not suitable for applications in magnetic fields over 0.1 T.
On the other hand, the MCP-PMT based on the independently biased design shows a significantly improved tolerance to magnetic fields. The performance was measured at various magnetic field strengths and applying various bias high voltages. At a fixed magnetic field strength, the signal amplitude increases with increasing bias high voltage. This behavior confirms our previous measurements of MCP-PMTs in a negligible magnetic field [@21]. At a fixed bias voltage of 3100 V, the signal amplitude of the MCP-PMT increases slightly as the magnetic field strength increases to 0.2 T, and then is seen to decrease as the magnetic field strength continues to increase, and eventually is reduced below 5 mV at a magnetic field strength of 0.7 T. With lower bias voltages, the signal amplitudes are seen to be reduced already at lower field strength. As these results show, the decrease in signal strength can to some extend be compensated by increased bias voltages.
Dependence on the tilt angle
----------------------------
The signal strength as function of tilt angle $\theta$ between the normal to the MCP-PMT window and the direction of the magnetic field, as shown in Figure \[fig:theta\], was investigated using the MCP-PMT based on the independently biased design. We applied a fixed high voltage of 3000 V and rotated the tilt angles $\theta$ from -90$^{\circ}$ to 90$^{\circ}$. Figure \[fig:6\] shows the signal amplitude as a function of the tilt angle $\theta$ for two magnetic field strengths of 0.25 and 0.5 Tesla, respectively. The signal amplitude shows a strong angle dependence with vanishing signals outside the range of -30$^{\circ}~\leq~\theta~\leq~$30$^{\circ}$ and two maxima at $\pm$8$^{\circ}$. The latter are related to the 8$^{\circ}$ bias angle of the MCP pores and their chevron configuration. When the direction of one MCP pore is aligned with the direction of the magnetic field, the MCP-PMT shows an enhanced magnetic field tolerance. The signal maximum at +(-)8$^{\circ}$ corresponds to the position where the direction of the pores of the top (bottom) MCP is aligned with the direction of the magnetic field.
![The electrical circuit of HV connections devised specifially to be able to vary the gap voltage between the photocathode and the top MCP.[]{data-label="fig:7"}](Figure7.jpg)
Dependence on the gap high voltage
----------------------------------
The MCP-PMTsignal amplitude as a function of applied HV to the gap between the photocathode and the top MCP was studied at different magnetic field strengths. Figure \[fig:7\] shows the circuit diagram which allowed to vary the applied gap voltage. While the HV$_{MCPs}$ was kept at a fixed value, the HV$_{Photocathode}$ was varied to adjust the gap voltage.
![Performance of the MCP-PMTs in terms of signal amplitude as a function of gap voltage applied between the photocathode and top MCP in different magnetic fields.[]{data-label="fig:8"}](Figure8.jpg)
Figure \[fig:8\] shows the signal amplitude versus gap voltage for a selection of magnetic field strengths. With low magnetic fields, the signal amplitude increases as the gap voltage increases and reaches a maximum at a gap voltage of $\sim$ 500 V. Further increasing the gap voltage beyond 500 V results in a decreasing signal amplitude. This effect is related to the energy of the primary electrons, as studied previously [@22]. As these studies showed, the yield of secondary emissions of ALD coated materials is highest when the primary electron energy is around 300 – 500 eV, corresponding to a gap HV at $\sim$ 500 V. With higher primary electron energies, the electrons penetrate deeper into the coating, thus again reducing the yield of secondary emission electrons. At high magnetic fields, the magnetic field strength becomes the main parameter affecting the secondary emission process. The secondary yield is not seen to decrease anymore with primary electron energy over 500 eV, resulting in a continuously increasing signal amplitude with increasing gap voltages.
Conclusions
===========
Two 6$\times$6 cm$^2$ MCP-PMTs based on the internal resistor chain design and the independently biased design were fabricated at Argonne National Laboratory and characterized in magnetic fields. The behavior of the MCP-PMT signal amplitude was investigated as a function of the magnetic field strength, the distribution of bias voltage, the tilt angle, and the gap voltage. It was found that the MCP-PMT based on the internal resistor chain design shows a magnetic field tolerance only up to 0.1 T. With the independently biased voltage design, the magnetic field tolerance of the MCP-PMT is significantly improved, up to 0.7 T. These findings indicate the importance of ensuring that both MCPs are operated at their optimal bias voltages when operated in high magnetic fields. As the magnetic field strength increases, the signal amplitude of the MCP-PMT decreases while being operated at a constant bias voltage. However, the reduction of signal amplitude can be compensated by increasing the operation voltage, extending the range of operability in high magnetic fields. Due to the original MCP bias angle of 8$^{\circ}$ and the chevron configuration, the pores of both MCPs can not be aligned simultaneously with the direction of the magnetic field. The MCP-PMT shows higher signal amplitudes when either MCP pores is aligned with the direction of the magnetic field, where the direction of the top MCP pores exhibits a stronger impact on the signal amplitude. Increasing the bias voltage applied on the gap between the photocathode and the top MCP results in a maximum signal amplitude with gap voltage around 500 V at low magnetic fields, while a continuously increasing signal amplitude is observed at high magnetic fields.
acknowledgment.tex
[00]{}
A. Accardi, et al., Electron ion collider: The next QCD frontier, Eur. Phys. J. A 52 (2016) 268.
A. Aprahamian, et al., Reaching for the horizon: The 2015 long range plan for nuclear science (2015).
A. Kiselev, BeAST Detector (Brookhaven eA Solenoidal Tracker), presentation on Electron Ion Collider User Group Meeting, Berkeley, CA (2016).
PHENIX Collaboration, Concept for an Electron Ion Collider (EIC) detector built around the BaBar solenoid, arXiv:1402.1209.
G. Wei, et al., Integration of the full-acceptance detector into the JLEIC, Proceedings of IPAC2017, THPAB084, Copenhagen, Denmark (2017).
J. Repond, TOPSiDE – Concept of an EIC Detector, Workshop on Streaming Readout, Boston, MA (2018).
J. L. Wiza, Microchannel plate detectors, Nucl. Instr. and Meth. A 162 (1979) 587.
B. Adams, et al., A brief technical history of the Large-Area Picosecond Photodetector (LAPPD) Collaboration, (2016) arXiv:1603.01843.
M. Minot, et al., Pilot production & commercialization of LAPPD, Nucl. Instr. and Meth. A 787 (2015) 78.
A. U. Mane, et al., An atomic layer deposition method to fabricate economical and robust large area microchannel plates for photodetectors, Physics Procedia, 37 (2012) 722.
M. Popecki et al., Microchannel plate fabrication using glass capillary arrays with atomic layer deposition films for resistance and gain, J. Geophys. Res. Space Phys., 121 (2016) 7449.
A. O’Mahony et al., Atomic layer deposition of alternative glass microchannel plates, J. Vac. Sci. Tech., 34 (2016) 01A128.
T. M. Conneely, J. S. Milnes, and J. Howorth, Extended lifetime MCP-PMTs: Characterisation and lifetime measurements of ALD coated microchannel plates, in a sealed photomultiplier tube, Nucl. Instr. and Meth. A 732 (2013) 388.
O. H. W. Siegmund, et al., Performance characteristics of atomic layer functionalized microchannel plates, Proc. SPIE 8859, (2013) 88590Y.
J. Xie, et al., Development of a small form-factor 6 x 6 cm$^2$ picosecond photodetector as a path towards the commercialization of large area devices, Proceeding of The Technology and Instrumentation in Particle Physics 2014, PoS (2014).
Incom, Inc.: http://www.incomusa.com/
B. Adams, et al., Measurements of the gain, time resolution, and spatial resolution of a large 20 x 20 cm$^2$ MCP-based picosecond photo-detector, Nucl. Instr. and Meth. A 732 (2013) 392.
4 Tesla Magnet Facility: https://www.anl.gov/hep/group/4-tesla-magnet-facility
DT5742 desktop digitizer: http://www.caen.it
DRS chip developed at Paul Scherrer Institute, Switzerland: https://www.psi.ch/drs
J. Wang et al., Development and testing of cost-effective, 6 x 6 cm$^2$ MCP-based photodetectors for fast timing applications, Nucl. Instr. and Meth. A 804 (2015) 84.
S. Jokela et al., Secondary electron yield of emissive materials for large-area micro-channel plate detectors: surface composition and film thickness dependencies, Physics Procedia 37 (2012) 740.
| {
"pile_set_name": "ArXiv"
} |
The effort devoted during the last decades to investigate off equilibrium systems has achieved a series of successes by virtue of a combination of experiments, numerical simulations, exact theoretical results and clever phenomenological arguments, but our comprehension of the area is far from complete. Among these systems granular materials (GM), i.e. assemblies of macroscopic particles which dissipate their energy through inelastic collisions and frictional forces, have acquired a special rank due to their complex phenomenology often intriguing and only partly understood [@review].
The existing theoretical approaches are based either on realistic descriptions of the grains or on idealized models, which in virtue of their major simplicity lend themselves to analytic solutions or to efficient computations. Furthermore such an idealized modeling can deliberately sacrifice likeness to reality, in order to identify the physical mechanism responsible for the salient features of the system. In such a spirit, we shall discuss a minimal model based on the simplest rule which describes the inelasticity of collisions and local momentum conservation.
Years ago, S. Ulam showed that an ensemble of elastic particles, starting from an arbitrary configuration, converged to a Maxwell equilibrium distribution, postulating a simple redistribution law of the kinetic energies of randomly selected pairs to simulate the effect of binary collisions in an elastic gas [@Ulam]. Ben Naim and Kaprivski (BK), recently, performed a variation over this theme, by letting the particles endowed with a scalar velocity to dissipate inelastically a fraction of the relative kinetic energy at each collision and found that the total kinetic energy decreases exponentially with time [@Bennaim]. Both models fulfill Boltzmann’s molecular chaos hypothesis, and consequently rule out the formation of dynamical correlations. On the other hand computer simulations have shown the appearance of a shear instability, i.e. vortices, during the cooling process, before the spontaneous formation of density clusters. In an interesting series of papers, Ernst and collaborators have put forward a mesoscopic theory of these phenomena, making a connection with phase ordering kinetics [@Ernst].
In the present work we shall introduce and study a new microscopic model which preserves the simplicity of the approach [*á la*]{} Ulam, and displays a complexity similar to that observed in granular systems. The focus of our study will be on the statistics of the velocity field and on its spatial and temporal correlations, stressing the analogies and the differences with related models aimed to describe off equilibrium systems. We introduce our dynamical model by associating a d-dimensional velocity field $\bf{v_i}$ with each node of a d-dimensional lattice; at each time step a nearest neighbor pair $(i,j)$ is randomly selected and the two velocities are updated according to the rule:
$$\begin{aligned}
{\bf v_j'}&=&{\bf v_j}+\Theta({\bf -(v_i-v_j)}\hat \sigma ) \frac{ 1+\alpha}{2}({\bf(v_i-v_j)}\hat \sigma ) \hat \sigma \nonumber\\
{\bf v_i'}&=&{\bf v_i}-\Theta({\bf-(v_i-v_j)}\hat \sigma )\frac{1+\alpha}{2}({\bf (v_i-v_j)}\hat \sigma ) \hat\sigma
\label{rule}\end{aligned}$$
where ${\hat \sigma}$ represents the unit vector pointing from site ${\bf i}$ to ${\bf j}$, $\Theta$ is the Heaviside function which enforces the kinematic constraint and $\alpha$ the normal restitution parameter. We shall measure time in the non dimensional number of collisions per particle. In each elementary collision (see Eq.(\[rule\])) the total linear and angular momentum are conserved, whereas a fraction $(1-\alpha^2)/4$ of the relative kinetic energy is dissipated. The inelasticity of the collisions has the effect of reducing the quantity $|({\bf v_i-v_j})\hat
\sigma|$, i.e. induce a partial allignement of the velocities. Hereafter we report the study performed in two dimensions on a triangular lattice[@1d].
The freely cooling process exhibits striking similarities with the quench from an initially stable disordered phase to a low temperature phase in a magnetic system: whereas in a standard quench process [@Bray] one considers the process by which a thermodynamic system, brought out of equilibrium by a sudden change of an external constraint, such as temperature or pressure, finds its new equilibrium state, in a GM one wants to study the relaxation of a fluidized state, after the external driving force (whose role is to reinject the energy dissipated by the collisions, keeping the system in a statistically steady state) is switched-off abruptly at some time $t=0$. The rotational symmetry of the order parameter ${\bf v_i}$ and the momentum conserving interaction determine the presence of many configurations having comparable dissipation rates. Due to their competition the system does not relax immediately towards a motionless state, but displays a phenomenology similar to that observed in a coarsening process.
One sees from Fig.\[fig1\] that during the initial stage, the total energy per particle $\epsilon (t)=\sum v_i(t)^2/N$ is dissipated at an exponential rate $\tau^{-1}= (1-\alpha^2)/4$. This can be deduced from Eq.(\[rule\]) imposing that each ’spin’ fluctuates independently of the others. For times larger than $t \sim t_c$, of the order of $\tau$, the system displays a crossover to a different regime, where the cooperative effects become dominant and the average energy per particle decay as $\epsilon(t) \sim t^{-1}$. Such a behavior agrees with inelastic hard spheres simulations (IHSS) reported in [@Ernst2]. As shown below, the crossover from one regime to the other is due to the formation of a macroscopic velocity field. This is analogous to the formation of magnetic domains in standard quench processes. After the formation stage these regions start to compete to homogenize, causing a conversion of kinetic energy into heat by viscous heating, i.e. act against the collisional cooling and lead to a slower decay of the energy [@Ernst].
Within the early regime the velocity distribution deviates sensibly from a Maxwell distribution (corresponding to the same average kinetic energy), but displays fatter tails, a phenomenon which mirrors the behavior of the BK [@Bennaim] model. The existence of these tails seems to be due to the lack of spatial correlations, intrinsically absent at all times in their model, whereas negligible in ours up to $t_c$. When the energy begins to decay as $t^{-1}$ the velocity distribution turns Gaussian. Interestingly, the smaller the inelasticity, the faster is the energy dissipation, a phenomenon observed in IHSS [@Mcnamara].
The most relevant information about the spatial ordering process is contained in the equal-time structure functions, i.e. the Fourier transforms of the velocity correlation function: $$S^{t,l}(k,t)=\sum_{\hat k} {\bf v^{t,l}}({\bf k},t){\bf v^{t,l}}(-{\bf k},t)$$ where the superscripts $t,l$ indicate the transverse and longitudinal components of the field with respect to the wave vector ${\bf k}$ and the sum $\sum_{\hat k}$ is over a circular shell of radius $k$. Such structure factors, if rewritten in terms of the variable $(k t^{1/2})$ display fairly good data collapse, apart from the large $k$ region, and identify two growing lengths $L^{t,l}(t)$ (see Fig.\[fig2\]). Considering the sum rule $\epsilon(t)=\sum_k [S^{t}(k,t)+S^{l}(k,t)]$ we observed that in the early ’exponential’ regime the contribution from the two terms is approximately equal, whereas for times larger than $t_c$ and $\alpha$ not too small most of the kinetic energy remains stored in the transverse field, while the longitudinal component decays faster, with an apparent exponent $t^{-2}$.
The findings concerning the energy decay, the distribution of the velocity field, and the growth of $L^{t,l}(t)$, lead to the conclusion, that, if the observation time is longer than the time between two collisions and if the spatial scale is larger than the lattice spacing, the system behaves as if its evolution were governed by a diffusive dynamics [@Ernst]. To be more precise, let us consider a vector field ${\vec \phi}(x,t)$ which evolves according to the law $ \partial_t {\vec \phi}=\nu \nabla^2 {\vec \phi} $ starting from a random uncorrelated initial condition. The explicit solution shows that ${\vec \phi}(x,t)$ is asymptotically Gaussian distributed, with a variance $<\vec \phi(x,t) \vec \phi(x,t)>\propto t^{-d/2}$. The structure factors $S^{t,l}(k,t)$ assume a scaling form $S^{t,l}(k,t)=s(kL(t))$ where $L(t)=\sqrt{t}$. Furthermore, we compared the two-time autocorrelation $C(t_1,t_2)=\sum_i {\bf
v}_i(t_1) {\bf v}_i(t_2)/N$ with $C_{\phi}(t_1,t_2)=<{\bf
\phi}(x,t_1){\bf \phi}(x,t_2)>$, whose expression reads: $$\frac{C_{\phi}(t_1,t_2)}{C_{\phi}(t_1,t_1)}=\frac{2}{(1+\frac{t_1}{t_2} )}$$
During a short time transient, the autocorrelation function of our model differs from $C_{\phi}$, since it depends on $t_1-t_2$, i.e. it is time translational invariant ([*TTI*]{}). Later, $C(t_1,t_2)$ reaches the “aging” regime and depends only on the ratio $x=t_1/t_2$. Something similar occurs in a coarsening process, where the autocorrelation of the local magnetization $a(t_w,t_w+\tau)$ reaches, for large $\tau$ (but $\tau << t_w$), a constant value $m_{eq}^2(T)$, that is the square of the equilibrium magnetization. Obviously, for $T\to 0$, $m_{eq}^2\to 1$ and the [*TTI*]{} transient regime disappears. The short time transient in our model is analogous to such a [*TTI*]{} regime, with the difference that the cooling process imposes a decreasing temperature $T(t_w)\to 0$, that progressively erodes the [*TTI*]{} regime. The same dependence on the [*TTI*]{} manifests itself in the angular auto-correlation: $$A(t,t_w)=\frac{1}{N}\sum_i \cos(\theta_i(t+t_w)-\theta_i(t_w)).$$ Again, for large waiting times $t_w$ this function assume the diffusive $t/t_w$ scaling form, but for a small fixed $t_w$, displays a minimum and a small peak before decreasing at larger $t$ (see Fig.\[fig3\]). The non-monotonic behavior of $A(t,t_w)$ suggests that the initial direction of the velocity induces a change in the velocities of the surrounding particles, which in turn generates, through a sequence of correlated collisions, a kind of retarded field oriented as the initial velocity. As $t_w$ increases the maximum is less and less pronounced.
In spite of these first results, that seem to give support to the idea that the model dynamics is purely diffusive [@persistence], the model is more complex. The main evidence stems from the following facts:
i\) the structure functions do not have the typical Gaussian tails of a diffusive system, due to the non linearity represented by the kinematic factor in Eq.(\[rule\]) and the shapes of $S^{t,l}(k,t)$ display three different regions: a long-wavelength region which is diffusive in character; an intermediate region where the structure functions decay as $k^{-\beta}$ with $\beta \sim 4$; a plateau region where $S^{t,l}$ decay in time with a power law $t^{-2}$ but remain nearly constant with respect to $k$ (for $r>0$);
ii\) the Fourier modes interact and an initial shear state, obtained assuming the initial configuration to be a plane wave, decays into shorter wavelength modes by a mechanism of period doubling; that is to say, contrary to the diffusion, plane waves are not eigenmodes.
The existence of the quasi-elastic plateaux is the fingerprint of localized fluctuations which, for small inelasticity, propagate and are damped less than exponentially. A small $\alpha$ determines a rapid locking of the velocities of neighboring elements to a common value, while in the case of $\alpha\to 1$, short range small amplitude disorder persists within the domains, breaking simple scaling of $S^{t,l}$ for large $k$ and having the effect of a self induced noise.
One can characterize such internal noise by means of an average local granular temperature $T_{\sigma}$, i.e. a measure of the variance of ${\bf v_i}$ with respect to the local average of ${\bf v}$ within a region of linear size $\sigma$. Obviously, since when $\sigma \to
\infty$ the local average tends to the global (zero) momentum, then $T_{\sigma} \to \epsilon$, as shown in the insets of Fig.\[fig1\]. For $\sigma < L(t)$, instead, $T_{\sigma} < \epsilon $. For quasi elastic systems $T_\sigma$ exhibits a plateau for $1\ll \sigma\ll
L(t)$ that identify the strength of the internal noise. The local temperature ceases to be well defined for smaller $\alpha$ due to the absence of scale separation between microscopic and macroscopic fluctuations in the strongly inelastic regime [@Mcnamara]. The existence of a $L^{-2}(t)k^{-4}$ region in the structure functions is consistent with Porod’s law [@Bray] and is the signature of the presence of vortices, a salient feature of the cooling process. Vortices form spontaneously and represent the boundaries between regions which selected different orientations of the velocities during the quench and are an unavoidable consequence of the conservation laws which forbid the formation of a single domain. With the random initial conditions adopted, vortices are born at the smallest scales and subsequently grow in size by pair annihilation, conserving the total charge. By locating the vortex cores, we measured the vortex density $\rho_v(t)$, which represents an independent measure of the domain growth, and in fact it decays asymptotically ($t \gg t_c$) as an inverse power of time, i.e. $L_v(t)=\rho_v^{-1/2} \propto t^{1/2}$. The vortex distribution turns out to be not uniform for $\alpha$ not too small. Its inhomogeneity is characterized by the correlation dimension $d_2$ : $H(R)=N_v^{-2} \sum_{i<j}\Theta(R-({\bf r_i-r_j }))
\sim R^{d_2}$ where the $r_i$ are the core locations. For $\alpha \to
1$ the vortices are clusterized ($d_2<2$) i.e. do not fill homogeneously the space, whereas at smaller $\alpha$ their distribution is homogeneous ($d_2 \to 2$).
Vortices are not the only topological defects of the velocity fields. In fact we observe shocks, similarly to recent experiments in rapid granular flows[@Swinney]. Shocks have a major influence on the statistics of velocity field, i.e. on the probability distributions of the velocity increments. The probability density function (p.d.f.) of the longitudinal increment $$\Delta_l({\bf R}) = \frac 1N \sum_{\bf i} ({\bf v_{i+ R}} - {\bf v_i})\cdot \frac {{\bf R}}R$$ is shown in Fig.\[fig4\] for $R=1$ (longitudinal velocity gradient) in the main frame, and for $R=40>L(t)$ in the inset. For small $R\ll
L(t)$ the longitudinal increment p.d.f. is skewed with an important positive tail, whereas for $R\gg L(t)$ it turns Gaussian. The distribution of transverse increments $({\bf v_{i+ R}} - {\bf
v_i})\times {\bf \hat R}$, instead, is always symmetric, but non Gaussian distributed for small $R$. A similar situation exists in fully developed turbulence [@multiscal].
To conclude, our model provides a link between the microscopic rules of granular dynamics and its hydrodynamical description. It allows to follow the cooling of a granular material and the build up of velocity correlations, by means of efficient numerical measures of structure factors, two-time correlations and topological defects. The data analysis reveals the presence of vortices, shocks and internal noise and suggests the existence of a scale separation only in the case of quasi elastic systems, which is instead suppressed for large inelasticities.
Even independently from the problem of granular flows, the model represents a simple but unusual phase ordering system. In fact, despite the apparently purely diffusive aspects shown by one-point quantities, it displays anomalous statistics of spatial properties for the order parameter field as witnessed by the velocity gradient p.d.f. and by the structure functions.
[**Acknowledgments.**]{} We thank E. Caglioti, M. Cencini, L.F. Rull, A. Vulpiani and S. Zapperi for several useful discussions.
H.M. Jaeger, S.R. Nagel and R.P. Behringer, Rev.Mod.Phys. [**68**]{}, 1259 (1996); T. Pöschel, S. Luding (Eds.) [*Granular Gases*]{}, Springer, Berlin (2001).
S. Ulam, Adv.Appl.Math. [**I**]{}, 7-21 (1980).
E. Ben-Naim, P.L. Krapivsky, Phys.Rev. E, [**61**]{} R5 (2000).
N. Sela and I. Goldhirsh, Phys.Fluids [**7**]{}, 507 (1995).
E. Ben-Naim, S.Y. Chen, G.D. Doolen, and S. Redner, [*Phys.Rev.Lett.*]{} [**83**]{} 4069-4072 (1999).
T.P.C. van Noje, M.E. Ernst, R. Brito and J.A.G. Orza, Phys.Rev.Lett. [**79**]{}, 411 (1997).
J.A.G. Orza, R. Brito, M.H. Ernst, [cond-mat/002383]{}.
S.McNamara, Phys.Fluids A 5, 3056 (1993).
The 1-d version of our model compares well with models of hard rods (see ), since in both cases each particle collides only with its neighbors (A.Baldassarri, U. Marini Bettolo and A.Puglisi, in preparation).
A.J. Bray, Adv.Phys. [**43**]{} , 357 (1994).
Even the persistence exponent $\theta$, defined through the decay $N_s \sim t^{-\theta}$ of the number of sites where a velocity component never changed its sign up to time $t$, agrees with the analogous exponent of the diffusion equation (we measured $\theta=0.18$).
R. Benzi, L. Biferale, G. Paladin, A. Vulpiani, and M. Vergassola, Phys.Rev.Lett. [**67**]{}, 2299 (1991).
E. Rericha et al. [cond-mat/0104474]{}.
S. Aumaitre, S. Fauve, S. McNamara and P. Poggi, Eur.Phys.J. B [**19**]{}, 449 (2001).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Neutral atoms have been observed to survive intense laser pulses in high Rydberg states with surprisingly large probability. Only with this Rydberg-state excitation (RSE) included is the picture of intense-laser-atom interaction complete. Various mechanisms have been proposed to explain the underlying physics. However, neither one can explain all the features observed in experiments and in time-dependent Schrödinger equation (TDSE) simulations. Here we propose a fully quantum-mechanical model based on the strong-field approximation (SFA). It well reproduces the intensity dependence of RSE obtained by the TDSE, which exhibits a series of modulated peaks. They are due to recapture of the liberated electron and the fact that the pertinent probability strongly depends on the position and the parity of the Rydberg state. We also present measurements of RSE in xenon at 800 nm, which display the peak structure consistent with the calculations.'
author:
- 'Shilin Hu$^{1*}$, Xiaolei Hao$^{2,*}$, Hang Lv$^{3,4}$, Mingqing Liu$^{5}$, Tianxiang Yang$^{3,4}$, Haifeng Xu$^{3,4,\S}$, Mingxing Jin$^{3,4}$, Dajun Ding$^{3,4}$, Qianguang Li$^{6}$, Weidong Li$^{2}$, Wilhelm Becker$^{7,8,\dag}$, and Jing Chen$^{5,9,\ddag}$'
date:
title: '[Quantum dynamics of atomic Rydberg excitation in strong laser fields ]{}'
---
Introduction
============
In strong-field atomic and molecular physics, Rydberg states attracted much attention in the 1990s [@Freeman1987PRL; @Jones1993; @Muller1988PRL; @Rottke1994PRA] but thereafter have been ignored for a long time. This is because, in an intense laser field, the atoms or molecules are so strongly disturbed that the electrons already in the ground state of the neutral atom or even the ion can be liberated very efficiently [@ap79; @wbecker1; @Becker2012RMP]. This appeared to imply that the weakly bound Rydberg states would lose any significance. Only recently, however, both theoretical and experimental works surprisingly found that quite a large portion of neutral atoms in Rydberg states can survive very strong laser fields [@wang2006CPL; @EichmannPRL08]. This has become the object of elaborate experimental and theoretical studies in the past few years [@Eichmann2009Nat; @NISS2009; @voklova2010JETP; @sandner2013PRL; @ZPIE2017; @Eichmann2013PRL; @huang2013PRA; @Azarm2013JPCS; @Eichmann2012PRA; @lin2014PRA; @Eichmann2015PRL; @Piraux2017PRA; @Popruzhenko2018]. Besides atoms, Rydberg state excitation (RSE) has also been observed in atomic fragments from Coulomb explosion of dimers [@wu2011PRL] and diatomic molecules [@Eichmann2009PRL; @Lv2016PRA].
Theoretically, a semiclassical two-step model has been proposed to explain the experimental observations [@wang2006CPL; @EichmannPRL08; @Eichmann2009Nat; @NISS2009]. At first, the electron is pumped into a continuum state from its initial bound state by the external field via tunneling [@keldysh; @ADK]. The subsequent description of the electron in the continuum follows completely classical lines [@Schafer1993; @Corkum1993PRL]. Owing to the presence of the attractive Coulomb field of the ion, the electron may be left with negative energy at the end of the laser pulse, which corresponds to capture into a Rydberg state [@wang2006CPL], also known as frustrated tunneling ionization (FTI) [@EichmannPRL08]. An alternative theoretical approach proceeds via the solution of the time-dependent one-electron Schrödinger equation (TDSE) [@voklova2010JETP; @lin2014PRA; @ZPIE2017]. This includes, of course, all quantum effects, but it is difficult to extract from it much physical insight into the details of the dynamics. In fact, two mechanisms have been proposed to explain the peak structure in the intensity dependence of the RSE population, which is clearly visible in the TDSE calculation. One is the Freeman-resonance perspective in which the Rydberg states are populated via a resonant multiphoton transition [@voklova2010JETP; @Freeman1987PRL]. The other one suggests that RSE is just the continuation of above-threshold ionization (ATI) to negative energies in the Rydberg quasi-continuum [@lin2014PRA]. On the other hand, closer inspection shows that the peaks in the RSE population as a function of intensity alternate between comparatively low and high yield, which was not addressed in these papers [@voklova2010JETP; @lin2014PRA]. Apparently, the mechanism via Freeman resonances is unable to explain this effect, and the existence of peak structures is clearly beyond the scope of the afore-mentioned semiclassical model.
In this paper, we formulate a quantum-mechanical model of the RSE process and compare its results with experiments and TDSE calculations (which are in a sense almost equivalent to real experimental data). Our quantum model well reproduces all the features observed in a TDSE simulation, including the dependence on the parities of the initial and final states and the just-mentioned fact that the peaks in the RSE intensity dependence alternate in height. The model is also applied to the investigation of RSE of He and Xe atoms. For He, we reproduce the experimentally observed distributions of the principal quantum number of the Rydberg states [@EichmannPRL08; @Eichmann2013PRL]. For Xe, we present to our knowledge the first measurements of RSE for 800-nm laser wavelength that display the peak structure. Our work provides a quantum picture of RSE in intense laser fields: first, the electron is pumped by the laser field into a continuum state. Subsequently, it evolves in the external field. Some of the electrons may be coherently captured into Rydberg states. The probability of the capture process strongly depends on the location and the parity of the final Rydberg state, resulting in all the quantum features observed. A schematic picture of the RSE process is presented in Fig. \[fig1\].\
![(color online). The Rydberg-state excitation process: The dashed gray curve denotes the laser electric field. The valence electron is liberated through tunneling in one of the various optical cycles of the pulse, whereupon it evolves in the laser field. The electrons captured into a certain Rydberg state in one direction (orbit 1 & orbit 3) or the other (orbit 2 & orbit 4) have different phases and will interfere, which leads to a peak structure in the intensity dependence with interval of $\Delta
U_p=\hbar\omega$. Interference between the two directions leads to a peak structure with interval of $\Delta
U_p=2\hbar\omega$ (see text for details).[]{data-label="fig1"}](fig1.eps){width="3.6"}
Quantum model
=============
The transition amplitude of the capture process is given by (atomic units $m=\hbar=e=1$ are used) $$M_{nlm}=\lim_{t\rightarrow\infty, t'\rightarrow -\infty}\big<\Psi_{nlm}(t)|U(t,t')|
\Psi_g(t')\big>, \label{eq1}$$ where $|\Psi_{nlm}\left( t\right) \rangle$ denotes a field-free Rydberg state with principal quantum number $n$, angular-momentum quantum number $l$, and magnetic quantum number $m$. $U(t,t')$ denotes the time-evolution operator with the Coulomb and the laser fields [@wbecker1], and $\Psi_{g}(\mathbf{r},t)=e^{iI_{p}t}\phi_{g}(\mathbf{r})$ indicates the wave function of the field-free ground state with the ionization potential $I_p$. It is assumed that the laser field is turned off in the distant past and future. The time-evolution operator $U(t,t')$ satisfies the Dyson equation $$U(t,t')=U_0(t,t')-i\int_{t'}^{t}d\tau U(t,\tau)H_{I}(\tau)U_{0}(\tau,t'),
\label{eq2}$$ where $U_{0}(t,t')$ is the time-evolution operator for only the Coulomb field [@wbecker1] and (in length gauge) $H_{I}(t)=\textbf{r}\cdot \textbf{E}(t)$. With the help of Eq. (\[eq2\]), the transition amplitude (\[eq1\]) can be rewritten as $$\begin{aligned}
\label{eq3}
M_{nlm}=&\lim_{t\rightarrow\infty}\Big\{\big<\Psi_{nlm}(t)|\Psi_g(t)\big> \nonumber\\
&-i\int_{-\infty}^{t}
d\tau\big<\Psi_{nlm}(t)|U(t,\tau)H_{I}(\tau)|\Psi_g(\tau)\big>\Big\}.\end{aligned}$$ Here, the first term on the right-hand side is zero due to the orthogonality of the Coulomb eigenstates. For the second term, we use a different form of the Dyson equation, $$U(t,t')=U_V(t,t')-i\int_{t'}^{t}d\tau U(t,\tau)VU_V(\tau,t'),
\label{eq4}$$ where $V$ is the binding (Coulomb) potential and $U_V(t,t')$ indicates the Volkov time-evolution operator $$U_{V}(t,t')=\int d^3\mathbf{k}|\Psi_\mathbf{k}^{(V)}(t)\rangle\langle\Psi_\mathbf{k}^{(V)}(t')|,
\label{eq5}$$ which we have expanded in terms of the Volkov states with the wave functions $$\begin{aligned}
\Psi_{\mathbf{k}}^{(V)}(\mathbf{r}, t)=&\frac{1}{ (2\pi)^{3/2}}\exp\Big\{i(\mathbf{k}+\mathbf{A}(t))
\cdot\mathbf{r}\nonumber\\
&-\frac{i}{2}\int^{t}dt'(\mathbf{k}+\mathbf{A}(t^{\prime}))^{2}\Big\},
\label{eq6}\end{aligned}$$ where $\mathbf{k}$ denotes the asymptotic (drift) momentum. The coordinate-space representation of the Volkov time-evolution operator is $$\label{eq7}
U_V(\mathbf{r}t,\mathbf{r}'t')=i\left(\frac{i}{2\pi(t-t')}\right)^\frac32 e^{-iS(\mathbf{r}t,\mathbf{r}'t')}$$ with $$\begin{aligned}
\label{volkovaction}
S(\mathbf{r}t,\mathbf{r}'t')=&\mathbf{A}(t)\cdot\mathbf{r}-\mathbf{A}(t')\cdot\mathbf{r}'+\frac12\int^t_{t'}d\tau\mathbf{A}^2(\tau) \nonumber\\
&-\frac{1}{2(t-t')}\Big[\mathbf{r}-\mathbf{r}'+\int^t_{t'}d\tau\mathbf{A}(\tau)\Big]^2.\end{aligned}$$
Using Eq. (\[eq4\]) in Eq. (\[eq3\]) we get $$\begin{aligned}
\label{eq9}
M_{nlm} =& \lim_{t\rightarrow\infty}(-i)\int_{-\infty}^{t}
d\tau\big<\Psi_{nlm}(t)|U_V(t,\tau)H_{I}(\tau)|\Psi_g(\tau)\big> \nonumber \\
&+\lim_{t\rightarrow\infty}(-i)^2\int_{-\infty}^{t}d\tau'\int^{\tau'}_{-\infty}d\tau \nonumber \\
& \times\big<\Psi_{nlm}(t)|U(t,\tau')VU_V(\tau',\tau)H_{I}(\tau)|
\Psi_g(\tau)\big>.\end{aligned}$$ This is analogous to the Born expansion of the ionization amplitude for above-threshold ionization. In that case, the final state is in the continuum and is approximated by a plane wave so that $\langle\Psi_{nlm}(t)|U_V(t,\tau)$ becomes a Volkov state $\langle\Psi_\mathbf{p}^V(\tau)|$. The first term then describes the direct electrons and the second term (with the exact propagator $U(t,t')$ replaced by the Volkov propagator $U_{V}(t,t')$) indicates single rescattering. For energies below $\approx 2U_p$, the first term is dominant. However, for RSE the first term in Eq. (\[eq5\]) goes to zero in the limit of $t\to\infty$. This is due to the factor of $(t-t')^{-3/2}$ in Eq. (\[eq7\]), which accounts for wave-function spreading, and the fact that the Rydberg state $|\Psi_{nlm}(t)\rangle$ is localized. (Note that the values of $\tau$ in Eq. (\[eq9\]) are restricted to the finite duration of the laser pulse.) Hence, for RSE we only consider the second term.
In order to make further progress we replace the bra $\langle\Psi_{nlm}(t)|U(t,\tau')$ in the second term by an approximate field-dressed Rydberg state with the wave function $$\Psi_{nlm}^d(\mathbf{r},\tau')=\phi_{nlm}(\mathbf{r})e^{-iE_{n}\tau'}e^{i\mathbf{r}%
\cdot\mathbf{A}(\tau')}e^{-i\int_{-\infty}^{\tau'}d\tau
A^{2}(\tau)/2}\label{eq10},$$ where $\phi_{nlm}(\mathbf{r})$ is a field-free Rydberg state corresponding to the energy level $E_{n}=-1/(2n^{2})$. The principal quantum number, angular quantum number, and magnetic quantum number are indicated by $n$, $l$, and $m$, respectively. The field-free Rydberg states are given by $$\begin{aligned}
\label{s2}
\nonumber \phi_{nlm}(\mathbf{r})&= N_{nl}R_{nl}(r)Y_{lm}(\theta,\varphi),\\
\nonumber N_{nl}&=\frac{(2\kappa_n)^{3/2}}{\Gamma(2l+2)}\sqrt\frac{\Gamma(n+l+1)}{2n\Gamma(n-l)},\\
R_{nl}(r)&=(2\kappa_nr)^le^{-\kappa_nr}\ _1F_1(-n+l+1,2l+2,2\kappa_nr),\end{aligned}$$ where $\kappa_n=1/n$, $Y_{lm}(\theta,\varphi)$ is a spherical harmonic function, and $_1F_1(x,y,z)$ is the confluent hypergeometric function. The approximation (\[eq10\]), called the Coulomb-Volkov state, has been frequently used to account for the Coulomb-field in noncontinuum states; see [@Reiss1977; @Jain1978] and many later references. The field-dressed state (\[eq10\]) does not exactly satisfy the Schrödinger equation. Namely, we have $$\begin{aligned}
%
i\frac{\partial}{\partial t}\Psi_{nlm}^d(\mathbf{r},t)=&[E_{n}\phi_{nlm}(\textbf{r})+\frac{A^{2}(t)}{2}\phi_{nlm}(\textbf{r})+\mathbf{r}%
\cdot\mathbf{E}(t)\phi_{nlm}(\textbf{r})] \nonumber \\
&\times e^{-iE_{n}t}e^{i\mathbf{r}\cdot\mathbf{A}(t)}e^{-i\int_{0}^{t}\frac{A^{2}(\tau)}{2}d\tau},
\label{eq12}%\end{aligned}$$ and $$\begin{aligned}
%
& [-\frac{1}{2}\nabla^{2}\!-\!\frac{1}{r}\!+\!\mathbf{r}\cdot\mathbf{E}(t)]\Psi
_{nlm}^d(\mathbf{r},t)\!=\![E_{n}\phi_{nlm}(\textbf{r})\!-\!i\nabla\phi_{nlm}(\mathbf{r})\cdot\mathbf{A}%
(t) \nonumber \\
&+\!\frac{A^{2}(t)}{2}\phi_{nlm}(\textbf{r})\!+\!\mathbf{r}\cdot\mathbf{E}
(t)\phi
_{nlm}(\textbf{r})]e^{-iE_{n}t}e^{i\mathbf{r}\cdot\mathbf{A}(t)}e^{-i\int_{0}^{t}%
\frac{A^{2}(\tau)}{2}d\tau},
\label{eq13}%\end{aligned}$$ so that $$\begin{aligned}
&\left[i\frac{\partial}{\partial t}+\frac12\mbox{\boldmath$\nabla$}^2 + \frac1r - \mathbf{r}\cdot\mathbf{E}(t)\right]\Psi_{nlm}^d(\mathbf{r},t) \nonumber \\
&=ie^{-iE_{n}t}e^{i\mathbf{r}\cdot\mathbf{A}(t)}e^{-i\int_{0}^{t}
\frac{A^{2}(\tau)}{2}d\tau}\mathbf{A}(t)\cdot\mbox{\boldmath$\nabla$} \phi_{nlm}(\mathbf{r}).\label{eq14}\end{aligned}$$ Hence, the approximation (10) would be exact if it were not for the term $-i\nabla\phi_{nlm}(\textbf{r})\cdot A(t)$ on the right-hand side of Eq. (\[eq14\]). A comparison of the four terms of Eq. (\[eq13\]) is shown in Table I. It can be seen that the disturbing term $\nabla\phi_{nlm}(\textbf{r})$ is several orders smaller than the other three terms in the region where the Rydberg state is concentrated (see Figs. 3 and 8 in the appendix) for the times of capture (for details, see the semiclassical analysis part of the appendix) shown in Fig. \[figr1\]. Therefore, Eq. (10) can be considered a good approximation to the dressed Rydberg state, and our final approximation to the Rydberg-capture amplitude is $$\begin{aligned}
M_{nlm}&=\!(-i)^2\!\int^\infty_{-\infty}\!dt\!\int^t_{-\infty}\!dt'\!\langle\Psi^d_{nlm}(t)|VU_V(t,t')H_I(t')|
\Psi_g(t')\rangle \nonumber \\
&=(-i)^2\int_{-\infty}^{\infty}dt\int_{-\infty}^{t}dt^{\prime}\int
d^{3}\mathbf{k}\nonumber\\
& \times\langle\Psi_{nlm}^d( t) \vert V(\mathbf{r}) |\Psi_{\mathbf{k}
}^{(V)}( t) \rangle
\langle\Psi_{\mathbf{k}%
}^{( V) }( t^{\prime}) \vert \mathbf{r'}%
\cdot\mathbf{E}( t^{\prime}) \vert \Psi_{g}(
t^{\prime}) \rangle.
\label{amplitude}\end{aligned}$$
![A typical trajectory of the electron in the laser pulse with intensity of $1\times 10^{14}$ W/cm$^2$. The numbers indicate the instants of time $t$ when the electron is captured by the Rydberg states of $n=6$ of the hydrogen atom with different angular momenta. []{data-label="figr1"}](fig2.eps "fig:")\
$t$ $l$ E$_n(n=6)$ $\frac{\nabla\phi_{nlm}(r)}{\phi_{nlm}(r)}\cdot\mathbf{A}(t)$ $\frac{A^2(t)}{2}$ $\mathbf{r}\cdot\mathbf{E}(t)$
-- ------- ------- -------------- ----------------------------------------------------------------- ---------------------- -------------------------------- --
1 2 -0.014 $-2.47\times10^{-7}$ 0.009 0.20
2 2 -0.014 $-9.91\times10^{-7}$ 0.147 0.35
3 3 -0.014 $-3.62\times10^{-7}$ 0.102 0.44
4 4 -0.014 $-4.16\times10^{-6}$ 0.129 0.75
5 5 -0.014 $2.21\times10^{-6}$ 0.080 -1.45
: \[tab:2\] Relative amplitudes of the four terms of Eq. (\[eq13\]). The quantities $t$ and $l$ indicate the capture moment and the angular momentum number shown in Fig. \[figr1\], respectively.
With the help of Eqs. (\[eq7\]), (\[eq10\]), and the binding potential $V(r)=-1/r$, the capture amplitude (\[amplitude\]) has the following form $$\label{Mn}
M_{nlm}\!=\!(-i)^2\!\int_{-\infty}^{\infty}\!dt\!\int_{-\infty}^{t}\!dt^{\prime}\!\int
d^{3}\!\mathbf{k}\!V_{nlm,\mathbf{k}}\!V_{\mathbf{k}g}\!\exp[iS_{n}(t,t^{\prime
},\mathbf{k})],$$ with $$\begin{aligned}
V_{\mathbf{k}g}&=\!\frac{1}{(2\pi)^{3/2}}\!\int
d^{3}\!\mathbf{r'}\exp\{-i[ \mathbf{k+A}( t^{\prime})
] \cdot\mathbf{r'}\} \mathbf{r}'\cdot\mathbf{E}(t') \psi_g(\mathbf{r'})\nonumber\\
&=\!\frac{1}{(2\pi)^{3/2}}\!(-i\frac{\partial}{\partial t'})
\int d ^{3}\mathbf{r'}\exp\{-i[ \mathbf{k+A}( t^{\prime})
] \cdot\mathbf{r'}\} \psi_g(\mathbf{r'}). \label{Vgk}\end{aligned}$$ $$\label{Vkn}
V_{nlm,\mathbf{k}}=-\frac{1}{(2\pi)^{3/2}}\int d^3\mathbf{r}\phi^*_{nlm}(\mathbf{r}) \frac{1}{r}\exp(i\mathbf{k}\cdot\mathbf{r}),$$ and the action $$\label{sact}
S_{n}\!\left( t,t^{\prime},\mathbf{k}\right)\!=\!\frac{1}{2}\!\int^{t}_{-\infty}\!d\tau\!\mathbf{A}^{2}\left(\tau\right)\nonumber
-\frac{1}{2}\!\int_{t^{\prime}}^{t}\!d\tau\!\left[ \mathbf{k+A}\left(
\tau\right) \right] ^{2}+E_{n}t+I_{p}t^{\prime}.$$
![Density distribution of the hydrogenic Rydberg state $n = 6$ for different angular quantum numbers $l$. The two numbers in each panel correspond to the principal quantum number $n$ and the angular-momentum quantum number $l$. []{data-label="fig33"}](fig3.eps "fig:")\
In Fig. \[fig33\], the probability densities of the relevant Rydberg states are plotted for $n=6$ and different values of $l$. It can be easily seen that for $l\neq 0$ the Rydberg state has two centers at $\mathbf{r}_{nlm}^+ \equiv
\mathbf{r}_{nlm}$ and $\mathbf{r}_{nlm}^-=-\mathbf{r}_{nlm}^+\equiv -\mathbf{r}_{nlm}$. Accordingly, we decompose the Rydberg-state wave function as $$\label{rr}
\phi_{nlm}(\mathbf{r})=\phi_{nlm+}(\mathbf{r})+\phi_{nlm-}(\mathbf{r}),$$ where the functions $\phi_{nlm\pm}(\mathbf{r})$ are concentrated around $\mathbf{r}=\pm\mathbf{r}_{nlm}$. Also, $\phi_{nlm-}(-\mathbf{r})=(-1)^l\phi_{nlm+}(\mathbf{r})$, which allows us to write
$$\begin{aligned}
\label{vnknew}
V_{nlm,\mathbf{k}} & = -\frac{1}{(2\pi)^{3/2}}\int \frac{d^3\mathbf{r}}{r}\left[\phi^*_{nlm+}(\mathbf{r})e^{i\mathbf{k}\cdot\mathbf{r}}+\phi^*_{nlm-}(\mathbf{r})e^{i\mathbf{k}\cdot\mathbf{r}}\right] \nonumber \\
& = -\frac{1}{(2\pi)^{3/2}}\int \frac{d^3\mathbf{r}}{r}\left[\phi^*_{nlm+}(\mathbf{r})e^{i\mathbf{k}\cdot\mathbf{r}}+\phi^*_{nlm-}(-\mathbf{r})e^{-i\mathbf{k}\cdot\mathbf{r}}\right] \nonumber \\
&=-\frac{1}{(2\pi)^{3/2}}\int \frac{d^3\mathbf{r}}{r}\phi^*_{nlm+}(\mathbf{r})\left[e^{i\mathbf{k}\cdot\mathbf{r}}+(-1)^l e^{-i\mathbf{k}\cdot\mathbf{r}}\right].\end{aligned}$$
With the abbreviation $\phi_{nlm+}(\mathbf{r})=\phi_{nlm+}(\mathbf{r}-\mathbf{r}_{nlm}+\mathbf{r}_{nlm})\equiv\tilde{\phi}_{nlm}(\rho)$, where $\rho=\mathbf{r}-\mathbf{r}_{nlm}$, we end up with $$\begin{aligned}
V_{nlm,\mathbf{k}}&=-\frac{1}{(2\pi)^{3/2}}\int d^3\bm{\rho}
\frac{1}{|\bm{\rho}+\mathbf{r}_{nlm}|}\tilde{\phi}_{nlm}^*(\bm{\rho})\Big[e^{i\mathbf{k}\cdot\bm{\rho}}e^{i\mathbf{k}\cdot\mathbf{r}_{nlm}}\nonumber \\
&+(-1)^le^{-i\mathbf{k}\cdot\bm{\rho}}e^{-i\mathbf{k}\cdot\mathbf{r}_{nlm}}\Big].\label{vnknew'}\end{aligned}$$ We can now proceed with the saddle-point evaluation as it is usually done by including the fast exponential dependence generated by the factors $\exp(\pm i\mathbf{k}\cdot\mathbf{r}_{nlm})$ into the action while treating the remaining $\mathbf{k}$ dependence as slow. This means that we replace $$\begin{aligned}
V_{nlm,\mathbf{k}}e^{iS_n(t,t',\mathbf{k})}&=V_{nlm,\mathbf{k}+}e^{i\big[S_n(t,t',\mathbf{k})+\mathbf{k}\cdot\mathbf{r}_{nlm}\big]} \nonumber \\
&+(-1)^lV_{nlm,\mathbf{k}-}e^{i\big[S_n(t,t',\mathbf{k})-\mathbf{k}\cdot\mathbf{r}_{nlm}\big]},\end{aligned}$$ where $$\label{vnk1}
V_{nlm,\mathbf{k}\pm}=-\frac{1}{(2\pi)^{3/2}}\int d^3\bm{\rho}\tilde{\phi}^*_{nlm}(\bm{\rho}) \frac{1}{|\bm{\rho}+\mathbf{r}_{nlm}|} e^{\pm i\mathbf{k}\cdot{\bm{\rho}}}.$$
Then, we search for those values of the variables $t$, $t'$, and $\mathbf{k}$ that render the action $S_n\pm\mathbf{k}\cdot\mathbf{r}_{nlm}$ stationary, which yields, in turn, $$\begin{aligned}
[\mathbf{k}+\mathbf{A}(t')]^2/2 & = -I_p \label{sadd1}, \\
[\mathbf{k}+\mathbf{A}(t)]^2/2 & =A^2(t)/2+E_n \label{sadd2}, \\
\mathbf{k}_{\textrm{st}}^{\pm}= -\frac{1}{t-t'}\int^t_{t'} & d\tau\mathbf{A}(\tau) \pm \frac{\mathbf{r}_{nlm}}{t-t'} = \mathbf{k}_0 \pm \frac{\mathbf{r}^+_{nlm}}{\tau}, \label{sadd3}\end{aligned}$$ where $\mathbf{k}_0=-\frac{1}{t-t'}\int^t_{t'}d\tau\mathbf{A}(\tau)$ and $ \tau=t-t'$. Equations (\[sadd1\]) and (\[sadd2\]) describe, respectively, energy conservation in the tunneling process and in the capture process, while Eq. (\[sadd3\]) determines the intermediate electron momentum. The latter takes into account that the electron is captured at one of the two positions $\pm\mathbf{r}_{nlm}$. The solutions $(t,t',\mathbf{k})$ of Eqs. (\[sadd1\])–(\[sadd3\]) define the quantum orbits.
Throughout the paper, we will refer to Eq. (\[amplitude\]) as the quantum model (QM), and the multiple integrals are performed by numerical integration with respect to $t$ and $t^{\prime}$ and by saddle-point integration with respect to $\textbf{k}$. At the end of the pulse, the population of the $n$th Rydberg state is defined as $P_n=\sum_{l,m}|M_{nlm}|^2$. In our simulation, magnetic quantum number $m=0$ and (field-free) energy $E_n=-1/(2n^2)$ are adopted. The ten-cycle laser electric field is $\mathbf{E}(t)=E_{0}\sin\omega
t\mathbf{\hat{e}_z}$ with the vector potential $\mathbf{A}(t)=E_{0}/\omega\cos\omega t\mathbf{\hat{e}_z}$ ($\mathbf{\hat{e}_z}$ is a unit polarization vector and the wavelength $\lambda = 2\pi c/\omega= 800$ nm). The 1$s$ atomic orbital is expressed as $\phi_{1s}(\mathbf{r})=\frac{1}{\sqrt{\pi}} \kappa^{3/2}e^{-\kappa
r}$ with $\kappa=\sqrt{2I_{p}}$. For Xe(5$p$), $\phi_{5p}(\mathbf{r})
=\frac{(2\xi)^{11/2}}{\sqrt{10!}}\sqrt{\frac{3}{4\pi}}
r^{4}e^{-\kappa r}\cos\theta$ with $\xi=0.65\sqrt{2I_{p}}$ and $\kappa=0.94\sqrt{2I_{p}}$. The ionization potential of the H, He, and Xe atoms are 0.5 a.u., 0.9 a.u., and 0.44 a.u., respectively. For the TDSE simulation, the details can be found in Ref. [@lin2014PRA].
Results and discussions
=======================
**I. Comparison of TDSE and model calculations**\
![ (color online). Populations of Rydberg states with even or odd parity versus laser intensity calculated via TDSE (a) and the quantum model (QM)(b) for an initial 1$s$ state. The positions of the $N$-photon channel closings are indicated on the top $x$ axis. For visual convenience, the peaks are separated by dashed lines and underlaid with alternatingly gray or pink backgrounds. Total Rydberg-state populations on a linear scale (sum over parities) as functions of the laser intensity simulated by (c): the TDSE and (d): the QM for an initial 1$s$ state.[]{data-label="fig2"}](fig4.eps){width="3.6"}
In Figs. \[fig2\](a) and \[fig2\](b), we show the probability of Rydberg states with principal quantum numbers $n>3$ and with opposite parities of the final Rydberg states ($l$ even or odd) calculated by the TDSE and the quantum model for an initial $1s$ state of the H atom as functions of the laser intensity. (For the TDSE, the pulse envelope is given by a cosine square function with full width at half maximum of 10 fs, and the details of the simulation can be found in [@lin2014PRA]). Peaks separated by an intensity interval of about 25TW/cm$^{2}$, corresponding to a shift of the ponderomotive energy $U_p$ by the energy of a laser photon, i.e., $\Delta U_{p}=\hbar\omega$ ($\approx 1.55$ eV), can be clearly seen in Fig. \[fig2\] for both the TDSE and the quantum model calculation. The results of the quantum model are almost quantitatively consistent with the TDSE results except for a small shift of the peaks. In particular, for initial states and Rydberg states with specified parities, two consecutive peaks are separated by about $\Delta U_{p}=2\hbar\omega$. These obvious quantum features can be easily understood in the multiphoton transition picture of RSE [@voklova2010JETP; @Freeman1987PRL]. Since the energies of the Rydberg states are near the threshold so that their Stark shifts are all close to the Stark shift $U_p$ of the continuum, multiphoton resonance between the Rydberg and the initial states occurs at intensities that are separated by $\Delta
U_{p}=\hbar\omega$ [@Freeman1987PRL]. In addition, the dipole selection rule only allows even-order (odd-order) multiphoton transitions between states of equal (opposite) parity, which gives rise to a $\Delta U_{p}=2\hbar\omega$ separation between consecutive peaks for a transition between two states with specified parity [@lin2014PRA; @Piraux2017PRA]. Apparently, these features are beyond the scope of the semiclassical picture of the RSE.
Furthermore, closer inspection shows that the peaks alternate in height as shown in Figs. \[fig2\](c) and \[fig2\](d), which display the population of the Rydberg states with both parities. This can also be observed in the results of Refs. [@voklova2010JETP; @lin2014PRA; @ZPIE2017] though it was not addressed there. This feature is difficult, if not impossible, to understand in the multiphoton-resonance picture. For a transition from the ground state to the Rydberg state, the electron has to absorb more than ten photons under typical laser conditions as considered here ($\hbar\omega =1.55$ eV and $U_p=6$ eV for $I=1\times 10^{14}$ W/cm$^2$). The density of the Rydberg states does not depend on the parity. Hence, no mechanism such as a selection rule can give rise to a structure that depends on whether an even or odd number of photons is absorbed in the process.\
**II. Analysis of the quantum model**\
In order to elucidate these intriguing features, we classify the quantum orbits in our model according to the length of their travel time, that is, the time difference $\Delta t
=t-t'$ between recapture and ionization. Specifically, with the number $J$ we denote the pair of orbits where $(J-1)T/2<\Delta t
<JT/2$ where $T=2\pi/\omega$. We call an orbit even or odd according to whether $J$ is even or odd. Figure \[fig3\](a) shows the results calculated using Eq. (\[amplitude\]) for the Rydberg state $n=6$ and $l=5$ for different electronic orbits. We take the Rydberg state $n=6$, since for this state the calculated results show a maximum in our laser intensity range [@NISS2009; @lin2014PRA]. We notice that the intensity dependence is a straight line for the orbit with $J=20$. For $J=18$, an oscillatory structure is beginning to develop, which becomes more and more apparent when $J$ becomes smaller (not shown here). Finally, for $J=1$ and 2, pronounced peaks have emerged. This is because the total pulse duration used in our calculation is $10\,T$. Namely, for the orbit with $J=20$ the electron must be freed in the earliest half cycle while for orbits with smaller $J$ electrons liberated during more and more subsequent half cycles contribute and interfere with each other, which results in the peak structure. Hence, the peak structure in the intensity dependence of the RSE can be attributed to interference between electron wave packets generated in different optical half cycles during the laser pulse, which is in agreement with Ref. [@ZPIE2017].
![ (color online). (a): Interference pattern generated by orbits with different values of $J$ as given (see text for details). (b): The population of RSE for different $l$ for a fixed principal quantum number $n=6$ in an 800-nm laser field. (c): The electron trajectory in the field $\mathbf{E}(t)=E_0\sin\omega t
\mathbf{e}_z$ corresponding to ionization at the phase $\omega
t_0=92^\circ$*after* the field maximum. The laser intensity is $1.14\times 10^{14}$ W/cm$^2$ and the transverse velocity is 0 a.u. The colored horizontal dashed lines indicate the boundaries (in the coordinate $z$) of those regions where the Rydberg states with $n=6$ and $l\le 5$ have their highest density as shown in Fig. 8 below. (d): The populations of specific $l$ within the semiclassical picture for $n=6$ at four different peak intensities (see the text for details).[]{data-label="fig3"}](fig5.eps){width="3.6"}
For a closer look, we consider Fig. \[fig1\]. For a sufficiently long (i.e., almost perodic) laser pulse, the ionization dynamics triggered by this pulse repeat themselves from one cycle to the next (see, e.g., orbits 1 and 3). Following a similar analysis in Ref. [@Milosev2007], if a saddle point solving Eqs. (\[sadd1\]) - (\[sadd3\]) with the Rydberg-state recombination site in the positive direction \[+ sign in Eq. (\[sadd3\])\] is given by $(\mathbf{k},t,t')$, then the saddle point in the next optical cycle with the same Rydberg state in the same direction is $(\mathbf{k},\,t+T,\, t'+T)$, and the corresponding action is related by $$S_n(t+T,t'+T,\mathbf{k})=S_n(t,t',\mathbf{k})+(E_n+I_p+U_p)T.\label{PD1}$$ So the contributions of these trajectories differ by the phase $(U_p+I_p+E_n)T=(U_p+I_p+E_n)2\pi/\omega$ (orbits 1 and 3). The same holds true for orbits 2 and 4, half a period later when the field has changed sign. Due to parity symmetry, a Rydberg state (except for $l=0$) has two opposite density maxima in the field direction, and the electron can be captured into one or the other. In the figure, orbits 1 and 3 go into one direction and orbits 2 and 4 into the other. If a saddle point solving Eqs. (\[sadd1\]) - (\[sadd3\]) with the Rydberg-state recombination location in the positive direction \[+ sign in Eq. (\[sadd3\])\] is written as $(\mathbf{k},t,t')$, then the saddle point solving these equations with the same recombination location in the negative direction \[- sign in Eq. (\[sadd3\])\] is $(-\mathbf{k},\,t+T/2,\, t'+T/2)$, and the corresponding action is given by $$S_n(t+T/2,t'+T/2,-\mathbf{k})=S_n(t,t',\mathbf{k})+(E_n+I_p+U_p)T/2.\label{PD}$$ So the contributions of 2 and 4 differ from those of 1 and 3 by the phase $(U_p+I_p+E_n)T/2=(U_p+I_p+E_n)\pi/\omega$. The interference of the contributions of orbits 1 and 2 with those of orbits 3 and 4, i. e., the contributions of two directions, gives rise to peaks in the RSE separated by $\Delta U_p=2\hbar\omega$, while the afore-mentioned periodicity from cycle to cycle yields peaks with the separation of $\Delta U_p=\hbar\omega$, which is common for all interferences in a field with period $T$. These two kinds of interference can be clearly seen in Fig. \[fig3\](a).
Another element to be considered are the phases due to the parities of the initial ground state and the final Rydberg state, which are $(-1)^p$ with $p=\pm 1$ the respective parities. Hence, if the initial and the final state have opposite parities, an additional phase of $\pi$ occurs. We notice that the interference between quantum orbits that are recaptured in opposite directions of the same Rydberg state plays the same role in the quantum model as the selection rule does in the multiphoton transition picture.
It is noteworthy that the peaks in Figs. \[fig2\](b) and \[fig2\](d) also closely satisfy the channel-closing (CC) condition $I_p+U_p=m\omega$. For example, the two peaks at $1.16\times
10^{14}$ W/cm$^2$ and $1.42\times 10^{14}$ W/cm$^2$ in Fig. \[fig2\](b) correspond to $U_p=6.9$ and 8.5 eV and, with the ionization potential $I_p=13.6$ eV and the negligible ionization potential of the highly excited Rydberg state, satisfy the CC condition for $m=13$ and $m=14$, in agreement with the Freeman resonance picture and the selection rule.
From the above analysis, as schematically illustrated in Fig. \[fig1\] and elaborated above, we find that there are essentially two types of interference of the quantum orbits in the excitation of a Rydberg state with specific $n$ and $l$ from the ground state: i) interference of wave packets released during different optical cycles and captured in one and the same half of the spatial region of the Rydberg state gives rise to peaks separated by a laser intensity difference corresponding to $\Delta
U_p=\hbar\omega$; ii) interference of orbits released in adjacent half cycles (and captured in opposite spatial regions of the Rydberg state) results in a peak structure with intensity difference corresponding to $\Delta U_{p}=2\hbar\omega$, and the peak positions depending on the relative parity of the Rydberg and the ground states.
Figure \[fig3\](b) displays the population of different orbital angular momenta $l$ for $n=6$. In the intensity range from $1\times 10^{14}$ W/cm$^2$ to $2.5\times 10^{14}$ W/cm$^2$ the Rydberg states with $l=4$ and 5 are predominantly populated. The maximal populations of $l=4$ and 5 as a function of laser intensity correspond to the low and high peaks in Figs. \[fig2\](b) and \[fig2\](d), respectively.
For a different view of these features, we now turn to a semiclassical picture (see the Appendix for the details). In Fig. \[fig3\](c), we present an example of an electron trajectory for ionization at $\omega t_0=92^\circ$ with zero initial longitudinal and transverse velocity. We assume that the probability of capture into the Rydberg state $(n,l,m)$ is maximal if the electron trajectory $(X(t),Z(t))$ passes the spatial region where its density $|\phi_{nlm}(\mathbf{r})|^2$ is highest at a time where its kinetic energy is very low. We determine this spatial region from a graph of the Rydberg-state wave function (see Fig. 8 of the Appendix). For the kinetic energy we require $E_\mathrm{kin}<0.05$ a.u., and there are two time domains which satisfy the capture condition for every optical cycle (see Fig. 7 of the Appendix). Figure \[fig3\](d) displays the angular-momentum distributions obtained this way. Clearly, regardless of intensity $l=5$ dominates all other angular momenta, which is in reasonable agreement with the results of the QM model in Fig. \[fig3\](b). In addition, it can be seen that the population of $l=5$ for 166 TW/cm$^2$ is higher than that of $l=4$ for 192 TW/cm$^2$ and also for 140 TW/cm$^2$, which confirms the alternating heights of the peaks in the quantum model displayed by Fig. \[fig3\](b). Therefore, the alternating heights of the RSE peaks shown in Fig. \[fig2\] are already engrained in this “simple-man model”. It should be noted that even though the population at 114 TW/cm$^2$ is higher than that of 140 TW/cm$^2$ in Figs. \[fig3\](b) and \[fig3\](d), this is not the case for the total population shown in Figs. \[fig2\](a) and \[fig2\](b). This can be attributed to the fact that the relative contributions from other Rydberg states with different $n$ change and affect the modulation when the intensity is low.\
**III. Comparison with experiments**\
Experiments show that for the He atom the population of the Rydberg states peaks at about $n=9$ at $I=1.8\times 10^{15}$ W/cm$^2$ with a slight tendency to shift to higher $n$ when the intensity increases to $I=2.9\times 10^{15}$ W/cm$^2$ [@Eichmann2015PRL]. For comparison, the focal average of the RSE calculated by our proposed quantum model is shown in Fig. \[fig4\](a) which shows that the maximum of the RSE shifts to higher energy for increasing intensity, in qualitative agreement with the experiment. For fixed-intensity RSE yields, however, our quantum model shows a reverse intensity dependence (see Fig. \[fig4\](b)), which reproduces the results of TDSE simulations [@lin2014PRA]. This apparent conflict can be resolved as follows: According to the physical picture underlying our quantum model, the electron after tunneling out from the ground state oscillates in the laser field while drifting towards the detector. When it reaches the position of the Rydberg state, provided its instantaneous kinetic energy is very low, it can be captured (see the typical orbits in Fig. \[fig1\]). Therefore, for small drift momentum, it can be expected that the RSE yield becomes maximal around principal quantum numbers $n$ roughly given by $n^2\approx
4\sqrt{U_p}/\omega$, so that the yield slightly shifts with increasing intensity as shown in Fig. \[fig4\](a). The positions of the maxima of the RSE yields in the fixed-intensity results of Fig. \[fig4\](b) and Ref. [@lin2014PRA] are determined by interference as was shown above. However, interference effects are largely smeared out by focal averaging, which restores agreement with the TDSE and the semiclassical calculations in Ref. [@Eichmann2015PRL]. It should be noted that this $n$ distribution cannot be explained by the Freeman resonance mechanism.
![ (color online). (a): *Focal-averaged* RSE vs. principal quantum number calculated by the quantum model for the He atom, for $I_0=10^{15}$ W/cm$^2$; (b): Calculated electron spectra below (multiplied by a factor of 0.2) and above the continuum threshold for the hydrogen atom and various fixed intensities; (c): Experimental yields of single ionization and RSE of the Xe atom; (d): Experimental and calculated ratios between the Rydberg excitation and the single ionization yields (see the text for more detail). []{data-label="fig4"}](fig6.eps){width="3.6"}
Moreover, we experimentally measured the single-ionization yields and the RSE of the Xe atom for principal quantum numbers between $n=20$ and $n=30$. The experimental setup is introduced in Ref. [@Lv2016PRA] (see the Appendix for more details). The results for Xe are presented in Figs. \[fig4\](c) and \[fig4\](d). The intensity-dependent ionization yield follows a smooth curve. However, the RSE yield does display some peak structure especially in the low intensity regime below $I=1.5\times 10^{14}$ W/cm$^2$ \[see Fig. \[fig4\](c)\]. This is especially evident in the ratio Xe$^*$/Xe$^+$, which exhibits peaks at $I_1=0.89\times 10^{14}$ W/cm$^2$, $I_2=1.18\times
10^{14}$ W/cm$^2$, and $I_3=1.45\times 10^{14}$ W/cm$^2$ \[see the blue arrows in Fig. \[fig4\](d)\]. This structure is qualitatively reproduced by the quantum model if focal averaging is included, indicating that this structure is the remainder of the peak structure after taking into account the intensity distribution in the focus.\
Conclusions and Perspectives
============================
In summary, we propose a new quantum model based in the spirit of the SFA to study the RSE process in an intense infrared laser field. In this quantum model, the electron is first pumped by the laser field into the continuum where subsequently it evolves in the laser field. Most of the liberated electrons end up as free electrons (ATI); however, some may be captured into a Rydberg state. The electronic wave packets liberated in different optical half cycles are responsible for peak structures in the intensity dependence of the RSE yield. The peaks exhibit a modulation, which can be attributed to a strong dependence of the capture probability on the spatial position and the parity of the Rydberg state. Our calculation also well reproduces the experimentally observed atomic RSE distribution and the intensity dependence of RSE of atoms. The quantum picture of the RSE in an intense laser field can be understood as a coherent recapture process accompanying above-threshold ionization.
We expect similar quantum effects for molecules. For example, RSE of the O$_2$ molecule is suppressed compared with that of its companion atom Xe, and this suppression is stronger than the corresponding suppression of ionization of O$_2$ compared with that of Xe [@Lv2016PRA]. Apparently, this cannot be described by the semiclassical model. Another such example will be two-center interference, which has been accepted to be essential in molecular ionization [@Becker2000PRL; @lin2012PRL]. With some extensions, our quantum model can be applied to investigate these intriguing phenomena and reveal the underlying physics. Work along these lines is in progress.\
Acknowledgment
==============
The authors acknowledge Xiaojun Liu for helpful discussions. This work was supported by the National Key program for S&T Research and Development (No. 2016YFA0401100), NNSFC (Nos. 11804405, 11425414, and 11534004), and Fundamental Research Fund of Sun Yat-Sen University (18lgpy77).
APPENDIX: Semiclassical analysis and experimental technique
============================================================
**Semiclassical analysis**\
In the simpleman picture where the ionic Coulomb potential is ignored, the equation of motion of the electron in the laser field after ionization is $$\tag{A1}
\ddot{\mathbf{r}}(t)=-\mathbf{E}(t). \label{smm}$$ Here the electric field is $\mathbf{E}(t)=E_{0}\sin\omega
t\mathbf{\hat{e}_z}$ with the vector potential $\mathbf{A}(t)=E_{0}/\omega\cos\omega t\mathbf{\hat{e}_z}$ ($\mathbf{\hat{e}_z}$ is a unit polarization vector), and the wavelength is $\lambda =$ 800 nm. The electron trajectory in the laser field starts at the tunnel exit $z_0=-I_p/E(t_0)$ with zero longitudinal and nonzero initial transverse velocity $\mathbf{v}(t_0)=(0,v_0)$. If we integrate the equation of motion, we obtain $$\begin{aligned}
\tag{A2}
dz(t)/dt = & A(t)-A(t_0), \nonumber \\
dx(t)/dt = & v_0. \label{velo} \nonumber\end{aligned}$$ Integrating again we get the trajectory of the electron: $$\begin{aligned}
\tag{A3}
z(t) & =z(t_0)-A(t_0)(t-t_0)+\int^t_{t_0}A(\tau)d\tau \nonumber \\
&=\! -\frac{I_p}{E_0\sin\omega t_0}\!-\!\frac{E_0}{\omega}\cos\omega t_0(t\!-\!t_0)\!+\!\frac{E_0}{\omega^2}(\sin\omega t\!-\!\sin\omega t_0), \nonumber \\
x(t) &= v_0(t-t_0). \label{traj} \nonumber\end{aligned}$$
The semiclassical physical picture of the RSE can be summarized as follows: the electron is released at the time $t_0$ into a continuum Volkov state. Subsequently, it evolves in the laser field and can be captured into the Rydberg state at time $t$ provided (i) it reaches the spatial region where the Rydberg state has a high density (here we use the condition $|\phi|^2 >
0.8|\phi|^2_{\textrm{max}}$) and (ii) its kinetic energy $E_{kin}$ is small (here we use $E_{kin}\leqslant 0.05$ a.u.). It should be noted that the result is not sensitive to these criteria.
![Kinetic energy of the electron during its evolution in a laser pulse of 10 optical cycles. The laser intensity is $1.14\times 10^{14}$ W/cm$^2$, and the initial phase and the transverse velocity are 92$^\circ$ and 0, respectively.[]{data-label="fig22"}](fig7.eps "fig:")\
In Fig. \[fig22\], we show the evolution of the kinetic energy of the electron under the given initial conditions and parameters. It can be seen that there are two time intervals in each optical cycle where the kinetic energy is small enough for capture, i. e., $E_{kin}\leqslant 0.05$ a.u., as defined above.
For Rydberg states of $n=6$ and different values of $l$ (see Fig. \[fig33\]), Fig. \[fig44\] identifies for these $l$ the regions where the density criterion $|\phi|^2 > 0.8|\phi|^2_{\textrm{max}}$ is satisfied. Clearly, the relevant region moves towards larger distances from the origin with increasing $l$.
![The regions in polar coordinates where the Rydberg-state densities $|\phi_{nlm}(\mathbf{r})|^2$ are larger than 80 percent of their maximal values for $n=6$ and angular momenta $l\le n-1$. Different colors denote different angular momenta.[]{data-label="fig44"}](fig8.eps "fig:")\
In our calculation, each electron trajectory is released via tunneling at some instant during the laser-pulse duration (a 10-cycle pulse with constant electric amplitude is considered). Its weight, which is dependent on the ionization rate and the initial transverse velocity (from -0.3 a.u. to 0.3 a.u.), is determined as in Ref. [@NISS2009]. Then the electron evolves in the laser field until the end of the laser pulse. If it reaches the regions shown in Fig. \[fig44\] and its kinetic energy is smaller than 0.05 a.u., it is considered to be captured by the Rydberg state $\phi_{nlm}$. The final Rydberg excitation probabilities shown in Fig. \[fig3\](d) of the main body of the paper are obtained from statistic of all the trajectories satisfying the afore-mentioned criteria of capture (about 10$^4$-10$^6$ trajectories for each state depending on $l$) in overall 2$\times10^7$ trajectories.\
**Experimental technique**\
In our experiments, we applied the delayed-static-field-ionization method to ionize the neutral Rydbergs, using a time-of-flight (ToF) mass spectrometer operated under a pulsed-electric-field mode. Experimentally, an effusive atomic or molecular beam through a leak valve interacted with a focused Ti:Sapphire femtosecond laser with a central wavelength of 800 nm and pulse duration of 50 fs. After the direct-ionized ions (Xe$^+$) were pushed away from the detector by an electric field, the remaining high-lying neutral Rydbergs (Xe$^*$) were field-ionized by another electric field with a delay time of typically 1.0 $\mu$s, and were detected by dual micro-channel plates at the end of the flight section of about 50 cm. In the case of detection of Xe$^+$, standard dc electric fields were applied in the ToF mass spectrometer. The voltages in both cases were kept the same to ensure identical detection efficiencies for Xe$^+$ and (Xe$^*$)$^+$. This allows us to detect the neutral Rydberg states with $20<n<30$, estimated by the saddle-point model of static-field ionization $[F=1/(9n^4)]$. The laser-pulse energy was controlled by a half-wave plate and a Glan prism, before being focused into the vacuum chamber with a 250 mm lens. The peak intensity of the focused laser pulse was calibrated by comparing the measured saturation intensity of Xe with that calculated by the ADK model [@ADK]. The scanning of the laser intensity was precisely controlled by simultaneously monitoring the pulse energy using a fast photodiode while rotating the half-wave plate. Each data point was an averaged result of $10^4$ laser shots with an intensity uncertainty of 1 TW/cm$^2$.\
[99]{}
R. R. Freeman, P. H. Bucksbaum, H. Milchberg, S. Darack, D. Schumacher, and M. E. Geusic, Above-threshold ionization with subpicosecond laser pulses, Phys. Rev. Lett. **59**, 1092 (1987).
H. G. Muller, H. B. van Linden van den Heuvell, P. Agostini, G. Petite, A. Antonetti, M. Franco, and A. Migus, Multiphoton ionization of Xenon with 100-fs Laser Pulses, Phys. Rev. Lett. **60**, 565 (1988).
R. R. Jones, D. W. Schumacher, and P. H. Bucksbaum, Population trapping in Kr and Xe in intense laser fields, Phys. Rev. A **47**, R49 (1993).
H. Rottke, B. Wolf-Rottke, D. Feldmann, K. H. Welge, M. Dörr, R. M. Potvliege, and R. Shakeshaft, Atomic hydrogen in a strong optical radiation field, Phys. Rev. A **49**, 4837 (1994).
P. Agostini, F. Fabre, G. Mainfray, G. Petite, and N. K. Rahman, Free-free transitions following six-photon ionization of xenon atoms, Phys. Rev. Lett. **42**, 1127 (1979).
For a review, see, e.g., W. Becker, F. Grasbon, R. Kopold, D. B. Milošević, G. G. Paulus, and H. Walther, Above-threshold ionization: from classical features to quantum effects, Adv. At. Mol. Opt. Phys. **48**, 35 (2002).
W. Becker, X. Liu, P. Ho, and J. H. Eberly, Theories of photoelectron correlation in laser-driven multiple atomic ionization, Rev. Mod. Phys. **84**, 1011 (2012).
B. B. Wang, X. F. Li, P. M. Fu, J. Chen, and J. Liu, Coulomb potential recapture effect in above-barrier ionization in laser pulses, Chin. Phys. Lett. **23**, 2729 (2006).
T. Nubbemeyer, K. Gorling, A. Saenz, U. Eichmann, and W. Sandner, Strong-field tunneling without ionization, Phys. Rev. Lett. **101**, 233001 (2008).
U. Eichmann, T. Nubbemeyer, H. Rottke, and W. Sandner, Acceleration of neutral atoms in strong short-pulse laser fields, Nature **461**, 1261 (2009).
N. I. Shvetsov-Shilovski, S. P. Goreslavski, S. V. Popruzhenko, and W. Becker, Capture into Rydberg states and momentum distributions of ionized electrons, Laser Phys. **19**, 1550 (2009).
E. A. Volkova, A. M. Popov, and O. V. Tikhonova, Ionization and stabilization of atoms in a high intensity, low frequency laser field, Sov. Phys. JETP **140**, 450 (2011).
A. Emmanouilidou, C. Lazarou, A. Staudte, and U. Eichmann, Routes to formation of highly excited neutral atoms in the breakup of strongly driven H$_2$, Phys. Rev. A **85**, 011402(R) (2012).
A. von Veltheim, B. Manschwetus, W. Quan, B. Borchers, G. Steinmeyer, H. Rottke, and W. Sandner, Frustrated tunnel ionization of noble gas dimers with Rydberg-electron shakeoff by electron charge oscillation, Phys. Rev. Lett. **110**, 023001 (2013).
K. Y. Huang, Q. Z. Xia, and L. B. Fu, Survival window for atomic tunneling ionization with elliptically polarized laser fields, Phys. Rev. A **87**, 033415 (2013).
A. Azarm, S. M. Sharifi, A. Sridharan, S. Hosseini, Q. Q. Wang, A. M. Popov, O. V. Tikhonova, E. A. Volkova, and S. L. Chin, Population trapping in Xe atoms, J. Phys. Conf. Ser. **414**, 012015 (2013).
Q. G. Li, X. M. Tong, T. Morishita, H. Wei, and C. D. Lin, Fine structures in the intensity dependence of excitation and ionization probabilities of hydrogen atoms in intense 800-nm laser pulses, Phys. Rev. A **89**, 023421 (2014).
H. Zimmermann, S. Patchkovskii, M. Ivanov, and U. Eichmann, Unified time and frequency picture of ultrafast atomic excitation in strong laser fields, Phys. Rev. Lett. **118**, 013003 (2017).
B. Piraux, F. Mota-Furtado, P. F. O’Mahony, A. Galstyan, and Yu. V. Popov, Excitation of Rydberg wave packets in the tunneling regime, Phys. Rev. A **96**, 043403 (2017).
S. V. Popruzhenko, Quantum theory of strong-field frustrated tunneling, J. Phys. B **51** 014002 (2018).
U. Eichmann, A. Saenz, S. Eilzer, T. Nubbemeyer, and W. Sandner, Observing Rydberg atoms to survive intense laser fields, Phys. Rev. Lett. **110**, 203002 (2013).
H. Zimmermann, J. Buller, S. Eilzer, and U. Eichmann, Strong-field excitation of Helium: bound state distribution and spin effects, Phys. Rev. Lett. **114**, 123003 (2015).
J. Wu, A. Vredenborg, B. Ulrich, L. Ph. H. Schmidt, M. Meckel, S. Voss, H. Sann, H. Kim, T. Jahnke, and R. Dörner, Multiple recapture of electrons in multiple ionization of the Argon dimer by a strong laser field, Phys. Rev. Lett. **107**, 043003 (2011).
B. Manschwetus, T. Nubbemeyer, K. Gorling, G. Steinmeyer, U. Eichmann, H. Rottke, and W. Sandner, Strong laser field fragmentation of H$_2$: Coulomb explosion without double ionization, Phys. Rev. Lett. **102**, 113002 (2009).
H. Lv, W. L. Zuo, L. Zhao, H. F. Xu, M. X. Jin, D. J. Ding, S. L. Hu, and J. Chen, Comparative study on atomic and molecular Rydberg-state excitation in strong infrared laser fields, Phys. Rev. A **93**, 033415 (2016).
M. V. Ammosov, N. B. Delone, and V. P. Krainov, Tunnel ionization of complex atoms and of atomic ions in an alternating electromagnetic field, Sov. Phys. JETP **64**, 1191 (1986).
L. V. Keldysh, Ionization in the field of a strong electromagnetic wave, Zh. Eksp. Teor. Fiz. **47**, 1945 (1964).
K. J. Schafer, B. Yang, L. F. DiMauro, and K. C. Kulander, Above threshold ionization beyond the high harmonic cutoff, Phys. Rev. Lett. **70**, 1599 (1993).
P. B. Corkum, Plasma perspective on strong-field multiphoton ionization, Phys. Rev. Lett. **71**, 1994 (1993).
H. D. Jones and H. R. Reiss, Intense-field effects in solids, Phys. Rev. B **16**, 2466 (1977).
M. Jain and N. Tsoar, Compton scattering in the presence of coherent electromagnetic radiation, Phys. Rev. A **18**, 538 (1978).
D. B. Milošević, E. Hasović, M. Busuladžić, A. Gazibegović-Busuladžić, and W. Becker, Intensity-dependent enhancements in high-order above-threshold ionization, Phys. Rev. A **76**, 053410 (2007).
J. Muth-Böhm, A. Becker, and F. H. M. Faisal, Suppressed molecular ionization for a class of diatomics in intense femtosecond laser fields, Phys. Rev. Lett. **85**, 2280-2283 (2000).
Z. Lin, X. Jia, C. Wang, Z. Hu, H. Kang, W. Quan, X. Lai, X. Liu, J. Chen, B. Zeng, W. Chu, J. Yao, Y. Cheng, and Z. Xu, Ionization suppression of diatomic molecules in an intense midinfrared laser field, Phys. Rev. Lett. **108**, 223001 (2012).
| {
"pile_set_name": "ArXiv"
} |
‘@=11 ‘@=12
The standard model (SM) of electroweak interactions and many of its extensions generically predict the existence of Higgs bosons. Detecting Higgs bosons and studying their properties in future collider experiments would provide crucial information for the mechanism of electroweak symmetry breaking and hopefully fermion flavor physics as well. These have been the most prominent issues in contemporary particle physics.
The upgraded Fermilab Tevatron will start its mission next year with c. m. energy $\sqrt s=2$ TeV and an annual luminosity $L\approx 2~\fbi$ per detector (Run IIa). Ultimately, one would hope to reach an integrated luminosity of $L\approx 15-30~\fbi$ (Run IIb). In terms of the search for the SM Higgs boson ($h$), the most promising processes beyond the LEP2 reach would be electroweak gauge boson-Higgs associated production [@scott] $
p\bar p \to W h,\ Zh.
$ The leptonic decays of $W,Z$ provide a good trigger and $h\to b\bar b$ may be reconstructible with adequate $b$-tagging and $b\bar b$ mass resolution, allowing a Higgs boson reach of $m_h\sim 120-130$ GeV [@runii]. For a heavier Higgs boson $m_h\approx 2M_W$, the leading production channel via gluon fusion $gg\to h$ and the relatively clean decay mode $h\to WW^* \to \ell\bar\nu \bar\ell \nu$ may be useful in digging out a weak Higgs boson signal [@tevprl]. It is believed that a SM-like Higgs boson may be observable up to a mass of about 180 GeV at a $3\sigma$ statistical level for $L\approx 25~\fbi$ [@runii]. In the minimal supersymmetric extension of the standard model (MSSM), the mass of the lightest CP-even Higgs boson is bounded by $m_h\lsim 130$ GeV [@mhbound]. When the CP-odd Higgs state ($A$) of the MSSM is heavy $m_A \gsim 2M_Z$, the lightest Higgs boson has SM-like properties and the conclusion for a light SM Higgs boson search remains valid in a large parameter region of the MSSM. The only exception is when $m_A\sim {\cal O}(M_Z)$ and $\tan\beta$ (ratio of the Higgs vacuum expectation values) is large, where the production of $b\bar b h,\ b\bar b A$ is enhanced by $\tan^2\beta$ and $h,A\to b\bar b, \tau\bar \tau$ may be accessible [@mssmh]. At the CERN Large Hadron Collider (LHC) with $\sqrt s=14$ TeV and $L\approx 100-300~\fbi$, one expects to fully cover the range of theoretical interest for the SM Higgs boson, or to discover at least one of the MSSM Higgs bosons [@LHC].
The Higgs sector is the least constrained in theories beyond the SM. It is thus prudent to keep an open mind when studying Higgs physics phenomenologically and experimentally. A particularly important question about the Higgs sector is its role in fermion flavor dynamics, [*i.e.*]{}, the generation of fermion masses and flavor mixings. There have been attempts to explain flavor mixings by a generalized Higgs sector with multiple Higgs doublets. It is argued [@model3] that the fermion flavor mixing structure due to the Higgs coupling at tree level can be of the form, $$\kappa_{ij}\ {\sqrt{m_i m_j} \over v}\ h^0 {\bar \psi_i} \psi_j\,,
\label{coupling}$$ where $i,j$ are generation indices and $v\approx 246$ GeV is the Higgs vacuum expectation value. $\kappa_{ij}$ is a product of the model parameter $\lambda_{ij}$ and the neutral Higgs mixing $\cos\alpha$ [@model3]. Although they are free parameters without a priori knowledge of a more fundamental theory, $\lambda_{ij}$ is naturally order of unity from a model-building point of view and $\cos\alpha=1$ corresponds to no Higgs mixing. Such Higgs-fermion couplings would yield flavor-changing neutral currents, and therefore lead to rich phenomenology [@pheno; @g2; @hdecay; @sher]. However, transitions involving the light generations are naturally suppressed and the largest couplings occur between the third and second generations.
In this Letter we explore the lepton flavor-changing coupling $\kappa_{\mu\tau}$ of a Higgs boson. This is particularly motivated by the favorable interpretation for $\nu_\mu-\nu_\tau$ flavor oscillation from recent atmospheric neutrino experiments [@superK]. If a large mixing between $\nu_\mu$ and $\nu_\tau$ exists as indicated by the Super-K experiment [@superK], then it will necessarily lead to the decay $h \to \mu \tau$. The branching fraction depends on the particular model of the Higgs sector, which can be parameterized by $\kappa_{ij}$. The current constraints on this coupling from low energy experiments are rather weak, giving $\lambda_{\mu\tau}<10$ derived from the muon anomalous magnetic moment [@g2]. Other low energy probes are not expected to be sensitive enough to reach the natural size $\lambda_{\mu\tau}\sim {\cal O}(1)$. The potentially interesting lepton flavor-changing decay modes for a Higgs boson were recently discussed [@hdecay], and their search at a muon collider [@hfactory] was studied [@sher]. In this work, we propose to look for the signal at the upgraded Tevatron and the LHC. The leading production mechanism for a neutral Higgs boson through gluon fusion is $$p p(\bar p)\to ggX\to h X\to \mu\tau X.
\label{process}$$ We find that due to the unique flavor-changing signature and the distinctive kinematics of the signal final state, the Tevatron Run II will have significant sensitivity to such a coupling, making this signal a possible Higgs discovery channel for $m_h \approx 100 - 140$ GeV if $\kmt\sim {\cal O}(1)$. At the LHC, the sensitivity is substantially improved leading to a probe for the coupling to a level of $\kappa_{\mu\tau}\sim 0.15$ and extending the mass coverage to 160 GeV.
0.1in
The dominant decay mode for a SM-like Higgs boson is $h\to b\bar b$ for $m_h<130$ GeV and $h\to WW^*$ for a heavier mass. The partial decay width for $h\to \mu\tau$ is given by $$\Gamma(h\to \mu\tau)={\kappa_{\mu\tau}^2\over 4\pi}
{m_\mu m_\tau\over v^2}\ m_h\,.$$ Here and henceforth $\mu\tau\equiv\mu^-\tau^+ + \mu^+\tau^-$. In comparison to the $\tau^+\tau^-$ mode in the SM, we have In Fig. \[BR\], we show these decay branching fractions versus the Higgs boson mass. The $\mt$ mode is plotted assuming $\kmt=1$, for which BR($h\to \mu\tau$) is at the $1\%$ level. For $\kmt \approx 3$, the $\mt$ mode can be as large as the SM $\tau^+\tau^-$ mode. For $m_h>140$ GeV, this mode dies away quickly due to the opening of the large $WW^*$ mode. This is the primary reason for the limitation to a low Higgs mass ($m_h<140$) at a muon collider [@sher; @hfactory].
In Fig. \[SIG\] we show the total cross section for $gg\to h$ as well as the final states from the $h$ decay versus $m_h$ at the (a) Tevatron and (b) LHC. The production is SM-like as we take $\kappa_{tt}=1$. We normalize our signal cross section to include next-to-leading order QCD corrections [@spira], and use the CTEQ4M distribution functions [@cteq4m]. The scales on the right-hand side give the number of events expected for 4 $\fbi$ at the Tevatron (the 2 $\fbi$ luminosity at the CDF and D0 detectors are combined) and 10 $\fbi$ at the LHC. We see that for the $\mh$ range of $110-140$ GeV and $\kmt=1$, there may be about $10-40$ events produced at the Tevatron and $100-4000$ events at the LHC.
0.1in
The signal final state $\mu\tau$ is quite unique: two flavor-changing charged leptons back-to-back in the transverse plane without much hadronic activity. To estimate the observability of the signal in hadron collider environments, we consider the $\tau$ to decay to an electron or (at least one charged) hadrons, excluding the mode to a muon. We do not require explicit $\tau$ tagging in the analysis. We simulate the detector coverage at the Tevatron (LHC) by imposing some “basic cuts” $$\begin{aligned}
p_T^\mu>20\ {\gev},\ p_T^\pm >10\ {\gev},\ |\eta|<2\ (2.5),
\label{Basic}\end{aligned}$$ where $p_T^\mu\ (p_T^\pm)$ is the transverse momentum for the muon (charged track and other observable hadrons from $\tau$ decay), and $\eta$ is their pseudo-rapidity. We further simulate the detector energy resolutions at the Tevatron [@runii] $$\begin{aligned}
\Delta E_j/E_j &=& 0.8/\sqrt E_j \quad {\rm for\ hadrons},\nonumber\\
\Delta E_e/E_e &=& 0.2/\sqrt E_e \quad {\rm for\ electrons},\end{aligned}$$ and at the LHC [@LHC] $$\begin{aligned}
\Delta E_j/E_j &=& 0.65/\sqrt E_j \oplus 0.05\quad {\rm for\
hadrons},\nonumber\\
\Delta E_e/E_e &=& 0.1/\sqrt E_e
\oplus 0.005\quad {\rm for\ electrons}.\end{aligned}$$ The muon is required to be well isolated and we neglect the $p_T^\mu$ smearing. We finally veto extra jets in the range $$p_T^j > 20\ {\gev},\ |\eta^j|<3$$ to maximally preserve the signal kinematics.
Although the lepton flavor-changing signal is quite spectacular, it is not background-free. The leading SM backgrounds include the Drell-Yan (DY) process $$\begin{aligned}
\label{DY}
pp(\bar p) &\to& Z(\gamma^*) \to \tau^+\tau^-\to \mu\nu_\mu\nu_\tau\ \tau, \end{aligned}$$ and $W^+W^-$ pair production ($WW$) $$\begin{aligned}
\label{WW}
pp(\bar p) &\to& W^+W^- \to \mu\nu_\mu\ \tau \nu_\tau. \end{aligned}$$ The background processes are calculated with the full SM matrix elements at tree level including spin correlations of gauge boson decays. QCD corrections as $K$-factors for the total production rates are also taken into account [@kDY]. With the basic cuts of Eq. (\[Basic\]), the backgrounds turn out to be very large. The results are given by the entries under “basic cuts” in Tables \[tabI\] and \[tabII\] for the Tevatron and LHC, respectively.
There are several distinctive kinematical features for the signal that we can exploit to discriminate it from the backgrounds. First, the missing neutrinos from $\tau$ decay are collimated along the charged track since the $\tau$’s are ultra-relativistic. Thus, for the signal, the missing transverse momentum ($p_T^{miss}$) is along the charged track direction and is essentially back-to-back with respect to the muon $\phi(\mu,\pm)\approx 180^\circ$. This is not the case for the $WW$ background. Secondly, the muons in the signal are stiff $p_T^\mu\sim m_h/2$ as a result of the two-body Higgs decay; while the secondary tracks and hadrons from $\tau$ decay are softer. If we define momentum imbalance $$\Delta p_T=p_T^\mu-p_T^\pm,
\label{Dpt}$$ we expect that it would be positive for the signal if the momentum measurements were perfect. This variable turns out to be very powerful in separating the DY background. We now define the “refined cuts” as $$\phi(\mu,\pm) > 160^\circ,\
\Delta p_T>0,\
p_T^\mu>m_h/5.
\label{refine}$$ The most important aspect for the signal observation is reconstruction of the Higgs boson mass. This is quite feasible for the signal under consideration. This can be done with the following steps: (1) define the missing transverse momentum $p_T^{miss}$ as the imbalance from the observable particles (which is $\Delta p_T$ in Eq. (\[Dpt\]) for the signal case); (2) reconstruct the $\tau$ transverse momentum ${{\vec p}_T}^\tau=
{{\vec p}_T}^\pm + {{\vec p}_T}^{miss}$, and the longitudinal component $p_z^\tau=p_z^\pm(1 + p_T^{miss}/p_T^\pm)$; (3) form the $\mt$ invariant mass $m_{\mt}^2=(p^\mu+p^\tau)^2$. This mass variable should be sharply peaked at $m_h$ for the signal, broadly peaked around $M_Z$ for the DY background, and rather smooth over a large range for the $WW$ background. Indeed, with the proper energy smearing, we find the reconstructed Higgs mass peak within a 5 GeV range. The results are summarized in Tables [\[tabI\]]{} and \[tabII\] for the Tevatron and LHC, respectively. The entries under “refined cuts” give the cross sections including the cuts of Eq. (\[refine\]). The signal-to-background ratio $S/B$ within a 5 GeV window for $m_{\mt}$ is shown next. The last rows illustrate the statistical significance $S/\sqrt{B}$ for the Tevatron with 20 $\fbi$ (CDF and D0 combined) and for the LHC with 10 $\fbi$.
$\sigma$ \[fb\]
----------------------------- ----------------- ----------------- ------------------ ------------------ ------------------
100 110 120 130 140
basic cuts
signal 6.5 5.0 3.6 2.3 1.3
DY
WW
refined cuts
signal 5.5 4.2 3.0 1.9 1.0
DY \[pb\] $7.6$ $6.6$ $5.6$ $4.7$ $3.8$
WW 60 59 58 57 55
$S/B$ ${5.4\over 25}$ ${4.1\over 14}$ ${2.9\over 9.0}$ ${1.9\over 6.4}$ ${1.0\over 4.9}$
$S/\sqrt{B}$ (20 fb$^{-1}$) 4.9 4.9 4.5 3.4 2.0
: Signal $h\to \mu\tau$ and SM background cross sections at the 2 TeV Tevatron for $\mh=100-140$ GeV and $\kmt=1$ after different stages of kinematical cuts. The signal statistical significance $S/\sqrt{B}$ is presented for 20 fb$^{-1}$.[]{data-label="tabI"}
$\sigma$ \[fb\]
----------------------------- ------------------- ------------------ ------------------ ----------------- ----------------- ----------------- ------------------
100 110 120 130 140 150 160
basic cuts
signal 230 200 160 120 69 32 6.6
DY
WW
refined cuts
signal 200 170 130 94 56 26 5.3
DY \[pb\] $48$ $42$ $36$ $30$ $24$ $19$ $14$
WW 700 700 690 680 670 650 630
$S/B$ ${190 \over 160}$ ${160 \over 91}$ ${130 \over 63}$ ${91 \over 47}$ ${54 \over 37}$ ${25 \over 30}$ ${5.1 \over 25}$
$S/\sqrt{B}$ (10 fb$^{-1}$) 47 54 52 42 28 15 3.2
: Signal $h\to \mu\tau$ and SM background cross sections at the 14 TeV LHC for $\mh=100-160$ GeV and $\kmt=1$ after different stages of kinematical cuts. The signal statistical significance $S/\sqrt{B}$ is presented for 10 fb$^{-1}$.[]{data-label="tabII"}
0.1in
So far, for our signal discussion, we have chosen the coupling parameter as $\kmt=1$ for illustration. From a model-building point of view, it is natural for $\kmt$ to be of order unity, while the upper bound from low energy constraint is about 10. Generically, the cross section scales like $\kmt^2$. We explored to what value of this coupling the signal would yield a 3$\sigma$ evidence statistically near the Higgs mass peak. Figure \[kreach\] shows $\kmt$ versus $m_h$ at the (a) Tevatron and (b) LHC for several luminosities. We see that at Run IIa where a luminosity of 4 $\fbi$ is expected combining CDF and D0 data, $\kmt\sim 1.2-1.8$ can be reached for $m_h\lsim 140$ GeV. With a higher luminosity of 30 $\fbi$ per detector, one can reach a coupling of $0.6-0.9$. At the LHC, the sensitivity is significantly improved and a signal for $\kmt\sim 0.15$ would even be observable with 100 $\fbi$. Assuming $\kmt\approx 1$, the reach could go beyond $m_h\approx 160$ GeV, in contrast to the accessible limit $m_h\lsim 140$ GeV at a muon collider [@sher]. Similarly, one can ask how much luminosity is needed to reach a certain level of observation. Note that the statistical significance scales like $S/\sqrt B \sim \kmt^2 \sqrt L$. The results are summarized in Fig. \[lreach\], where a 2$\sigma\ (95\%$ confidence level exclusion), 3$\sigma$ and 5$\sigma$ signals are illustrated versus $m_h$ at the (a) Tevatron and (b) LHC for $\kmt=1$. Due to the large number of signal events near the $m_h$ peak at the LHC (see Table \[tabII\]), the statistical accuracy of determining a coupling $\kmt\sim {\cal O}(1)$ can be at a few percent level with only $L=10\ \fbi$. Note that strictly speaking, all the bounds quoted here apply to the product $\kappa_{tt} \kmt$. We have implicitly assumed $\kappa_{tt}=1$ throughout.
In summary, we have studied the observability for a lepton flavor-changing decay of a Higgs boson $h\to \mu\tau$ at the upgraded Tevatron and the LHC. At the Tevatron, the unique signature may serve as the Higgs discovery channel, yielding a 3$\sigma$ signal for $m_h\sim 110$ GeV and $\kmt\sim 1.2$ with 4 $\fbi$ (CDF and D0 combined), surpassing expectations for Higgs boson searches in the SM and in a large parameter region of the MSSM. The sensitivity will be greatly improved at the LHC, probing as small a coupling as $\kmt\sim 0.15$ or determining $\kmt\sim {\cal O}(1)$ better than a few percent accuracy, and extending the reach to $m_h\sim 160$ GeV, beyond the coverage at a muon collider. [*Acknowledgments*]{}: This work was supported in part by a DOE grant No. DE-FG02-95ER40896 and in part by the Wisconsin Alumni Research Foundation.
-0.2in
[99]{}
A. Stange, W. Marciano and S. Willenbrock, Phys. Rev. [**D49**]{}, 1354 (1994); Phys. Rev. [**D50**]{}, 4491 (1994).
Physics at Run II: Supersymmetry/Higgs workshop http://fnth37.fnal.gov/susy.html, [hep-ph/0010338]{}.
T. Han and R.-J. Zhang, Phys. Rev. Lett. [**82**]{}, 25 (1999); T. Han, A. Turcot and R.-J. Zhang, Phys. Rev. [**D59**]{}, 093001 (1999).
H. Haber and R. Hempfling, Phys. Rev. Lett. [**66**]{}, 1815 (1991); M. Carena, M. Qurios and C. Wagner, Nucl. Phys. [**B461**]{}, 407 (1996); H. Haber, R. Hempfling and A. H. Hoang, Z. Phys. [**C57**]{}, 539 (1997); S. Heinemeyer, W. Hollik and G. Weiglein, Phys. Rev. [**D58**]{}, 091701 (1998); Phys. Lett. [**B440**]{}, 296 (1998); R.-J. Zhang, Phys. Lett. [**B447**]{}, 89 (1998); J.R. Espinosa and R.-J. Zhang, [hep-ph/0003246]{}.
J. Dai, J. Gunion and R. Vega, Phys. Lett. [**B387**]{}, 801 (1996); C. Balazs, J. Diaz-Cruz, H. He, T. Tait and C.-P. Yuan, Phys. Rev. [**D59**]{}, 055016 (1999); M. Carena, S. Mrenna and C. Wagner, Phys. Rev. [**D62**]{}, 055008 (2000).
CERN/LHCC/94–38 (1994); ATLAS Collaboration, TDR, CERN/LHCC/99-44 (1999).
T.P. Cheng and M. Sher, Phys. Rev. [**D35**]{}, 3484 (1987); A. Antaramian, L. Hall and A. Rasin, Phys. Rev. Lett. [**69**]{}, 1871 (1992).
M. Sher and Y. Yuan, Phys. Rev. [**D44**]{}, 1461 (1991); W.-S. Hou, Phys. Lett. [**296**]{}, 179 (1992); M. Luke and M.J. Savage, Phys. Lett. [**307**]{}, 387 (1993); L. Hall and S. Weinberg, Phys. Rev. [**D48**]{}, R979 (1993); D. Chang, W.-S. Hou and W.-Y. Keung, Phys. Rev. [**D48**]{}, 217 (1993); D. Atwood, L. Reina and A. Soni, Phys. Rev. Lett. [**75**]{}, 3800 (1995); Phys. Rev. [**D53**]{}, 1199 (1996); Phys. Rev. [**D55**]{}, 3156 (1997).
S. Nie and M. Sher, Phys. Rev. [**D 58**]{}, 097701 (1998).
J.L. Diaz-Cruz and J.J. Toscano, [hep-ph/9910233]{}.
M. Sher, Phys. Lett. [**B487**]{}, 151 (2000).
Super-Kamiokande Collaboration (Y. Fukuda [*et al.*]{}), Phys. Rev. Lett. [**81**]{}, 1562 (1998); [hep-ex/0009001]{}.
V. Barger, M. Berger, J. Gunion and T. Han, Phys. Rev. Lett. [**75**]{}, 1462 (1995); Phys. Rept. [**286**]{}, 1 (1997).
D. Graudenz, M. Spira and P.M. Zerwas, Phys. Rev. Lett. [**70**]{}, 1372 (1993); M. Spira, A. Djouadi, D. Graudenz and P. M. Zerwas, Nucl. Phys. [**B453**]{}, 17 (1995).
H.L. Lai [*et al.*]{}, Phys. Rev. [**D55**]{}, 1280 (1997).
R. Hamberg, W.L. van Neerven and T. Matsuura, Nucl. Phys. [**B359**]{}, 343 (1991); W.L. van Neerven and E.B. Zijlstra, Nucl. Phys. [**B382**]{}, 11 (1992); J. Ohnemus, Phys. Rev. [**D50**]{}, 1931 (1994).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Mr.'
bibliography:
- 'refs.bib'
nocite: '[@*]'
title: Social Media Analysis for Crisis Informatics in the Cloud
---
macros.tex introduction.tex background.tex problem\_statement.tex system\_requirements.tex approach.tex evaluation.tex results.tex related\_work.tex future\_work.tex discussion.tex conclusion.tex
appendixA.tex
| {
"pile_set_name": "ArXiv"
} |
=cmbx10 at 15pt =cmbx10 at 12pt =cmr8
[Centre de Physique Théorique[^1] - CNRS - Luminy, Case 907]{}
[F-13288 Marseille Cedex 9 - France ]{}
[SPIN EFFECTS IN LEPTON-NUCLEON REACTIONS [^2]]{}
[**Jacques SOFFER**]{}
[**Abstract**]{}
We summarize some theoretical issues, which have been considered of special importance in the discussions of the above working group session.
Number of figures : 1
July 1995
CPT-95/P.3210
anonymous ftp or gopher: cpt.univ-mrs.fr
Introduction
============
In this summary talk we would like to cover the following topics. First we will comment on the interpretation of new results from SLAC and SMC on measurements of the spin-dependent structure functions $g_1(x)$ with different targets. Testing the validity of sum rules is an important issue for perturbative QCD which is also related, in particular, to the determination of the value of the proton spin carried by the strange quarks $\Delta s$. There are some arguments, based on $SU(3)$ breaking effects, which introduce uncertainties, such that $\Delta s$ might turn out to be close to zero. The same conclusion can be obtained by using positivity arguments and the fact that there are very few strange quarks in the proton. Of course this is also related to the total light quark contribution to the proton spin $\Delta\Sigma$, whose value is now, more accurately known, larger than before and getting closer to the naive expectation. However one should keep in mind that all these experiments measure indeed $g_1(x)$, not its first moment, and one should, in the first place, try to understand properly the useful information contained in these $x$-distributions. Next we will recall the importance of the other polarized structure function, so-called $g_2(x)$, its origin, its properties and what we expect to learn from its measurement which is now under way, at the very early stage. Finally, we will review briefly the spin programme which will be undertaken in a few years time at BNL, by using RHIC as a polarized proton-proton collider. We will recall the main motivations of the project and, because the center-of-mass energy will be reaching up to $500$ GeV, it will be possible, for the first time, to test the spin sector of perturbative QCD. We will also indicate that it will allow to pin down, in a unique way, some polarized parton distributions which are not directly accessible in polarized deep inelastic scattering experiments.
Spin content of the nucleon and the $g_1(x)$ structure function
===============================================================
Seven years ago, the first results of the EMC at CERN[@[1]] on the $g_1^p(x)$ structure function on a proton target were rather surprizing, because they uncovered a very serious defect in the Ellis-Jaffe sum rule [@[2]]. This defect was confirmed by recent experiments from SMC and SLAC [@[3]] and has been interpreted as due to a large negative value for $\Delta s$, the contribution of the proton spin carried by the strange quarks and a small value of $\Delta\Sigma=\Delta u +\Delta d+\Delta s$, total contribution of the quark spins. Such a large negative $\Delta s$ was not found by the E142 at SLAC [@[3]] in the measurement of the neutron structure function $g_1^n(x)$ directly with a polarized $He^3$ target, which gives a result consistent with $\Delta s=0$. Of course the mean value of $Q^2$ is smaller than for the earlier EMC and the more recent SMC proton experiments, and one can be tempted to blame non-perturbative effects for the difference between the proton and the neutron cases. One more piece of data coming from SMC and SLAC on deuteron targets [@[3]], leads also to a negative value for $\Delta s$. At this stage it is important to recall that the experimental determination of $\Delta s$ and $\Delta\Sigma$ relies on exact $SU(3)$ flavor symmetry which is a questionable assumption. If one ignores $SU(3)$ flavor breaking, it was shown [@[4]] that there is strong correlation between $\Delta\Sigma$ and $\Delta s$, e.g. $\Delta\Sigma=0.337+0.57\Delta s$ for the proton case, which disappears when one takes into account $SU(3)$ breaking effects. In a recent work [@[5]], one postulates that $F$ and $D$ parameters are related [*only*]{} to the valence quarks and the $SU(3)$ symmetry breaking for the decay $\Sigma^-\to n$ is realized by a suppression of the strange pair production in both $n$ and $\Sigma^-$ seas and described by one parameter $\varepsilon$. In this case one can show [@[4]], by studying the dependence of $\Delta\Sigma$ and $\Delta
s$ on $\varepsilon$, both for proton and deuteron, that $\Delta\Sigma$ remains around $0.3$ and is almost insensitive to the value of $\varepsilon$, whereas $\Delta s$ varies strongly between $-0.1$ and $-0.02$. In a different approach [@[6]] where one reanalyzes hyperon beta decay to extract the $F/D$ ratio, one is assuming that the $SU(3)$ breaking can be evaluated just in terms of mass difference of the baryons. Instead of the value $F/D=0.575\pm 0.016$ generally used, one finds a smaller value with a large uncertainty $F/D=0.49\pm 0.08$. In this case there is no violation of the proton Ellis-Jaffe sum rule and again $\Delta s$ is consistent with zero. Finally there is another independent argument for $\Delta s$ small based on positivity [@[7]], namely one should have $|\Delta s(x)|\leq s(x)$ for all $x$. Using the data on $s(x)$ extracted from charm production in neutrino deep inelastic scattering, one finds that the strange quark distribution is essentially dominated by the Pomeron which goes like $1/x$ and is spin independent. As a result it follows that $\Delta s$ is consistent with zero, but perhaps this analysis should be reconsidered by using more recent CCFR data [@[8]]. To conclude this discussion we think it would be extremely useful to have a direct measurement of $\Delta s$ and any suggestion is very welcome[^3].
\#1\#2[0.7\#1]{}
Fig. 1 - Experimental test of eq.(2) (see text) taken from ref.\[13\]
1truecm
As indicated before, we believe that the $x$-dependence of the $g_1(x)$ structure functions contains relevant information which is somehow washed out when one considers only first moment sum rules[@[11]]. Moreover testing sum rules involves necessarily some uncertain extrapolations of the data due to the limited kinematic range accessible in any experiment. We will now briefly illustrate this point as follows. The $g_1(x)$ structure functions for proton, neutron and deuteron are expressed in terms of three distributions $\Delta u(x),
\Delta d(x),\Delta s(x)$ (here $u$ means $u+\bar u$ etc...). It is well known that $g_1^p(x)-g_1^n(x)$ is independent of $\Delta s(x)$ and similarly one can eliminate $\Delta d(x)$ by considering $4g^p_1(x)-g^n_1(x)$. Let us now assume $\Delta s(x)\equiv 0$ and, as explained in Ref.[@[12]], let us postulate the following simple relation between unpolarized and polarized distributions
$$\Delta u(x)=u(x)-d(x)\ .$$
It implies $$5xg^p_1(x)-4/(2-3\omega_D)xg^d_1(x)=5(F_2^p(x)-F_2^d(x))$$ a simple relationship between two $g_1$ and two unpolarized structure functions $F_2(x)$ directly measured by experiment. We have used the standard relations between deuteron, proton and neutron and for the polarized case and $\omega_D$ is the $D$-state probability in the deuteron. We show in Fig.1 an experimental test of eq.(2). We have tested eq.(2) by using for the l.h.s., the
SLAC data on the $g_1$’s at $Q^2=3GeV^2$ (full circles) and for the r.h.s., the NMC parametrization for $F^p_2$ and $F^d_2$ (full line and dash lines for the estimated errors)[@[14]]. In the small $x$ region we have also included the preliminary SMC data[@[3]] (full squares). The test is indeed very well satisfied and gives, within the present experimental errors, a fairly good support to eq.(1) and $\Delta s(x)\equiv 0$. Moreover if one takes the first moment of both sides of eq.(2) using also ref.[@[15]], one finds for the l.h.s. $0.588\pm 0.054$ and for the r.h.s. $0.587\pm 0.065$ which are in remarkable agreement.
Let us end by making a few remarks on the prospects of the $g_1$ structure functions. First in testing the sum rules, one is relying on the crucial assumption that the measured asymmetry $A_1(x,Q^2)$ is independent of $Q^2$, which allows to rescale the $g_1(x,Q^2)$ obtained at different $Q^2$, to a single $Q^2$ value. This has to be more firmly established. Second, the Bjorken sum rule is now satisfied up to 13%[@[3]] and this important test must be confirmed at a higher level of accuracy. Of course this is connected to the behaviour of $g_1(x)$ in the very small $x$ region which seems to be different for proton and neutron according to the recent SMC data[@[3]]. This puzzling situation must be resolved urgently.
The $g_2$ structure function
============================
In deep inelastic $\mu(e)$ scattering with the lepton beam longitudinally polarized and the target with the spin transverse with respect to the beam direction, one can measure a spin asymmetry which is related to a ”transverse” spin-dependent structure function $g_T(x,Q^2)$. It turns out that $g_T$ has a simple expression in terms of $g_1$ and another structure function, so called $g_2(x,Q^2)$, since we have $g_T=g_1 + g_2$. The basic properties of $g_2$ have been nicely summarized in the contribution of X. Ji[@[16]]. In the simple parton model, as a consequence of helicity conservation, one finds that $g_T$ must vanish and therefore, in the scaling limit, one has $g_2(x)=-g_1(x)$. However from an analysis in terms of operator product expansion (OPE) on the light cone, one finds that the situation is not so simple and $g_2$ can be decomposed into two sets of operators. The first set is twist-2 operators, which are the same as those of the decomposition of $g_1$. The second set is twist-3 operators, which involve quark-gluon correlation functions. Therefore one can write
$$g_2(x,Q^2)=g_2^{WW}(x,Q^2)+\bar g_2(x,Q^2)\ ,$$
where $g^{WW}_2$ and $\bar g_2$ correspond to twist-2 and twist-3 respectively. It was shown that $g^{WW}_2$ is fully determined by $g_1$ because one has[@[17]]
$$g_2^{WW}(x,Q^2)=-g_1(x,Q^2)+\int^{1}_{x}\frac{dy}{y}g_1(x,Q^2),$$
but since there is a priori no theoretical reason to expect $\bar
g_2<<g^{WW}_2$, it is important to measure $g_2$.
Very preliminary results were obtained recently both on proton and deuteron targets at SLAC[@[3]] and, although the errors are large, they seem to indicate that for proton, $\bar g_2$ is small except in a region around $x=0.01$. There is also the Burkhardt-Cottingham (BC) sum rule[@[18]] which says that $$\int^{1}_{0}dx g_2(x,Q^2)=0$$ and this important result has to be checked experimentally. The $Q^2$ dependence of the BC sum rule and its possible violation in the very low $Q^2$ region is an interesting problem which has new implications[@[19]]. In perturbative QCD, the validity of the BC sum rule was checked[@[20]] at order $\alpha_s$ on a quark target of mass $m$ to all orders in $m^2/Q^2$, a result which might help to clarify the nucleon target case. Finally, it is worth recalling the relevance of the second moment of $g_2$ and, in particular, the twist-3 coefficient $d^{(2)}$ for which there are different theoretical predictions[@[16]] which will have to be confronted to future accurate data.
Spin physics at RHIC
====================
Before one can come up with a realistic picture of the nucleon, many fundamental questions remain to be answered and, in particular as we have seen above, in the area of the polarized parton distributions. Polarized deep inelastic scattering provides some valuable insight in this direction, but polarized hadron-hadron collisions at high energy give access to new spin-dependent observables which contain also, in some cases, a far more unique information. A Relativistic Heavy Ion Collider (RHIC) is now under construction at Brookhaven National Laboratory and more than four years ago, it was already realized that one should propose a very exciting physics programme, provided this machine could be ever used as a polarized $pp$ collider. Of course all these considerations relie on the foreseen keys parameters of this new facility, i.e. a luminosity up to $2.10^{32} cm^{-2} sec^{-1}$ and an energy of $50-250$ GeV per beam with a polarization, either longitudinal or transverse, of $70\%$. Since then, the RHIC Spin Collaboration (RSC) has produced a letter of intent[@[21]] and has undertaken several serious studies in various areas, leading to a proposal[@[22]] which is now fully approved. Both detectors STAR and PHENIX of the heavy ion programme will be involved and the first data taking is expected by 1999. Some of the physics topics which will be explored at RHIC have been reviewed in the working group by A. Schäfer[@[23]] who also presented many interesting results from a dedicated Monte Carlo code. The magnitude and the sign of the polarized gluon distribution $\Delta G(x,Q^2)$ has to be determined because it is believed to affect the quark helicity distributions $\Delta q(x,Q^2)$, via the axial anomaly[@[24]], and it is also needed to perform their $Q^2$ evolution. So far nothing is known about $\Delta G(x,Q^2)$ experimentally, but it can be measured by the double helicity asymmetry $A_{LL}$, for example, in direct $\gamma$ production and in jet production. Direct photon production in $pp$ collisions is largely dominated, at leading order, by the Compton diagram $qg\to q\gamma$ and therefore $A_{LL}$ provides an almost direct measurement of $\Delta
G(x,Q^2)/G(x,Q^2)$. The number of subprocess contributing to jet production is larger, because we have $gg$, $gq$ and $qq$ as initial states. However for one-jet production, gluon-gluon scattering dominates the $p_T$ spectrum in the low $p_T$ region, i.e. for $10<p_T<20$ GeV/c at the RHIC energy, and this also allows to extract $\Delta G$. In the large $p_T$ region (i.e. $p_T>50$ GeV/c), the cross section is dominated by quark-quark scattering, so $A_{LL}$ will be driven by $(\Delta u/u)^2$. Another very efficient way to isolate $\Delta q/q$ or $\Delta \bar q/\bar q$ for both $u$ and $d$ quarks, is by measuring the single (or double) helicity parity-violating asymmetry in weak boson production[@[25]]. This is due to the anticipated high luminosity (i.e. $800 pb^{-1}$ in three months running time) which will allow to collect a very copious number of $W^{\pm}$ and $Z$ bosons. In addition the polarized antiquark distribution can be very well measured through $A_{LL}$ in lepton pair production, described by the standard Drell-Yan mechanism and, for example, if $\Delta\bar q=0$ one will observe $A_{LL}=0$.
Finally, it is worth mentioning that with the option of transversely polarized protons at RHIC, we will be able to measure directly, for the first time, the leading twist-2 quark transversity distribution, so called $h_1^q(x)$[@[26]]. Clearly there is one such distribution for each flavor ($u,d,$ etc...) and for both quark and antiquark, but nothing is known experimentally about them. At the level of the parton model, there is a useful bound[@[27]], implied by positivity, which reads
$$q(x)+\Delta q(x)\geq 2|h^q_1(x)|\ .$$
Once more, lepton pair production and $Z$ production are the best hadronic probes for these new transversity distributions which are simply related to the double transverse spin asymmetry $A_{TT}$[@[12]].
[**Acknowledgements**]{}
I am glad to thank the organizers of this Workshop, in particular Madame V. Brisson. It was a pleasure to set up the programme of the Working Group III with E. Hughes and I also thank X. Ji, J. Lichtenstadt, Ph. Ratcliffe and A. Schäfer for their contributions.
[100]{} J. Ashman et al., Phys. Lett. B206 (1988) 364 ; Nucl. Phys. B328 (1989) 1.
J. Ellis and R. Jaffe, Phys. Rev. D9 (1974) 1444.
E. Hughes, these proceedings; R. Voss, these proceedings.
J. Lichtenstadt and H.J. Lipkin, preprint TAUP-2244-95 and WIS-95/15/ Mar-PH; J. Lichtenstadt, these proceedings.
H.J. Lipkin, Phys. Lett. B337 (1994) 157.
B. Ehrnsperger and A. Schäfer, Phys. Lett., B348 (1995) 619.
G. Preparata, P.G. Ratcliffe and J. Soffer, Phys. Lett., B273 (1991) 306 and references therein; P.G. Ratcliffe, these proceedings.
S. Rabinowitz et al., Phys. Rev. Lett., 70 (1993) 134.
L.A. Ahrens et al., Phys. Rev., D35 (1987) 785.
G.T. Garvey, W.C. Louis and D.H. White, Phys. Rev., C48 (1993) 761.
M. Karliner, these proceedings.
C. Bourrely and J. Soffer, Phys. Rev., D51 (1995) 2108; Nucl. Phys., B445 (1995) 341.
C. Bourrely and J. Soffer, preprint CPT-95/P.3224.
P. Amaudruz et al., (New Muon Collaboration), Phys. Lett., B295 (1992) 159.
M. Arneodo et al., (New Muon Collaboration), Phys. Rev., D50 (1994) R1.
X. Ji, these proceedings and references therein.
S. Wandzura and F. Wilczek, Phys. Lett., 72B (1977) 195.
H. Burkhardt and W.N. Cottingham, Ann. Phys., 56 (1970) 453.
J. Soffer and O. Teryaev, Phys. Rev., D51 (1995) 25.
G. Altarelli et al., Phys. Lett., B334 (1994) 187.
G. Bunce et al., Particle World, vol. 3 (1992) 1.
Proposal on Spin Physics using RHIC Polarized Collider, (14 august 1992), experiment R5 approved october 1993.
A. Schäfer, these proceedings.
G. Altarelli and G.G. Ross, Phys. Lett. B212 (1988) 391; A.V. Efremov and O.V. Teryaev, Dubna preprint E2-88-287, Czech. Hadron Symp. 302 (1988).
C. Bourrely and J. Soffer, Nucl. Phys., B423 (1994) 329.
R.L. Jaffe and X. Ji, Phys. Rev. Lett., 67 (1991) 552; X. Ji, Phys. Lett., B284 (1992) 137.
J. Soffer, Phys. Rev. Lett., 74 (1995) 1292.
[^1]: Unité Propre de Recherche 7061
[^2]: Plenary talk presented at the Workshop on Deep Inelastic Scattering and QCD, Paris, April 24-28, 1995 (to be published in the Proceedings).
[^3]: $\Delta s$ can also be measured in elastic $\nu p$ scattering by separating the axial form factor contribution $G_A(Q^2)$ extrapolated to $Q^2=0$. The existing data [@[9]] yields a large negative $\Delta s$ with large errors which are due to several uncertainties in getting $G_A(0)$[@[10]].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $A$ be an amenable separable [$C^*$-algebra]{} and $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with continuous scale. We show that two essential extensions $\tau_1$ and $\tau_2$ of $A$ by $B$ are approximately unitarily equivalent if and only if $$[\tau_1]=[\tau_2]\,\,\, {\rm in}\,\,\, KL(A, M(B)/B).$$ If $A$ is assumed to satisfy the Universal Coefficient Theorem, there is a bijection from approximate unitary equivalence classes of the above mentioned extensions to $KL(A, M(B)/B).$ Using $KL(A, M(B)/B),$ we compute exactly when an essential extension is quasidiagonal. We show that quasidiagonal extensions may not be approximately trivial. We also study the approximately trivial extensions.'
author:
- |
[**Dedicated to Lawrence G. Brown on his 60th birthday**]{}\
Huaxin Lin\
Department of Mathematics\
University of Oregon\
Eugene, Oregon 97403-1222
title: |
Extensions by simple $C^*$-algebras\
– Quasidiagonal extensions [^1]
---
\[section\] \[thm\][Lemma]{} \[thm\][Proposition]{} \[thm\][Definition]{} \[thm\][Corollary]{} \[thm\][Example]{} \[thm\][Problem]{} \[thm\][Remark]{} \[thm\]
[****]{}
The study of $C^*$-algebra extensions of $C(X)$ by compact operators was motivated by the understanding of essentially normal operators on an infinite dimensional Hilbert space. The Brown-Douglas-Fillmore Theory for essentially normal operators gives the classification of essentially normal operators up to unitary equivalence ([@BDF1]). The original BDF-theory quickly developed into [$C^*$-algebra]{} extension theory ([@BDF2] and [@BDF3]) and the KK-theory of Kasparov. Applications of this development can be found not only in operator theory and operator algebras but also in both geometry and non-commutative geometry.
Let $0\to B\to E\to A\to 0$ be an essential extension of $A$ by $B.$ This is determined by a monomorphism $\tau: A\to M(B)/B.$ While $KK$-theory gives the classification of extensions up to stable unitary equivalence, it does not give much information on essential extensions when $B\not ={\cal K},$ where ${\cal K}$ is the compact operators on $l^2.$ The example in 1.1 below shows that a non-trivial extension $\tau$ may have $[\tau]=0$ in $KK^1(A, B).$ Other examples (see \[IIEx\]) show that there may be infinitely many non-equivalent trivial extensions. Extensions by simple [$C^*$-algebra]{}s have been studied in a few special cases (see [@Ln3], [@Ln4], [@Ln5] and [@Lnamj]).
In this paper, we study approximately unitary equivalence classes of essential extensions of separable amenable [$C^*$-algebra]{}s by $\sigma$-unital simple [$C^*$-algebra]{}s. One of the reasons that BDF-theory was successful is that the Calkin algebra $M({\cal K})/{\cal K}$ is simple (and purely infinite). We will restrict ourselves to the case that $M(B)/B$ is simple. It is shown in [@Ln1] and more recently in [@Lnnew] that, for a non-unital and $\sigma$-unital simple [$C^*$-algebra]{} $A\not\cong {\cal K},$ $M(A)/A$ is simple if and only if $A$ has a continuous scale. Furthermore, in [@Lnnew] it is shown that when $M(A)/A$ is simple it is necessarily purely infinite simple.
With the Busby invariant, to study essential extensions of $A$ by $B$ it is sufficient to study monomorphisms from $A$ to $M(B)/B.$ With recent development in classification of simple amenable [$C^*$-algebra]{}s, we know a great deals concerning monomorphisms from one amenable (simple) [$C^*$-algebra]{} to a separable amenable purely infinite simple [$C^*$-algebra]{} (see for example, [@Ph1], [@Lnpro] and [@Lnsemi]). However, $M(B)/B$ is not amenable and we do not assume that $A$ is simple. One of the main results of this article is the following: Two essential extensions are approximately unitarily equivalent if they induce the same element in $KL(A, M(B)/B).$ If furthermore $A$ satisfies the Universal Coefficient Theorem, then there is a bijection between ${\bf Ext}_{ap}(A,B),$ the approximate unitary equivalence classes of essential extensions, and $KL(A, M(B)/B).$
However, unlike the classical case, the zero element in $KL(A,M(B)/B)$ does not in general give an approximately trivial extension. On the contrary, at least in some cases, $[\tau]=0$ in $KL(A,M(B)/B)$ never gives an approximately trivial extension and only when $[\tau]\not=0$ in $KL(A, M(B)/B)$ may approximately trivial extensions occur. To make matters worse, there may not be any essential trivial extensions of $A$ by $B$ even though we can use the above mentioned bijection to classify extensions. This leads us to study quasidiagonal extensions.
Quasidiagonality was defined by P. R. Halmos ([@H]) in 1970. The [$C^*$-algebra]{} version soon appeared. L. G. Brown, R. G. Douglas and P. A. Fillmore ([@BDF2]) first recognized that the study of quasidiagonal extensions might be approached by $K$-theory. They noticed that limits of trivial extensions correspond to the quasidiagonal extensions. L. G. Brown pursued this further in (see [@Br2]). Further developments in the study of quasidiagonality can be found in [@Sa], [@V1] and [@V2]. C. L. Schochet proved that stable quasidiagonal extensions are the same as limits of stable trivial extensions and can be characterized by $Pext(K_*(A), K_*(B))$ if $A$ is assumed to be quasidiagonal relative to $B$ and it satisfies the Universal Coefficient Theorem. These results might lead one to believe that quasidiagonal extensions are the same as limits of trivial extensions in greater generality. However, in this paper we show this fails when $B$ is neither isomorphic to ${\cal K}$ nor purely infinite simple.
One should note that the existence of quasidiagonal extensions implies that $B$ has at least one approximate identity consisting of projections. Suppose that $B$ is a non-unital and $\sigma$-unital simple [$C^*$-algebra]{} with real rank zero, stable rank one and weakly unperforated $K_0(B).$ If $A$ is a quasidiagonal [$C^*$-algebra]{}, then there exists an essential quasidiagonal extension of $A$ by $B.$ This condition is necessary if we assume that $B$ is also a quasidiagonal [$C^*$-algebra]{}. Using $K$-theory and the classification result mentioned above, we give a necessary and sufficient condition for essential extensions to be quasidiagonal for a large class of amenable quasidiagonal [$C^*$-algebra]{}s $A.$ We also give a necessary condition for essential extensions to be approximately trivial for amenable [$C^*$-algebra]{}s which satisfy the UCT. As a consequence, a large class of quasidiagonal extensions are [*not*]{} the limits of trivial extensions.
The essential extensions of a separable amenable [$C^*$-algebra]{} $A$ by $B$ (where $B$ is a non-unital and $\sigma$-unital [$C^*$-algebra]{} with a continuous scale) is proved in this paper to be determined by $KL(A, M(B)/B).$ However, to determine which elements in $KL(A, M(B)/B)$ give an approximately trivial extension is still a difficult task. As mentioned above, for example, the zero element in $KL(A,M(B)/B)$ does not usually give an approximately trivial extension. When $A$ is stably finite, both $K_0(A)$ and $K_0(M(B))$ have nice order while $K_0(M(B)/B)$ has no useful order. Even if we know which [homomorphism]{} $\beta: K_0(A)\to K_0(M(B)/B)$ can be lifted to a [homomorphism]{} from $K_0(A)$ to $K_0(M(B)),$ the lifting may not be positive. In this paper, at least for some special cases, we give a precise condition for an element in $KL(A, M(B)/B)$ to be represented by approximately trivial extensions.
This paper is organized as follows.
Section 1: Preliminaries\
This section is a preparation for the rest of the paper which contains a computation of $K$-theory for $M(B)$ and $M(B)/B$ for $\sigma$-unital simple [$C^*$-algebra]{} with real rank zero, stable rank one, weakly unperforated $K_0(B)$ and with a continuous scale. We also point out that $M(B)/B$ is simple (and purely infinite) if and only if $B$ has a continuous scale (if $B\not\cong {\cal K}$).
Section 2: Monomorphisms from $A\otimes {\cal O}_2$ into a purely infinite simple [$C^*$-algebra]{}\
This section studies [homomorphism]{}s from $A\otimes {\cal O}_2$ into a purely infinite simple [$C^*$-algebra]{}.
Section 3: Approximately unitarily equivalent extensions\
We show that, if $B$ is a non-unital and $\sigma$-unital simple [$C^*$-algebra]{} with a continuous scale, two monomorphisms from $A$ to $M(B)/B$ are approximately unitarily equivalent if and only if they induce the same element in $KL(A, M(B)/B).$
Section 4: ${\bf Ext}_{ap}(A,B)$\
In this section, under the assumption that $A$ satisfies the UCT, we give a bijection $\Gamma: {\bf Ext}_{ap}(A,B)\to KL(A, M(B)/B).$
Section 5: Examples\
In this section, we present a few examples which show that the bijection $\Gamma$ may not answer all questions about these extensions. For example, we show that the zero element in $KL(A,M(B)/B)$ does not represent an approximately trivial extension in general.
Section 6: Quasidiagonal extensions — general and infinite cases\
This section discusses quasidiagonal extensions. Without assuming the UCT, we give a general $K$-theoretical necessary condition for an essential extension to be quasidiagonal. We also show that for any separable exact [$C^*$-algebra]{} $A,$ there exist quasidiagonal extensions of $A$ by any $\sigma$-unital purely infinite simple [$C^*$-algebra]{}s.
Section 7: Quasidiagonal extensions — finite case\
Let $A$ be a separable amenable [$C^*$-algebra]{} and $B$ be a $\sigma$-unital [$C^*$-algebra]{} admitting an approximate identity consisting of projections and having the property (SP). We show that, if $A$ is a quasidiagonal [$C^*$-algebra]{}, then there exists an essential quasidiagonal extension. If, in addition that $B$ is also assumed to be a quasidiagonal [$C^*$-algebra]{}, then the condition that $A$ is quasidiagonal is also necessary. When $B$ is a $\sigma$-unital simple [$C^*$-algebra]{} with real rank zero, stable rank one, weakly unperforated $K_0(B)$ and with a continuous scale, we present a $K$-theoretical necessary and sufficient condition for an essential extension to be quasidiagonal for a class of separable quasidiagonal amenable [$C^*$-algebra]{}s.
Section 8: Approximately trivial extensions\
In the last section, we give a general $K$-theoretical necessary condition for essential extensions to be approximately trivial. Combining this condition with the results in section 7, we show that there are essential quasidiagonal extensions that are [*not*]{} approximately trivial extensions. We also show how to use the bijection $\Gamma$ to determine which essential extensions are approximately trivial at least in some special cases.
[**Acknowledgement**]{} Part of this work was done when the author was visiting East China Normal University. This work was partially supported by a grant from National Science Foundation of U. S. A.
Preliminaries
=============
Throughout this paper, we will use the following conventions:
1\) An ideal of a [$C^*$-algebra]{} is always a closed two-sided ideal.
2\) By a unital [$C^*$-subalgebra]{} $C$ of a unital [$C^*$-algebra]{} $A$ we mean $C\subset
A$ and $1_C=1_A.$
3\) If $p$ and $q$ are two projections in a [$C^*$-algebra]{} $A,$ we say $p$ is equivalent to $q$ if there exists a partial isometry $v\in A$ such that $v^*v=p$ and $vv^*=q.$
4\) Let $A$ and $B$ be two [$C^*$-algebra]{}s and $L_1, L_2: A\to B$ be two maps. Let ${\varepsilon}>0$ and ${\cal F}\subset A$ be a subset. We write $$L_1\approx_{{\varepsilon}} L_2\,\,\,\, {\rm on}\,\,\, {\cal F}$$ if $$\|L_1(a)-L_2(b)\|<{\varepsilon}\,\,\,\,{\rm for \,\,\, all}\,\,\, a\in {\cal
F}.$$
Suppose that $A$ and $B$ are unital and $L_1(1)$ and $L_2(1)$ are projections. If there is an isometry $s\in B$ such that $s^*L_2(1)s=L_1(1),$ $sL_1(1)s^*=L_2(1)$ and $${\rm ad}\, s\circ L_2\approx_{{\varepsilon}} L_1\,\,\,\, {\rm on}\,\,\, {\cal F},$$ we will write $$L_2 {\sim}_{{\varepsilon}} L_1\,\,\,\, {\rm on}\,\,\, {\cal F}.$$
5\) A separable [$C^*$-algebra]{} $A$ is said to be ${\it amenable}$ (or nuclear), if for any ${\varepsilon}>0$ and finite subset ${\cal F}\subset
A,$ there exists a finite dimensional [$C^*$-algebra]{} $C$ and two [contractive completely positive linear map]{}s $L_1: A\to C$ and $L_2: C\to A$ such that $$L_2\circ L_1\approx_{{\varepsilon}} {{\rm id}}_A \,\,\,{\rm on}\,\,\, {\cal F}.$$
6\) Let $A$ be a separable amenable [$C^*$-algebra]{}. We say $A$ satisfies the Universal Coefficient Theorem (UCT) and write $A\in {\cal N},$ if for any $\sigma$-unital [$C^*$-algebra]{} $C,$ the map $\gamma: KK(A,C)\to
{\rm Hom}(K_*(A), K_*(C))$ is surjective and the map $\kappa: {\rm
ker}\gamma\to ext(K_*(A), K_{*}(C))$ is an isomorphism, i.e., there is a short exact sequence: $$0\to ext(K_*(A), K_*(C)) \to KK(A,C)
{\stackrel{\gamma}{\to}}{\rm Hom}(K_*(A), K_*(C))\to 0.$$ If $h: A\to C$ is a [homomorphism]{} then $h$ gives an element $[h]$ in $KK(A,C).$
7\) An extension $0\to B\to E\to A\to 0$ of [$C^*$-algebra]{}s is said to be essential if $$\{e\in E: eb=be=0 \,\,\,{\rm for\,\,\, all}\,\,\, b\in B\}=\{0\}.$$ If $E$ is an essential extension of $A$ by $B$ as above, then it is determined by a monomorphism $\tau: A\to M(B)/B$ and $E=\pi^{-1}\circ \tau(A),$ where $\pi: M(B)\to M(B)/B$ is the quotient map.
We start with the following essential extensions: $$0\to A\to E\to C(D)\to 0,$$ where $D$ is the unit disk and $A=B\otimes {\cal K}$ and where $B$ is a separable unital simple AF-algebra with a unique tracial state. For example $B$ may be a UHF-algebra. Let $I$ be the unique proper ideal of $M(A)$ which contains $A$ (3.2 in [@Ell1]). Denote $J=\pi(I),$ where $\pi: M(A)\to M(A)/A$ is the quotient map. Let $p\in
M(A)/A\setminus J$ be a projection such that $1-p\in J$ is a non-zero projection. To see such a projection exists, one takes a projection $q\in M(A)\setminus A$ with finite trace. Then $q\in
I.$ Define $p=1-\pi(q).$ It follows from 1.17 (4) in [@Ln5] that $K_1(M(A)/I))={\mathbb R}.$ It is known that $M(A)/I$ is purely infinite and simple (see [@Zh2]). Thus there is a unitary $u\in p(M(A)/A)p$ such that ${\bar \pi}(u)$ is not in $U_0(M(A)/I),$ where ${\bar \pi}: M(A)/A\to M(A)/I$ is the quotient map. Let $y\in (1-p)M(A)/A(1-p)$ with the spectrum $sp(y)=D.$ Set $x=u+y.$ Define $\tau: C(D)\to M(A)/A$ by $\tau(f)=f(x)$ for $f\in C(D).$ It is easy to see that $\tau$ is not trivial nor it is approximately trivial. However, it is known that $Ext(C(D),A)=KK^1(C(D), A)=\{0\}.$ So certainly in this case $KK^1(C(D), A)$ can not be used to understand extensions of $C(D)$ by $A.$ Clearly the complicity of the extension is caused by the fact that $M(A)/A$ is not simple. One can easily imagine that when the ideal structure of $M(A)/A$ is more complicated, equivalent classes of extensions will be hard to compute if it is even possible to compute. The success of the BDF-theorem for the classification of extensions by ${\cal K}$ depends on the fact that $M({\cal K})/{\cal K},$ the Calkin algebra, is simple. In this paper, we will therefore consider only those essential extensions by a simple [$C^*$-algebra]{} $A$ such that $M(A)/A$ is simple.
So the question is:
When is $M(B)/B$ simple?
Let $B$ be a $\sigma$-unital simple [$C^*$-algebra]{}. Recall ([@Ln1]) that $B$ is said to have a [*continuous scale*]{} if for any approximate identity $\{e_n\}$ of $B$ satisfying $e_{n+1}e_n=e_ne_{n+1}=e_n$ and any nonzero positive element $a\in
B_+,$ there exists an integer $N>0$ such that $$(e_m-e_n) \lesssim a,\,\,\, {\rm for}\,\,\, m>n\ge N$$ i.e., there exists a sequence of elements $r_n\in B$ such that $r_k^*ar_k\to e_m-e_n$ for all $m>n\ge N.$ It should be noted that if $p$ and $q$ are projections and $p\lesssim q,$ then $p$ is equivalent to a projection $q'\le q.$
It is proved in [@Ln1] that, for non-unital separable simple [$C^*$-algebra]{} $B\not\cong {\cal K},$ $M(B)/B$ is simple if $B$ has a continuous scale. Recently we have proven the following:
[([@Lnnew])]{}\[0Tadd\] Let $A\not\cong {\cal K}$ be a non-unital and $\sigma$-unital simple [$C^*$-algebra]{}. The following are equivalent:
[(1)]{} $A$ has a continuous scale;
[(2)]{} $M(A)/A$ is simple,
[(3)]{} $M(A)/A$ is a purely infinite simple [$C^*$-algebra]{}.
Clearly every (non-unital) $\sigma$-unital purely infinite simple [$C^*$-algebra]{} has a continuous scale. Essential extensions of separable [$C^*$-algebra]{}s $A$ which satisfy the UCT by a non-unital separable purely infinite simple [$C^*$-algebra]{} $B$ is classified by $KK^1(A, B)$ by Kirchberg’s absorbing theorem ([@K1]).
In this paper we will focus on essential extensions by a $\sigma$-unital simple [$C^*$-algebra]{} with real rank zero, stable rank one, weakly unperforated $K_0$ and a continuous scale.
Suppose that $B$ is a non-unital separable simple [$C^*$-algebra]{} with real rank zero, stable rank one and weakly unperforated $K_0(B).$ Fix any nonzero projection $e\in B.$ Denote by $T$ the set of those quasi-traces $\tau$ on $B$ such that $\tau(e)=1.$ Note that $T$ is (weak $*$-) compact convex set. Let $a\in M(B)_+.$ Define $\hat{a}(\tau)=\tau(a)$ for $\tau\in T.$ Then $\hat{a}$ is a lower semi-continuous affine function on $T.$ If $a\in A,$ then $\hat{a}$ is continuous.
To see examples of simple [$C^*$-algebra]{}s with continuous scale, we quote the following result ([@Lnnew]). It also justifies the terminology “continuous scale".
\[0Tcs\] Let $A$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with real rank zero, stable rank one and weakly unperforated $K_0(A).$ Let $1$ be the identity of $M(A).$ Then $A$ has a continuous scale if and only if $\hat{1}(\tau)=\tau(1)$ for $\tau\in T$ is a continuous function on $T.$
It is also proved in [@Lnnew] that given any separable simple [$C^*$-algebra]{} $A,$ there is a non-unital hereditary [$C^*$-subalgebra]{} $B$ such that $B$ has a continuous scale. In particular, $M(B)/B$ is a purely infinite simple [$C^*$-algebra]{}. Note that $B\otimes {\cal K}\cong A \otimes
{\cal K}$ and $B$ may not have any non-trivial projections. Furthermore, $B$ may contain both infinite and finite projections (given by Rørdam ([@Ro5]).
\[0Daff\]
Let $T$ be a compact convex set. A function $f$ defined on $T$ is said to be affine if\
$f(a\xi+(1-a)\zeta)=af(\xi)+(1-a)f(\zeta)$ for all $\xi,\zeta\in T$ and $0\le a\le 1.$ We denote $$\mathit{Aff}(T)=\{f\in C(T): f\,\,\,{\rm is\,\,\, affine}\}.$$ If $f, g\in \mathit{Aff}(T)$ and $f(t)>g(t)$ for all $t\in T,$ we will write $f\gg g.$ Denote $$Aff(T)_{++}=\{f\in Aff(T): f\gg 0, {\rm or}\,\,\,
f=0\}.$$
Let $B$ be a simple [$C^*$-algebra]{} with real rank zero, stable rank one and weakly unperforated $K_0(B).$ Fix a nonzero projection $e\in B.$ Denote by $T$ the set of those traces $\tau$ defined on $B$ such that $\tau(e)=1.$ Define $$\rho_B: K_0(B)\to Aff(T)$$ by $\rho_B([p])(\tau)=\tau(p)$ for projections $p\in M_m(B),$ $m=1,2,...$ It is known that $\rho_B$ is a positive [homomorphism]{} from $(K_0(B), K_0(B)_+)$ to $(Aff(T),Aff(T)_{++})$ (see [@BH]). In fact (by [@BH]), $[p]\ge [q]$ and $[p]\not=[q]$ if and only if $\tau(p)>\tau(q)$ for all $\tau\in T.$
The following was first proved in [@Ln2] in 1991.
\[0T1\] Let $e\in B$ be a nonzero projection. Let $$T=\{ \tau: \tau(e)=1, \tau\,\,\, {\rm is\,\,\, a\,\,\,
trace\,\,\, defined \,\,\,
on}\,\,\, B\}.$$ Then
[(1)]{} $(K_0(M(B)), K_0(M(B))_+)=(Aff(T), Aff(T)_{++});$
[(2)]{} two projections $p$ and $q$ in $M(B)\setminus B$ are equivalent if $\tau(p)=\tau(q)$ for all $\tau\in T;$
[(3)]{} for any $f\in Aff(T)_{++},$ there is a projection $p\in M_k(M(B))\setminus M_k(B)$ (for some $k\ge 1$) such that $\hat{p}=f;$ and
[(4)]{} $K_1(M(B))=\{0\}$ and $U(M(B))=U_0(M(B)).$
Since $B$ has real rank zero, we obtain an approximate identity $\{e_n\}$ for $B$ consisting of projections (with $e_0=0$). Let $p\in M(B)$ be a projection. It follows from Theorem 4.1 in [@Zh1] that we may assume that $p=\sum_{n=1}^{\infty}p_n,$ where $p_n\le
e_{n+1}-e_{n}.$ Since $\hat{1}$ is continuous on $T,$ by the Dini Theorem, $\rho_B(e_n)$ converges to $\hat{1}$ uniformly on $T.$ This implies that $\hat{p}$ is also continuous. Define $\rho: K_0(M(B))\to Aff(T)$ by defining $\rho([p])=\hat{p}.$ It is clear that $\rho$ is a well-defined [homomorphism]{}.
We now prove (2). It follows from Theorem 4.1 in [@Zh1] again that we may assume that $p=\sum_{n=1}^{\infty}p_n$ and $q=\sum_{n=1}^{\infty}q_n,$ where $p_n, q_n\le e_{n+1}-e_{n}$ and the sum converges in the strict topology. Without loss of generality, we may assume that $p_n$ and $q_n$ are not zero. Since $B$ is simple, we have $\widehat{p_1}\ll \hat{p}=\hat{q}.$ Since $\sum_{n=1}^{\infty}\hat{q_n}$ converges uniformly on $T,$ there is $n_1>1$ such that $\sum_{j=1}^{n_1}\widehat{q_j}\gg\widehat{p_1}.$ It follows from III2.2 and III2.3 in [@BH] that there is a partial isometry $v_1\in B$ such that $$v_1^*v_1=p_1{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}v_1v_1^*\le \sum_{j=1}^{n_1}q_j.$$ There is $m_1>1$ such that $$\tau(\sum_{j=2}^{m_1}p_j)>\tau(\sum_{j=1}^{n_1}q_j -v_1v_1^*)\,\,\,\,
{\rm for\,\,\,all}\,\,\,\tau\in T.$$ It follows that there is a partial isometry $u_1\in B$ such that $$u_1^*u_1=\sum_{j=1}^{n_1}q_j -v_1v_1^* {\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}u_1u_1^*\le
\sum_{j=2}^{m_1}p_j.$$ Put $w_1=v_1+u_1^*.$ Then $$\sum_{j=1}^{m_1}p_j >w_1^*w_1\ge p_1{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\sum_{j=1}^{n_1}q_j=w_1w_1^*\ge q_1.$$ By induction one constructs a sequence of partial isometries $w_k\in B$ ($w_0=0$) such that $$\sum_{j=1}^{m_k}p_j-\sum_{j=1}^{k-1}w_j^*w_j >w_k^*w_k\ge
\sum_{j=1}^{m_{k-1}}p_j-\sum_{j=1}^{k-1}w_j^*w_j {\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\sum_{j=1}^{n_k}q_j-\sum_{j=1}^{k-1}w_jw_j^*=w_kw_k^*\ge
\sum_{j=1}^{n_{k-1}}q_j -\sum_{j=1}^{k-1}w_jw_j^*,$$ where $\{n_k\}$ and $\{m_k\}$ are increasing sequences of positive integers.
Set $W=\sum_{k=1}^{\infty}w_k.$ One checks that the sum converges in the strict topology and $W$ is a partial isometry in $M(B).$ One then verifies that $$W^*W=p{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}WW^*=q.$$
This proves (2).
Note that (2) also implies that $\rho$ is injective. In fact, if $p\in M(B)\setminus B$ and $q\in B$ are two projections such that $\tau(p)=\tau(q)$ for all $\tau\in T,$ then $p\oplus 1$ and $q\oplus 1$ are both in $M(B)\setminus B.$ So, by (2), $p\oplus 1$ and $q\oplus 1$ are equivalent. Therefore $\rho$ is injective.
To see $\rho$ is surjective, let $f\in Aff(T).$ We need to show that $f$ is in the image of $\rho.$ It is clear that it suffices to prove the case for $f\gg 0.$ So we may assume that $f\gg 0.$ We claim that there exists a sequence of positive functions $\{f_n\}$ in $\rho_B(K_0(B))$ such that $f_n\ll f_{n+1}$ such that $f_n\to f$ uniformly on $T.$
Let $d_0=\inf\{f(t): t\in T\}.$ Then $d_0>0.$
Let $d_0/2>{\varepsilon}>0.$ Since $\rho_A(K_0(B))$ is dense in $Aff(T)$ (see [@BH]), there is $g_1\in \rho_B(K_0(B)_+)$ such that $$\|g_1-(f-{\varepsilon})\|<{\varepsilon}/4.$$ Therefore $(1-{\varepsilon}/2)f\ll g_1\ll f.$ By applying the same argument to the function $f-g_1,$ we obtain $g_2\in \rho_B(K_0(B)_+)$ such that $(1-{\varepsilon}/4)(f-g_1)\ll g_2\ll f-g_1.$ Note that $g_1+g_2\in
\rho_B(K_0(B)_+).$ From this the claim follows.
Now we will show that $f$ is in the image of $\rho.$ By replacing $f$ by $f-f_n$ if necessary, we may assume that $f\ll
\widehat{e_1}.$ There are projections $r_n\in B$ such that $\rho(r_n)=f_n-f_{n-1}$ (with $f_0=0$). Since $f\ll \hat{e_1}$ and $f_n$ converges to $f$ uniformly on $T,$ we may assume that $$\sum_{k=1}^{n_1}(f_k-f_{k-1})\ll \rho(e_1){\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\sum_{k=n_1+1}^{\infty} (f_k-f_{k-1})\ll \rho(e_2-e_1).$$ Thus we may assume that $$\sum_{k=1}^{n_1}r_k\le e_1{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\sum_{k=n_1+1}^m r_k\le e_2-e_1$$ for all $m>n_1.$ We obtain $n_2>n_1$ such that $$\sum_{k=n_2}^{\infty} (f_k-f_{k-1})\ll \rho(e_3-e_2)$$ Therefore we also assume that $$\sum_{k=n_1+1}^{n_2} r_k\le e_2-e_2{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\sum_{k=n_2+1}^mr_k\le e_3-e_2$$ for all $m>n_2.$ By induction, we obtain an increasing sequence of integers $\{n_k\}$ such that $$\sum_{i=n_k+1}^{n_{k+1}}r_i\le e_{k+1}-e_k,\,\,\, k=1,2,....$$ Note that this implies that $\sum_{n=1}^{\infty} r_n$ converges in the strict topology to a projection $p\in M(B).$ It is easy to see that $\hat{p}=f.$ This proves (1) as well as (3).
Finally we note $pBp$ has real rank zero for all projection $p\in M(B)$ and by [@Lnep] $cer(pAp)\le \pi.$ Thus (4) has been proved in Lemma 3.3 in [@Lngw].
\[0T2\] Let $B$ be as in \[0T1\]. Then
[(1)]{} $K_1(M(B)/B)={\rm ker}\rho_B,$ and
[(2)]{} there is a short exact sequence $$0\to Aff(T)/\rho_A(K_0(B))\to K_0(M(B)/B)\to K_1(B)\to 0.$$
From the following six-term exact sequence $$\begin{array}{ccccc}
K_0(B) & {\stackrel{\rho_B}{\to}} &K_0(M(B)) & \to & K_0(M(B)/B)\\
\uparrow && &&\downarrow\\
K_1(M(B)/B)& \leftarrow & K_1(M(B)) & \leftarrow & K_1(B)\\
\end{array}$$ we obtain, by \[0T1\], $$\begin{array}{ccccc}
K_0(B)& {\stackrel{\rho_B}{\to}} & Aff(T)& \to & K_0(M(B)/B)\\
\uparrow && &&\downarrow\\
K_1(M(B)/B)& \leftarrow & 0 & \leftarrow & K_1(B)\\
\end{array}$$ This six-term exact sequence unsplices into $$K_1(M(B)/B)={\rm ker}\rho_B {\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}0\to Aff(T)/\rho_B(K_0(B))\to
K_0(M(B)/B)\to K_1(B)\to 0.$$
[It should be noted that $Aff(T)$ as an ordered group does not depend on the choice of the non-zero projection $e.$ In what follows when we write $Aff(T)$ it is understood that the projection $e$ is fixed.]{}
The following fact will be used in this paper.
\[0Pph\] Let $G$ be a dense ordered subgroup of ${\mathbb R}$ containing $1$ and let $T$ be a Choquet simplex. Suppose that $h: G\to Aff(T)$ is a positive [homomorphism]{} with $h(1)=a.$ Then $$h(z)=za\,\,\, for\,\,\, all\,\,\, z\in G.$$
Since $G$ is dense in ${\mathbb R},$ there exists a sequence $g_n>0$ in $G$ such that $g_n\to 0.$ We may assume that $ng_n<1$ for all $n.$ Therefore $$nh(g_n)\le a\,\,\, n=1,2,....$$ It follows that $h(g_n)\to 0.$ Let $f_n>0$ in $G$ such that $f_n\to 0.$ Thus, for each $m,$ there exists $N(m)$ such that $f_n\le g_m$ if $n\ge N(m).$ This implies that $h(f_n)\to 0.$ Thus $h$ is continuous. For each nonzero integer $m,$ define ${\tilde h} (1/m)=a/m.$ Then one checks that ${\tilde h}$ is a positive [homomorphism]{} from ${\mathbb Q}G$ to $Aff(T).$ The same argument above shows that ${\tilde h}$ is also continuous. Fix $z\in G.$ Suppose that $r_n\in {\mathbb Q}$ such that $r_n\to z.$ Then ${\tilde h}(r_n)\to h(z),$ or $r_na\to h(z).$ Therefore $h(z)=az.$
The following example shows that even in the case that $M(B)/B$ is simple, in general, $KK^1(A,B)$ can not be used to give a meaningful description of extensions of $A$ by $B.$
\[0E2\]
Let $A$ be a unital simple AF-algebra and $B$ be a $\sigma$-unital simple [$C^*$-algebra]{} with real rank zero, stable rank one, weakly unperforated $K_0(B)$ and a continuous scale. Let $r$ be an irrational number and $D_r=\{m+nr: m,n\in {\mathbb Z}\} $ ($={\mathbb Z}\oplus {\mathbb Z}r$). Suppose $K_0(A)=D_r\oplus
{\mathbb Z}.$ Define $$K_0(A)_+=\{ x+z: x\in D_r, x>0\}\cup\{(0,0)\}.$$ Suppose that there is a group [homomorphism]{} $\theta: K_0(A)\to
Aff(T)/\rho_B(K_0(B))$ such that $\theta((0,1))\not=0,$ where $(0,1)$ denotes the generator of the last summand ${\mathbb Z}$ of $K_0(A).$ This gives a group [homomorphism]{} $\alpha: K_0(A)\to
K_0(M(B)/B)$ which maps $(0,1)$ to $r.$ Let $\Phi: Aff(T)\to
Aff(T)/\rho_B(K_0(B))$ be the quotient map. Then $\theta=\Phi\circ
\alpha$ gives one such [homomorphism]{}. Since $M(B)/B$ is a purely infinite simple [$C^*$-algebra]{}, it is easy to construct a [homomorphism]{} $\tau: A\to M(B)/B$ such that $\tau_{*0}=\theta$ (see for example \[IITM\]). This $\tau$ gives an essential extension of $A$ by $B$ which gives an element in $KK^1(A,B).$ Let $E$ be the [$C^*$-algebra]{} determined by $\tau$ we have the following commutative diagram: $${\small \put(-160,0){$K_0(B)$} \put(20,0){$K_0(E)$}
\put(180,0){$K_0(A)$} \put(-85,-40){$K_0(B)$}
\put(0,-40){$K_0(M(B))$} \put(105,-40){$K_0(M(B)/B))$} \put(-85,
-70){$K_1(M(B)/B)$} \put(27,-70){$0$} \put(115,-70){$K_1(B)$}
\put(-160,-110){$K_1(A)$} \put(15,-110){$K_1(E)$}
\put(180,-110){$K_1(B)$} \put(-120, 2){\vector(1,0){95}}
\put(70,1){\vector(1,0){95}} \put(-123,-3){\vector(1,-1){30}}
\put(30,-3){\vector(0,-1){25}} \put(180,-2){\vector(-1,-1){30}}
\put(-45,-38){\vector(1,0){35}} \put(70,-38){\vector(1,0){25}}
\put(-147, -90){\vector(0,1){85}} \put(-75,-60){\vector(0,1){15}}
\put(125, -45){\vector(0,-1){15}} \put(190,-7){\vector(0,-1){85}}
\put(15,-68){\vector(-1,0){30}} \put(95,-68){\vector(-1,0){25}}
\put(-123,-102){\vector(1,1){30}} \put(175,
-105){\vector(-1,1){30}} \put(30,-100){\vector(0,1){25}} \put(-5,
-108){\vector(-1,0){100}} \put(170,-108){\vector(-1,0){95}}
\put(-111,-13){$=$} \put(151, -13){$\tau_{*0}$} \put(160,
-88){$=$} }$$
Since the image of $\tau_{*0}$ is in $Aff(T)/\rho_B(K_0(B)),$ by (2) in \[0T2\] and from the above diagram one concludes that the map from $K_0(A)$ to $K_1(B)$ is zero. Since $K_1(A)=\{0\},$ one further concludes that the map from $K_1(A)$ to $K_0(B)$ is also zero. By the Universal Coefficient Theorem, one computes that $[\tau]\in ext_{\mathbb
Z}(K_0(A), K_0(B)).$ Using the map $\alpha$ or using the fact that $K_0(A)$ is finitely generated free group, $[\tau]=0$ in $KK^1(A,B).$ However, there is no [homomorphism]{} $h: A\to M(B)$ such that $(\pi\circ h)_{*0}=\theta.$ Otherwise, since $h_{*0}$ is positive, it maps ${\rm ker}\rho_A={\mathbb Z}$ into zero. It would imply that $\tau_{*0}$ maps ${\rm ker}\rho_A$ to zero. But we constructed otherwise. Therefore $\tau$ is not trivial. Furthermore it can not be even approximately trivial (see \[IITapptr\] below). This shows that even in the case that $M(B)/B$ is simple, $KK^1(A,B)$ can not be used to give a good description of extensions of $A$ by $B.$
Monomorphisms from $A\otimes {\cal O}_2$ into a purely infinite simple [$C^*$-algebra]{}
========================================================================================
\[IIDfilter\]
Recall that a family ${\omega}$ of subsets of ${\mathbb N}$ is an [*ultrafilter*]{} if
\(i) $X_1,...,X_n\in \omega$ implies $\cap_{i=1}^n X_i\in \omega,$
\(ii) $\O\not\in \omega,$
\(iii) if $X\in\omega$ and $X\subset Y,$ then $Y\in\omega$ and
\(iv) if $X\subset {\mathbb N}$ then either $X$ or ${\mathbb
N}\setminus X$ is in $\omega.$
An ultrafilter is said to be [*free*]{}, if $\cap_{X\in \omega} X
=\emptyset.$ The set of free ultrafilters is identified with elements in $\beta {\mathbb N}\setminus {\mathbb N},$ where $\beta{\mathbb N}$ is the Stone-Cech compactification of ${\mathbb N}.$
A sequence $\{x_n\}$ (in a normed space for example) is said to converge to $x_0$ along $\omega,$ written $\lim_{\omega}x_n=x_0,$ if for any ${\varepsilon}>0$ there exists $X\in\omega$ such that $\|x_n-x_0\|<{\varepsilon}$ for all $n\in X.$
Let $\{B_n\}$ be a sequence of [$C^*$-algebra]{}s. We write $l^{\infty}(\{B_n\})$ for the [$C^*$-algebra]{} $\prod_{n=1}^{\infty} B_n.$ Fix an ultrafilter $\omega.$ The ideal of $l^{\infty}(\{B_n\})$ which consists of those sequences $\{a_n\}$ in $l^{\infty}(\{B_n\})$ such that $\lim_{\omega}\|a_n\|=0$ is denoted by $c_{\omega}(\{B_n\}).$ Define $$q_{\omega}(\{A_n\})=l^{\infty}(\{B_n\})/c_{\omega}(\{B_n\}).$$ If $B_n=B, n=1,2,...,$ we use $l^{\infty}(B)$ for $l^{\infty}(\{B_n\}),$ $c_{\omega}(B)$ for $c_{\omega}(\{B_n\})$ and $q_{\omega}(A)$ for $q_{\omega}(\{A_n\}),$ respectively.
\[NIL1\] Let $A$ be a separable [$C^*$-algebra]{} and $\{B_n\}$ be a sequence of unital [$C^*$-algebra]{}s. Let $\omega\in \beta{\mathbb N}\setminus {\mathbb N}.$ Suppose that $\psi_m, \phi_m: A\to B_m$ are two bounded sequences of maps such that $\psi=\pi\circ \{\psi_m\}, \phi=\pi\circ \{\phi_m\}: A\to q_{\omega}(\{B_n\})$ are two [homomorphism]{}s, where $\pi: l^{\infty}(\{B_m\})\to q_{\omega}(\{B_n\})$ is the quotient map.
[(1)]{} Suppose that there are isometries $u_n\in q_{\omega}(\{B_n\})$ such that $$\lim_{n\to\infty}\|u_n^*\psi(a) u_n -\phi(a)\|=0\,\,\,{\rm for \,\,\, all}\,\,\,
a\in A.$$ Then there is an isometry $w\in q_{\omega}(\{B_n\})$ such that $$w^*\psi(a)w=\phi(a)\,\,\,{\rm for \,\,\, all}\,\,\,
a\in A.$$
[(2)]{} Suppose that $\psi$ and $\phi$ are approximately unitarily equivalent in $q_{\omega}(\{B_n\}).$ Then they are unitarily equivalent.
Suppose that there is a sequence of isometries $u_n\in
q_{\omega}(\{B_n\})$ such that $$\lim_{n\to\infty}\|u_n^*\psi(a)u_n-\phi(a)\|=0$$ for all $a\in A.$ Let $\{a_n\}$ be a dense sequence of $A.$ By passing to a subsequence if necessary, we may assume that $$\|u_n^*\psi(a_j)u_n-\phi(a_j)\|<1/n,\,\,\,j=1,2,...,n.$$ It follows from 6.2.4 in [@Rob] that there exists, for each $n,$ a sequence of isometries $u_m^{(n)}\in B_n$ such that $\pi(\{u_m^{(n)}\}=u_n,$ where $\pi: l^{\infty}(\{B_n\})\to q_{\omega}(\{B_n\})$ is the quotient map. For each $n,$ there exists $X_n\in \omega$ such that for $m\in X_n,$ $$\|(u_m^{(n)})^*\psi_m(a_j)u_m^{(n)}-\phi_m(a_j)\|\le 1/n,j=1,2,...,n.$$ Since $\omega$ is free, there is for each $j,$ $Y_j'\in \omega$ such that $j\not\in Y_j'.$ Let $Y_j=\cap_{1\le k\le j} Y_j'.$ Then $Y_j\in \omega$ and $\{1,2,...,j\}\cap Y_j=\emptyset.$ Let $Z_k'=X_k\cap Y_k.$ Then $Z_k'\in \omega$ and $\cap_{k\ge N}Z_k'=\emptyset,$ $N=1,2,....$ Put $Z_k=\cap_{1\le j\le k}Z_j'.$ Then $Z_k\in \omega$ and $Z_1\supset Z_2\supset \cdots \supset Z_k\supset \cdots.$ Moreover, $Z_k\subset X_k,$ $k=1,2,....$
Define $l(m)$ as follows. If $m\in Z_k\setminus Z_{k+1},$ define $l(m)=k,$ $k=1,2,...;$ and if $m\not\in Z_1,$ define $l(m)=m,$ $m=1,2,....$ Put $w_m=u_m^{(l(m))}\in B_m$ and $w=\pi(\{w_m\}).$ Then, for any ${\varepsilon}>0$ and $j,$ let $k>0$ be an integer such that $1/k<{\varepsilon}$ and $j\le k.$ If $m\in Z_k,$ then $m\in Z_{k'}\setminus Z_{k'+1}$ for some $k'\ge k.$ Thus $w_m=u_m^{(k')}$ and $m\in X_{k'}.$ Therefore $$\|w_m^*\psi_m(a_j)w_m-\phi_m(a_j)\|=\|u_m^{(k')}\psi_m(a_j)u_m^{(k')}-\phi_m(a_j)\|<1/k'\le 1/k<{\varepsilon}$$ for all $j=1,2,...,k.$ This implies that $$\lim_{\omega}\|w_m^*\psi_m(a_j)w_m-\phi_m(a_j)\|=0$$ for all $j.$ Hence $$w^*\psi(a_j)w=\phi(a_j)\,\,\,j=1,2,....$$ Since $\{a_n\}$ is dense in $A,$ $$w^*\psi(a)w=\phi(a)\,\,\,{\rm for \,\,\, all}\,\,\, a\in A.$$ This proves (1).
To prove (2), we note that if $u_n$ are unitaries, so is $w.$
[([@KP], Proposition 1.4)]{}\[NIL3\] Let $A$ be a unital separable [$C^*$-subalgebra]{} of a unital purely infinite simple [$C^*$-algebra]{} $B$ and let $\phi: A\to B$ be an amenable [contractive completely positive linear map]{}. Then for any finite subset ${\cal F}\subset A$ and any ${\varepsilon}>0$ there is a non-unitary isometry $s\in B$ such that $$\|s^*as-\phi(a)\|<{\varepsilon}\,\,\,\,for\,\,\, all\,\,\, a\in {\cal F}.$$
By the assumption that $\phi$ is amenable there are [contractive completely positive linear map]{}s $\sigma: A\to M_n({\mathbb C})$ (for some integer $n>0$) and $\eta: M_n({\mathbb C})\to B$ such that $$\phi\approx_{{\varepsilon}} \eta\circ \sigma\,\,\,\,{\rm on}\,\,\, {\cal F}.$$ We may therefore assume that $\phi=\eta\circ \sigma.$ It is well known (see for example, 2.3.5 in [@Lnb]) that there exists a [contractive completely positive linear map]{} ${\tilde \sigma}: B\to M_n({\mathbb C})$ such that ${\tilde \sigma}|_{A}=\sigma.$ Define ${\tilde \phi}=\eta\circ {\tilde \sigma}.$ Since now $B$ is purely infinite and ${\tilde \phi}$ is amenable, the lemma follows immediately from 1.4 in [@KP].
\[NIL4\] Let $B$ be a unital purely infinite simple [$C^*$-algebra]{} and $A$ be a separable amenable [$C^*$-algebra]{}. Let $\phi,\psi: A\to B$ be two monomorphisms. Then there are two sequences of isometries $s_n$ and $w_n$ in $B$ such that $$\lim_{n\to\infty}\|s_n^*\phi(a)s_n-\psi(a)\|=0 \,\,\, and \,\,\,
\lim_{n\to\infty}\|w_n^*\psi(a)w_n-\phi(a)\|=0$$ for all $a\in A.$
The proof of the following proposition is exactly the same as that of 6.2.6 in [@Rob].
\[NIPps\] Let $B_n$ be a sequence of purely infinite simple [$C^*$-algebra]{}s. Then $q_{\omega}(\{B_n\})$ is a purely infinite simple [$C^*$-algebra]{} for every free ultrafilter $\omega.$
[(3.4 in [@KP])]{}\[NIP1\] Let $B_n$ be a sequence of unital purely infinite simple [$C^*$-algebra]{}s and $A$ be a unital separable amenable simple [$C^*$-algebra]{}. Suppose that $j: A\to q_{\omega}(\{B_n\})$ is a monomorphism. Then the relative commutant $j(A)'$ in $q_{\omega}(\{B_n\})$ is a unital purely infinite simple [$C^*$-algebra]{}.
Let $a\in j(A)'$ be a nonzero positive element with $\|a\|=1.$ It suffices to show that there is an isometry $s\in j(A)'$ such that $s^*as=1.$ Let $X=sp(a)\subset [0,1]$ and define two [homomorphism]{}s $\phi,\psi: C(X)\otimes A\to q_{\omega}(B)$ by $$\phi(f\otimes b)=f(a)b{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\psi(f\otimes b)=f(1)b$$ for $f\in C(X)$ and $b\in A.$ Since $A$ is an amenable simple [$C^*$-algebra]{} $\phi$ is a monomorphism. It follows from \[NIPps\] that $q_{\omega}(B)$ is a purely infinite simple [$C^*$-algebra]{}. Therefore by \[NIL3\] there is a sequence of isometries $s_n\in q_{\omega}(B)$ such that $$\lim_{n\to\infty}\|s_n^*\phi(x)s_n-\psi(x)\|=0$$ for all $x \in C(X)\otimes A.$ It follows from \[NIL1\] that there is an isometry $s\in q_{\omega}(B)$ such that $$s^*\phi(x)s=\psi(x)\,\,\,{\rm for\,\,\, all}\,\,\, x\in C(X)\otimes A.$$ In particular, $$s^*as=s^*\phi(\imath\otimes 1)s=\psi(\imath\otimes 1)=1,$$ where $\imath$ is the identity function $\imath(t)=t.$ We also have $$s^*bs=s^*\phi(1\otimes b)s=\psi(1\otimes b)=b.$$ Hence $s\in j(A)'$ by Lemma 6.3.6 of [@Rob].
\[NIL5\] Let $B$ be a unital purely infinite simple [$C^*$-algebra]{} and $A$ be a unital separable amenable [$C^*$-algebra]{}. Suppose that $\phi, \psi: A\otimes {\cal O}_2\to B$ are two monomorphisms. Then $\psi$ and $\phi$ are approximately unitarily equivalent.
Let $p_1=\phi(1_{A\otimes {\cal O}_2})$ and $p_2=\psi(1_{A\otimes {\cal O}_2}).$ It follows from 3.6 in [@Ro1] that $\phi|_{1\otimes {\cal O}_2}$ and $\psi|_{1\otimes {\cal O}_2}$ are approximately unitarily equivalent. It follows that $p_1$ and $q_1$ are equivalent in $B.$ Therefore we may assume, without loss of generality, that $p_1=p_2.$ By replacing $B$ by $p_1Bp_1,$ we may further assume that both $\phi$ and $\psi$ are unital.
Let $\Psi, \Phi: A\otimes {\cal O}_2\to l^{\infty}(B)$ be defined by $\Psi=\{\psi(a),\psi(a),...,\psi(a),...)$ and $\Phi(a)=\{\phi(a),\phi(a),...,\phi(a),...)$ for $a\in A,$ respectively. Fix a free ultrafilter $\omega.$ Put ${\bar
\phi}=\pi\circ \Phi$ and ${\bar \psi}=\pi\circ \Psi,$ where $\pi:
l^{\infty}(B)\to q_{\omega}(B)$ is the quotient map. It follows from 3.6 in [@Ro1] that ${\bar \phi}|_{1\otimes {\cal O}_2}$ and ${\bar \psi}|_{1\otimes {\cal O}_2}$ are approximately unitarily equivalent. It follows from \[NIL1\] that they are unitarily equivalent in $q_{\omega}(B).$ Without loss of generality, we may assume that $${\bar \phi}|_{1\otimes {\cal O}_2}={\bar \psi}|_{1\otimes {\cal
O}_2}.$$ Let $D={\bar \phi}(1\otimes {\cal O}_2).$ Then $D\cong {\cal O}_2.$ By [@Ro2] ${\cal O}_2\otimes {\cal O}_2\cong {\cal O}_2.$ Let $\imath: {\cal O}_2\to {\cal O}_2\otimes {\cal O}_2$ be defined by $\imath(a)=a\otimes 1$ and let $\lambda: {\cal O}_2\otimes {\cal O}_2
\to {\cal O}_2$ be an isomorphism. Then $\lambda\circ \imath$ and ${{\rm id}}_{{\cal O}_2}$ are approximately unitarily equivalent.
Let $\gamma: A\otimes {\cal O}_2\to A\otimes {\cal O}_2\otimes {\cal O}_2
\to A\otimes {\cal O}_2\otimes 1$ be the [homomorphism]{} induced by $\lambda\circ \imath$ above. Then ${\bar \phi}\circ \gamma$ is approximately unitarily equivalent to ${\bar \phi}$ and ${\bar \psi}\circ \gamma$ is approximately unitarily equivalent to ${\bar \psi}.$ To prove that ${\bar \psi}$ and ${\bar \phi}$ are approximately unitarily equivalent it suffices to show that ${\bar \phi}\circ \gamma$ and ${\bar \psi}\circ \gamma$ are approximately unitarily equivalent.
There is a unital [$C^*$-subalgebra]{} $D_1 \subset D$ of $q_{\omega}(B)$ which is isomorphic to ${\cal O}_2$ and its commutant contains both the images of ${\bar \phi}\circ \gamma$ and ${\bar \psi}\circ \gamma.$
Let $D_1'$ be the commutant in $q_{\omega}(B).$ Then by \[NIP1\] $D'$ is purely infinite simple. It follows from Corollary \[NIL4\] that there are isometries $s_n, w_n\in D_1'$ such that $$\lim_{n\to\infty}\|s_n^*{\bar \psi}\circ \gamma(a)s_n-{\bar \phi}(a)\|=0{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\lim_{n\to\infty}\|w_n^*{\bar \phi}\circ \gamma(a)w_n-{\bar \psi}(a)\|=0$$ for all $a\in A.$ Since $((D_1)')'$ contains a unital subalgebra $D_1$ which is isomorphic to ${\cal O}_2,$ by Lemma 6.3.7 in [@Rob], ${\bar \phi}\circ \gamma$ and ${\bar \psi}\circ \gamma$ are approximately unitarily equivalent. It follows that ${\bar \phi}$ and ${\bar \psi}$ are approximately unitarily equivalent in $q_{\omega}(B).$ It follows from 6.2.5 in [@Rob] that $\phi$ and $\psi$ are approximately unitarily equivalent.
Approximately unitarily equivalent extensions
=============================================
The purpose of this section is to prove Theorem \[IabTTad\] and Theorem \[IabT1\]. The statements have been proved for the case that the target [$C^*$-algebra]{} is a separable amenable purely infinite simple [$C^*$-algebra]{}. The problem we deal with in this section is to show a certain absorption property in the absence of “approximate divisibility" for $M(B)/B.$
\[IabL1\] Let $B$ be a non-unital and $\sigma$-unital [$C^*$-algebra]{} and $A$ be a separable [$C^*$-algebra]{}. Let $\tau: A\to M(B)/B$ be a [homomorphism]{}. Then there is a sequence of non-zero mutually orthogonal elements $a_n\in \tau(A)',$ where $\tau(A)'$ is the commutant of $\tau(A)$ in $M(B)/B.$
Let $D$ be a separable [$C^*$-algebra]{} in $M(B)$ such that $\tau(A)\subset \pi(D),$ where $\pi: M(B)\to M(B)/B$ is the quotient map. It follows from Lemma 3.1 in [@Lnnew] that there exists an approximate identity $\{e_n\}$ such that $e_{n+1}e_n=e_ne_{n+1}=e_n,$ $n=1,2,...$ and $$\|e_kd-de_k\|\to 0\,\,\,{\rm as}\,\,\, k\to \infty$$ for all $d\in D.$ Fix a subsequence $X\subset {\mathbb N},$ then $a_X=\sum_{n\in X}(e_{n+1}-e_n)$ is a positive element in $M(A).$ Since $\lim_{k\to\infty}\|e_kd-de_k\|=0$ for each $d\in A,$ $\pi(a_X)\pi(d)=\pi(d)\pi(a_X).$ In other words $a_X\in \tau(A)'.$ Suppose that $X$ and $Y$ are two disjoint subset of ${\mathbb N}$ such that for any $n\in x$ and $m\in Y,$ $|n-m|\ge 2.$ By the assumption that $e_{n+1}e_n=e_ne_{n+1}=e_n,$ we conclude that $a_Xa_Y=a_Ya_X=0.$ From this it is easy to see that there exists a sequence of nonzero mutually orthogonal elements in $\tau(A)'.$
\[IabL2\] Let $A$ be a unital separable amenable [$C^*$-algebra]{}, $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with a continuous scale and let $\omega\in \beta {\mathbb N}\setminus {\mathbb N}$ be a free ultrafilter. Suppose that $\tau: A\to M(B)/B$ is an essential unital extension. Let $\tau_{\infty}: A\to l^{\infty}(M(B)/B)$ be defined by $\tau_{\infty}(a)=(\tau(a),\tau(a),...)$ and let $\psi=\Phi\circ \tau_{\infty},$ where $\Phi: l^{\infty}(M(B)/B) \to
q_{\omega}(M(B)/B).$ Then there is a unital [$C^*$-subalgebra]{} $C\cong {\cal
O}_{\infty}$ in the commutant of $\psi(A)$ in $q_{\omega}(M(B)/B).$
Let $J: M(B)/B\to q_{\omega}(M(B)/B)$ be defined by $J(b)=\Phi((b,b,...,b,...))$ for $b\in B.$ By (the proof of ) 7.4 in [@Lnpro], there exists a unital separable purely infinite simple [$C^*$-algebra]{} $D\subset M(B)/B$ such that $\tau(A)\subset D.$ It follows from \[IabL1\] that there is a sequence of nonzero mutually orthogonal positive elements $\{a_n\}$ in $D',$ the commutant of $D$ in $M(B)/B.$ Let $X=sp(a_1).$ Without loss of generality we may assume that $\|a_1\|=1$ and $1\in X.$ Define $L_1, L_2:
A\otimes C(X)\to q_{\omega}(M(B)/B)$ by $L_1(x\otimes f)=\psi(x)f(J(a))$ and $L_2(x\otimes f)=\psi(x)f(1)$ for $x\in A$ and $f\in C(X).$ Define $L_3: D\otimes C(X)\to q_{\omega}(M(B)/B)$ by $L_3(y\otimes
f)=yf(J(a))$ for $y\in D.$ Since $\psi$ is injective and $D$ is purely infinite simple, one concludes that $L_3$ is injective. Consequently, $L_1$ is injective. Now we apply an argument in [@KP]. Since $A\otimes C(X)$ is amenable, the Choi-Effros lifting theorem ([@CE]) provides unital completely positive lifting $\rho,\sigma:
A\otimes C(X)\to l^{\infty}(B)$ of $L_1$ and $L_2.$ Write $$\rho(a)=(\rho_1(a),\rho_2(a),...,\rho_n(a),...) {\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\sigma(a)=(\sigma_1(a), \sigma_2(a),...,\sigma_n(a),...)$$ for $a\in A\otimes C(X),$ where $\rho_k$ and $\sigma_k$ are unital, completely positive maps from $A\otimes C(X)$ into $B_k.$ It follows from 6.3.5 (iii) in [@Rob] that there are non-unitary isometries $s_k\in B_k$ ($k=1,2,...$) such that $$\lim_{n\to\infty}\|s_n^*\rho_n(a)s_n-\sigma_n(a)\|=0\,\,\,\,{\rm for\,\,\,
all}\,\,\, a\in A\otimes C(X).$$ Put $t_1=\pi(s_1,s_2,...,s_n,...))\in l^{\infty}( B).$ Then $t_1$ is a non-unitary isometry. It follows that $$t_1^*\psi(a)t_1=t_1^*L_1(a\otimes 1)t_1=
L_2(a\otimes 1)=\psi(a)$$ for all $a\in A.$ It follows from Lemma 6.3.6 in [@Rob] that $t_1\in (\psi(A))'.$ Furthermore, $$t_1^*J(a_1)t_1=t_1^*L_1(1\otimes \imath)t_1=L_2(1\otimes \imath)=1,$$ where $\imath$ is the function $\imath(t)=t.$ Let $t_1t_1^*=q_1.$ Then $q_1\in \psi(A)'$ and $q_1\in
{\overline{J(a_1)\psi(A)'J(a_1)}}.$ We repeat the above argument for $a_2,a_3,....$ Then we obtain a sequence of isometries $t_1,t_2,...$ in $\psi(A)'$ such that $t_n^*t_n=1$ and $$\sum_{i=1}^nt_it_i^*\in \overline{(\sum_{i=1}^nJ(a_i) )\psi(A)'
(\sum_{i=1}^nJ(a_i)) }.$$ It follows that $\sum_{i=1}^n a_i\le 1.$ Therefore we obtain a unital [$C^*$-subalgebra]{} $C\cong {\cal O}_{\infty}$ in $\psi(A)'.$
\[Iab3\] Let $A$ be a unital separable amenable [$C^*$-algebra]{} and $C$ be a unital separable amenable purely infinite simple [$C^*$-algebra]{}. Suppose that $A\otimes C$ is a unital [$C^*$-subalgebra]{} of a unital [$C^*$-algebra]{} $B.$ Then there is an embedding $j: A\otimes C\to A\otimes {\cal O}_2\to B$ satisfying the following: for any ${\varepsilon}>0,$ any finite subset ${\cal F}\subset A\otimes C$ and any integer $n>0,$ there exists a partial isometry $u\in M_{n+1}(B)$ such that $u^*u=1,$ $uu^*=1\oplus j(1_{A\otimes C})\oplus j(1_{A\otimes C})\oplus\cdots
\oplus j(1_{A\otimes C})$ [( ]{}where $j(1_{A\otimes C})$ repeats $n$ times[)]{} and $$u^*({{\rm id}}_{A\otimes C}\oplus j\oplus j\oplus \cdots \oplus j)u
\approx_{{\varepsilon}} {{\rm id}}_{A\otimes C}
\,\,\,\, on\,\,\, {\cal F},$$ where $j$ repeats $n$ times.
Let ${\varepsilon}>0$ and ${\cal G}\subset C$ be a finite subset. It follows from [@KP] that there is [homomorphism]{} $\imath: C\to {\cal O}_2\to C$ satisfying the following: for any ${\varepsilon}>0,$ any finite subset ${\cal G}\subset C$ and any integer $n$ there exists a partial isometry $w\in M_{n+1}(C)$ such that $w^*w=1_C,$ $ww^*=p=1_C\oplus
\imath(1_C)\oplus \imath(1_C)\oplus \cdots \oplus \imath(1_C)$ (where $\imath (1_C)$ repeats $n$ times) and $$w^*({{\rm id}}_C\oplus \imath\oplus \imath\oplus \cdots \oplus \imath)w\approx_{{\varepsilon}/2}
{{\rm id}}_C\,\,\,\,\,{\rm on}\,\,\, {\cal G},$$ where $\imath$ repeats $n$ times. Let $u=1\otimes w$ and define $j: A\otimes C\to B$ by $j(a\otimes b)=a\otimes \imath(b))$ for $a\in A$ and $b\in C.$ One checks that the lemma follows.
\[Iab5\] Let $A$ be a unital separable amenable [$C^*$-algebra]{} and $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with a continuous scale. Suppose that $h: A\to M(B)/B$ is a unital injective [homomorphism]{} and $\omega\in
\beta{\mathbb N}\setminus {\mathbb N}$ is a free ultrafilter. Let $\pi:l^{\infty}(M(B)/B)\to q_{\omega}(M(B)/B)$ be the quotient map. Define $H_0: A\to l^{\infty}(M(B)/B)$ by $H_0(a)=(h(a),h(a),...,h(a),...)$ and $H=\pi\circ H_0.$ Then, there exists an injective [homomorphism]{} $j: A\to A\otimes {\cal O}_2
\to q_{\omega}(M(B)/B)$ satisfying the following: For any ${\varepsilon}>0,$ any finite subset ${\cal F}\subset A$ and any integer $n>0$ there exist an isometry $u\in M_{n+1}(q_{\omega}(M(B)/B)$ with $u^*u=1,$ $uu^*=1\oplus
j(1_A)\oplus\cdots \oplus j(1_A)$ [(]{} where $j(1_A)$ repeats $n$ times[)]{} such that $$u^*(H\oplus j\oplus j\oplus\cdots \oplus j)u
\approx_{{\varepsilon}} H\,\,\,\,{\rm on}\,\,\,{\cal F},$$ where $j$ repeats $n$ times. Moreover, there is $q\in
q_{\omega}(M(B)/B)$ such that $[q]=[H(1_A)]$ and $qj(a)=j(a)q$ for all $a\in A$ and $qjq$ is an injective full [homomorphism]{}.
Fix a free ultrafilter $\omega\in \beta{\mathbb N}\setminus
{\mathbb N}.$ We identify $h(A)$ with $A.$ It follows from \[IabL2\] that $H(A)'$ contains a unital [$C^*$-subalgebra]{} $C\cong {\cal
O}_{\infty}.$ Thus we obtain an injective [homomorphism]{} $\Psi: A\otimes
{\cal O}_{\infty} \to q_{\omega}(M(B)/B).$ Thus the first part of the lemma follows from this and \[Iab3\].
To prove the very last part of the lemma, we may assume that $[\imath(1_C)]\not=[1_C],$ where $\imath$ is as in \[Iab3\]. There is a projection $q\in C$ such that $q\le \imath(1_C)$ and $[q]=[1_C].$ Then $qj(a)=j(a)q$ for all $a\in A.$ Since $\Psi$ is injective, for any nonzero element $a\in A$ and $b\in C,$ $ab=0$ implies that $b=0.$ Thus $qjq$ is injective. To see that $qjq$ is full we note that $q_{\omega}(M(B)/B)$ is purely infinite (see 6.26 in [@Rob]).
\[IIDuct\]
Let $A$ be a separable amenable [$C^*$-algebra]{} and $B$ be a $\sigma$-unital [$C^*$-algebra]{}. Then $KL(A,B)$ is defined to be $KK(A,B)/{\cal T}(A,B),$ where ${\cal T}(A,B)$ is the subgroup of stable approximately trivial extensions (see [@Lnuct]). When $A$ is in ${\cal N},$ then $KL(A,B)=KK(A,B)/Pext(K_*(A), K_*(B))$ (see [@Ro3]).
Let $C_n$ be a commutative [$C^*$-algebra]{} with $K_0(C_n)={\mathbb
Z}/n{\mathbb Z}$ and $K_1(C_n)=0.$ Suppose that $A$ is a [$C^*$-algebra]{}. Then set $K_i(A, {\mathbb Z}/k{\mathbb Z})=K_i(A\otimes C_k)$ (see [@Sch1]). One has the following six-term exact sequence (see [@Sch1]): $$\begin{array}{ccccc}
K_0(A) & \to & K_0(A, {\mathbb Z}/k{\mathbb Z}) & \to & K_1(A)\\
\uparrow_{{\bf k}} & & & & \downarrow_{{\bf k}}\\
K_0(A) & \leftarrow & K_1(A,{\mathbb Z}/k{\mathbb Z}) & \leftarrow
&
K_1(A)\,.\\
\end{array}$$ In [@DL], $K_i(A, {\mathbb Z}/n{\mathbb Z})$ is identified with $KK^i({\mathbb I}_n, A)$ for $i=0,1$.
As in [@DL], we use the notation $${\underline K}(A)=\oplus_{i=0,1, n\in {\mathbb Z}_+}
K_i(A;{\mathbb Z}/n{\mathbb Z}).$$
By ${ \rm Hom}_{\Lambda}({\underline K}(A),{\underline K}(B))$ we mean all [homomorphism]{}s from ${\underline K}(A)$ to ${\underline K}(B)$ which respect the direct sum decomposition and the so-called Bockstein operations (see [@DL]). It follows from the definition in [@DL] that if $x\in KK(A,B),$ then the Kasparov product $KK^{i}({\mathbb I}_n, A)\times x$ gives an element in $KK^i({\mathbb I}_n, B)$ which we identify with ${ \rm
Hom}(K_i(A,{\mathbb Z}/n{\mathbb Z}), K_0(B,{\mathbb Z}/n{\mathbb
Z})).$ Thus one obtains a map $\Gamma: KK(A,B)\to { \rm
Hom}_{\Lambda}(\underline{K}(A),\underline{K}(B)).$ It is shown by Dadarlat and Loring ([@DL]) that if $A$ is in ${\cal N}$ then, for any $\sigma$-unital [$C^*$-algebra]{} $B,$ the map $\Gamma$ is surjective and ${\rm ker}\,\Gamma=Pext(K_*(A),K_*(B)).$ In particular, $$\Gamma: KL(A,B)\to{ \rm Hom}_{\Lambda}({\underline
K}(A),{\underline K}(B))$$ is an isomorphism.
We will use the following theorem. It is a consequence of the uniqueness theorem in 5.6.4 in [@Lnb] which first appeared in the (preprint) of [@Lnjot]. It is proved in [@Lnuct].
[(Theorem 3.9 in [@Lnuct])]{}\[Q\] Let $A$ be a separable unital amenable [$C^*$-algebra]{} and let $B$ a unital [$C^*$-algebra]{}. Suppose that $h_1, h_2: A\to B$ are two unital [homomorphism]{}s such that $$[h_1]=[h_2]\,\,\,\,{\rm in}\,\,\, KL(A,B).$$ Suppose that $h_0: A\to B$ is a full unital monomorphism. Then, for any ${\varepsilon}>0$ and finite subset ${\cal F}\subset A,$ there is an integer $n$ and a unitary $W\in U(M_{n+1}(B))$ such that $$\|W^*{\rm diag}(h_1(a), h_0(a),\cdots, h_0(a))W -{\rm
diag}(h_2(a), h_0(a),\cdots, h_0(a))\|<{\varepsilon}$$ for all $a\in {\cal F}.$
\[IabTTad\] Let $A$ be a separable amenable [$C^*$-algebra]{} and $B$ be a non-unital and $\sigma$-unital simple [$C^*$-algebra]{} with continuous scale. Suppose that $\tau_1, \tau_2: A\to M(B)/B$ be two essential extensions. Then $\tau_1$ and $\tau_2$ are approximately unitarily equivalent if and only if $$[\tau_1]=[\tau_2]\,\,\,{\rm in}\,\,\, KL(A, M(B)/B).$$
Fix an ultrafilter $\omega\in \beta{\mathbb N}\setminus {\mathbb N}.$ Let $\pi: l^{\infty}(M(B)/B)\to q_{\omega}(M(B)/B)$ denote the quotient map. Define $\Psi_i: A\to l^{\infty}(M(B)/B)$ by $\Psi_i(A)=
(\tau_i(a),\tau_i(a),...)$ for $a\in A.$ Set $H_i=\pi\circ \Psi_i.$ We will show that for any ${\varepsilon}>0$ and any finite subset ${\cal F}\subset A$ there is a unitary $w\in q_{\omega}(M(B)/B)$ such that $${\rm ad} w\circ H_1\approx_{{\varepsilon}/2} H_2\,\,\,\,{\rm on}\,\,\, {\cal F}.$$ There are unitaries $u_n\in M(B)/B$ such that $\pi((u_1,u_2,...,u_n,...))=w.$ Therefore $$\liminf_{\omega}\|u_n^*h_1(a)u_n-h_2(a)\|\le {\varepsilon}/2$$ for $a\in {\cal F}.$ Hence, there exists a subset $X\subset \omega$ such that for all $n\in X,$ $$\|u_n^*h_1(a)u_n-h_2(a)\|<{\varepsilon}\,\,\,\,\,{\rm for\,\,\, all}\,\,\, a\in {\cal F}.$$ Then the theorem follows.
Let $j_1$ and $q$ be as in \[Iab5\] associated with $H_1$ and $j_2$ be as in \[Iab5\] associated with $H_2.$ It follows from Theorem \[Q\] that there is a unitary $z\in
M_{K+1}(q_{\omega}(M(B)/B))$ for some integer $K>0$ such that $${\rm ad}\,z\circ (H_1\oplus qj_1q\oplus qj_1q\oplus\cdots \oplus qj_1q)
\approx_{{\varepsilon}/5} H_2\oplus qj_1q\oplus qj_1q\oplus\cdots \oplus qj_1q
\,\,\,{\rm on}\,\,\, {\cal F}.$$ Therefore (by adding $(1-q)j_1(1-q)\oplus\cdots\oplus (1-q)j_1(1-q)$) there is a unitary $v\in M_{K+1}(q_{\omega}(M(B)/B))$ such that $${\rm ad}\,v\circ(H_1\oplus j_1\oplus j_1\oplus \cdots \oplus j_1)\approx_{{\varepsilon}/5}
H_2\oplus j_1\oplus j_1\oplus \cdots \oplus j_1\,\,\,\,{\rm on}\,\,\,{\cal F}.$$ In particular, we may assume that $v^*(1\oplus j(1_A)\oplus\cdots
\oplus j(1_A))v=1.$ It follows from \[Iab5\] and \[NIL5\] that $$H_1\oplus j_1\oplus j_1\oplus \cdots \oplus j_1\, {\sim}_{{\varepsilon}/5}\,
H_1\,\,\,\,{\rm on}\,\,\, {\cal F}.$$ By \[NIL5\], $$j_1\,{\sim}_{{\varepsilon}/5}\, j_2.$$ By \[Iab5\] and \[NIL5\] again, $$H_2\oplus j_1\oplus j_1\oplus \cdots \oplus j_1\,
{\sim}_{2{\varepsilon}/5}\, H_2\,\,\,\,{\rm on}\,\,\, {\cal F}.$$ Combining these inequalities, we obtain a unitary $w\in q_{\omega}(M(B)/B)$ such that $${\rm ad}\circ w \, H_1\approx_{{\varepsilon}} H_2\,\,\,{\rm on}\,\,\,{\cal F}.$$
When $A$ satisfies the UCT, we have the following approximate version of Theorem \[IabTTad\]. This statement is very close to that of Theorem 6.3 of [@Lnpro].
\[IabTadd\] Let $A$ be a separable amenable [$C^*$-algebra]{} in ${\cal N}.$ For any ${\varepsilon}>0$ and finite subset ${\cal F}\subset A,$ there exists a finite subset ${\cal P}\subset {\underline{K}}(A)$ satisfying the following: if $h_1, h_2, h_3: A\to C$ are three [homomorphism]{}s, where $C$ is a unital purely infinite simple [$C^*$-algebra]{} such that $$[h_1]|_{\cal P}=[h_2]|_{\cal P}$$ then there is an integer $n>0$ and a unitary $u\in M_{n+1}(C)$ such that $${\rm ad}\circ (h_1\oplus h_3\oplus h_3\oplus \cdots h_3)\approx_{{\varepsilon}}
h_2\oplus h_3\oplus h_3\oplus \cdots h_3)\,\,\,{\rm on}\,\,\,{\cal F},$$ where $h_3$ repeats $n$ times.
Let $\{{\cal P}_n\}$ be a sequence of finite subsets of ${\underline{K}}(A)$ such that $\cup_{n=1}^{\infty} {\cal
P}_n={\underline{K}}(A).$ Suppose that there are three sequences of [homomorphism]{}s $\phi_n, \psi_n, f_n: A\to C_n,$ where $C_n$ is a sequence of unital purely infinite simple [$C^*$-algebra]{}s such that $$[\phi_n]|_{{\cal P}_n}=[\psi_n]|_{{\cal P}_n},\,\,\,n=1,2,....$$ It suffices to show that, there exists $N>0$ and $K>0$ such that when $n\ge N$ there are unitaries $u_n\in M_{K+1}(C_n)$ satisfying the following: $${\rm ad} u_n(\phi_n\oplus f_n\oplus f_n\oplus \cdots\oplus
f_n)\approx_{{\varepsilon}} (\psi_n\oplus f_n\oplus f_n\oplus\cdots\oplus
f_n) \,\,\,\,{\rm on}\,\,\,{\cal F},$$ where $f_n$ repeats $K$ times. Let $H_1=\{\phi_n\}, H_2=\{\psi_n\}$ and $H_3=\{f_n\}$ be [homomorphism]{}s from $A$ into $l^{\infty}(\{C_n\})$ and let ${\bar H_i}=\pi\circ H_i,$ where $\pi: l^{\infty}(\{C_n\})\to
l^{\infty}(\{C_n\})/c_0(\{C_n\})$ is the quotient map, $i=1,2,3.$ So it suffices to show that there exists $K>0$ such that there is a unitary $U\in M_{K+1}(l^{\infty}(\{C_n\})/c_0(\{C_n\})$ such that $${\rm ad} U({\bar H}_1\oplus {\bar H}_3\oplus {\bar H}_3\oplus\cdots \oplus
{\bar H}_3)
\approx_{{\varepsilon}} {\bar H}_2\oplus {\bar H}_3\oplus {\bar H}_3\oplus\cdots \oplus
{\bar H}_3
\,\,\,{\rm on}\,\,\, {\cal F},$$ where ${\bar H}_3$ repeats $K$ times.
Since each $C_n$ is a unital purely infinite simple [$C^*$-algebra]{}, it follows from 6.5 [@Lnuct] that ${\bar H_3}: A\to l^{\infty}(\{C_n\})/c_0(\{C_n\})$ is full. So the theorem follows from Theorem \[Q\] if we can show that $$[{\bar H}_1]=[{\bar H}_2] \,\,\,{\rm in }\,\,\,
KL(A, l^{\infty}(\{C_n\})/c_0(\{C_n\}).$$ It follows from Corollary 2.1 in [@GL] that, if each $C_n$ is purely infinite and simple, $$K_i(l^{\infty}(\{C_n\}))=\prod_n K_i(C_n),\,\,\,i=0,1,$$ $$K_i(l^{\infty}(\{C_n\}),{\mathbb Z}/k{\mathbb Z}))\subset
\prod_n K_i(C_n, {\mathbb Z}/k{\mathbb Z}),i=0,1, k=2,3,...$$ and $$K_i(l^{\infty}(\{C_n\})/c_0(\{C_n\})=\prod_nK_i(C_n)/\oplus_nK_i(C_n), \,\,\,
i=0,1,$$ $$K_i(l^{\infty}(\{C_n\})/c_0(\{C_n\}), {\mathbb Z}/k{\mathbb Z}))
\subset \prod_n K_i(C_n, {\mathbb Z}/k{\mathbb Z})/\oplus_n
K_i(C_n, {\mathbb Z}/k{\mathbb Z}),\,\,\, k=2,3,....$$ Thus, since $[H_1]|_{{\cal P}_n}=[H_2]|_{{\cal P}_n}$ for each $n,$ we conclude from the above computation that $$[{\bar H}_1]=[{\bar H}_2] \,\,\,\,{\rm in}\,\,\,
Hom_{\Lambda}({\underline{K}}(A), {\underline{K}}(l^{\infty}(\{C_n\})/c_0(\{C_n\})).$$ Therefore the theorem follows.
The following is an approximate version of Theorem \[IabTTad\].
\[IabT1\] Let $A$ be a unital separable amenable [$C^*$-algebra]{} in ${\cal N}$ and $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with a continuous scale. Suppose that $h_1, h_2: A\to M(B)/B$ are two monomorphisms. For any ${\varepsilon}>0$ and any finite subset ${\cal F}\subset A,$ there exists a finite subset ${\cal P}\subset {\underline{K}(A)}$ satisfying the following: if $$[h_1]|_{{\cal P}}=[h_2]|_{{\cal P}}$$ then there exists a unitary $u\in M(B)/B$ such that $${\rm ad}\circ h_1\approx_{{\varepsilon}} h_2\,\,\,\,\,{\it on}\,\,\, {\cal F}.$$
The proof is exactly the same as that of \[IabTTad\] but applying \[IabTadd\] instead of \[Q\].
\[Iabcor\] Let $A$ be a separable amenable [$C^*$-algebra]{} in ${\cal N}$ and $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with a continuous scale. Let $\tau_1, \tau_2: A\to M(B)/B$ be two essential extensions. Then there exists a sequence of unitaries $u_n\in M(B)/B$ such that $$\lim_{n\to\infty}{\rm ad} u_n\circ \tau_1(a)=\tau_2(a)\,\,\,\,{\it
for\,\,\, all}\,\,\, a\in A$$ if and only if $[\tau_1]=[\tau_2]$ in $KL(A, M(B)/B).$
${\bf Exp}_{ap}(A,B)$
=====================
\[IVD1\] [An essential extension $\tau: A\to M(C)/C$ is said to be [*approximately trivial*]{} if there is a sequence of trivial extensions $\tau_n: A\to M(C)/C$ such that $\tau(a)=\lim_{n\to\infty}\tau_n(a)$ for all $a\in A.$ Denote by ${\bf Ext}_{ap}(A,B)$ the set of approximately unitarily equivalent classes of essential extensions. ]{}
Let $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with a continuous scale and $A\in {\cal N}.$ In this section we will classify essential extensions of $A$ by $B:$ $$0\to B\to E\to A\to 0$$ up to approximately unitary equivalence.
\[IIL2p\] Let $B$ be a unital purely infinite simple [$C^*$-algebra]{} and let $G_i$ be a countable subgroup of $K_i(B)$ $(i=0,1).$ There exists a unital separable purely infinite simple [$C^*$-algebra]{} $B_0\subset B$ such that $K_i(B_0)\supset G_i$ and $j_{*i}={{\rm id}}_{K_i(B_0)},$ where $j:
B_0\to B$ is the embedding.
Since $B$ is purely infinite, all elements in $K_0(B)$ and in $K_1(B)$ can be represented by projections and unitaries in $B,$ respectively. Let $p_1,...,p_n,...$ be projections in $B$ and $u_1,u_2,...,u_n,...$ be unitaries in $B$ such that $\{p_n\}$ and $\{u_n\}$ generates of $G_0$ and $G_1,$ respectively. Let $B_1$ be a unital separable purely infinite simple [$C^*$-algebra]{} containing $\{p_n\}$ and $\{u_n\}$ (see the proof of 7.4 in [@Lnpro]). Note that $K_i(B_1)$ is countable. The embedding $j_1: B_1\to B$ gives [homomorphism]{}s $(j_1)_{*i}: K_0(B_1)\to K_i(B).$ Let $F_{1,i}$ be the subgroup of $K_0(B_1)$ generated by $\{p_n\}$ and $\{u_n\},$ respectively. It is clear that $(j_1)_{*i}$ is injective on $F_{1,i},$ $i=0,1.$ In particular, the image of $(j_1)_{*i}$ contains $G_i,$ $i=0,1.$ Let $N_{1,i}'={\rm ker}(j_1)_{*i}$ and let $N_{1,i}$ be the set of all projections (if $i=0$), or unitaries (if $i=1$) in $B_1$ which have images in $N_{1, i}'.$ Let $\{p_{1,n}\}$ be a dense subset of $N_{1,0}$ and $\{u_{1,n}\}$ be a dense subset of $N_{1,1},$ respectively. Fix a nonzero projection $e\in
B_1$ such that $[e]=0$ in $K_0(B).$ For each $p_{1,n},$ there exists a partial isometry $w_{1,n}\in B$ such that $e=w_{1,n}^*w_{1,n}$ and $w_{1,n}w_{1,n}^*=p_{1,n},$ $n=1,2,....$ For each $u_{1,n},$ there are unitaries $z_{1,n,k}\in B,$ $
k=1,2,...,m(n)$ such that $$\|z_{1,n,1}-1\|<1/2, \|z_{1,n,m(n)}-u_{1,n}\|<1/2 {\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\|z_{1,n,k}-z_{1,n,k+1}\|<1/2,$$ $k=1,2,...,m(n),$ $n=1,2,....$ Let $B_2$ be a separable unital purely infinite simple [$C^*$-algebra]{} containing $B_1$ and all $\{w_{1,n}\}$ and $\{z_{1,n,k}\}.$ Note that if $p\in B_1$ is a projection and $[p]\in N_{1,0}$ then $[p]=0$ in $K_0(B_2).$ Similarly, if $u\in B_1$ and $[u]\in N_{1,1},$ then $[u]=0$ in $B_2.$ Suppose that $B_l$ has been constructed. Let $j_l: B_l\to
B$ be the embedding. Let $N_{l, i}={\rm ker}(j_l)_{*i},$ $i=0,1.$ As before, we obtain a unital separable purely finite simple [$C^*$-algebra]{} $B_{l+1}\supset B_l$ such that every projection $p\in B_l$ with $[p]\in N_{l,0}$ has the property that $[p]=0$ in $K_0(B_{l+1}),$ and every unitary $u\in B_l$ with $[u]\in N_{l,1}$ has the property that $[u]=0$ in $K_1(B_{l+1}).$ Let $B_0$ be the closure of $\bigcup_{l=1}^{\infty} B_l.$ Since each $B_l$ is purely infinite and simple, so is $B_0.$ Note also that $B_0$ is separable. Let $j: B_0\to B$ be the embedding.
We claim that $j_{*i}$ is injective. Suppose that $p\in B_0$ is a projection such that $[p]\in {\rm ker}j_{*0}$ and $[p]\not=0$ in $B_0.$ Without loss of generality, we may assume that $p\in B_l$ for some large integer $l.$ Then $[p]$ must be in the ${\rm ker}(j_l)_{*0}.$ By the construction, $[p]=0$ in $K_0(B_{l+1}).$ This would imply that $[p]=0$ in $K_0(B_0).$ Thus $j_{*0}$ is injective. An exactly same argument shows that $j_{*1}$ is also injective. The lemma then follows.
\[IILsix\] Let $B$ be a unital purely infinite simple [$C^*$-algebra]{}. Suppose that $G_i\subset K_i(B)$ and $F_i(k)\subset K_i(B,{\mathbb Z}/k{\mathbb Z})$ are countable subgroups such that the image of $F_i(k)\subset K_i(B,{\mathbb Z}/k{\mathbb Z})$ in $K_{i-1}(B)$ is contained in $G_{i-1}$ $(i=0,1, k=2,3,...)$. Then there exists a separable unital purely infinite simple [$C^*$-algebra]{} $C\subset B$ such that $K_i(C)\supset G_i,$ $K_i(C,{\mathbb Z}/k{\mathbb Z})
\supset F_i(k)$ and the embedding $j: C\to B$ induces an injective map $j_{*i}: K_0(C)\to K_i(B)$ and an injective map $j_*: K_i(C, {\mathbb Z}/k{\mathbb Z})\to K_i(B,{\mathbb
Z}/k{\mathbb Z}),$ $k=2,3,....$
It follows from \[IIL2p\] that there is a separable unital purely infinite simple [$C^*$-algebra]{} $C_1$ such that $K_0(C_1)\supset G_0$ and $K_1(C_1)\supset G_1$ and $j$ induces an identity map on $K_0(C_1)$ and $K_1(C_1),$ where $j: C\to B$ is the embedding. Fix $k,$ let $\{x\in K_i(C_1): kx=0\}=\{g_1^{(i)},g_2^{(i)},...,\}.$ Suppose that $\{s_1^{(i)},s_2^{(i)},...,\}$ is a subset of $K_i(B,{\mathbb Z}/k{\mathbb Z})$ such that the map from $K_i(B,{\mathbb Z}/k{\mathbb Z})$ to $K_{i-1}(B)$ maps $s_j^{(i)}$ to $g_j^{(i)}.$ For each $z^{(i)}\in K_i(C_1, {\mathbb
Z}/k{\mathbb Z}),$ there is $s_j^{(i)}$ such that $z^{(i)}-s_j^{(i)}\in K_i(B)/kK_i(B).$ Since $K_i(C_1)$ is countable, the set of all possible $z^{(i)}-s_j^{(i)}$ is countable. Thus one obtains a countable subgroup $G_i^{(')}$ which contains $K_i(C_1)$ such that $G_i^{(')}/kK_i(B)$ contains the above the mentioned countable set as well as $F_i(k)\cap (
K_i(B)/kK_i(B))$ for each $k.$ Since countably many countable sets is still countable, we obtain a countable subgroup $G_i^{(2)}\subset K_i(B)$ such that $G_i^{(2)}$ contains $G_i^{(')}$ and $kK_i(B)\cap G_i^{(2)}= kG_i^{(2)},$ $k=1,2,...,$ and $i=0,1.$ Note also $F_i(k)\cap (K_i(B)/kK_i(B))\subset
G_i^{(2)}/kK_i(B).$ By applying \[IIL2p\], we obtain a separable purely infinite simple [$C^*$-algebra]{} $C_2\supset C_1$ such that $K_i(C_2)\supset G_i^{(2)}$ and embedding from $C_2$ to $B$ gives an injective map on $K_i(C_2),$ $i=0,1.$ Repeating what we have done above, we obtain an increasing sequence of countable subgroups $G_i^{(n)}\subset K_i(B)$ such that $G_i^{(n)}\cap
kK_i(B)=kG_i^{(n)}$ for all $k$ and $i=0,1$ and an increasing sequence of separable purely infinite simple [$C^*$-subalgebra]{}s $C_n$ such that $K_i(C_n)\supset G_i^{(n)}$ and the embedding from $C_n$ into $B$ gives an injective map on $K_i(C_n),$ $i=0,1,$ and $n=1,2,....$ Moreover $F_i^{(k)}\cap (K_i(B)/kK_i(B))\subset
K_i(C_n)/kK_i(B).$ Let $C$ denote the closure of $\bigcup_nC_n$ and $j: C\to B$ be the embedding. Then $C$ is a separable purely infinite simple [$C^*$-algebra]{} and $j_{*i}$ is an injective map, $i=0,1.$ We claim that $K_i(C)\cap kK_i(B)=kK_i(C),$ $k=1,2,...,$ and $i=0,1.$ Note that $K_i(C)=\cup_nG_i^{(n)}.$ Since $G_i^{(n)}\cap
kK_i(B)=kG_i^{(n)} \subset kK_i(C),$ we see that $K_i(C)\cap
kK_i(B)=kK_i(C),$ $i=0,1.$ Thus $K_i(C)/kK_i(C)=K_i(C)/kK_i(B).$ Since $K_i(C)/kK_i(B)\supset F_i^{(k)}\cap(K_i(B)/kK_0(B)),$ we conclude also that $K_i(C,{\mathbb Z}/k{\mathbb Z})$ contains $F_i(k).$ Since $j_{*0}$ is injective, $j$ induces an injective map from $K_0(C)/kK_0(C)$ into $K_0(B)/kK_0(B)$ for all integer $k\ge 1.$ Using this fact and the fact that $j_{*i}: K_i(C)\to
K_i(B)$ is injective by chasing the following commutative diagram, $${\small \put(-160,0){$K_0(C)$} \put(0,0){$K_0(C,{\bf Z}/k{\bf
Z})$} \put(180,0){$K_1(C)$} \put(-85,-40){$K_0(B)$}
\put(0,-40){$K_0(B,{\bf Z}/k{\bf Z})$} \put(105,-40){$K_1(B)$}
\put(-85, -70){$K_0(B)$} \put(0,-70){$K_1(B, {\bf Z}/k{\bf Z})$}
\put(105,-70){$K_1(B)$} \put(-160,-110){$K_0(C)$}
\put(0,-110){$K_1(C,{\bf Z}/k{\bf Z})$} \put(180,-110){$K_1(C)$}
\put(-120, 2){\vector(1,0){95}} \put(70,1){\vector(1,0){95}}
\put(-123,-3){\vector(1,-1){30}} \put(30,-3){\vector(0,-1){25}}
\put(180,-2){\vector(-1,-1){30}} \put(-45,-38){\vector(1,0){35}}
\put(70,-38){\vector(1,0){25}} \put(-147, -90){\vector(0,1){85}}
\put(-75,-60){\vector(0,1){15}} \put(115, -45){\vector(0,-1){15}}
\put(190,-7){\vector(0,-1){85}} \put(-7,-68){\vector(-1,0){35}}
\put(95,-68){\vector(-1,0){25}} \put(-123,-102){\vector(1,1){30}}
\put(175, -105){\vector(-1,1){30}} \put(30,-100){\vector(0,1){25
}} \put(-5, -108){\vector(-1,0){100}}
\put(170,-108){\vector(-1,0){95}} \put(-112,-12){$j_{*0}$}
\put(15, -15){$j_* $} \put(150, -14){$j_{*1}$}
\put(-130,-92){$j_{*0}$} \put(15, -88){$j_*$} \put(160,
-88){$j_{*1}$} }$$ one sees that $j$ induces an injective map from $K_i(C, {\mathbb
Z}/k{\mathbb Z}) $ to $K_i(B, {\mathbb Z}/k{\mathbb Z}).$
\[IIT2\] Let $A$ be a unital separable amenable [$C^*$-algebra]{} in ${\cal N}$ and $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with a continuous scale. Then, for any $x\in KL(A, M(B)/B),$ there exists a monomorphism $h: A\to M(B)/B$ such that $[h]=x.$
Put $Q=M(B)/B.$ Since $A$ satisfies the UCT, we may view $x$ as an element in ${\rm Hom}_{\Lambda}(\underline{K}(A), \underline{K}(Q)).$ Note that $K_i(A)$ is a countable abelian group ($i=0,1$). Let $G_{0}^{(i)}=\gamma(x)(K_i(A)),$ $i=0,1,$ where $\gamma: {\rm Hom}_{\Lambda}(\underline{K}(A), \underline{K}(Q))
\to {\rm Hom}(K_*(A),K_*(Q))$ is the surjective map. Then $G_{0}^{(i)}$ is a countable subgroup of $K_i(Q),$ $i=0,1.$ Consider the following commutative diagram: $${\small \put(-160,0){$K_0(A)$} \put(0,0){$K_0(A,{\bf Z}/k{\bf
Z})$} \put(180,0){$K_1(A)$} \put(-85,-40){$K_0(Q)$}
\put(0,-40){$K_0(Q,{\bf Z}/k{\bf Z})$} \put(105,-40){$K_1(Q)$}
\put(-85, -70){$K_0(Q)$} \put(0,-70){$K_1(Q, {\bf Z}/k{\bf Z})$}
\put(105,-70){$K_1(Q)$} \put(-160,-110){$K_0(A)$}
\put(0,-110){$K_1(A,{\bf Z}/k{\bf Z})$} \put(180,-110){$K_1(A)$}.
\put(-120, 2){\vector(1,0){95}} \put(70,1){\vector(1,0){95}}
\put(-123,-3){\vector(1,-1){30}} \put(30,-3){\vector(0,-1){25}}
\put(180,-2){\vector(-1,-1){30}} \put(-45,-38){\vector(1,0){35}}
\put(70,-38){\vector(1,0){25}} \put(-147, -90){\vector(0,1){85}}
\put(-75,-60){\vector(0,1){15}} \put(115, -45){\vector(0,-1){15}}
\put(190,-7){\vector(0,-1){85}} \put(-7,-68){\vector(-1,0){35}}
\put(95,-68){\vector(-1,0){25}} \put(-123,-102){\vector(1,1){30}}
\put(175, -105){\vector(-1,1){30}} \put(30,-100){\vector(0,1){25}}
\put(-5, -108){\vector(-1,0){100}}
\put(170,-108){\vector(-1,0){95}} \put(-111,-14){$\gamma(x)$}
\put(15, -15){$\times x $} \put(148, -14){$\gamma(x)$}
\put(-130,-90){$\gamma(x)$} \put(15, -88){$\times x$} \put(160,
-88){$\gamma(x)$} }$$ It follows from \[IILsix\] that there is a unital purely infinite simple [$C^*$-algebra]{} $C\subset Q$ such that $K_i(C)\subset G_0^{(i)},$ $K_i(C)\cap kK_i(Q)=kK_i(C),$ $k=1,2,...,$ and $i=0,1,$ and the embedding $j: C\to Q$ induces injective maps on $K_i(C)$ as well as on $K_i(C,{\mathbb Z}/k{\mathbb Z})$ for all $k$ and $i=0,1.$ Moreover $K_i(C,{\mathbb Z}/k{\mathbb Z})\supset
(\times x)(K_i(A,{\mathbb Z}/k{\mathbb Z}))$ for $k=1,2,...$ and $i=0,1.$ We have the following commutative diagram: $${\small \put(-160,0){$K_0(A)$} \put(0,0){$K_0(A,{\bf Z}/k{\bf
Z})$} \put(180,0){$K_1(A)$} \put(-85,-40){$K_0(C)$}
\put(0,-40){$K_0(C,{\bf Z}/k{\bf Z})$} \put(105,-40){$K_1(C)$}
\put(-85, -70){$K_0(C)$} \put(0,-70){$K_1(C, {\bf Z}/k{\bf Z})$}
\put(105,-70){$K_1(C)$} \put(-160,-110){$K_0(A)$}
\put(0,-110){$K_1(A,{\bf Z}/k{\bf Z})$} \put(180,-110){$K_1(A)$}
\put(-120, 2){\vector(1,0){95}} \put(70,1){\vector(1,0){95}}
\put(-123,-3){\vector(1,-1){30}}
\put(180,-2){\vector(-1,-1){30}} \put(-45,-38){\vector(1,0){35}}
\put(70,-38){\vector(1,0){25}} \put(-147, -90){\vector(0,1){85}}
\put(-75,-60){\vector(0,1){15}} \put(115, -45){\vector(0,-1){15}}
\put(190,-7){\vector(0,-1){85}} \put(-7,-68){\vector(-1,0){35}}
\put(95,-68){\vector(-1,0){25}} \put(-123,-102){\vector(1,1){30}}
\put(175, -105){\vector(-1,1){30}}
\put(-5, -108){\vector(-1,0){100}}
\put(170,-108){\vector(-1,0){95}} \put(-112,-12){$\gamma(x)$}
\put(145, -12){$\gamma(x)$}
\put(-135,-92){$\gamma(x)$} \put(160, -88){$\gamma(x)$} }$$ We will add two more maps on the above diagram. From the fact that the image of $K_i(A, {\mathbb Z}/k{\mathbb Z})$ under $\times x$ is contained in $K_i(C,{\mathbb Z}/k{\mathbb
Z}),$ ($k=2,3,...,$ $i=0,1$), we obtain two maps $\beta_i:K_i(A, {\mathbb Z}/k{\mathbb
Z})\to K_i(C,{\mathbb Z}/k{\mathbb
Z}),$ $k=2,3,...,$ $i=0,1$ such that $j_*\circ \beta_i=\times x$ and obtain the following commutative diagram: $${\small \put(-160,0){$K_0(A)$} \put(0,0){$K_0(A,{\bf Z}/k{\bf
Z})$} \put(180,0){$K_1(A)$} \put(-85,-40){$K_0(Q)$}
\put(0,-40){$K_0(C,{\bf Z}/k{\bf Z})$} \put(105,-40){$K_1(C)$}
\put(-85, -70){$K_0(C)$} \put(0,-70){$K_1(C, {\bf Z}/k{\bf Z})$}
\put(105,-70){$K_1(C)$} \put(-160,-110){$K_0(A)$}
\put(0,-110){$K_1(A,{\bf Z}/k{\bf Z})$} \put(180,-110){$K_1(A)$}
\put(-120, 2){\vector(1,0){95}} \put(70,1){\vector(1,0){95}}
\put(-123,-3){\vector(1,-1){30}} \put(30,-3){\vector(0,-1){25}}
\put(180,-2){\vector(-1,-1){30}} \put(-45,-38){\vector(1,0){35}}
\put(70,-38){\vector(1,0){25}} \put(-147, -90){\vector(0,1){85}}
\put(-75,-60){\vector(0,1){15}} \put(115, -45){\vector(0,-1){15}}
\put(190,-7){\vector(0,-1){85}} \put(-7,-68){\vector(-1,0){35}}
\put(95,-68){\vector(-1,0){25}} \put(-123,-102){\vector(1,1){30}}
\put(175, -105){\vector(-1,1){30}} \put(30,-100
){\vector(0,1){25}} \put(-5, -108){\vector(-1,0){100}}
\put(170,-108){\vector(-1,0){95}} \put(-112,-12){$\gamma(x)$}
\put(15, -15){$\beta_0 $} \put(145, -12){$\gamma(x)$}
\put(-135,-92){$\gamma(x)$} \put(15, -88){$\beta_1$} \put(160,
-88){$\gamma(x)$} }$$ Consider the following commutative diagram: $$\begin{array}{ccccccc}
\to & K_i(A,{\mathbb Z}/mn{\mathbb Z}) & \to & K_i(A, {\mathbb Z}/n{\mathbb Z})&
\to & K_{i-1}(A, {\mathbb Z}/m{\mathbb Z}) &\to \\
& \downarrow & & \downarrow && \downarrow\\
\to & K_i(Q,{\mathbb Z}/mn{\mathbb Z}) & \to & K_i(Q, {\mathbb Z}/n{\mathbb Z})&
\to & K_{i-1}(Q, {\mathbb Z}/m{\mathbb Z}) &\to \cr
\end{array}$$ Since $j_*\circ \beta_i=\times x$ and all vertical maps in the following diagram is injective $$\begin{array}{ccccccc}
\to & K_i(C,{\mathbb Z}/mn{\mathbb Z}) & \to & K_i(C, {\mathbb Z}/n{\mathbb Z})&
\to & K_{i-1}(C, {\mathbb Z}/m{\mathbb Z}) &\to \\
& \downarrow & & \downarrow && \downarrow\\
\to & K_i(Q,{\mathbb Z}/mn{\mathbb Z}) & \to & K_i(Q, {\mathbb Z}/n{\mathbb Z})&
\to & K_{i-1}(Q, {\mathbb Z}/m{\mathbb Z}) &\to \cr,
\end{array}$$ we obtain the following commutative diagram: $$\begin{array}{ccccccc}
\to & K_i(A,{\mathbb Z}/mn{\mathbb Z}) & \to & K_i(A, {\mathbb Z}/n{\mathbb Z})&
\to & K_{i-1}(A, {\mathbb Z}/m{\mathbb Z}) &\to \\
& \downarrow & & \downarrow && \downarrow\\
\to & K_i(C,{\mathbb Z}/mn{\mathbb Z}) & \to & K_i(C, {\mathbb Z}/n{\mathbb Z})&
\to & K_{i-1}(C, {\mathbb Z}/m{\mathbb Z}) &\to \cr
\end{array}$$ Thus we obtain an element $y\in KL(A, C)$ such that $y\times
[j]=x.$ Since $A$ satisfies the UCT, one checks that $KL(A,
C)=KL(A\otimes {\cal O}_{\infty}, C).$ It follows from 6.6 and 6.7 in [@Lnsemi] that there exists a [homomorphism]{} $\phi: A\otimes {\cal
O}_{\infty}\to C\otimes {\cal K}$ such that $[\phi]=y.$ Define $\psi=\phi|_{A\otimes 1}.$ It is then easy to check that $[\psi]=y.$ Since $A$ is unital, we may assume that the image of $\psi$ is in $M_m(C)$ for some integer $m\ge 1.$ Since $C$ is a unital purely infinite simple [$C^*$-algebra]{}, $1_m$ is equivalent to a projection in $C.$ Thus we may further assume that $\psi$ maps $A$ into $C.$ Put $h_1=j\circ \psi.$ To obtain a monomorphism, we note that there is an embedding $\imath: A\to {\cal O}_2$ (see Theorem 2.8 in [@KP]). Since $M(B)/B$ is purely infinite, we obtain a monomorphism $\psi: {\cal O}_2\to M(B)/B.$ Let $e=\psi(1_{{\cal
O}_ 2}).$ There is a partial isometry $w\in M_2(M(B)/B)$ such that $w^*w=1_{M(B)/B}$ and $ww^*=1\oplus e.$ Define $h=w^*(h_1\oplus
\psi\circ \imath)w.$ One checks that $[h]=[h_1]$ and $h$ is a monomorphism.
\[IITM\] Let $A$ be a separable amenable [$C^*$-algebra]{} in ${\cal N}$ and $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with a continuous scale. Then there is a bijection: $$\Gamma: {\bf Ext}_{ap}(A,B) \to KL(A, M(B)/B).$$
This follows immediately from \[Iabcor\].
\[IIC1\] Let $A$ be a unital separable amenable [$C^*$-algebra]{} satisfying the UCT and $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with a continuous scale. Let $\tau$ be a unital essential extension and $\psi: A\to
M(B)$ be a [contractive completely positive linear map]{} such that $\pi\circ \psi=\tau.$ Suppose that $[\tau]=[t]$ in $KL(A, M(B))$ for some trivial extension $t.$ Then, there exists a sequence of monomorphisms $h_n: A\to M(B)$ such that $$\lim_{\to\infty} \pi\circ (h_n(a)-\psi(a))=0$$ for all $a\in A.$
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx AF xxxxxxxxxxxxxx
\[IIEx\] [Let $B$ be a non-unital separable simple [$C^*$-algebra]{} with finite trace and $K_0(B)={\mathbb Q}.$ So $B$ has a continuous scale and $K_0(M(B)/B)={\mathbb R}/{\mathbb Q}.$ Let $\xi\not=0,1$ in ${\mathbb R}/{\mathbb Q}.$ Suppose that $A$ is a unital separable amenable [$C^*$-algebra]{} which satisfies the UCT and suppose that there are two nonzero elements $g_1, g_2$ in $K_0(A)$ such that $[1_A]=g_1$ and the subgroup generated $g_1$ and $g_2$ is not cyclic. Since ${\mathbb R}/{\mathbb Q}$ is divisible, there is a group [homomorphism]{} $\alpha: K_0(A)\to K_0(M(B)/B)$ such that $\alpha(g_1)=1$ and $\alpha(g_2)=\xi.$ It follows from \[IITM\] that there is an essential unital extension $\tau_{\xi}:
A\to M(B)/B$ such that $(\tau_{\xi})_{*0}=\alpha$ and $(\tau_{\xi})_{*1}=0.$ Since $K_0(B)={\mathbb Q}$ and $K_1(B)=0,$ we compute that $[\tau_{\xi}]=0$ in $KK(A, B)$ for any such $\xi.$ However, $[\tau_{\xi}]\not=[\tau_{\xi'}]$ in $KL(A,M(B)/B)$ if $\xi\not=\xi'.$ This shows that there are uncountably many non-equivalent essential extensions which represent the same element in $KK^1(A,B).$ This example shows how $KL(A, M(B)/B))$ can be used to compute ${\bf Ext}_{ap}(A,B),$ while $KK^1(A,B)$ fails.]{}
Examples
========
Theorem \[IITM\] provides a complete classification of ${\bf Ext}_{ap}(A,B).$ However, it is not immediately clear which elements in $KL(A,M(B)/B)$ give an approximate trivial extension or a quasidiagonal extension. It turns out it is rather a complicated problem. First of all from item (1) below, it could be the case that there are no essential extensions which are approximately trivial. Second, item (2) and item (3) below show that that $[\tau]=0$ in $KL(A,M(B)/B)$ does not imply that $\tau$ is an approximately trivial extension. In this section we will discuss these problems.
[*In this section $B$ is a non-unital and $\sigma$-unital simple [$C^*$-algebra]{} with real rank zero, stable rank one, weakly unperforated $K_0(B)$ and with a continuous scale.*]{}
We will show the following:
\(1) There are $A$ and $B$ such that there are no trivial essential extensions of $A$ by $B.$
\(2) There are essential extensions $\tau$ such that $[\tau]=0$ in $KL(A, M(B)/B)$ which are not limits of trivial extensions.
\(3) For the same $A$ and $B$ as in (2), there are trivial essential extensions $\tau$ such that $[\tau]\not=0.$
\[IIe1\]
Let $A$ be a unital separable amenable [$C^*$-algebra]{} and $B$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} (with a continuous scale). It is possible that there are no essential trivial extensions of the form: $$0\to B\to E\to A\to 0.$$ For example, let $A={\cal O}_n$ ($n\ge 2$) and $B$ be any non-unital AF-algebra with a continuous scale. There are many extensions of $A$ by $B.$ This is because $M(B)/B$ is purely infinite simple and one can easily find monomorphisms from ${\cal
O}_n$ into $M(B)/B.$ But none of them are splitting. In fact there is no monomorphism $h: A\to M(B).$ Since $M(B)$ admits a tracial state, $h(A)$ would have a tracial state too. But this is impossible.
From this example, one sees clearly that for many [$C^*$-algebra]{}s $A$ there is no single essential trivial extension of $A$ by $B.$ Therefore some restriction on $A$ is needed to guarantee that there are trivial essential extensions.
\[IIILemb\] Let $A$ be a unital AF-algebra such that there is a positive [homomorphism]{} $\alpha: K_0(A)\to Aff(T).$ Then there exists a [homomorphism]{} $h:
A\to M(B)$ such that $h_{*0}=\alpha$ and $h(A)\cap B=\{0\}.$
It is easy to see and known that the lemma holds for the case that $A$ is finite dimensional. Let $F$ be a finite dimensional [$C^*$-algebra]{}. Suppose that $h_1, h_2: F\to M(B)$ are two [homomorphism]{}s such that $h_i(F)\cap B=\{0\}.$ Suppose also that $(h_1)_{*0}=(h_2)_{*0}.$ Then by (2) in \[0T1\], it is standard to see that $h_1$ and $h_2$ are unitarily equivalent.
Now let $A$ be the closure of $\bigcup_{n=1}^{\infty} A_n,$ where $A_n\subset A_{n+1}$ and ${\rm dim} A_n<\infty.$ Denote by $j_n$ the embedding from $A_n$ to $A.$ Let $\alpha_n=\alpha\circ (j_n)_{*0}.$ Let $h_1: A_1\to M(A)$ be such that $h_1(A_1)\cap B=\{0\}$ and $(h_1)_{*0}=\alpha_1.$ Suppose that $h_m: A_m\to M(B)$ has been defined such that $h_m|_{A_j}=h_j$ for $j<m,$ $h_m(A_m)\cap B=\{0\}$ and $(h_m)_{*0}=\alpha_m.$ Let $\phi_{m+1}: A_{m+1}\to M(B)$ be such that $\phi_{m+1}(A_{m+1})\cap M(B)=\{0\}$ and $(\phi_{m+1})_{*0}
=\alpha_{m+1}.$ Let $\imath_{m}: A_m\to A_{m+1}$ be the embedding. Then we have $\alpha_m=(h_m)_{*0}=(\phi_{m+1}\circ \imath_{m}).$ From what we have shown there is a unitary $u_{m+1}\in M(B)$ such that $${\rm ad}\circ \phi_{m+1}\circ \imath_{m}=h_m.$$ Put $h_{m+1}={\rm ad}\circ \phi_{m+1}.$ Then we have the following commutative diagram: $$\begin{array}{ccccccccc}
A_1 &\to &A_2& \to & A_3& \to &\cdots & A\\
\downarrow_{h_1} & &\downarrow_{h_2} &&\downarrow_{h_3} &\cdots\\
M(B)&{\stackrel{{{\rm id}}_M(B)}{\to}} & M(B)&{\stackrel{{{\rm id}}_M(B)}{\to}}&
M(B) & {\stackrel{{{\rm id}}_M(B)}{\to}} &\cdots &M(B)\cr
\end{array}$$ It follows that there is a monomorphism $h: A\to M(B)$ such that $h(A)\cap B=\{0\}.$
\[IIITtriv\] Let $A$ be a separable amenable [$C^*$-algebra]{} satisfying the UCT. Suppose that $A$ can be embedded into a unital simple AF-algebra. Then for any $B$ there exists an essential trivial extension $\tau$ of $A$ by $B.$
Suppose that $C$ is a unital simple AF-algebra and $j: A\to C$ is an embedding. Let $t$ be a normalized trace on $C.$ Define $\beta:
K_0(C)\to Aff(T)$ by $\beta([p])=t(p)[1_{M(B)/B}]$ for projection $p\in C.$ Then $\beta$ is a positive [homomorphism]{}. It follows from \[IIILemb\] that there is a monomorphism $h: C\to M(B)$ such that $h_{*0}=\beta$ and $h(C)\cap B=\{0\}.$ Define $\phi: A\to
M(B)$ by $\phi=h\circ j.$ One sees that $\phi$ give an essential trivial extension of $A$ by $B.$
Suppose that there are trivial essential extensions of $A$ by $B.$ One would like to know when an extension trivial, or when an extension is the limit of trivial extensions.
\[IIe2\] There are essential extensions $\tau$ which are not approximately trivial but $[\tau]=0$ in\
$KL(A,M(B)/B).$
Let $A$ be a unital separable amenable [$C^*$-algebra]{} and let $\tau$ be an essential extension of $A$ by $B$ such that $[\tau]=0$ in $KL(A,M(B)/B).$ Such $\tau$ exists (by \[IITM\] or by first mapping $A$ to ${\cal O}_2$ and then mapping ${\cal O}_2$ into $M(B)/B$).
To be more precise, we let $A$ be the unital simple AF-algebra with $K_0(A)=D_{\theta},$ where $\theta$ is an irrational number and $$D_{\theta}=\{ m+n\theta: m, n\in {\mathbb Z}\}.$$ with usual order inherited from ${\mathbb R}. $ We may assume that $[1_A]=1.$ Let $B$ be a non-unital (non-zero) hereditary [$C^*$-subalgebra]{} of the UHF-algebra with $K_0(B)={\mathbb Z}[1/2].$ Note that $B$ has a unique normalized trace. So it has a continuous scale. We further assume that $[1_{M(B)/B}]=0.$ So there is an essential extension $\tau$ of $A$ by $B$ such that $[\tau]=0.$ However, there is no (non-zero) positive [homomorphism]{} $\alpha: K_0(A)\to K_0(B).$ If there is a trivial extension $\tau$ of $A$ by $B$ with $[\tau]=0$ in $KL(A,M(B)/B),$ then $\tau_{*0}: K_0(A)\to K_0(M(B)/B)$ is zero. It follows \[IIILemb\] that $$K_0(M(B)/B)=Aff(T(B))/K_0(B)={\mathbb R}/{\mathbb Z}[1/2].$$ If $\tau$ is trivial, there would be a monomorphism $h: A\to M(B)$ such that $h|_{*0}$ maps $K_0(A)$ to $K_0(B)\subset Aff(T(B))$ positively. However there is no positive [homomorphism]{} from $D_{\theta}$ into ${\mathbb Z}[1/2].$ In fact any positive [homomorphism]{} from $D_{\theta}$ into ${\mathbb R}$ has to be the form (see \[0Pph\]) $$h_{*0}(r)=(h)_{*0}(1)r \,\,\,{\rm for\,\,\, all }\,\,\, r\in
D_{\theta}.$$ So $\tau$ can never be trivial. Furthermore, $\tau$ cannot be approximately trivial. To see this, assume that $\tau_n: A\to M(B)/B$ are trivial extensions such that $$\lim_{n\to\infty}\tau_n(a)=\tau(a)$$ for all $a\in A.$ Let $G_0\subset K_0(A)$ which contains $1$ and $\theta.$ Thus, for all large $n,$ $(\tau_n)_{*0}(\theta)=0.$ Suppose that $h_n: A\to M(B)$ such that $\pi\circ h_n=\tau_n,$ $n=1,2,...,$ where $\pi: M(B)\to M(B)/B$ is the quotient map. Thus $(h_n)_{*0}$ is a positive [homomorphism]{} from $D_{\theta}$ into ${\mathbb
R}.$ From the above expression of $h_{*0}$ (see \[0Pph\]) we see that $(h_n)_{*0}$ cannot map both $1$ and $\theta$ into rational numbers. In other words, such an $h_n$ does not exist. Hence $\tau$ is not approximately trivial.
\[IIIEx0not\] Nevertheless, there are essential trivial extensions of $A$ by $B$ such that $[\tau]\not=0$ in $KL(A,M(B)/B).$
[Let $s$ be the unique normalized trace on $B.$ Suppose that $[1_A]=1$ in $D_{\theta}.$ Let $\beta: D_{\theta}\to {\mathbb R}$ the usual embedding. It follows from \[IIILemb\] that there is a monomorphism $h: A\to M(B)$ such that $h_{*0}=\beta$ and $h(A)\cap
B=\{0\}.$ Let $\tau=\pi\circ h,$ where $\pi: M(B)\to M(B)/B$ be the quotient map. Then $\tau$ is a trivial essential extension. However, $\tau_{*0}: {\mathbb Q}\to {\mathbb R}/{\mathbb Z}[1/2]$ ($\, \cong K_0(M(B)/B))$ is not zero. Therefore $[\tau]\not= 0$ in $KL(A, M(B)/B).$]{}
Quasidiagonal extensions — general and infinite cases
=====================================================
\[IVDquas\]
Let $A$ be a separable [$C^*$-algebra]{}, $C$ be a non-unital but $\sigma$-unital [$C^*$-algebra]{} and $\tau: A\to M(C)/C$ be an essential extension. Let $\pi: M(C)\to M(C)/C$ be the quotient map. Set $E=\pi^{-1}(\tau(A)).$ The extension $\tau$ is said to be [*quasidiagonal*]{} if there exists an approximate identity $\{e_n\}$ of $C$ consisting of projections such that $$\lim_{n\to\infty}\|e_nb-be_n\|=0$$ for all $b\in E.$
Suppose that there is a bounded linear map $L: A\to M(B)$ such that $\pi\circ L=\tau.$ Then $$\|e_nL(a)-L(a)e_n\|\to 0\,\,\,\, {\rm as}\,\,\,n\to\infty$$ for all $a\in A.$
In this section and the next, we will study quasidiagonal extensions. The first question is when quasidiagonal extensions exist.
\[IVqkadd\] Let $A$ be a separable amenable [$C^*$-algebra]{} and $B$ be a non-unital and $\sigma$-unital [$C^*$-algebra]{}. Suppose that $\tau: A\to M(B)/B$ is an essential quasidiagonal extension. Then for each finitely generated subgroup $G$ of $\underline{K}(A)$ there exists a [homomorphism]{} $\alpha: G\to \underline{K}(M(B))$ such that $$\pi_*\circ \alpha|_G=(\tau_*)|_{G},$$ where $\pi: M(B)\to M(B)/B$ is the quotient map.
Suppose that $\tau: A\to M(B)/B$ is a quasidiagonal essential extension of $A$ by $B.$ Let $L: A\to M(B)$ be a [contractive completely positive linear map]{} such that $\pi\circ L=\tau.$ Since $\tau$ is quasidiagonal, there exists an approximate identity $\{e_n\}$ for $B$ such that $$\|e_nL(a)-L(a)e_n\|\to 0,\,\,\,{\rm as}\,\,\, n\to\infty$$ for all $a\in A.$ Define $L_n: A\to M(A)$ by $L_n(a)=(1-e_n)L(a)(1-e_n)$ for $a\in A.$ Then $\{L_n\}$ is a sequence of asymptotically multiplicative [contractive completely positive linear map]{}s. It follows that for each finitely generated subgroup $G$ of $\underline{K}(A),$ there exists $N>0$ such that $\{L_n\}$ gives a [homomorphism]{} $\alpha_n: G\to
\underline{K}(M(B))$ for all $n\ge N.$ Since $\pi\circ L_n=\tau$ for all $n,$ it follows that $\pi_*\circ \alpha|_{G}=(\tau_*)|_{G}.$
\[IVCadd1\] Let $A$ be a separable amenable [$C^*$-algebra]{} and $B$ be a $\sigma$-unital and stable [$C^*$-algebra]{}. Suppose that $\tau: A\to M(B)/B$ is an essential quasidiagonal extension. Then $\tau$ induces the zero map from $\underline{K}(A)$ to $\underline{K}(M(B)/B).$ Furthermore the six-term exact sequence in $K$-theory associated with the extension splits into two pure extensions of groups: $$0\to K_i(B)\to K_i(E)\to K_i(A)\to 0\,\,\,\, i=0,1.$$
This follows from the fact that when $B$ is stable, $K_i(M(B))=0$ ($i=0,1$) and \[IVqkadd\].
\[IVLqalim\] Let $A$ be a separable amenable [$C^*$-algebra]{} and $C$ be a non-unital but $\sigma$-unital [$C^*$-algebra]{}. Suppose that $\tau: A\to M(C)/C$ is an essential extension such that there exists a sequence of quasidiagonal extensions $\tau_n: A\to M(C)/C$ such that $$\lim_{n\to\infty} \tau_n(a)=\tau(a)$$ for all $a\in A.$ Then $\tau$ is quasidiagonal.
Let $\{a_n\}$ be a dense sequence in the unit ball of $A.$ Suppose that $$\|\tau_n(a)-\tau(a)\|<1/2^{n+3}$$ for all $a\in \{a_1,a_2,...,a_n\},$ $n=1,2,....$ Let $L_n: A\to M(B)$ be a [contractive completely positive linear map]{} such that $\pi\circ L_n=\tau_n.$ There exists an approximate identity $\{e_k^{(n)}\}$ for $B$ consisting of projections such that $$\lim_{n\to\infty}\|e_k^{(n)}L_n(a)-L_n(a)e_k^{(n)}\|=0$$ for all $a\in A.$ Let $L: A\to M(B)$ be a [contractive completely positive linear map]{} such that $\pi\circ L=\tau.$ Suppose that $b$ is a strictly positive element for $B.$ We may assume that $$\|(1-e_1^{(1)})(L_1(a)-L(a))\|<1/2^{3}{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\|(L_1(a)-L(a))(1-e_1^{(1)})\|<1/2^{3}$$ for $a=a_1.$ Put $q_1=e_1^{(1)}.$ Note that $$\|q_1L(a_1)-L(a_1)q_1\|<1/2.$$ By changing notation if necessary, we may assume that $$\|(1-e_2^{(2)})(L_2(a)-L(a))\|<1/2^{2+2}{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\|(L_2(a)-L(a))(1-e_2^{(2)})\|<1/2^{2+2}$$ for $ a\in \{a_1,a_2\}$ as well as $$\|(1-e_2^{(2)})q_1\|<1/2^{2+3}{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\|(1-e_2^{(2)})b\|<1/2^3.$$ There is a projection $q_2\ge q_1$ such that $$\|e_2^{(2)}-q_2\|<1/2^3.$$ Note also $$\|q_2L(a)-L(a)q_2\|<1/2^2\,\,\,{\rm for}\,\,\, a\in \{a_1,a_2\}.$$ We also have $$\|(1-q_2)b\|<1/2^2.$$ We may assume that $$\|(1-e_3^{(3)})(L_3(a)-L(a))\|<1/2^{3+2} {\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\|(L_3(a)-L(a))(1-e_3^{(3)}\|<1/2^{3+2}$$ for $a\in \{a_1,a_2,a_3\}$ as well as $$\|(1-e_3^{(3)})q_i\|<1/2^{3+3}, i=1,2 {\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\|(1-e_3^{(3)})b\|<1/2^4.$$ There exists a projection $q_3\ge q_2$ such that $$\|e_3^{(3)}-q_3\|<1/2^4.$$ Thus we have $$\|q_3L(a)-L(a)q_3\|<1/2^3\,\,\,{\rm for}\,\,\, a\in \{a_1,a_2,a_3\}.$$ We also have $$\|(1-q_3)b\|<1/2^3.$$ We continue in this fashion. It follows that we obtain an increasing sequence of projections $\{q_n\}$ in $B$ such that $$\|q_nL(a)-L(a)q_n\|<1/2^n\,\,\,\, a\in \{a_1,a_2,...,a_n\}$$ and $$\|(1-q_n)b\|<1/2^n,$$ $n=1,2,....$ It remains to show that $\{q_n\}$ is an approximate identity for $B.$ Since $\lim_{n\to\infty}\|(1-q_n)b\|=0,$ one concludes that for any positive function $f\in C_0((0,\|b\|]),$ $$\lim_{n\to\infty}\|(1-q_n)f(b)\|=0.$$ For any $a\in A$ and ${\varepsilon}>0$ there exists a positive function $f\in C_0((0,\|b\|])$ such that $$\|f(b)a-a\|<{\varepsilon}/3.$$ Choose $N>0$ such that $$\|(1-q_n)f(b)\|<{\varepsilon}/3(\|a\|+1)\,\,\,{\rm for \,\,\, all}\,\,\, n\ge N.$$ Then, for $n\ge N,$ $$\|(1-q_n)a\|\le \|q_na-q_nf(b)a\|+\|q_nf(b)a-f(b)a\|+\|f(b)a-a\|
<{\varepsilon}.$$ It follows that $\{q_n\}$ is an approximate identity for $B.$
Results in this paper can be also used to prove the following.
[(Brown-Salinas-Schochet)]{}\[IVTbss\] Let $A$ be a separable amenable [$C^*$-algebra]{} in ${\cal N}$ and $B$ be a $\sigma$-unital stable [$C^*$-algebra]{}. Suppose that there exists an essential quasidiagonal extension of $A$ by $B.$ The zero element in $KL(A, M(B)/B)$ corresponds to the set of stably quasidiagonal extensions as well as stably approximately trivial extensions.
It is proved in [@Lnuct] that stably approximately trivial extensions correspond to the zero element in $KL(A, M(B)/B)$ without assuming $A$ satisfies the UCT. Corollary \[IVCadd1\] proves that quasidiagonal extensions give a zero element in $KL(A, M(B)/B).$ It follows from 3.9 in [@Lnuct] that extensions which represent the same element in $KL(A,M(B)/B)$ are stably approximately unitarily equivalent. Then by \[IVLqalim\] every (stably) approximately trivial extension is stably quasidiagonal.
\[IVTinq\] Let $A$ be a separable exact [$C^*$-algebra]{} and $B$ be a $\sigma$-unital purely infinite simple [$C^*$-algebra]{}. Then there are essential quasidiagonal extensions
Let $e\in B$ be a nonzero projection such that $[e]=0$ in $K_0(B).$ Then by [@Br1] $eBe\otimes {\cal K}\cong B\otimes {\cal K}.$ It follows from a result of S. Zhang that $B\cong B\otimes {\cal K}$ ([@Zh3]). Thus we obtain an approximate identity $\{e_n\}$ of $B$ such that each $e_n$ is a projection and $[e_n]=0.$ Since $A$ is exact, by Theorem 2.8 in [@KP], there exists a monomorphism $\imath: A\to {\cal O}_2.$ Since $e_nBe_n$ is purely infinite and $[e_n]=0,$ there is an embedding $\phi_n: {\cal O}_2\to e_nBe_n.$ Now define $$\psi(a)=\sum_{n=1}^{\infty} \phi_n\circ \imath(a)\,\,\,{\rm for}\,\,\,
a\in A.$$ Then $\psi$ is an injective [homomorphism]{} from $A$ into $M(B)$ such that $\psi(A)\cap B=\{0\}.$ Let $\tau=\pi\circ \psi.$ Then $\tau$ is an essential quasidiagonal extension of $A$ by $B.$
The following follows from \[IVTinq\], Theorem 1.4 in [@Sch2] and Kirchberg’s absorbing theorem [@K1]. It also follows from \[IVTinq\] and \[IVCadd1\].
\[IVTqsin\] Let $A$ be a separable amenable [$C^*$-algebra]{} in ${\cal N}$ and $B$ be a non-unital and $\sigma$-unital purely finite simple [$C^*$-algebra]{}. Suppose that $\tau: A\to M(B)/B$ is an essential extension. Then $\tau$ is quasidiagonal extension if and only if it is an approximately trivial extension, and, if and only if $\tau$ induces a zero element in $KL(A, M(B)/B).$
In the next section we will discuss the case that $B$ is not purely infinite.
Quasidiagonal extensions – finite case
======================================
\[IVDcq\]
Recall that a separable [$C^*$-algebra]{} is said to be quasidiagonal if there exists a faithful representation $\phi: A\to B(H)$ for some separable Hilbert space $H$ such that $$\|p_n\phi(a)-\phi(a)p_n\|\to 0\,\,\,{\rm as}\,\,\,n\to\infty$$ for all $a\in A,$ where $\{p_n\}$ is an approximate identity of ${\cal K}$ consisting of finite rank projections.
All AF-algebras are quasidiagonal. All commutative [$C^*$-algebra]{}s are quasidiagonal. All AH-algebras are quasidiagonal. All residually finite dimensional [$C^*$-algebra]{}s are quasidiagonal. Inductive limits of quasidiagonal [$C^*$-algebra]{}s are quasidiagonal.
Recall that a [$C^*$-algebra]{} $A$ is said to have the property (SP) if every non-zero hereditary [$C^*$-subalgebra]{} contains a nonzero projection. One should note that every [$C^*$-algebra]{} with real rank zero has the property (SP) but the converse is not true.
\[IVTqa\] Let $A$ be a separable quasidiagonal amenable [$C^*$-algebra]{} and $C$ be a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} which admits an approximate identity consisting of projections and has the property (SP). Then there exists an (essential) quasidiagonal extension $\tau: A\to M(C)/C.$
We may assume that $C\not={\cal K}.$ There is a sequence of [contractive completely positive linear map]{}s $L_n: A\to F_n,$ where $F_n$ are finite dimensional [$C^*$-algebra]{}s, such that $$\|L_n(ab)-L_n(a)L_n(b)\|\to 0,\,\,\,{\rm as}\,\,\, n\to\infty$$ for all $a, b\in A.$ Let $\{a_n\}$ be a dense sequence in the unit ball of $A.$ By passing to a subsequence, we may assume that $$\|L_n(ab)-L_n(a)L_n(b)\|<1/2^{n+1}$$ for all $a, b\in \{a_1,...,a_n\}.$
Let $\{e_n\}$ be an approximate identity for $B$ consisting of projections. We may assume that $e_{n+1}-e_n\not=0$ for all $n.$ It is known (see 3.5.7 in [@Lnb]) that there is a monomorphism $j_n: F_n\to (e_{n+1}-e_n)C(e_{n+1}-e_n).$ Define $$L(a)=\sum_{n=1}^{\infty} j_n\circ L_n(a)\,\,\,{\rm for}\,\,\, a\in A.$$ Note that the sum converges in the strict topology. One checks that $L: A\to M(C)$ is a (completely) positive linear contraction. Note also, for any $a, b\in \{a_1,a_2,...,a_n\},$ $$\|\sum_{k=n}^{n+m} (L_n(ab)-L_n(a)L_n(b))\|<1/2^{n}\,\,\,{\rm for\,\,\, all}
\,\,\, m>0.$$ This implies $$L(ab)-L(a)L(b)=\sum_{n=1}^{\infty} j_n\circ (L_n(ab)-L_n(a)L_n(b))\in C$$ for all $a, b\in A.$ Let $\tau=\pi\circ L,$ where $\pi: M(C)\to M(C)/C$ is the quotient map. Then $\tau$ is an essential quasidiagonal extension.
\[IVTqa2\] Let $A$ be a separable amenable [$C^*$-algebra]{} and let $C$ be as in \[IVTqa\]. Suppose that, in addition, $C$ is also a quasidiagonal [$C^*$-algebra]{}. Then there is an (essential) quasidiagonal extension of $A$ by $B$ if and only if $A$ is quasidiagonal.
It suffices to show the “only if" part. Suppose that $L: A\to M(C)$ is a bounded linear map such that $\pi\circ L: A\to M(C)/C$ is a monomorphism and $$\lim_{n\to\infty}\|e_nL(a)-L(a)e_n\|=0\,\,\,{\rm for\,\,\, all}\,\,\, a
\in A,$$ where $\{e_n\}$ is an approximate identity consisting of projections. Let $\{a_n\}$ be a dense sequence of $A.$ We may assume that $$\|e_nL(a)-L(a)e_n\|<1/2^n
\,\,\,{\rm for}\,\,\,a\in \{a_1,a_2,...,a_n\},$$ $n=1,2,....$ Since $\pi\circ L$ is a monomorphism, we may further assume that $$\|e_nL(a)e_n\|\ge \|a\|-1/2^n\,\,\,{\rm for}\,\,\, a\in \{a_1,a_2,...,
a_n\},$$ $n=1,2,....$ Since $C$ is quasidiagonal, it follows that there exists a finite dimensional [$C^*$-algebra]{} $F_n$ and a [contractive completely positive linear map]{} $\phi_n: e_nCe_n\to F_n$ such that $$\|\phi_n(b)\|\ge \|b\|-1/2^n{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\|\phi_n(bc)-\phi_n(b)\phi_n(c)\|<1/2^n$$ for $b\in e_nL(a_i)e_n,$ $i=1,2,...,n$ and $n=1,2,....$ Define $L_n: A\to F_n$ by $L_n(a)=\phi_n(e_nL(a)e_n)$ for $a\in A.$ Then $$\|L_n(a_i)\|\ge \|a_i\|-1/2^{n-1}, i=1,2,...,n {\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}\lim_{n\to\infty}\|L_n(ab)-L_n(a)L_n(b)\|=0$$ for all $a, b\in A.$ It follows from Theorem 1 in [@V2] that $A$ is quasidiagonal.
\[II1Llift\] Let $A$ be a finite dimensional [$C^*$-algebra]{}. Let $\tau: A\to M(B)/B$ be a monomorphism such that ${\rm im}\tau_{*0}\subset
Aff(T)/\rho_B(K_0(B)).$ Then $\tau$ is trivial and there is a monomorphism $h: A\to M(B)$ such that $\pi\circ h=\tau.$
Suppose that $A=M_{r(1)}\oplus\cdots M_{r(k))}.$ So $K_0(A)$ is $k$ copies of ${\mathbb Z}.$ Let $e_i$ be a minimal projection in $M_{r(i)},$ $i=1,2,...,k.$ There are $x_i\in
Aff(T(B))/\rho(K_0(B))$ such that $[\tau(e_{i})]=x_i,$ $i=1,2,...,k$ and $\sum_{i=1}^{k}r(i)x_i=[\tau(1_A)].$ It follows from Lemma 1.3 in [@Lnamj] that there are projections $q_i\in M(B)$ such that $[\pi(q_i)]=r(i)[\tau(e_i)].$ Thus we obtain a positive [homomorphism]{} $\alpha: K_0(A)\to Aff(T)$ such that $\pi_{*0}\circ \alpha=\tau_{*0}.$ It follows from \[IIILemb\] that there is a monomorphism $h: A\to M(B)$ such that $h(A)\cap
B=\{0\}$ and $h_{*0}=\alpha.$ It then follows that $ [\pi\circ
h]=[\tau]$ in $KL(A, M(B)/B).$ Since $A$ is finite dimensional, it follows that $\pi\circ h$ is unitarily equivalent to $\tau.$ This implies that $\tau$ is trivial.
\[IVLqanes\] Let $A$ be a separable amenable [$C^*$-algebra]{}. Suppose that $\tau: A\to M(B)/B$ is an essential quasidiagonal extension. Then $\tau_{*1}=0,$ ${\rm im}\tau_{*0}\subset Aff(T)/\rho_B(K_0(B))$ and $[\tau]|_{K_i(A, {\mathbb Z}/k{\mathbb Z})}=0$ for $i=0,1$ and for all $k\ge 2.$
Since $K_1(M(B))=\{0\}$ (see \[0T1\]), by \[IVqkadd\], $\tau_{*1}=0.$ By \[IVqkadd\] and \[0T2\], ${\rm im}\tau_{*0}\subset Aff(T)/\rho_B(K_0(B)).$ Since $K_0(M(B))=Aff(T)$ is torsion free, $\tau_{*0}|_{{\rm tor}(K_0(A))}=0.$ Let $C_k$ be as in \[IIDuct\]. Then $L\otimes {{\rm id}}_{C(C_k)}: A\otimes
C(C_k)\to M(B)\otimes C(C_k)$ lifts $\tau\otimes {{\rm id}}_{C(C_k)}.$ Since $K_0(M(B))=Aff(T)$ is divisible and $K_1(M(B))=0,$ $K_0(M(B),{\mathbb Z}/k{\mathbb Z})=\{0\}$ for all $k\ge 2.$ It follows from \[IVqkadd\] that $[\tau]|_{K_0(A, {\mathbb Z}/k{\mathbb Z})}=0.$
Note also that since $K_0(M(B))$ is torsion free and $K_1(M(B))=0,$ $K_1(M(B),{\mathbb Z}/k{\mathbb Z})=\{0\}.$ The same argument above also shows that $[\tau]|_{K_1(A, {\mathbb
Z}/k{\mathbb Z})}=0,$ $k=2,3,....$
It should be noted that, since $Aff(T)/\rho_B(K_0(B))$ is divisible, $K_0(M(B)/B)/kK_0(M(B)/B)=K_0(B)/kK_0(M(B)/B).$ Therefore one sees that, for any nonzero [homomorphism]{} $\gamma: K_0(A)
\to Aff(T)/\rho_B(K_0(B)),$ there is $\alpha\in
{\rm Hom}_{\Lambda}(\underline{K}(A), \underline{K}(M(B)/B))$ such that $\alpha|_{K_0(A)}=\gamma$ but $[\alpha]|_{K_0(A, {\mathbb Z}/k{\mathbb Z})}=0,$ $k=1,2,....$ Furthermore, if $K_1(B)$ is also divisible (or $K_1(B)=\{0\}$), one computes that $\tau_{*1}=0$ and ${\rm
im}\tau_{*0}\subset Aff(T)/\rho_B(K_0(B))$ imply that $$[\tau]|_{K_0(A,{\mathbb Z}/k{\mathbb Z})}=0,\,\,\,k=2,3,...,$$ by using the six-term exact sequence in \[IIDuct\]. One should note that $[\tau]|_{K_1(A, {\mathbb Z}/k{\mathbb
Z})}=0$ for $k=2,3,...,$ implies that $\tau_{*0}({\rm
tor}(K_0(A)))=0.$ On the other hand, if $\tau_{*0}({\rm tor}(K_0(A))=0$ and ${\rm
ker}\rho_B$ is divisible (or ${\rm ker}\rho_B(K_0(B))=\{0\}$), then $$[\tau]|_{K_1(A, {\mathbb Z}/k{\mathbb Z})}=0,\,\,\,k=2,3,....$$
\[IVCqa\] Let $A$ be the closure of $\cup_{n=1}^{\infty} A_n,$ where each $A_n$ is a separable amenable [$C^*$-algebra]{} in ${\cal N}$ and let $j_n:
A_n\to A$ be the embedding. Suppose that $\tau: A\to M(B)/B$ is an essential extension such that $\tau\circ j_n$ is a quasidiagonal extension for each $n.$ Then $\tau$ is also a quasidiagonal extension.
The proof is almost the exactly the same as that of \[IVLqalim\].
\[IVDEm\]
Denote by ${\cal C}_{afem}$ the class of separable [$C^*$-algebra]{}s $A$ satisfying the following: there is an embedding $j: A\to C$ such that $j_{*0}:K_0(A)/{\rm tor}(K_0(A))\to K_0(C)$ is injective, where $C$ is a unital AF-algebra.
Clearly every AF-algebra is in ${\cal C}_{afem}.$ It is easy to see that [$C^*$-algebra]{}s of the form $C(X)\otimes M_n$ are in ${\cal
C}_{afem},$ where $X$ is a finite CW complex. But much more is true.
Recall that a [$C^*$-algebra]{} $A$ is called [*residually finite dimensional*]{} if there is a separating family $\Pi$ of finite dimensional irreducible representations of $A,$ i.e., for any $a\in A,$ there is $\phi\in \Pi$ such that $\phi(a)\not=0.$
The following is a modification of a Dadarlat’s construction.
\[IVemb\] Let $A$ be a separable amenable residually finite dimensional [$C^*$-algebra]{} in ${\cal N}.$ Then there exists a separable unital simple AF-algebra $C$ and an embedding $j: A\to C$ such that $j_{*0}$ induces an injective map from $K_0(A)/{\rm tor}(K_0(A))$ into $K_0(C).$ In particular, $A\in {\cal N}\cap {\cal C}_{afem}.$
Fix a separating sequence of finite dimensional irreducible representations $\{t_n\}.$ For convenience, we assume that each $t_n$ repeats infinitely many times in the sequence. Suppose that $t_n(A)$ has rank $k(n).$ For each $n,$ define $\psi_n: A\to M_{k(n)}$ by the composition: $A{\stackrel{t_n}{\to}} M_{k(n)}{\stackrel{{{\rm id}}\otimes 1_A}{\to}}
M_{k(n)}(A).$ We define a [homomorphism]{} $h_1: A\to M_{I(2)}(A),$ where $I(2)=1+k(1),$ by $$h_1(a)={\rm diag}(a, \psi_1(a))\,\,\,{\rm for}\,\,\, a\in A.$$ Suppose that $h_m: M_{I(m)}(A)\to M_{I(m+1)}(A)$ is defined. Defined $h_{m+1}: M_{I(m+1)}(A)\to M_{I(m+2)}(A)$ by $$h_{m+1}(a)={\rm diag}(a, {\bar \psi}_1(a), {\bar \psi}_2(a),...,
{\bar \psi}_{m+1}(a))\,\,\,\,{\rm for}\,\,\, a\in M_{I(m+1)}(A),$$ where $I(m+2)=I(m+1)(1+\sum_{i=1}^{m+1}k(i))$ and ${\bar
\psi}_i=\psi_i\otimes {{\rm id}}_{I(m+1)},$ $i=1,2,...,m+1.$ Set $B=\lim_{m\to\infty} (M_{I(m)}(A), h_m).$ It is shown (see 3.7.8 and 3.7.9 of [@Lnb]) that $B$ is a unital separable amenable simple [$C^*$-algebra]{} with $TR(B)=0.$ Since each $M_{I(m)}(A)$ satisfies the UCT, so does $B.$ It follows from [@Lncl] that $B$ is isomorphic to a unital simple AH-algebra with real rank zero and with no dimension growth. Let $C$ be a unital simple AF-algebra with $K_0(C)=K_0(B)/{\rm tor}(K_0(B)).$ It follows from [@EG] that there exists a monomorphism $\phi$ from $B$ into $C$ such that $\phi_{*0}$ is the quotient map from $K_0(B)$ onto $K_0(B)/{\rm tor}(K_0(B))=K_0(C).$
Thus it remains to show that $h_{1,\infty}: A\to B$ induces an injective map $(h_{1,\infty})_{*0}$ on $K_0(A).$
It suffices to show that $(h_m)_{*0}$ is injective. Suppose that $p$ and $q$ are two projections in $M_{I(m)}(A)$ such that $h_m(p)$ and $h_m(q)$ are equivalent. Then ${\bar \psi}_j\circ
h_m(p)$ and ${\bar \psi}_j\circ h_m(q)$ are equivalent (in a matrix algebra) for each $j.$ Therefore $${\rm diag}({\bar \psi}_1(p),...,{\bar \psi}_{m+1}(p),{\bar
\psi}_1(p),...,{\bar \psi}_{m+1}(p)){\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}{\rm diag}({\bar
\psi}_1(q),,...,{\bar \psi}_{m+1}(q),{\bar \psi}_1(q),,...,{\bar
\psi}_{m+1}(q))$$ are equivalent. Let $t$ be any finite dimensional representation of $M_{I(m)}(A).$ Then $(t\oplus t)\circ h_m(p)$ and $(t\oplus
t)\circ h_m(q)$ are equivalent (in a matrix algebra). From the above, it follows that $t(p)\oplus t(p)$ and $t(q)\oplus t(q)$ are equivalent in a matrix algebra. Thus $t(p)$ and $t(q)$ are equivalent in the matrix algebra. This in turn implies that $${\rm diag}(0, {\bar \psi}_1(p),{\bar \psi}_2(p),...,{\bar \psi}_m(p))
{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}{\rm diag}(0, {\bar \psi}_1(q), {\bar \psi}_2(q),...,{\bar \psi}_m(q))$$ are equivalent. Consequently $$[p]=[q] \,\,\,{\rm in}\,\,\, K_0(M_{I(m)}(A)).$$ This implies that $(h_m)_{*0}$ is injective for each $m.$
\[II1Tqd\] Let $A$ be a separable amenable [$C^*$-algebra]{} in ${\cal N}\cap {\cal
C}_{afem}.$ Suppose that $\tau$ is an essential extension.
Then $\tau$ is quasidiagonal if and only if $\tau_{*1}=0,$ ${\rm im} \tau_{*0}\subset
Aff(T(B))/\rho(K_0(B))$ and $[\tau]|_{K_i({\mathbb Z}/k{\mathbb
Z})}=0,$ $i=0,1$ and $k=2,3,....$
The “if only" part follows from \[IVLqanes\]. For the “if" part, we first assume that $A$ is an AF-algebra. We may write $A=\overline{\cup_{n=1}^{\infty} A_n},$ where $A_n\subset A_{n+1}$ and each $A_n$ is a finite dimensional [$C^*$-algebra]{}. Let $j_n: A_n\to A$ be the embedding. It follows from \[IVCqa\] that it suffices to show that $\tau\circ j_n$ are quasidiagonal. Note that $(\tau\circ
j_n)_{*0} \subset Aff(T)/\rho_B(K_0(B)).$ It follows from \[II1Llift\] that each $\tau\circ j_n$ is in fact trivial and therefore quasidiagonal (since $A_n$ is finite dimensional).
For the general case, let $C$ be a unital AF-algebra and $j: A\to
C$ be an embedding such that $j_{*0}$ induces an injective [homomorphism]{} from $K_0(A)/{\rm tor}(K_0(A))$ into $K_0(C).$ Let $\tau$ be as in the theorem. Since $Aff(T)/\rho_B(K_0(B))$ is divisible, there exists a [homomorphism]{} $\alpha: K_0(C)\to Aff(T)/\rho_B(K_0(B))$ such that $\alpha\circ j_{*0}=\tau_{*0}.$ It follows from \[IITM\] that there is an essential extension $t: C\to M(B)/B$ such that $t_{*0}=\alpha.$ From what we have shown, $t$ is quasidiagonal. Let $\tau_0=t\circ j.$
Since $C$ is an AF-algebra, $K_0(C,{\mathbb Z}/k{\mathbb
Z})=K_0(C)/kK_0(C),$ $k=2,3,....$ On the other hand, ${\rm
im}t_{*0}\subset Aff(T)/\rho_B(K_0(B))$ and $Aff(T)/\rho_B(K_0(B))$ is divisible, so one computes that $[t]|_{K_0(C, {\mathbb Z}/k{\mathbb Z})}=0$ for all $k,$ by using the six-term exact sequence in \[IIDuct\]. Thus $[\tau_0]|_{K_0(A,{\mathbb Z}/k{\mathbb Z})}=0$ for all $k.$ We also have $(\tau_0)_{*1}=0.$
Since $C$ is an AF-algebra, $K_1(C,{\mathbb Z}/k{\mathbb
Z})=\{0\}$ for $k=2,3,....$ Since $\tau_0$ factors through $C,$ we conclude that $[\tau_0]|{K_1(A,{\mathbb Z}/k{\mathbb Z})}=0.$ Furthermore $(\tau_0)_{*0}=\tau_{*0}.$ We then conclude that $$[\tau]=[\tau_0] \,\,\,{\rm in}\,\,\, KL(A, M(B)/B).$$ Therefore, by \[IabTTad\], $\tau$ and $\tau_0$ are approximately unitarily equivalent. We have shown that $t$ is a quasidiagonal extension. So is $\tau_0,$ by \[IVCqa\]. It follows that $\tau$ is a quasidiagonal extension.
For the last theorem in this section, one should note that every strong NF-algebra is an inductive limit of amenable residually finite dimensional [$C^*$-algebra]{}s (see 6.16 in [@BK]).
\[IVTqaT\] Let $A$ be the closure of $\cup_{n=1}^{\infty} A_n,$ where each $A_n$ is a separable amenable residually finite dimensional [$C^*$-algebra]{} in ${\cal N}.$ Let $\tau: A\to M(B)/B$ be an essential extension. Then $\tau$ is quasidiagonal if and only if $\tau_{*1}=0,$ ${\rm
im}\tau_{*0}\subset Aff(T)/\rho_B(K_0(B))$ and $[\tau]|_{K_i(A,
{\mathbb Z}/k{\mathbb Z})}=0,$ $i=0,1$ and $k=2,3,...$
It follows from \[IVLqanes\] that we only need to prove the “if" part of the theorem.
Fix an integer $n\ge 1.$ Let $\phi_n: A_n\to A$ be the embedding. Put $\tau_n=\tau\circ \phi_n.$ So $\tau_n$ is an essential extension. Then $(\tau_n)_{*1}=0,$ ${\rm
im}(\tau_n)_{*0}\subset Aff(T)/\rho_B(K_0(B))$ and $[\tau_n]|_{K_i(A_n,{\mathbb Z}/k{\mathbb Z})}=0$ for $i=0,1$ and for $k=2,3,....$ It follows from \[II1Tqd\] that $\tau_n$ is quasidiagonal. Therefore the theorem follows from \[IVCqa\].
Approximately trivial extensions
================================
Let $$0\to {\cal K}\to E\to A\to 0$$ be an essential extension for a amenable quasidiagonal [$C^*$-algebra]{} $A.$ It is shown that the extensions is quasidiagonal if and only if it is approximately trivial (see [@Br2] and [@Sch2]).
In this section, we will show that there are quasidiagonal extensions that are not approximately trivial. The obstruction of a quasidiagonal extension to be approximately trivial can be computed. We will also discuss when an essential extension is approximately trivial.
[*Throughout this section $B$ is always a non-unital but $\sigma$-unital simple [$C^*$-algebra]{} with real rank zero, stable rank one, weakly unperforated $K_0(B)$ and a continuous scale.*]{}
If $A$ is a unital separable quasidiagonal [$C^*$-algebra]{} then $A$ admits at least one tracial state. Let $T_A$ denote the tracial state space. Let $\rho_A: K_0(A)\to Aff(T_A)$ be defined by $\rho_A([p])(t)=t(p)$ for projection $p\in M_k(A),$ $k=1,2,...$ This map $\rho_A$ is a positive [homomorphism]{}. In general, ${\rm ker}\rho_A$ is not zero.
\[IITapptr\] Let $A$ be a unital separable amenable [$C^*$-algebra]{} in ${\cal N}.$ Let $\tau:
A\to M(B)/B$ be an essential extension of $A$ by $B.$ If $\tau$ is approximately trivial, then $\tau_{*1}=0,$ ${\rm
im}\tau_{*0}\subset Aff(T)/\rho_B(K_0(B)),$ $\tau_{*0}|_{{\rm
ker}\rho_A}=0$ and $[\tau]|_{K_i(A,{\mathbb Z}/k{\mathbb Z})}=0$ for $i=0,1$ and for $k=2,3,....$
Suppose that there are monomorphisms $t_n: A\to M(B)$ such that $$\lim_{n\to\infty}\pi\circ t_n(a)=\tau(a)\,\,\,{\rm for\,\,\,all}\,\,\, a\in A,$$ where $\pi: M(B)\to M(B)/B$ is the quotient map. We obtain a positive [homomorphism]{} from $K_0(A)$ into $Aff(T).$ This implies that $$(t_n)_{*0}({\rm ker}\rho_A))=0.$$ Since $K_0(B)=Aff(T)$ is a torsion free divisible group and $K_1(M(B))=0,$ we compute that, by using the six-term exact sequence in \[IIDuct\], $$K_i(M(B),{\mathbb Z}/k{\mathbb Z})=\{0\},\,\,\, i=0,1, k=2,3,...$$ Thus $$(t_n)_{*1}=0{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}[t_n]|_{K_i(A, {\mathbb Z}/k{\mathbb
Z})}=0,\,\,\, i=0,1,\,\,\,k=2,3,...,\,\,\,n=1,2,....$$ For any finite subset ${\cal P}\subset {\underline{K}}(A),$ there is an integer $n_0\ge 1$ such that $$[\tau_n]|_{\cal P}=[\tau]|_{\cal P}\,\,\,{\rm for\,\,\,all}\,\,\,
n\ge n_0.$$ Thus $$(\pi\circ t_n)_{*0}({\rm ker}\rho_A)=0, \,(\pi\circ
t_n)_{*1}=0{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}[\pi\circ t_n]|_{K_i(A, {\mathbb Z}/k{\mathbb
Z})}=0,\,\,\,i=0,1,\, k=2,3,...,\,n=1,2,....$$ Therefore $$\tau_{*0}({\rm
ker}\rho_A)=0,\,\tau_{*1}=0{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}[\tau]|_{K_i(A, {\mathbb
Z}/k{\mathbb Z})}=0,\,\,\,i=0,1,\, k=2,3,...,\,n=1,2,....$$
We also have ${\rm im}(\pi\circ t_n{*0})\subset
Aff(T)/\rho_B(K_0(B)).$ It follows that $${\rm im}\tau_{*0}\subset Aff(T)/\rho_B(K_0(B)).$$
\[IICqdnotat\] Let $A$ be a unital separable AF-algebra such that ${\rm ker}\rho_A\not=0.$ Then there are quasidiagonal extensions of $A$ by $B$ that are not approximately trivial.
Let $g_0\in {\rm ker}\rho_A$ be a nonzero element. Take any nonzero $x\in Aff(T)/\rho_B(K_0(B))$ and define a nonzero [homomorphism]{} $\alpha_0: {\mathbb Z}g_0\to {\mathbb Z}x\in
Aff(T)/\rho_B(K_0(B))$ by $\alpha_0(mg_0)=mx$ for $m\in {\mathbb Z}.$ Since $ Aff(T)/\rho_B(K_0(B))$ is divisible, we obtain a [homomorphism]{} $\alpha: K_0(A)\to
Aff(T)/\rho_B(K_0(B))$ such that $\alpha(g_0)=\alpha_0(g_0).$ It follows from \[IIILemb\] that there is a monomorphism $\tau:
A\to M(B)/B$ such that $\tau_{*0}=\alpha.$ It follows from \[II1Tqd\] that $\tau$ is quasidiagonal. But by \[IITapptr\] $\tau$ is not approximately trivial.
[From \[IICqdnotat\], one sees that it is typical rather than unusual that quasidiagonal extensions are different from approximately trivial extensions. The assumption that $A$ is AF is certainly not necessary. ]{}
\[IVLgroup\] Let $G$ be an unperforated ordered group with the Riesz interpolation property. Suppose that $G_0\subset G$ is a countable ordered subgroup. Then there exists a countable subgroup $G_1\supset G_0$ which satisfies the Riesz interpolation property and is unperforated. If $G$ is simple, we may further assume that $G_1$ is also simple.
Since $G_0$ is countable, there exists a countable ordered subgroup $F_1$ of $G$ such that $G_0\subset F_1$ and if $g_1, g_2, g_3, g_4\in G_0$ with $g_1, g_2\le g_3, g_4$ then there is $g\in F_1$ such that $$g_1, g_2\le h \le g_3, g_4.$$ If a countable ordered subgroup $F_n$ has been constructed, we have a countable ordered subgroup $F_{n+1}$ such that if $x_1,x_2\le y_1,y_2$ are in $F_n$ there exists $g\in F_{n+1}$ such that $$x_1, x_2\le g\le y_1, y_2.$$ Set $G_1=\cup_{n=1}^{\infty} F_n.$ Then $G_1$ is a countable ordered subgroup of $G$ containing $G_0.$ From the construction, it is also clear that $G_1$ has the Riesz interpolation property.
Now we further assume that $G$ is simple. Let $g\in (G_1)_+$ be a nonzero positive element and $f\in G_1.$ Since $G$ is simple there is an integer $n\ge 1$ such that $ng\ge f.$ This implies that $G_1$ is also simple. Since $G$ is unperforated, it follows that $G_1$ is also unperforated.
\[IVLabemb\] Let $A$ be a unital separable commutative [$C^*$-algebra]{}. Then there exists a monomorphism $h: A\to M(B)$ such that $\pi\circ h$ is an essential extension and ${\rm im} h_{*0}\subset\rho_B(K_0(B)).$
Let $G\subset \rho_B(K_0(B))$ be a countable simple ordered group with the Riesz interpolation property. We may also assume that $G\not\cong {\mathbb Z}.$ It follows [@EHS] that there is a unital non-elementary simple AF-algebra $C$ such that $K_0(C)$ is order isomorphic to $G.$ Thus we obtain an order isomorphism $\alpha: K_0(C)\to G\subset \rho_B(K_0(B))\subset Aff(T).$ It follows from p. 67 in [@AS] that there is monomorphism $j: A\to C.$ It follows from \[IIILemb\] that there is a [homomorphism]{} $\phi: C\to
M(B)$ such that $\phi(A)\cap B=\{0\}$ and $\phi_{*0}=\alpha.$ Define $h=\phi\circ j.$
\[IVTabeappr\] Let $A$ be a separable unital commutative [$C^*$-algebra]{}. Then $\tau: A\to
M(B)/B$ is approximately trivial if and only if $\tau_{*1}=0,$ $\tau_{*0}({\rm ker}\rho_A)=0,$ ${\rm im}\tau_{*0}\subset
Aff(T)/\rho_B(K_0(B))$ and $[\tau]|_{K_i(A,{\mathbb Z}/k{\mathbb
Z})}=0,$ $i=0,1,$ $k=2,3,....$
We may write $A=C(X),$ where $X$ is a compact metric space. There are finite CW complexes $X_n$ such that $X=\lim_{\leftarrow n}
X_n.$ We also write $C(X)=\lim_{n\to\infty} C(X_n)$ and $\psi_n:
C(X_n)\to C(X)$ is the induced [homomorphism]{}. We will show that if $\tau_{*1}=0,$ $\tau_{*0}({\rm ker}\rho_A)=0,$ ${\rm
im}\tau_{*0}\subset Aff(T)/\rho_B(K_0(B))$ and $[\tau]|_{K_i(A,{\mathbb Z}/k{\mathbb Z})}=0,$ $i=0,1$ and $k=2,3,...,$ then $\tau$ is approximately trivial. Let $e_1^{(n)},e_2^{(n)},...,e_{k(n)}^{(n)}$ be projections corresponding to each summand of $C(X_n)$ which corresponds to each connected component of $X_n.$ So $K_0(C(X_n))/{\rm
ker}\rho_C$ is generated by $\{e_1^{(n)},e_2^{(n)},...,e_{k(n)}^{(n)}\}.$ Denote by $G_n$ the subgroup of $K_0(C(X))$ generated by $[\psi_n(e_1^{(n)})],
[\psi_n(e_2^{(n)})], ..., [\psi_n(e_{k(n)}^{(n)})].$ Let $f_n:
X\to X_n$ denote the continuous map induced by the inverse inductive limit $X=\lim_{\leftarrow n} X_n.$ Let $\xi_1^{(n)},\xi_2^{(n)},...,\xi_{k(n)}^{(n)}$ be points in $X_n$ which lie in different components. Let $x_1^{(n)},x_2^{(n)},...,x_{k(n)}^{(n)}$ be points in $X$ such that $f_n(x_i^{(n)})=\xi_i^{(n)}.$ Since $\rho_B(K_0(B))$ is dense in $Aff(T),$ we can find mutually orthogonal projections $p_1^{(n)},p_2^{(n)},..., p_{k(n)}^{(n)}$ in $M(B)$ such that $\sum_{i=1}^{k(n)}p_i^{(n)}\le 1_{M(B)},$ $1_{M(B)}-\sum_{i=1}^{k(n)}p_i^{(n)}\not\in B$ and $[\pi(p_i^{(n)})]=[\tau\circ \psi_n(p_i^{(n)})],$ $i=1,2,...,k(n).$ Define a [homomorphism]{} $\phi_n: C(X)\to M(B)$ by $$\phi_n(f)=\sum_{i=1}^{k(n)}f(x_i^{(n)})p_i^{(n)}\,\,\,{\rm for}\,\,\,
f\in C(X).$$ Let $P_n\le 1_{M(B)}-\sum_{i=1}^{k(n)}p_i^{(n)}$ be a projection in $M(B)\setminus B$ such that $[P_n]\in \rho_B(K_0(B)).$ It follows from \[IVLabemb\] that there is a monomorphism $h_n^{(0)}: C(X)\to P_nM(B)P_n$ such that ${\rm
im}(h_n^{(0)})_{*0} \subset \rho_B(K_0(B))$ and $\pi\circ
h_n^{(0)}$ is injective. Now define $h_n=\phi_n+h_n^{(0)}.$ Then $h_n: C(X)\to M(B)$ give an essential trivial extension of $C(X)$ by $B.$ Note that $(\pi\circ h_n)_{*1}=0,$ $[\pi\circ
h_n]|_{K_i({\mathbb Z}/k{\mathbb Z})}=0,$ $i=0,1,$ $k=2,3,...,$ and $$(\pi\circ h_n)_{*0}|_{G_n}=\tau_{*0}|_{G_n}, n=1,2,...,$$ and ${\rm im}(\pi\circ h_n)_{*0}) \subset Aff(T)/\rho_B(K_0(B)).$ Since $\cup_{n=1}^{\infty} G_n=K_0(C(X)),$ one checks that, for any finite subset ${\cal P}\subset {\underline{K}}(C(X)),$ there exists an integer $n,$ such that $$[\pi\circ h_n]|_{\cal P}=[\tau]_{\cal P}.$$ It follows from \[IabT1\] that there exists a sequence of unitaries $u_n\in M(B)/B$ such that $$\lim_{n\to\infty} {\rm ad}\,u_n\circ \pi\circ
h_n(a)=\tau(a)\,\,\,a\in C(X).$$ This implies that $\tau$ is approximately trivial.
\[IIDpl\]
Let $
Pl(K_0(A), Aff(T)/\rho_B(K_0(B)))
$ be the set of those elements $\alpha$ in $ {\rm Hom}(K_0(A), Aff(T)/\rho_B(K_0(B))$ such that there exists a positive [homomorphism]{} $\beta: K_0(A)\to Aff(T)$ such that $\Phi\circ \beta=\alpha,$ where $\Phi: Aff(T)\to Aff(T)/\rho_B(K_0(B))$ is the quotient map.
Denote by $ Apl(K_0(A), Aff(T)/\rho_B(K_0(B))) $ the set of those elements $\alpha $ in $ {\rm Hom}(K_0(A), Aff(T)/\rho_B(K_0(B))$ satisfying the following: there exists an increasing sequence of finitely generated subgroups $\{G_n\}\subset K_0(A)$ such that $\cup_{n=1}^{\infty} G_n=K_0(A)$ and a sequence of [homomorphism]{} $\alpha_n\in Pl(K_0(A),
Aff(T)/\rho_B(K_0(B)))$ such that $$(\alpha_n)|_{G_n}=\alpha|_{G_n},$$ $n=1,2,....$
One should note that if $\alpha\in Apl(K_0(A). Aff(T)/\rho_B(K_0(B))$ then $\alpha|_{{\rm ker}\rho_A}=0.$ In fact, if $x\in {\rm ker}\rho_A,$ then $x\in G_n$ for some $n.$ But $\alpha_n(x)=0$ for all $n.$ Therefore $\alpha(x)=0.$
\[IITnappext\] Let $A$ be a separable amenable [$C^*$-algebra]{} and let $\tau: A\to M(B)/B$ be an essential approximately trivial extension. Then $$\tau_{*1}=0, [\tau]|_{K_i(A,{\mathbb Z}/k{\mathbb Z})}=0{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}$$ $$\tau_{*0}\in Apl(K_0(A), Aff(T)/\rho_B(K_0(B))).$$
Suppose that $t_n: A\to M(B)/B$ be a sequence of trivial extensions such that $$\lim_{n\to\infty}t_n(a)=\tau(a)\,\,\,\,{\rm for\,\,\,all}\,\,\, a\in A.$$ There is a sequence of monomorphism $h_n: A\to M(B)$ such that $\pi\circ h_n=t_n,$ where $\pi: M(B)\to M(B)/B$ is the quotient map, $n=1,2,....$ Since $K_1(M(B))=0$ and $K_0(M(B))$ has no torsion, we conclude (by using the six-term exact sequence in \[IIDuct\]) that $$[t_n]|_{K_1(A)}=0{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}[t_n]|_{K_i(A, {\mathbb Z}/k{\mathbb Z})}=0
\,\,\,i=0,1,\, k=1,2,....$$ It follows that $$[\tau]|_{K_1(A)}=0 {\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}[t]|_{K_i(A, {\mathbb Z}/k{\mathbb Z})}=0,
\,\,\,i=0,1,\, k=1,2,....$$ Let $\{G_n\}$ be a sequence of finitely generated groups of $K_0(A)$ such that $\cup_{n=1}^{\infty} G_n=K_0(A).$ For each $n,$ there is $m(n)$ such that $$(t_m)|_{G_n}=\tau|_{G_n}$$ for all $m\ge m(n),$ since $\lim_{n\to\infty}t_n(a)=\tau(a)$ for all $a\in A.$ However, $(t_m)_{*0}\in Pl(K_0(A), Aff(T)/\rho_B(K_0(B))).$ Therefore $\tau_{*0}\in Apl(K_0(A), Aff(T)/\rho_B(K_0(B))).$
\[IITappext\] Let $A$ be a separable amenable [$C^*$-algebra]{} in ${\cal N}$ which can be embedded into a unital AF-algebra $C$ such that $K_0(A)/{\rm ker}\rho_A=K_0(C)/{\rm ker}\rho_C$ (as ordered groups). Let $\tau: A\to M(B)/B$ be an essential extension. Then $\tau$ is approximately trivial if and only if $\tau_{*1}=0,$ $[\tau]|_{K_i(A, {\mathbb Z}/k{\mathbb Z})}=0,$ $i=0,1,$ $k=1,2,...,$ and $\tau_{*0}\in Apl(K_0(A), Aff(T)/\rho_B(K_0(B))).$
The “onlyi if part" follows from \[IITnappext\]. Suppose that $\alpha_n\in Pl(K_0(A), Aff(T)/\rho_B(K_0(B)))$ is such that $$(\alpha_n)|_{G_n}=\tau_{*0}|_{G_n},$$ where $\cup_{n=1}^{\infty}G_n=K_0(A)$ and $G_n$ is finitely generated. Suppose also $h_n: K_0(A)\to K_0(M(B))=Aff(T)$ is a positive [homomorphism]{} such that $\Phi\circ h_n=\alpha_n.$ Note that $h_n|_{{\rm ker}\rho_A}=0.$ Let ${\tilde h}_n: K_0(A)/{\rm ker}\rho_A\to K_0(M(B))$ denote the induced positive [homomorphism]{}. Let $C$ be the unital simple AF-algebra such that there exists an embedding $j: A\to C$ such that $j_{*0}$ induces an order isomorphism from $K_0(A)/{\rm ker}\rho_A$ onto $K_0(C)/{\rm ker}\rho_C$ with $K_0(C)=K_0(A)/{\rm ker}\rho_A$ and $[1_C]=[1_A].$ There is a [homomorphism]{} $\psi_n: C\to M(B)$ such that $(\psi_n)_{*0}={\tilde h_n}.$ Put $\phi_n=\psi_n\circ j.$ Let $t_n=\pi\circ \phi_n.$ It is clear that $[t_n]_{K_1(A)}=0$ and $(t_n)_{*0}=\alpha_n.$ We note that since $K_0(M(B))=Aff(T)$ is divisible, $K_0(M(B))/kK_0(M(B))=0.$ We also have $K_1(M(B))=0.$ This implies that $K_0(M(B),{\mathbb Z}/k{\mathbb Z})=0$ for all $k.$ Therefore $$[t_n]|_{K_0(A, {\mathbb Z}/k{\mathbb Z})}=0\,\,\,\, k=1,2,....$$ Since $(\phi_n)_{*1}=0$ and $K_0(M(B))$ has no torsion, we compute that $$[t_n]|_{K_1(A)}=0{\,\,\,\,\,\, {\rm and} \,\,\,\,\,\,}[t_n]|_{K_1(A, {\mathbb Z}/k{\mathbb Z})}=0$$ for all $k.$ For $\tau,$ we note that $\tau_{*0}\in APl(K_0(A),
Aff(T)/\rho_B(K_0(B)))$ implies that $\tau_{*0}|_{{\rm
ker}\rho_A}=0.$ In particular, $\tau_{*0}|_{{\rm tor}(K_0(A))}=0.$ Since $\tau_{*1}=0,$ it follows that $$[\tau]|_{K_i(A, {\mathbb Z}/k{\mathbb Z})}=0\,\,\,i=0,1\,\,\,k=1,2,....$$ Therefore, for any finite subset ${\cal P}\subset
{\underline{K}(A)},$ there exists $n$ such that $$[\tau]|_{{\cal P}}=[t_m]|_{{\cal P}}$$ for all $m\ge n.$ It follows from \[IabT1\] that there are unitaries $u_n\in M(B)/B$ such that $$\lim_{n\to\infty} {\rm ad}\, u_n\circ t_n(a)=\tau(a)$$ for all $a\in A.$
\[IIE5\]
There are many examples where a [homomorphism]{} $\alpha: K_0(A)\to
Aff(T)/\rho_B(K_0(B))$ can not be lifted to a positive [homomorphism]{} $\beta: K_0(A)\to Aff(T).$ In \[IIe2\], an example is given where even if $\alpha=0,$ one can not get a nonzero positive [homomorphism]{} $\beta: K_0(A)\to Aff(T)$ such that $\Phi\circ \beta=\alpha.$ To show other complications, let $\rho_B(K_0(B))={\mathbb Z}[1/2]$ and let $K_0(A)={\mathbb
Z}\oplus {\mathbb Z}\sqrt{2}.$ Define a [homomorphism]{} $\alpha: K_0(A)\to
Aff(T)/\rho_B(K_0(B))= {\mathbb R}/{\mathbb Z}[1/2]$ so that $\alpha(1)=1$ and $\alpha(\sqrt{2}) =\pi.$ If there is a nonzero [homomorphism]{} $\beta: K_0(A)\to {\mathbb R}$ such that $\Phi\circ \beta(1)=\alpha(1)$ and $\Phi\circ
\beta(\sqrt{2})=\alpha(\sqrt{2}),$ then $\beta(1)=x$ and $\beta(\sqrt{2})=y$ with $x\in {\mathbb Z}[1/2]$ and $y=\pi+z,$ where $z\in {\mathbb Z}[1/2].$ If, in addition, $\beta$ were positive, then $\beta(\sqrt{2})=x\sqrt{2}.$ But that would imply that $x\sqrt{2}=\pi+z.$ This is impossible since $x, z\in {\mathbb Z}[1/2].$ Therefore $\alpha\not\in Apl(K_0(A),
Aff(T)/\rho_B(K_0(B))).$
This example also shows that there are very few elements in $APl(K_0(A), Aff(T)/\rho_B(K_0(B)))$ or in\
$Pl(K_0(A), Aff(T)/\rho_B(K_0(B))).$
Nevertheless, we have the following:
\[IIP1\] Let $A$ be a unital separable [$C^*$-algebra]{} with $K_0(A)={\mathbb Z}[1/p],$ where $p$ is a prime number. Suppose that $\tau_{*0}:K_0(A)
\to Aff(T)/\rho_B(K_0(B))$ is a [homomorphism]{}. Then the following hold:
[(1)]{} $\tau_{*0} \in Apl(K_0(A), Aff(T)/\rho_B(K_0(B))$ if $\tau_{*0}\not=0,$
[(2)]{} if $\rho_B(K_0(B))$ is divisible by $p,$ then every such nonzero $\tau_{*0}$ is in $Pl(K_0(A), Aff(T)/\rho_B(K_0(B)).$
[(3)]{} if $\rho_B(K_0(B))$ is finitely generated (such as ${\mathbb Z}\oplus {\mathbb Z}_{\theta}$ for some irrational number $\theta$), then $$Apl(K_0(A), Aff(T)/\rho_B(K_0(B))\not=Pl(K_0(A),Aff(T)/\rho_B(K_0(B)),$$
[(4)]{} $\tau_{*0}=0$ is in $Apl(K_0(A), Aff(T)/\rho_B(K_0(B)))$ if and only if there is a sequence of nonzero positive elements $ \eta_n\in \rho_B(K_0(B))$ such that $\eta_n\le {\widehat{1_{M(B)}}}$ and $\eta_n$ is divisible by $p^n,$ $n=1,2,....$
To prove (1), let $\tau_{*0}: K_0(A)\to Aff(T)/\rho_B(K_0(B))$ be a nonzero [homomorphism]{}. There exists $N$ such that $\xi_n=\tau_{*0}(1/p^n)\not=0$ for all $n\ge N.$ Since $\rho_B(K_0(B))$ is dense in $Aff(T),$ one can choose $r_n\gg 0$ in $Aff(T)$ such that $\Phi(r_n)=\tau_{*0}(1/p^n),$ where $\Phi: Aff(T)\to Aff(T)/\rho_B(K_0(B))$ is the quotient map. Since $\rho_B(K_0(B))$ is dense, it is easy to find $r_n\gg 0$ in $Aff(T)$ such that $p^nr_n\le 1_{M(B)}.$ Define $\beta_n(z)=p^nr_nz$ for $z\in K_0(A)$ ($n\ge N$). Here we identify $z$ with the real number $z$ (so $rz\in Aff(T)$). Write $$G_n=\{m/p^{n+N}: m\in {\mathbb Z}\},\,\,\,n=1,2,....$$ Then $G_n$ is finitely generated and $\bigcup_{n=1}^{\infty} G_n=K_0(A).$ Moreover, $$(\Phi\circ \beta_n)|_{G_n}=\tau_{*0}|_{G_n},\,\,\, n=1,2,....$$ This proves (1).
To prove (2), we first note that $Aff(T)/\rho_B(K_0(B))$ has no $p$-torsion. Suppose $x\in Aff(T)/\rho_B(K_0(B))$ is a nonzero element so that $px=0.$ Let $y\in Aff(T)$ so that $\Phi(y)=x.$ Then $py\in \rho_B(K_0(B)).$ Since $\rho_B(K_0(B))$ is divisible by $p,$ there is $z\in \rho_B(K_0(B))$ such that $p(y-z)=0.$ Since $Aff(T)$ is torsion free, $y=z,$ or $x=0.$
We will show that $\Phi(p^N_Nz)=\tau_{*0}(z)$ for $z\in K_0(A)={\mathbb Z}[1/p].$ Note we have shown above that for any $z\in {\mathbb Z}(1/p^N),$ $\Phi(p^Nr_Nz)=\tau_{*0}(z).$ Suppose that $x=\tau_{*0}(1/p^{N+1}).$ Then $px=\tau_{*0}(1/p^N)
=\Phi(r_N).$ Thus $px=p\Phi\circ\beta_N(1/p^{N+1}).$ This implies that $p(x-\xi_N)=0$ in $Aff(T)/\rho_B(K_0(B)).$ Since $Aff(T)/\rho_B(K_0(B))$ has no $p$-torsion, $x=\xi_N.$ Therefore $\Phi\circ \beta_N(z)=\tau_{*0}(z)$ for all $z\in {\mathbb Z}(1/p^{N+1}).$ By induction, we verify that $\Phi\circ \beta_N=\tau_{*0}.$ It is clear that $\beta_N$ is positive.
For (3), we note that $ext_{\mathbb Z}({\mathbb Z}[1/p], {\mathbb Z})\not=\{0\}$ (see Theorem 99.1 and p. 179 and in [@Fu], and also [@Rot]). Therefore, since $\rho_B(K_0(B))$ is a finite sum of ${\mathbb Z},$ we conclude that $ext_{\mathbb Z}({\mathbb Z}[1/p], \rho_B(K_0(B)))\not=\{0\}.$ Consider the following exact sequence: $$\cdots {\rm Hom}({\mathbb Z}[1/p], Aff(T))\to {\rm Hom}({\mathbb Z}[1/p], Aff(T)/\rho_B(K_0(B)))
\to ext({\mathbb Z}[1/p], \rho_B(K_0(B)))\to$$ $$ext({\mathbb Z}[1/p],
Aff(T))\to ext({\mathbb Z}[1/p], Aff(T)/\rho_B(K_0(B)))\to\cdots .$$ Since $Aff(T)$ is divisible, $ext({\mathbb Z}[1/p],Aff(T))=\{0\}.$ This implies that the map from\
${\rm Hom}({\mathbb Z}[1/p], Aff(T)/\rho_B(K_0(B)))$ to $ext({\mathbb Z}[1/p], \rho_B(K_0(B)))$ is surjective. Thus we obtain $\tau_{*0}: {\mathbb Z}[1/p]\to Aff(T)/\rho_B(K_0(B))$ such that it gives a non-splitting extension of ${\mathbb Z}[1/p]$ by $\rho_B(K_0(B)).$ This $\tau_{*0}$ can not be lifted to a [homomorphism]{} from $K_0(A)$ into $Aff(T).$ In particular $\tau_{*0}\not\in
Pl(K_0(A), Aff(T)/\rho_B(K_0(B)).$ It follows from (1) that $\tau_{*0}\in
Apl(K_0(A), Aff(T)/\rho_B(K_0(B)).$ Thus (3) follows.
To see (4), suppose that there exists a sequence of nonzero positive elements $\eta_n\in \rho_B(K_0(B))$ such that $\eta_n\le {\widehat{1_{M(B)}}}$ and $\eta_n=p^n\zeta_n$ for some $\zeta_n\in \rho_B(K_0(B)),$ $n=1,2,....$ Since $\rho_B(K_0(B))$ is weakly unperforated, $\zeta_n\gg 0.$ Fix $n,$ define $\beta_n: {\mathbb Z}[1/p]\to Aff(T)$ by $\beta_n(z)=p^nz \zeta_n$ for all $z\in {\mathbb Z}[1/p]\,{\rm (}
\subset {\mathbb R} {\rm )}.$ Then $\beta_n$ is a positive [homomorphism]{}. Let $G_n$ be as in the proof of (1). Then $\beta_n(G_n)\subset \rho_B(K_0(B)),$ $n=1,2,....$ So $(\Phi\circ \beta_n)|_{G_n}=0,$ $n=1,2,....$ Therefore $\tau_{*0}=0$ is in $Apl(K_0(A), Aff(T)/\rho_B(K_0(B))).$
Conversely, if $\tau_{*0}=0$ is in $Apl(K_0(A), Aff(T)/\rho_B(K_0(B))).$ Suppose that $F_n\subset {\mathbb Z}[1/p]$ is an increasing sequence of finitely generated subgroups such that $\bigcup_{n=1}^{\infty}F_n={\mathbb Z}[1/p]$ and there is a sequence of positive [homomorphism]{}s $\alpha_n: {\mathbb Z}[1/p]
\to Aff(T)$ such that $\alpha_n(F_n)\subset \rho_B(K_0(B)),$ $n=1,2,....$ Replacing $\alpha_n$ by $t_n\alpha_n$ for some positive real numbers $t_n,$ we may assume that $\alpha_n(1)\le {\widehat{1_{M(B)}}},$ $n=1,2,....$ Without loss of generality, we may also assume that $1, 1/p^n\subset F_n,$ $n=1,2,....$ Let $\eta_n=\alpha_n(1).$ Since $\alpha_n(1/p^n)\subset \rho_B(K_0(B)),$ $\eta_n$ is divisible by $p^n,$ $n=1,2,....$
Note that $\rho_B(K_0(B))$ may not have any nonzero elements to be divided by $p.$ In these cases, the condition in (4) in the previous theorem and the next corollary never holds. In other words, when $[\tau]=0$ in $KL(A, M(B)/B),$ $\tau$ is never approximately trivial.
\[IIICl\] Let $A$ be a unital separable amenable [$C^*$-algebra]{} in ${\cal N}.$ Suppose that there is a monomorphism $j:A \to C$ for some unital simple AF-algebra $C$ with $C={\mathbb Z}[1/p].$ Suppose also that $j_{*0}$ maps $K_0(A)/{\rm ker}\rho_A$ injectively to ${\mathbb Z}[1/p].$ Let $\tau: A\to M(B)/B$ be an essential extension.
[(1)]{} If $[\tau]\not=0,$ $\tau_{*1}=0,$ $\tau_{*0}|_{{\rm ker}\rho_A}=0,$ ${\rm im}\tau_{*0}\subset Aff(T)/\rho_B(K_0(B))$ and $[\tau]|_{K_i(A,{\mathbb Z}/k{\mathbb Z})}=0,$ $i=0,1$ and $k=2,3,...$ Then $\tau$ is approximately trivial.
[(2)]{} If there is no positive [homomorphism]{} from $K_0(A)$ into $\rho_B(K_0(B)),$ then no essential extension with $[\tau]=0$ in $KL(A, M(B)/B)$ can be trivial. Furthermore, if $A=C,$ then there exists an essential trivial extension $\tau$ with $[\tau]=0$ in $KL(A, M(B)/B)$ if and only if there is a positive [homomorphism]{} $\alpha: K_0(A)\to \rho_B(K_0(B)).$
[(3)]{} Suppose further that $K_0(A)/{\rm ker}\rho_A={\mathbb Z}[1/p].$ Then $[\tau]=0$ implies that $\tau$ is approximately trivial if and only if there exists a sequence of nonzero positive elements $\eta_n\in \rho_B(K_0(B))$ such that $\eta_n\le {\widehat{1_{M(B)}}}$ and $\eta_n$ is divisible by $p^n,$ $n=1,2,....$
Suppose $j: A\to C$ is the embedding. Since $Aff(T)/\rho_B(K_0(B))$ is divisible, there exists a [homomorphism]{} $\alpha: K_0(C)\to Aff(T)/\rho_B(K_0(B))$ such that $\alpha|_{j_{*0}(K_0(A))}=\tau_{*0}$ which is nonzero. Let $\psi: C\to M(B)/B$ be an essential extension such that $\psi_{*0}=\alpha.$ Since $Aff(T)/\rho_B(K_0(B))$ is divisible and $K_1(C)=\{0\},$ one computes that $[\tau]|_{K_0(C, {\mathbb Z}/k{\mathbb Z})}=0$ for all $k.$ Since $K_0(C)$ has no torsion and $K_1(C)=0,$ $[\tau]|_{K_1(C, {\mathbb Z}/k{\mathbb Z})}=0$ for all $k.$ It follows that $$[\tau]=[\psi\circ j]\,\,\, {\rm in}\,\,\, KL(A, M(B)/B).$$ So $\tau$ and $\psi\circ j$ are approximately unitarily equivalent. Thus it suffices to show that $\psi\circ j$ is approximately trivial. It follows from \[IIP1\] that $\psi_{*0}\subset Apl(K_0(C), Aff(T)/\rho_B(K_0(B)).$ It follows from \[IITappext\] that $\psi$ is approximately trivial. Hence $\psi\circ j$ is approximately trivial.
For (2), we assume that $[\tau]=0.$ Suppose that $\tau$ is trivial and $h: A\to M(B)$ is a monomorphism such that $\pi\circ h=\tau.$ Then $h_{*0}$ gives a positive [homomorphism]{} from $K_0(A)$ into $\rho_B(K_0(B)).$
Suppose now that $A=C.$ If there is a $\alpha: K_0(A)\to \rho_B(K_0(B))\subset Aff(T),$ then by \[IIILemb\] there exists a monomorphism $h:A\to M(B)$ such that $h_{*0}=\alpha$ and $h(A)\cap B=\{0\}.$ Thus $\tau=\pi\circ h$ is trivial and $[\tau]=0.$
Now consider (3). Suppose that $[\tau]=0$ in $KL(A, M(B)/B)$ and suppose also that there is a [homomorphism]{} $h_n: A\to M(B)$ such that $$\lim_{n\to\infty}\pi\circ h_n(a)=\tau(a)\,\,\,{\rm for\,\,\, all}\,\,\, a\in A.$$ Then $\tau_{*0}\in Apl(K_0(A), Aff(T)/\rho_B(K_0(B))).$ Thus the “if only" part follows from (4) in \[IIP1\]. On the other hand, if those $\eta_n$ exists, by (4) in \[IIP1\], $\tau_{*0}\in Apl(K_0(A), Aff(T)/\rho_B(K_0(B))).$ Thus the “if" part follows from \[IITappext\].
[BK1]{}
C. A. Akemann and G. K. Pedersen, [*Central sequences and inner derivations of separable $C\sp{*} $-algebras*]{}, Amer. J. Math. [**101**]{} (1979), 1047–1061.
C. A. Akemann and F. W. Shultz, [*Perfect $C\sp *$-algebras*]{}, Mem. Amer. Math. Soc. 55 (1985), no. 326.
E. M. Alfsen, [*Compact Convex Sets and Boundary Integrals*]{}, Springer-Verlag, New York 1971.
B. Blackadar and D. Handelman, [*Dimension functions and traces on [$C^*$-algebra]{}s*]{}, J. Funct. Anal., [**45**]{} (1982), 297-340.
B. Blackadar and E. Kirchberg, [*Generalized inductive limits of finite-dimensional $C\sp *$-algebras*]{}, Math. Ann. [**307**]{} (1997), 343–380.
L. G. Brown, [*Stable isomorphism of hereditary subalgebras of $C\sp*$-algebras*]{}, Pacific J. Math. [**71**]{} (1977), 335–348.
L. G. Brown, [*The universal coefficient theorem for $Ext$ and quasidiagonality*]{}, Operator Algebras and Group Representations, vol. 17, Pitman Press, Boston, 1983, pp. 60-64.
L. G. Brown and M. Dadarlat, [*Extensions of [$C^*$-algebra]{}s and quasidiagonality*]{}, J. London Math. Soc. [**53**]{} (1996), 582-600.
L. G. Brown, R. G. Douglas and P. A. Fillmore, [*Unitary equivalence modulo the compact operators and extensions of $C\sp{*} $-algebras*]{}, Proceedings of a Conference on Operator Theory (Dalhousie Univ., Halifax, N.S., 1973), pp. 58–128. Lecture Notes in Math., Vol. 345, Springer, Berlin, 1973.
L. G. Brown, R. G. Douglas and P. A. Fillmore, [*Extensions of [$C^*$-algebra]{}s, operators with compact self-commutators and $K$-homology*]{}, Bull. Amer. Math. Soc. [**79**]{} (1973), 973-978.
L. G. Brown, R. G. Douglas and P. A. Fillmore, [*Extensions of $C\sp*$-algebras and $K$-homology*]{}, Ann. of Math. [**105**]{} (1977), 265–324.
M. D. Choi and E. Effros, [*The completely positive lifting problem for $C\sp*$-algebras*]{}, Ann. of Math. [**104**]{} (1976) 585–609.
M. Dadarlat, [*Reduction to dimension three of local spectra of real rank zero $C\sp *$-algebras*]{}, J. Reine Angew. Math. 460 (1995), 189–212.
M. Dadarlat and S. Eilers, [*On the classification of nuclear [$C^*$-algebra]{}s*]{}, preprint 1998.
M. Dadarlat and T. Loring, [*A universal multi-coefficient theorem for the Kasparov groups*]{}, Duke J. Math., [**84**]{} (1996), 355–377.
E. Effros, D. Handelman and C.-L. Shen, [*Dimension groups and their affine representations*]{}, Amer. J. Math. [**102**]{} (1980), 385–407.
G. A. Elliott, [*On the classification of [$C^*$-algebra]{}s of real rank zero*]{}, J. Reine Angew. Math. [**443**]{} (1993), 179–219.
G. A. Elliott, [*Dimension groups with torsion*]{}, Inter. J. Math. [**1**]{} (1990), 361–380.
G. A. Elliott, [*A classification of certain simple [$C^*$-algebra]{}s, in Quantum and Non-Commutative Analysis (H. Araki et. al. eds), Kluwer, Dordrecht*]{}, 1993, pp. 373–385.
G. A. Elliott and D. E. Evans, [*The structure of the irrational rotation [$C^*$-algebra]{}s*]{}, Ann. of Math. [**138**]{} (1993), 477–501.
G. A. Elliott and G. Gong, [*On the classification of [$C^*$-algebra]{}s of real rank zero, II*]{}, [*Ann. of Math.*]{} [**144**]{} (1996), 497–610.
Fuchs, Infinite Abelian Groups, Academic Press, Orlando/San Diego/New York, 1973.
G. Gong, [*On the inductive limits of matrix algebras over higher dimensional spaces*]{}, Part I & II: Math. Scand. [**80**]{} (1997) 40-55 & 56-100.
G. Gong, [*On the classification of simple inductive limit [$C^*$-algebra]{}s, I: The reduction theorem*]{}, preprint.
G. Gong and H. Lin, [*Almost multiplicative morphisms and $K$-theory*]{}, Internat. J. Math. [**11**]{} (2000), 983–1000.
K. Goodearl and D. Handelman, [*Rank functions and $K_0$ of regular rings*]{}, J. Pure Appl. Algebra, [**7**]{} (1976), 195-216.
E. Kirchberg, [Classification of purely infinite simple [$C^*$-algebra]{}s by Kasparov groups]{}, third draft, 1996.
E. Kirchberg and N. C. Phillips, [*Embedding of exact $C\sp *$-algebras in the Cuntz algebra ${\cal O}_2$*]{}, J. Reine Angew. Math. [**525**]{}, (2000), 17–53.
P. R. Halmos, [The problems in Hilbert space]{}, Bull. Amer. Math. Soc. [**76**]{} (1970), 887-933.
H. Lin, [*Simple C\*-algebras with continuous scales and simple corona algebras*]{}, Proc. Amer. Math. Soc. [**112**]{} (1991) 871-880
H. Lin, [*Notes on $K$-theory of multiplier algebras*]{}, preprint 1991.
H. Lin, [*Exponential rank of $C\sp *$-algebras with real rank zero and the Brown-Pedersen conjectures*]{}, J. Funct. Anal. [**114**]{}, (1993), 1–11.
H. Lin, [*Generalized Weyl-von Neumann theorem (II)*]{}, Math. Scand. [**77**]{} (1995), 128-147.
H. Lin, [*$C^*$-algebra Extensions of $C(X)$*]{}, Memoirs Amer. Math. Soc., [**115**]{} (1995), no. 550.
H. Lin, [*Extensions by $C^*$-algebras with real rank zero II*]{}, Proc. London Math. Soc., 71 (1995), 641-674.
H. Lin, [*Extensions by $C^*$-algebras with real rank zero III*]{}, Proc. London Math. Soc., [**76**]{} (1998), 634-666.
H. Lin, [*Extensions of $C(X)$ by simple $C^*$-algebras of real rank zero*]{}, Amer. J. Math. [**119**]{} (1997), 1263-1289.
H. Lin, [*Classification of simple [$C^*$-algebra]{}s with unique traces*]{}, Amer. J. Math., [**120**]{} (1998), 1289-1315.
H. Lin, [*Stable approximate unitary equivalence of [homomorphism]{}s*]{}, J. Operator Theory, [**47**]{} (2002), 343–378.
H. Lin, [*Tracially AF [$C^*$-algebra]{}s*]{}, Trans. Amer. Math. [**353**]{} (2001), 693–722 .
H. Lin, [*Classification of simple TAF [$C^*$-algebra]{}s*]{}, Canad. J. Math. [**53**]{} (2001), 161–194..
H. Lin, [*Locally type I simple tracially AF [$C^*$-algebra]{}s*]{}, preprint 1998.
H. Lin, [*Tracial topological rank of [$C^*$-algebra]{}s*]{}, Proc. London Math. Soc. [**83**]{} (2001), 199–234.
H. Lin, [*Simple [$C^*$-algebra]{}s with locally bounded irreducible representation*]{}, J. Funct. Anal. [**187**]{} (2001), 42–69.
H. Lin, [*An Introduction to the Classification of Amenable $C^*$-Algebras*]{}, World Scientific, 2001.
H. Lin, [*Classification of simple [$C^*$-algebra]{}s with tracial topological rank zero*]{}, Duke Math. J., to appear.
H. Lin, [*An Approximate Universal Coefficient Theorem*]{}, preprint 2001.
H. Lin, [*A separable version of Brown-Douglas-Fillmore Theorem and weak stability*]{}, Trans. Amer. Math.Soc., to appear.
H. Lin, [*Semiprojectivity for purely infinite simple [$C^*$-algebra]{}s*]{}, preprint 2002.
H. Lin, [*Simple corona algebras*]{}, Proc. Amer. Math. Soc., to appear.
Q. Lin and N. C. Phillips, [*[$C^*$-algebra]{}s of minimal diffeomorphisms*]{}, preprint.
N. C. Phillips, [*A classification theorem for nuclear purely infinite simple $C\sp *$-algebras*]{}, Doc. Math. [**5**]{} (2000), 49–114.
N. C. Phillips, [*Real rank of [$C^*$-algebra]{}s of minimal diffeomorphisms*]{}, preprint.
M. Pimsner and D. Voiculescu, [*Exact sequences for $K$-groups and $Ext$-groups of certain cross-products of [$C^*$-algebra]{}s*]{}, J. Operator Theory [**4**]{} (1980), 93-118.
M. Rørdam, [*Classification of inductive limits of Cuntz algebras*]{} J. Reine Angew. Math. [**440**]{} (1993), 175–200.
M. Rørdam, [*A short proof of Elliott’s theorem: ${\cal O}_2\otimes{\cal O}_2\cong{\cal O}_2$*]{}, C. R. Math. Rep. Acad. Sci. Canada [**16**]{} (1994), 31–36.
M. Rørdam, [*Classification of certain infinite simple $C\sp *$-algebras. III, Operator algebras and their applications*]{} (Waterloo, ON, 1994/1995), 257–282, Fields Inst. Commun., 13, Amer. Math. Soc., Providence, RI, 1997.
M. Rørdam, [*Classification of nuclear, simple $C\sp *$-algebras. Classification of nuclear $C\sp *$-algebras, Entropy in operator algebras,*]{} 1–145, Encyclopaedia Math. Sci., 126, Springer, Berlin, 2002.
M. Rørdam, [*Simple [$C^*$-algebra]{}s with finite and infinite projections*]{}, preprint. J. Rotman, [*Torsion free and mixed abelian groups*]{}, Illinois J. Math, [**5**]{} (1961), 131-143. N. Salinas, [*Relative quasidiagonality and $KK$-theory*]{}, Houston J. Math. [**18**]{} (1992), 97-116. C. Schochet, [*Topological methods for [$C^*$-algebra]{}s IV: mod p homology*]{}, Pacific J. Math., [**114**]{} (1984), 447-468. C. Schochet, [*The fine structure of the Kasparov groups II: Relative quasidiagonality*]{}, preprint D. Voiculescu, [*A note on quasi-diagonal [$C^*$-algebra]{}s and homotopy*]{}, Duke Math. J. [**62**]{} (1991), 267-271. D. Voiculescu, [*Around quasidiagonal operators*]{}, Integral Equat. Operator Theory [**17**]{} (1993), 137-148.
S. Zhang, [*On the structure of projections and ideals of corona algebras*]{}, Canad. J. Math. [**41**]{} (1989), 721–742.
S. Zhang, [*Diagonalizing projections in multiplier algebras and in matrix over a [$C^*$-algebra]{}*]{}, Pacific J. Math., [**145**]{} (1990), 181-200. S. Zhang, [*A property of purely infinite simple $C\sp *$-algebras*]{}, Proc. Amer. Math. Soc. [**109**]{} (1990), 717–720.
[^1]: Research partially supported by NSF grants DMS 0097903. AMS 1991 Subject Classification Numbers: Primary 46L05, 46L35. Key words: Extensions, Simple $C^*$-algebras.\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider spatially homogeneous, anisotropic cosmological models in $5D$ whose line element can be written as $dS^2 = {\cal{A}}(u, v)du dv - {\cal{B}}_{i j}(u, v)dx^{i}dx^{j}$, $(i, j = 1, 2, 3)$, where $u$ and $v$ are light-like coordinates. In the case where ${\cal{B}}_{i j}$ is diagonal, we construct three families of analytic solutions to the $5D$ vacuum field equations $R_{AB} = 0$ $ (A, B = 0, 1, 2, 3, 4)$. Among them, there is a family of self-similar homothetic solutions that contains, as a particular case, the so-called light-like Kasner universes. In this work we provide a detailed study of the different types of $4D$ scenarios that can be embedded in such universes. For the sake of generality of the discussion, and applicability of the results, in our analysis we consider the two versions of non-compactified $5D$ relativity in vogue, viz., braneworld theory and induced matter theory. We find a great variety of cosmological models in $4D$ which are anisotropic versions of the FRW ones. We obtain models on the brane with a non-vanishing cosmological term $\Lambda_{(4)}$, which inflate [*à la*]{} de Sitter without satisfying the classical false-vacuum equation of state. Using the symmetry of the solutions, we construct a class of non-static vacuum solutions on the brane. We also develop [*static*]{} pancake-like distributions where the matter is concentrated in a thin surface (near $z = 0$), similar to those proposed by Zel’dovich for the shape of the first collapsed objects in an expanding anisotropic universe. The solutions discussed here can be applied in a variety of physical situations.'
author:
- |
J. Ponce de Leon[^1]\
Laboratory of Theoretical Physics, Department of Physics\
University of Puerto Rico, P.O. Box 23343, San Juan,\
PR 00931, USA
title: '$4D$ spacetimes embedded in $5D$ light-like Kasner universes'
---
PACS: 04.50.+h; 04.20.Cv
[*Keywords:*]{} Cosmological models; Kasner universes; 5D Models; Braneworld theory; Induced matter theory; Exact solutions; General Relativity.
Introduction
============
In recent years there has been an increased interest in theories that envision our spacetime as embedded in a universe with more than four large dimensions. There are several reasons that justify this interest, among them that extensions of four-dimensional general relativity to five and more dimensions seem to provide the best route to unification of gravity with the interactions of particle physics [@Davidson; @Owen]-[@particle; @physics]. In $5D$ there are two versions of relativity where the extra dimension is not assumed to be compactified. These are membrane theory [@algebraically] and space-time-matter (or induced matter) theory [@general]. They lead to a great variety of models both in the cosmological context and in the description of local self-gravitating objects (see, e.g., [@review], [@available]). Most of these models have been obtained in coordinates where the metric in $5D$ can be written as[^2] $$\label{General line element in 5D without off-diagonal terms}
dS^2 = g_{\mu\nu}(x^{\rho}, \; \psi)dx^{\mu}dx^{\nu} + \epsilon \Phi^2(x^{\rho}, \; \psi)d\psi^2,$$ in such a way that our $4D$ spacetime can be recovered by going onto a hypersurface $\Sigma_{\psi}: \psi = \psi_{0} = $ constant, which is orthogonal to the $5D$ unit vector $$\label{unit vector n}
{\hat{n}}^{A} = \frac{\delta^{A}_{4}}{\sqrt{\epsilon g_{44}}}, \;\;\;n_{A}n^{A} = \epsilon,$$ along the extra dimension, and $g_{\mu\nu}$ can be interpreted as the metric of the spacetime.
In this framework, the effective equations for gravity in $4D$ are obtained from dimensional reduction of the Einstein field equations in $5D$. The reduction is based on Campbell’s theorem [@Campbell], [@Seahra] and consists in isolating the $4D$ part of the relevant $5D$ geometric quantities and use them to construct the $4D$ Einstein tensor ${^{(4)}G}_{\alpha \beta}$. The crucial result is that, even in the case where the energy-momentum tensor (EMT) in $5D$ is zero, to an observer confined to making physical measurements in our ordinary spacetime, and not aware of the extra dimension, the spacetime is not empty but contains (effective) matter whose EMT, ${^{(4)}T}_{\alpha\beta}$, is determined by the Einstein equations in $4D$, namely $$\begin{aligned}
\label{4D Einstein with T and K}
{^{(4)}G}_{\alpha\beta} = 8 \pi \;{^{(4)}T}_{\alpha\beta} =
- \epsilon\left(K_{\alpha\lambda}K^{\lambda}_{\beta} - K_{\lambda}^{\lambda}K_{\alpha\beta}\right) + \frac{\epsilon}{2} g_{\alpha\beta}\left(K_{\lambda\rho}K^{\lambda\rho} - (K^{\lambda}_{\lambda})^2 \right) - \epsilon E_{\alpha\beta}, \end{aligned}$$ where $K_{\mu\nu}$ is the extrinsic curvature $$\label{extrinsic curvature}
K_{\alpha\beta} = \frac{1}{2}{\cal{L}}_{\hat{n}}g_{\alpha\beta} = \frac{1}{2\Phi}\frac{\partial{g_{\alpha\beta}}}{\partial \psi};$$ $E_{\mu\nu}$ is the projection of the bulk Weyl tensor ${^{(5)}C}_{ABCD}$ orthogonal to ${\hat{n}}^A$, i.e., “parallel" to spacetime, viz., $$\label{Weyl Tensor}
E_{\alpha\beta} = {^{(5)}C}_{\alpha A \beta B}{\hat{n}}^A{\hat{n}}^B
= - \frac{1}{\Phi}\frac{\partial K_{\alpha\beta}}{\partial \psi} + K_{\alpha\rho}K^{\rho}_{\beta} - \epsilon \frac{\Phi_{\alpha;\beta}}{\Phi},$$ and $\Phi_{\alpha} \equiv \partial \Phi/\partial x^{\alpha}$. Before going on, it is worthwhile to emphasize that the above dimensional reduction of the field equations in $5D$ is a standard technique that leads to the same effective matter content in $4D$, i.e., ${^{(4)}T}_{\alpha\beta}$, regardless of whether the line element (\[General line element in 5D without off-diagonal terms\]) is interpreted within the context of brane theory with ${\bf Z}_2$ symmetry [@Shiromizu] or space-time-matter (STM) theory [@Wesson; @and; @JPdeL]. In this sense these two approaches to $5D$ relativity are mathematically equivalent. However, they are different as regards physical interpretation and motivation [@physical; @motivation]. In brane theory there is a singular hypersurface that defines spacetime, and the properties of matter in that hypersurface are, in general, [*not identical*]{} to the ones of induced matter calculated in STM from the effective EMT defined by (\[4D Einstein with T and K\]).
In the cosmological realm, nearly all models assume spatial homogeneity and isotropy, which means that the line element in $5D$ is taken to be an extended version of the conventional Friedmann-Roberson-Walker (FRW) metric in $4D$, namely $$\label{Usual cosmological metric in 5D}
dS^2 = n^2(t, \psi)dt^2 - a^2(t, \psi)\gamma_{i j} dx^{i}dx^{j} + \epsilon \Phi^2(t, \psi) d\psi^2,\;\;\;i, j = 1, 2, 3,$$ where $\gamma_{i j}$ is a maximally symmetric $3$-dimensional metric, with curvature index $k = 0, \pm 1$. In these coordinates the full integration of the vacuum Einstein field equations in $5D$ requires the specification of [*two*]{} additional assumptions. One of them is usually an assumption of geometric nature, e.g., that $\Phi = 1$, or $n = 1$. The second one is usually an equation of state for the matter quantities in $4D$ [@Binetruy], [@JPdeL-isotropicCosm].
Observations indicate that on large scales ($\gg 100$ Mpc) the universe is homogeneous and isotropic and well described by spatially-flat FRW cosmologies. However, there is no reason to expect that such features should hold at the early stages of the evolution of the universe. Rather, it is generally accepted that anisotropy could have played a significant role in the early universe and that it has been fading away in the course of cosmic evolution. In the framework of 4-dimensional spacetime, a prototype for anisotropic vacuum cosmologies is provided by the Kasner metric [@Kasner], which mimics the behavior of more general solutions near the singularity during some finite periods of time[^3]. Various higher dimensional extensions of the vacuum Kasner model have been discussed in the literature [@Kokarev], [@Paul].
In this paper we consider spatially homogeneous but anisotropic cosmological models whose metric in $5D$ has the form $$\label{light-like 5D metrics}
dS^2 = {\cal{A}}(u, v)du dv - {\cal{B}}_{i j}(u, v)dx^{i}dx^{j},$$ where ${\cal{A}}$ and ${\cal{B}}$ are some functions of the “light-like" coordinates $u$ and $v$. These metrics are different from (\[Usual cosmological metric in 5D\]) in various aspects: (i) they do not contain the time or extra dimension in an explicit way; (ii) the hypersurfaces of constant $u$ or $v$ are three-dimensional instead of $4D$; (ii) [*[a priori]{}*]{} it is not clear how to define the $5D$ unit vector ${\hat{n}}^{A}$ along the extra dimension, which in turn is needed for defining the spacetime sections and for constructing the appropriate projected quantities in $4D$.
Here, for the case where ${\cal{B}}_{i j}$ is diagonal we construct three families of analytic solutions to the $5D$ vacuum field equations $R_{AB} = 0$ $ (A, B = 0, 1, 2, 3, 4)$. The simplest one is a family of self-similar solutions[^4], which contains as a particular case the so-called light-like Kasner universes. From a physical point of view, self-similar homothetic models are interesting because they may serve as asymptotic regimes, i.e., near the initial cosmological singularity and at late times, for many homogeneous and inhomogeneous cosmological models [@Coley]. The other two families of solutions are obtained under the assumption that some of the metric coefficients are separable functions of their arguments.
In view of their potential relevance to the “similarity hypothesis" [@Coley], in this work we focus our attention to the family of self-similar $5D$ spacetimes mentioned above. The main question under study is what kind of $4D$ scenarios can be embedded in such spacetimes. For the sake of generality of the discussion, and applicability of our results, in our analysis we consider both versions of non-compactified $5D$ relativity, viz., induced-matter and brane theory. Unfortunately, the expressions obtained in $4D$ as projections of the $5D$ solutions are quite complex and cumbersome. Therefore, to obtain manageable mathematical expressions in $4D$, in sections $3$, $4$ and $5$ we simplify the algebra (but not the physics) by restricting our discussion to the subset of light-like Kasner solutions.
Our solutions generalize a number of isotropic cosmological models and give back previous ones in the literature (see e.g., [@JPdelW]-[@JPdeL-JCAP] and references therein). Although we are not discussing particular applications here, they could be useful in the study of generalizations of Mixmaster or Belinskii-Khalatnikov-Lifshitz oscillations in theories with a single extra dimension [@mixmaster], [@BKL], [@Halpern]-[@Henneaux]. They could also be applied to studying conjectures about isotropic Big Bang singularities in braneworlds [@Maartens]-[@Coley2]. Certain cosmological models, such as the cyclic universe model, also require an understanding of the behavior of Kasner-like solutions and are based on brane-type models [@Erickson].
The paper is organized as follows. In section $2$ we derive our self-similar solution (the other two families of solutions are presented in the Appendix) and introduce a timelike coordinate $t$ and a spacelike coordinate $\psi$ along the extra dimension. This is equivalent to introducing two additional degrees of freedom, which are expressed in terms of two functions of $t$ and $\psi$. We will see that, as in the familiar FRW picture (\[Usual cosmological metric in 5D\]), these two functions can be related to the specific choice for embedding $\Sigma_{\psi}$ in $5D$ and to the physics in $4D$. In section $3$, within the context of STM we show that the light-like Kasner solutions generate a great variety of cosmological models in $4D$, including the de Sitter, Milne and power-law FRW models. In section $4$, within the context of the braneworld paradigm we find that they embed $4D$ cosmological models with a non-vanishing cosmological term $\Lambda_{(4)}$, which in principle can be either constant or time-dependent. In the case of constant $\Lambda_{(4)}$ the $3D$ space exponentially inflates, regardless of the specific embedding. We also show that by virtue of the symmetry $(x, y, z) \leftrightarrow \psi$, they generate a class of non-static vacuum solutions on the brane. In section $5$, once again using symmetry properties, we demonstrate that the Kasner-like metric (\[Lightlike Kasner solution\]) can be used to generate static pancake-like distributions of matter in $4D$, similar to those proposed by Zel’dovich for the shape of the first collapsed objects in an expanding anisotropic universe [@Zel'dovich]. In section $6$ we present a summary of our results.
Cosmological models in $5D$. Light-like coordinates
===================================================
Solving the field equations. Part I
-----------------------------------
In this work we obtain three families of solutions to the field equations $R_{A B} = 0$. However, to facilitate the discussion, in this subsection we present only one of them. Specifically, we present the family of solutions that we will use throughout the paper, which (as we will see in sections $3$-$5$) may be interpreted or used as $5D$ embeddings for a number of $4D$ universes. The derivation of the other two families of solutions, whose $4D$ interpretation is not discussed here, is deferred to the Appendix (Solving the field equations. Part II).
To simplify the shape of the field equations let us momentarily denote ${\cal{B}}_{11} = e^{\lambda(u, v)}$, ${\cal{B}}_{22} = e^{\mu(u, v)}$ and ${\cal{B}}_{33} = e^{\sigma(u, v)}$. From $R_{xx} = 0$, $R_{yy} = 0$ $R_{zz} = 0$ we obtain the equations $$\begin{aligned}
\label{Rxx, Ryy, Rzz}
4 \lambda_{u v} + \lambda_{u}\left( \sigma_{v} + 2 \lambda_{v} + \mu_{v}\right) + \lambda_{v}\left(\mu_{u} + \sigma_{u}\right) &=& 0, \nonumber \\
4 \mu_{u v} + \mu_{u}\left( \lambda_{v} + 2 \mu_{v} + \sigma_{v}\right) + \mu_{v}\left(\sigma_{u} + \lambda_{u}\right) &=& 0, \nonumber \\
4 \sigma_{u v} + \sigma_{u}\left( \mu_{v} + 2 \sigma_{v} + \lambda_{v}\right) + \sigma_{v}\left(\lambda_{u} + \mu_{u}\right) &=& 0.\end{aligned}$$ Here the subscripts $u$, $v$ indicate partial derivatives with respect to those arguments. The above equations show cyclic permutation symmetry, i.e., starting from any of them by means of the transformation $\lambda \rightarrow \mu \rightarrow \sigma \rightarrow \lambda$ we obtain the other two.
$\bullet$ Self-similar solutions: First we solve the field equations under the assumption that the metric (\[light-like 5D metrics\]) possesses self-similar symmetry. This assumption is motivated by a number of studies suggesting that self-similar models play a significant role at asymptotic regimes [@Coley]. From a mathematical point of view, it means that by a suitable transformation of coordinates all the dimensionless quantities can be put in a form where they are functions only of a single variable (say $\zeta$) [@Sedov]. In our particular case, this implies that $\lambda = \lambda(\zeta)$, $\mu = \mu(\zeta)$, $\sigma = \sigma(\zeta)$, where $\zeta$ is some function of $u$ and $v$, viz., $$\label{z}
\zeta = \zeta(u, v).$$ With this assumption the first equation in (\[Rxx, Ryy, Rzz\]) reduces to $$\label{equation for lambda in the self-similar solution}
2 \; \frac{\lambda_{\zeta \zeta}}{\lambda_{\zeta}} + \left(\lambda_{\zeta} + \mu_{\zeta} + \sigma_{\zeta}\right) + 2\; \frac{{\zeta}_{u v}}{{\zeta}_{u}{\zeta}_{v}} = 0.$$ The assumed symmetry requires $\left({{\zeta}_{u v}}/{{\zeta}_{u}{\zeta}_{v}}\right)$ to be some function of $\zeta$, say $ Z(\zeta) = \left({{\zeta}_{u v}}/{{\zeta}_{u}{\zeta}_{v}}\right)$. Integrating we get $$\label{lambdaz}
\lambda_{\zeta} = 2 \alpha \; e^{- \left(\lambda + \mu + \sigma\right)/2} e^{- \int{Z(\zeta) d\zeta}} \equiv 2 \alpha \left(\frac{f_{\zeta}}{f}\right),$$ where $\alpha$ is an arbitrary constant of integration. Similar equations, with new constants, e.g., $\beta$ and $\gamma$, are obtained for $\mu$ and $\sigma$ by means of a cyclic transformation. What this means is that $$\label{lambdaz, muz, sigma z}
\frac{\lambda_{\zeta}}{2 \alpha} = \frac{\mu_{\zeta}}{2 \beta} = \frac{\sigma_{\zeta}}{2 \gamma} = \frac{f_{\zeta}}{f},$$ which upon integration yields $$\label{elambda, emu, esigma}
e^{\lambda} = C_{1}f^{2 \alpha}, \;\;\;e^{\mu} = C_{2} f^{2 \beta}, \;\;\;e^{\sigma} = C_{3} f^{2 \gamma},$$ where $C_{1}$, $C_{2}$, $C_{3}$ are constants of integration. A single differential equation for $f(\zeta) = f(u, v)$ can be easily obtained by substituting (\[elambda, emu, esigma\]) into any (\[Rxx, Ryy, Rzz\]), namely $$\label{equation for f}
f f_{u v} + \left(a - 1\right) f_{u} f_{v} = 0.$$
[*Notation*]{}: Here and henceforth we denote $$\label{notation}
a \equiv \alpha + \beta + \gamma, \;\;\;b \equiv \alpha^2 + \beta^2 + \gamma^2 - \alpha - \beta - \gamma, \;\;\;c \equiv \alpha \beta + \alpha\gamma + \beta \gamma,$$ where $\alpha, \beta, \gamma$ are arbitrary parameters.
A simple integration gives $$\label{f(u, v)}
f = \left[h(u) + g(v)\right]^{1/a},$$ where $h(u)$ and $g(v)$ are arbitrary functions of their arguments. Clearly, in the present case ${\cal{B}}_{i j}$ are power-law type solutions of the similarity variable $\zeta = \left[h(u) + g(v)\right]$.
To simplify the discussion, and eliminate spurious degrees of freedom, we now make the coordinate transformation $h(u) = c_{1} \bar{u}$, $g(v) = c_{2} \bar{v}$, where $c_{1}$ and $c_{2}$ are constants. In these new coordinates $${\cal{A}}(u, v)du dv \rightarrow \bar{{\cal{A}}}(\bar{u}, \bar{v})d\bar{u} d\bar{v},$$ where $\bar{{\cal{A}}}(\bar{u}, \bar{v}) = \left[{\cal{A}}(u, v)/h_{u}g_{v}\right]$ with $u$ and $v$ expressed in terms of $\bar{u}$, $\bar{v}$. Then relabeling the coordinates (dropping the overbars) the metric becomes
$$dS^2 = {\cal{A}}(u, v) du d v - C_{1}\left[c_{1} u + c_{2}v\right]^{2\alpha/a}dx^2 - C_{2}\left[c_{1} u + c_{2}v\right]^{2\beta/a}dy^2 - C_{3}\left[c_{1} u + c_{2}v\right]^{2\gamma/a}dz^2.$$
For this metric the field equations $R_{u u} = 0$ and $R_{vv} = 0$ yield the equations $$\begin{aligned}
\label{Ruu, Rvv}
a^2 \left[c_{1} u + c_{2}v\right]{\cal{A}}_{u} &+& 2 c \; c_{1} {\cal{A}} = 0, \nonumber \\
a^2 \left[c_{1} u + c_{2}v\right]{\cal{A}}_{v} &+& 2 c \; c_{2} {\cal{A}} = 0,\end{aligned}$$ which have a unique solution given by ${\cal{A}} = C_{0} \left(c_{1} u + c_{2} v\right)^{- 2c/a^2}$, where $C_{0}$ is a constant of integration. Now, it is easy to verify that $R_{u v} = 0$ is identically satisfied. In summary, the final form of the self-similar solution is given by[^5] $$\label{new solution}
{\cal{A}} = \left(c_{1} u + c_{2} v\right)^{- 2c/a^2}, \;\;\;{\cal{B}}_{11} = {\cal{A}}^{- \alpha a/c}, \;\;\;{\cal{B}}_{22} = {\cal{A}}^{- \beta a/c}, \;\;\;{\cal{B}}_{33} = {\cal{A}}^{- \gamma a/c}, \;\;\;{\cal{B}}_{i j} = 0, \;\;\;i \neq j.$$
We note that this solution admits a homothetic Killing vector in $5D$ for any values of $\alpha$, $\beta$ and $\gamma$, namely, $$\label{Lie derivative of the self-similar metric}
{\cal{L}}_{{\zeta}}{g_{A B}} = 2 g_{A B}, \;\;\;\mbox{with}\;\;\; {\zeta}^C = \left[ {\eta}_{0} u,\; {\eta}_{0} v,\; (1 - \alpha {\eta}_{0}/a) x,\; (1 - \beta {\eta}_{0}/a) y,\; (1 - \gamma {\eta}_{0}/a) z \right],$$ where $g_{A B}$ is the metric (\[new solution\]), ${\cal{L}}_{\zeta}$ denotes the Lie derivative along the $5D$ vector ${\zeta}^C$ and ${\eta}_{0} \equiv a^2/(a^2 - c)$. In addition, by setting one of the constants equal to zero, say $c_{2} = 0$, and making the coordinate transformation $u^{- 2c/a^2}du \rightarrow d\bar{u} $, it reduces to $$\label{Lightlike Kasner solution}
dS^2 = d\bar{u} dv -A\bar{u}^{p_{1}}dx^2 - B \bar{u}^{p_{2}}dy^2 - C \bar{u}^{p_{3}}dz^2,$$ where $A, B, C$ are constants with the appropriate units, and $p_{1}, p_{2}, p_{3}$ denote $$\label{introduction of p1, p2 and p3}
p_{1} = \frac{2 \alpha (\alpha + \beta + \gamma)}{\alpha^2 + \beta^2 + \gamma^2}, \;\;\;p_{2} = \frac{2 \beta (\alpha + \beta + \gamma)}{\alpha^2 + \beta^2 + \gamma^2}, \;\;\;p_{3} = \frac{2 \gamma (\alpha + \beta + \gamma)}{\alpha^2 + \beta^2 + \gamma^2},$$ which satisfy the relation $\sum_{i = 1}^{3}\left(p_{i} - 1\right)^2 = 3$ for [*any*]{} values of $\alpha, \beta $ and $\gamma$. The metric (\[Lightlike Kasner solution\]) is usually called light-like Kasner solution. In this case the $5D$ homothetic vector is given by ${\zeta}^C = \left[\bar{u},\; v,\; (1 - p_{1}/2) x,\; (1 - p_{2}/2) y,\; (1 - p_{3}/2) z\right]$.
Introducing the timelike and “extra" coordinates
------------------------------------------------
In order to be able to apply the standard dimensional reduction (\[4D Einstein with T and K\]) to metrics (\[light-like 5D metrics\]) one has to introduce coordinates that are adapted to the spacetime sections $\Sigma_{\psi}$. With this aim we make the coordinate transformation
$$\label{definition of F and V}
u = F(t, \psi), \;\;\;v = V(t, \psi),$$
where $t$ is assumed to be the timelike coordinate; $\psi$ the “extra" coordinate; $F$ and $V$ are, in principle, arbitrary differentiable functions of their arguments, except from the condition that the Jacobian of the transformation must be nonzero.
With this transformation we obtain
$$\label{dudv}
du dv = \left(\dot{F}\dot{V} dt^2 + F' V' d\psi^2\right) + \left(\dot{F} V' + F' \dot{V}\right)dt d\psi,$$
where dots and primes denote derivatives with respect to $t$ and $\psi$, respectively. We can choose the coordinates $t, \psi$ in such a way that the $5D$ metric be diagonal. This requires $$\label{diagonal 5D metric}
V' = - \frac{F' \dot{V}}{\dot{F}}.$$ As a consequence, the line element (\[light-like 5D metrics\]) becomes $$\label{light-like metric in t-psi coordinates}
dS^2 = \bar{{\cal{A}}}(t, \psi)\dot{F}\dot{V}dt^2 - {\bar{{\cal{B}}}}_{i j}(t, \psi)dx^{i} dx^{j} - \bar{{\cal{A}}}(t, \psi)\frac{F'^2 \dot{V}}{\dot{F}} d\psi^2,$$ where $\bar{{\cal{A}}}(t, \psi) \equiv {\cal{A}}(F, V)$ and ${\bar{{\cal{B}}}}_{i j}(t, \psi) \equiv {{\cal{B}}}_{i j}(F, V)$. A couple of points should be noticed here. Firstly, that the physical requirement $g_{00} > 0$ demands $\psi$ to be spacelike. Secondly, that the line element (\[light-like metric in t-psi coordinates\]) contains two arbitrary functions, which are not present in the original solution (\[new solution\]). The question is, why? Is this a mathematical, or gauge, artifact?
The answer to this question is that the arbitrary functions in (\[light-like metric in t-psi coordinates\]) are [*not*]{} gauge artifacts. They reflect the physical reality that there are many ways of embedding a $4D$ spacetime in $5D$ while satisfying the field equations. If we choose some particular embedding we obtain a differential constraint connecting $V$ and $F$, which allows us to obtain one of them in terms of the other, e.g., $V$ in terms of $F$. Then, the remaining unknown function, e.g. $F$, can be determined from the physics in $4D$. As an illustration of the former assertion, let us consider two common embeddings that arise from the choice of the coordinate/reference system.
#### Gaussian normal coordinate system:
A popular choice in the literature is to use the five degrees of coordinate freedom to set $g_{4 \mu} = 0$ and $g_{44} = - 1$. This is the so-called ‘Gaussian normal coordinate system’ based on $\Sigma_{\psi}$. Consequently, in such coordinates $\dot{V} = ({\dot{F}}/{\bar{{\cal{A}}} F'^2})$ and (\[diagonal 5D metric\]) becomes $V' = - (1/{\bar{{\cal{A}}} F'})$. Now the condition $(\partial \dot{V}/\partial \psi) = (\partial {V'}/\partial t)$ yields $$\label{obtaining V from F in Gaussian coord.}
a^2\left(c_{1}F + c_{2}V\right) F'' - 2c c_{1} F'^2 = 0.$$ If $c_{2} \neq 0$, this equation gives $V(t, \psi)$ for any smooth function $F(t, \psi)$, and the metric (\[light-like metric in t-psi coordinates\]) becomes
$$\label{the metric in Gaussian coordinates}
dS^2 = \left(\frac{\dot{F}}{F'}\right)^2 dt^2 - {\bar{{\cal{B}}}}_{i j}(t, \psi)dx^{i} dx^{j} - d\psi^2.$$
If $c_{2} = 0$, then (\[obtaining V from F in Gaussian coord.\]) is an equation for $F$. Integrating it we find $$\label{F in Gaussian coordinates, c2 = 0}
F = \left[l(t) + \psi h(t)\right]^{a^2/(a + b)},$$ where $l(t)$ and $h(t)$ are arbitrary differentiable functions.
#### Synchronous reference system:
The choice $g_{00} = 1$ is usual in cosmology: it corresponds to the so-called synchronous reference system where the coordinate $t$ is the proper time at each point. Thus, setting $\dot{V} = (1/{\bar{{\cal{A}}}}\dot{F})$, the line element (\[light-like metric in t-psi coordinates\]) becomes
$$\label{cosmological homogeneous solution in synchronous coordinates }
dS^2 = dt^2 - {\bar{{\cal{B}}}}_{i j}(t, \psi)dx^{i} dx^{j} - \left(\frac{F'}{\dot{F}}\right)^2d\psi^2.$$
In these coordinates (\[diagonal 5D metric\]) reduces to $V' = - (F'/{\bar{{\cal{A}}}}{\dot{F}}^2)$ and $(\partial \dot{V}/\partial \psi) = (\partial {V'}/\partial t)$ yields $$\label{obtaining V from F in Synchronous coord.}
a^2\left(c_{1}F + c_{2}V\right) \ddot{F} - 2c c_{1} \dot{F}^2 = 0.$$ Thus, for $c_{2} = 0$ we get $$\label{F in the synchronous coordinates}
F = \left[M(\psi) + t N(\psi)\right]^{a^2/(a + b)},$$ where $M$ and $N$ are arbitrary differentiable functions of $\psi$. For any other $c_{2} \neq 0$, we obtain $V$ from (\[obtaining V from F in Synchronous coord.\]) after choosing some smooth function $F(t, \psi)$.
Thus, in principle the function $V$ can be determined if we know $F$. At this point the question arises of whether we can single out the function $F$ from “physical" considerations in $4D$. Further analysis of the field equations shows that if we assume an equation of state for the matter in $4D$, then we obtain an extra differential equation connecting $V$ and $F$, which in addition to (\[obtaining V from F in Gaussian coord.\]) or (\[obtaining V from F in Synchronous coord.\]), allows to express the solution (\[light-like metric in t-psi coordinates\]) in terms of $t$ and $\psi$. This is what is required for the $4 + 1$ dimensional reduction of the $5D$ solutions. The general calculations are straightforward, but the equations are notational cumbersome in both STM and braneworld theory. On the other hand, (\[F in Gaussian coordinates, c2 = 0\]) and (\[F in the synchronous coordinates\]) indicate that a great algebraic simplification is attained if $c_{2} = 0$. In fact, in this case we can re-scale the function $F$, as $F \rightarrow \bar{F}^{a^2/(a + b)}$, after which the solution (\[light-like metric in t-psi coordinates\]) with $c_{2} = 0$ reduces to $$\label{Kasner-like metric in terms of F-bar}
dS^2 = \dot{\bar{F}}\dot{V} dt^2 - A {\bar{F}}^{p_{1}}dx^2 - B {\bar{F}}^{p_{2}}dy^2 - C {\bar{F}}^{p_{3}}dz^2 - \frac{{\bar{F}}'^2 \dot{V}}{\dot{\bar{F}}} d\psi^2,$$ where $p_{1}, p_{2}, p_{3}$ are the parameters introduced in (\[introduction of p1, p2 and p3\]). In sections $3$, $4$ and $5$ we use this line element, which we call Kasner-like, for illustrating the fact that physics in $4D$ determines $F$.
Cosmological models in $4D$. The STM approach
=============================================
The aim of this section is to determine $F$ within the context of induced matter theory. To this end we assume an equation of state for the effective matter quantities. Our results show that the light-like Kasner metrics (\[Lightlike Kasner solution\]) can be used, or interpreted, as $5$-dimensional embeddings for a number of cosmological models in $4D$ that are spatially anisotropic extensions of the FRW ones.
For the Kasner-like metric (\[Kasner-like metric in terms of F-bar\]), the components of the effective EMT induced on $\Sigma_{\psi}: \psi = \psi_{0} = $ constant are given by (in what follows we simplify the notation by omitting the bar over $F$ in (\[Kasner-like metric in terms of F-bar\]) and the index $^{(4)}$ in $^{(4)}T_{\mu\nu}$)
$$\begin{aligned}
\label{EMT}
8 \pi G T_{0}^{0} &=& \frac{a^2 c \dot{F}}{(a + b)^2 \dot{V} F^2}, \nonumber \\
8\pi G T_{1}^{1} &=& \frac{a \left\{(\gamma + \beta)(a + b) F \left[\frac{\ddot{F}}{\dot{F}} - \frac{\ddot{V}}{\dot{V}}\right] - 2 c(\alpha - \beta - \gamma )\dot{F}\right\}}{2(a + b)^2 \dot{V} F^2}, \nonumber \\
8\pi G T_{2}^{2} &=& \frac{a \left\{(\gamma + \alpha)(a + b) F \left[\frac{\ddot{F}}{\dot{F}} - \frac{\ddot{V}}{\dot{V}}\right] - 2 c(\beta - \alpha - \gamma )\dot{F}\right\}}{2(a + b)^2 \dot{V} F^2}, \nonumber \\
8\pi G T_{3}^{3} &=& \frac{a \left\{(\alpha + \beta)(a + b) F \left[\frac{\ddot{F}}{\dot{F}} - \frac{\ddot{V}}{\dot{V}}\right] - 2 c(\gamma - \alpha - \beta )\dot{F}\right\}}{2(a + b)^2 \dot{V} F^2}.\end{aligned}$$
Certainly the specific shape of the EMT depends on the embedding. However, there are a number of relationships, between the components of the EMT, which are “embedding-independent". These are $$\begin{aligned}
\label{rel between the stresses}
(\gamma - \beta)T_{1}^{1} + (\alpha - \gamma)T_{2}^{2} + (\beta - \alpha)T_{3}^{3} = 0,\end{aligned}$$ and $$\begin{aligned}
\label{rel between T0 and the stresses}
(\alpha + \gamma)T_{1}^{1} - (\beta + \gamma)T_{2}^{2} &=& (\beta - \alpha)T_{0}^{0},\nonumber \\
(\alpha + \beta)T_{1}^{1} - (\beta + \gamma)T_{3}^{3} &=& (\gamma - \alpha)T_{0}^{0},\nonumber \\
(\alpha + \beta)T_{2}^{2} - (\alpha + \gamma)T_{3}^{3} &=& (\gamma - \beta)T_{0}^{0}.\end{aligned}$$
Let us notice some particular cases: (i) If two of the parameters are equal to each other (axial symmetry), say $\alpha = \beta$, then $T_{1}^{1} = T_{2}^{2}$; (ii) If $T_{1}^{1} = T_{2}^{2}$ but $\alpha \neq \beta$, then $T_{1}^{1} = T_{2}^{2} = T_{3}^{3} = - T_{0}^{0}$; (iii) If $\alpha = - \beta$, then $T_{3}^{3} = - T_{0}^{0}$; (iv) In the case of isotropic expansion $(\alpha = \beta = \gamma)$ then $T_{1}^{1} = T_{2}^{2} = T_{3}^{3}$ (but not necessarily $T_{1}^{1} = T_{2}^{2} = T_{3}^{3} = - T_{0}^{0}$).
Perfect Fluid
-------------
Let us consider the case where the effective EMT behaves like a perfect fluid. From (\[EMT\]) we find that $T_{1}^{1} = T_{2}^{2} = T_{3}^{3}$ requires $$\label{perfect fluid case}
(a + b)\left(\frac{\ddot{F}}{\dot{F}} - \frac{\ddot{V}}{\dot{V}}\right) + 4 c \left(\frac{\dot{F}}{F}\right) = 0,$$ which implies $\dot{V} \propto \dot{F} F^{4 c/(a + b)}$. Substituting this into (\[diagonal 5D metric\]) and using $(\partial \dot{V}/\partial \psi) = (\partial {V'}/\partial t)$ we find that $F$ must satisfy the equation[^6] $$\label{equation for F, perfect fluid}
(a + b)F \dot{F}' + 4 c \dot{F}F' = 0,$$ from which we get $$\label{F for perfect fluid}
F = \left[f(t) + g(\psi)\right]^{(a + b)/(4 c + a + b)},$$ where $f(t)$ and $g(\psi)$ are arbitrary functions of their arguments. The effective energy density $\rho^{(eff)} \equiv T_{0}^{0}$ and pressure $p^{(eff)} \equiv - T_{1}^{1} = - T_{2}^{2} = - T_{3}^{3}$ are given by $$\label{density and pressure, perfect fluid}
\rho^{(eff)} = p^{(eff)}, \;\;\;8\pi G \rho^{(eff)} = \frac{c a^2}{(a + b)^2 F^{2a^2/(a + b)}}.$$
Ultra-relativistic matter and radiation
---------------------------------------
It is well-known that in the case of radiation as well as for ultra-relativistic matter (i.e., particles with finite rest mass moving close to the speed of light) the trace of the EMT vanishes identically. From (\[EMT\]) we find that $T = T_{0}^{0} + T_{1}^{1} + T_{2}^{2} + T_{3}^{3} = 0$ requires $$\label{ultra-relativistic matter}
(a + b)\left(\frac{\ddot{F}}{\dot{F}} - \frac{\ddot{V}}{\dot{V}}\right) + 2 c \left(\frac{\dot{F}}{F}\right) = 0.$$ This equation is the analogue of (\[perfect fluid case\]). Following the same procedure as above we find $F = \left[f(t) + g(\psi)\right]^{(a + b)/(2 c + a + b)}$. Therefore, the solution for radiation-like matter resembles that of perfect fluid in the sense that the effective stresses $p_{x}^{(eff)} \equiv - T_{1}^{1}$, $p_{y}^{(eff)} \equiv - T_{2}^{2}$, $p_{z}^{(eff)} \equiv - T_{3}^{3}$ are proportional to the energy density, viz., $$\label{effective stresses, radiation-like solution}
p_{x}^{(eff)} = n_{x}\rho^{(eff)}, \;\;\;p_{y}^{(eff)} = n_{y}\rho^{(eff)}, \;\;\;p_{z}^{(eff)} = n_{z}\rho^{(eff)},$$ where $n_{x}, n_{y}$ and $n_{z}$ are constants satisfying $n_{x} + n_{y} + n_{z} = 1$. If we average over the three spatial directions, this is equivalent to saying that the equation of state is ${\bar{p}}^{(eff)} = \rho^{(eff)}/3$, where ${\bar{p}}^{(eff)} \equiv - T_{i}^{i}/3$.
Barotropic linear equation of state
------------------------------------
For the sake of generality, and in order to keep contact with isotropic FRW cosmologies, let us study the scenario where the effective mater is barotropic, that is the ratio $\bar{p}^{(eff)}/\rho^{(eff)}$ is constant. Thus we set $$\label{barotropic equation of state}
\bar{p}^{(eff)} = n \rho^{(eff)},\;\;\;n = \mbox{constant},$$ which for $n = 1$ and $n = 1/3$ gets back the above-discussed perfect fluid and radiation-like scenarios, respectively. Substituting into (\[EMT\]) we obtain an equation similar to (\[perfect fluid case\]) and (\[ultra-relativistic matter\]), but with the coefficient $(1 + 3n)$ in front of the term $c\dot{F}/F$. Consequently, $\dot{V} \propto \dot{F} F^{[c(1 + 3n)/(a + b)]}$. The condition $(\partial \dot{V}/\partial \psi) = (\partial {V'}/\partial t)$ then requires $$\label{F, barotropic case}
F = \left[f(t) + g(\psi)\right]^{\frac{(a + b)}{a + b + (3n + 1)c}}.$$ Substituting this expression into (\[Kasner-like metric in terms of F-bar\]) and making the coordinate transformation $(df/dt)dt \rightarrow d\tilde{t}$, $(d g/d\psi)d\psi \rightarrow d\tilde{\psi}$, the line element in $5D$ can be written as $$\label{5D line element for the barotropic case}
dS^2 = \frac{D d\tilde{t}^2}{\tilde{H}^{(3n + 1)c}} - A \tilde{H}^{2 a \alpha } dx^2 - B \tilde{H}^{2 a \beta } dy^2 - C \tilde{H}^{2 a \gamma } dz^2 - \frac{D d\tilde{\psi}^2}{\tilde{H}^{(3n + 1)c}},$$ where $\tilde{H} \equiv \left(\tilde{t} + E \tilde{\psi}\right)^{\frac{1}{a + b + (3n + 1)c}}$; $E$ is an arbitrary constant for $n = 1/3$, but $E = \pm 1$ for any other $n \neq 1/3$; $D$ is a positive constant introduced for dimensional considerations.
### Kasner universe in $5D$
We immediately note that the case where $E = 0$, which requires $n = 1/3$, gives back the well-known Kasner universe in $5D$. In fact, setting $\tilde{t} \propto \tau^{(a + b + 2c)/(a + b + c)}$ the line element (\[5D line element for the barotropic case\]) reduces to $$\label{Kasner solution in 5D}
dS^2 = d\tau^2 - A \tau^{2q_{1}} dx^2 - B \tau^{2 q_{2}}dy^2 - C \tau^{2 q_{3}}dz^2 \pm D \tau^{2 q_{4}}d\psi^2,$$ where $$\label{defifition of the q's}
q_{1} = \frac{\alpha a}{(a + b + c)}, \;\;\;q_{2} = \frac{\beta a}{(a + b + c)}, \;\;\;q_{3} = \frac{\gamma a}{(a + b + c)}, \;\;\; q_{4} = - \frac{c}{a + b + c},$$ satisfy $\Sigma_{i = 1}^4{q_{i}} = \Sigma_{i = 1}^4{q^2_{i}} = 1$, typical of the Kasner universe in $5D$. In order to avoid misunderstandings, it may be useful to reiterate our terminology: (i) the light-like Kasner metric is (\[Lightlike Kasner solution\]), which depends on the light-like variable $u$; (ii) by Kasner-like metric we refer to (\[Kasner-like metric in terms of F-bar\]), which depends on one arbitrary function of $t$ and $\psi$, and (iii) Kasner metric is the usual name given to (\[Kasner solution in 5D\]), which depends only on $\tau$.
In general, for the $4$-dimensional interpretation of (\[5D line element for the barotropic case\]) we should notice that on every hypersurface $\tilde{\psi} = \tilde{\psi}_{0} = $ constant (${\psi} = {\psi}_{0} = $ constant) the proper time $\tau$ is given by
$$\label{proper time}
d\tau = \pm \; \frac{\sqrt{D} d\tilde{t}}{(\tilde{t} + E \tilde{\psi}_{0})^m}, \;\;\;\;m \equiv \frac{(3n + 1)c}{2 \left[a + b + (3n + 1)c\right]}.$$
Bellow we consider several cases.
### Anisotropic Milne universe
If $m = 0$, then[^7]
$$\label{Milne universe}
n = - \frac{1}{3}.$$
In terms of the proper time $\tau = \sqrt{D}(\tilde{t} + E \tilde{\psi}_{0}) $, the metric induced on $4$-dimensional hypersurfaces $\Sigma_{\psi}$ can be written as $$\label{anisotropic milne universe }
ds^2 = dS^2_{\Sigma_{|_{\psi}}} = d\tau^2 - \bar{A} \tau^{p_{1}}dx^2 - \bar{B} \tau^{p_{2}}dy^2 - \bar{C} \tau^{p_{3}}dz^2,$$ where $p_{i}$ are the parameters defined in (\[introduction of p1, p2 and p3\]) and $\bar{A}, \bar{B}$ and $\bar{C}$ are some new constants. In addition, $n_{i}$, the ratios of the anisotropic stresses to the energy density (\[effective stresses, radiation-like solution\]) are given by $$\label{the n's for Milne universe}
n_{x} = \frac{(\alpha - \beta - \gamma)}{a}, \;\;\; n_{y} = \frac{(\beta - \alpha - \gamma)}{a}, \;\;\;n_{z} = \frac{(\gamma - \alpha - \beta)}{a}, \;\;\;8\pi G \rho^{(eff)} = \frac{a^2 c}{(a + b)^2 \tau^2}$$ For $\alpha = \beta = \gamma $ we find $n_{x} = n_{y} = n_{x} = - 1/3$ and consequently we recover Milne’s universe, as expected.
### Anisotropic de Sitter universe
If $m = 1$, then $$\label{m = 1}
n = - \frac{1}{3} - \frac{2(\alpha^2 + \beta^2 + \gamma^2)}{3(\alpha \beta + \alpha \gamma + \beta \gamma)}.$$ From (\[proper time\]) we get $(\tilde{t} + E\tilde{\psi}_{0} ) \propto e^{\pm \tau/\sqrt{D}}$. Taking the negative sign, the induced metric in $4D$ can be expressed as $$\label{barotropic case, 4D metric}
ds^2 = dS^2_{\Sigma_{|_{\psi}}} = d\tau^2 - \bar{A}e^{p_{1}\tau/\sqrt{D}}dx^2 - \bar{B}e^{p_{2}\tau/\sqrt{D}}dy^2 - \bar{C}e^{p_{3}\tau/\sqrt{D}}dz^2.$$ For this metric we find $$\label{matter quantities for the de Sitter-like solution}
n_{x} = - \frac{(\beta^2 + \gamma^2 + \beta \gamma)}{c}, \;\;\;n_{y} = - \frac{(\alpha^2 + \gamma^2 + \alpha \gamma)}{c}, \;\;\;n_{z} = - \frac{(\alpha^2 + \beta^2 + \alpha\beta)}{c}, \;\;\;8 \pi G \rho^{(eff)} = \frac{a^2 c}{(a + b)^2 D}.$$
In the case of isotropic expansion $(\alpha = \beta = \gamma)$ these equations yield $n_{x} = n_{y} = n_{z} = n = - 1$ and (\[barotropic case, 4D metric\]) reduces to the familiar de Sitter metric with cosmological constant $\Lambda_{(4)} = 3/D$. An interesting conclusion here is that an anisotropic universe can enter a phase of exponential expansion (inflation), without satisfying the classical “false-vacuum" equation $p = - \rho$ (see (\[m = 1\])).
### Anisotropic power-law FRW universe
For $m \neq 1$, from (\[proper time\]) we obtain $$\label{proper time, general case}
(\tilde{t} + E \tilde{\psi}_{0}) = \left[\frac{(1 - m)}{\sqrt{D}}(\tau - \tau_{0})\right]^{1/(1 - m)},$$ where $\tau_{0}$ is a constant of integration. Thus, the induced metric in $4D$ becomes $$\label{non-empty Kasner-like universe}
ds^2 = dS^2_{\Sigma_{|_{\psi}}} = d\tau^2 - \bar{A}\tau^{\alpha \kappa}dx^2 - \bar{B}\tau^{\beta \kappa}dy^2 - \bar{C}\tau^{\gamma \kappa}dz^2,$$ where we have set $\tau_{0} = 0$; $\bar{A}, \bar{B}, \bar{C}$ are constants with the appropriate units, and $\kappa$ is given by
$$\label{definition of kappa}
\kappa \equiv \frac{4 a}{2(a + b) + (3n + 1)c} = \frac{4(\alpha + \beta + \gamma)}{2(\alpha^2 + \beta^2 + \gamma^2) +(3n + 1)(\alpha \beta + \alpha \gamma + \beta \gamma)}.$$
We note that the denominator of $\kappa$ is non-zero because by assumption here $m \neq 1$, see (\[m = 1\]). The effective matter quantities are $$\label{matter quantities for the barotropic case}
\bar{p}^{(eff)} = n \rho^{(eff)}, \;\;\;8 \pi G \rho^{(eff)} = \frac{\kappa^2 c}{4 \tau^2}$$ We also find $$\label{the n's for the barotropic case}
n_{x} = \frac{2\alpha +(3n - 1)(\beta + \gamma)}{2 a}, \;\;\;n_{y} = \frac{2\beta +(3n - 1)(\alpha + \gamma)}{2 a}, \;\;\;n_{z} = \frac{2\gamma +(3n - 1)(\alpha + \beta)}{2 a}.$$ We note that for $c = 0$, the space is empty $(\rho^{(eff)} = 0)$, and the line element (\[non-empty Kasner-like universe\]) yields the well-known Kasner solution in $4D$. Besides, for $n = 1$ and $n = 1/3$ the above expressions reduce to those obtained for perfect fluid and radiation-like matter discussed in sections $3.1$ and $3.2$, respectively.
### Isotropic expansion: spatially flat FRW universe
The above expressions evidence the fact that for anisotropic expansion the effective EMT behaves like a perfect fluid for $n = 1$. In contrast, isotropic expansion allows perfect fluid for any value of $n$. In this case the $5D$ metric (\[5D line element for the barotropic case\]) can be written as (we omit the tilde over $t$ and $\psi$) $$\label{barotropic case with isotropic expansion}
dS^2 = \frac{ D dt^2}{(t + E \psi)^{(3n + 1)/(3n + 2)}} - C (t + \psi)^{2/(2 + 3n)}\left[dx^2 + dy^ 2 + dz^2\right] - \frac{ D d\psi^2}{(t + E \psi)^{(3n + 1)/(3n + 2)}}.$$ For $n \neq - 1$, on every hypersurface $\Sigma_{\psi}$ it reduces to $$\label{metric for the flat FRW model}
ds^2 = d\tau^2 - C \tau^{4/3(n + 1)}\left[dx^2 + dy^2 + dz^2\right],$$ which is the familiar flat FRW model with perfect fluid $$\label{matter for FRW flat model}
p = n \rho, \;\;\; 8\pi G \rho= \frac{4}{3(n + 1)^2 \tau^2}.$$ For $n = - 1$ we recover the de Sitter spacetime as shown in (\[barotropic case, 4D metric\]).
To finish this section we would like to emphasize that although the metrics with $a = 0$ and $c = 0$ correspond to empty space (Ricci-flat in $4D$), they are different in nature. For $a = 0$ the spacetime is Minkowski (Riemann-flat) in $5D$ and $4D$, while for $c = 0$ the components of the Riemann tensor are nonzero in $5D$ and in the $4D$ subspace $\Sigma_{\psi}$.
Cosmological models in $4D$. The braneworld approach
====================================================
The preceding discussion shows that, in the framework of STM the Kasner-like metric (\[Kasner-like metric in terms of F-bar\]) embeds a large family of $4D$ cosmological models that are anisotropic versions of the FRW ones. However, one could argue that the effective matter quantities (\[EMT\]) do not have to satisfy the regular energy conditions [@Bronnikov], or any physically motivated equation of state, because they involve terms of geometric origin[^8].
In this section we will see that the $5D$ metric (\[Kasner-like metric in terms of F-bar\]) can be completely determined if one imposes an equation of state on the matter in the brane. Although the concept is the same as in section $3$, the physics here is [*different*]{}. Namely, in this approach the spacetime is a singular hypersurface and, for the $5D$ Kasner metrics under consideration, there is an effective non-vanishing cosmological term in $4D$ (the brane). As a consequence, the time evolution as well as the interpretation of the solutions in $4D$ is distinct from the one obtained, under similar conditions, in the framework of STM.
The braneworld paradigm
-----------------------
In order to make the paper self-consistent, and set the notation, we give a brief sketch of the technical details that we need in our discussion. In the simplest RS2 braneworld scenario our universe is identified with a [*fixed*]{} singular hypersurface ${\Sigma_{\psi_{b}}}$ (called [*brane*]{}) embedded in a $5$-dimensional bulk with ${\bf Z}_{2}$ symmetry with respect to the brane. The discontinuity of the extrinsic curvature across $\Sigma_{\psi_{b}}$ is related to the presence of matter on the brane, which is described by an EMT that we denote as $\tau_{\mu\nu}$. Thus, now the Einstein field equations in $5D$ are $G_{AB} = k_{(5)}^2 T_{AB}^{(brane)}$, where $k_{(5)}^2$ is a constant with the appropriate units and $T_{AB}^{(brane)} = \delta_{A}^{\mu}\delta_{B}^{\nu}\tau_{\mu\nu} \delta(\psi)/\Phi$.
Israel’s boundary conditions [@Israel] relate the jump of $K_{\mu\nu}$ to $\tau_{\mu\nu}$, namely, $$({K_{\mu \nu}}_{|\Sigma^{+}_{\psi_{b}}} - {K_{\mu \nu}}_{|\Sigma^{-}_{\psi_{b}}}) = - \epsilon \frac{k_{(5)}^2}{2}(\tau_{\mu\nu} - \frac{1}{3}\tau g_{\mu\nu}).$$ Now, the assumed ${\bf{Z}}_{2}$ symmetry implies ${K_{\mu \nu}}_{|\Sigma^{+}_{\psi_{b}}} = - {K_{\mu \nu}}_{|\Sigma^{-}_{\psi_{b}}}$. Consequently,
$$\label{emt on the brane in terms of K}
\tau_{\mu\nu} = - \frac{2\epsilon}{k_{(5)}^2}\left(K_{\mu\nu} - g_{\mu\nu} K\right),$$
where the extrinsic curvature $K_{\mu\nu}$ has to be evaluated on ${\Sigma^{+}_{\psi_{b}}}$. From $G_{\nu 4} = 0$ it follows that $\tau^{\mu}_{\nu;\mu} = 0$. Thus $\tau_{\mu\nu}$ represents the total, vacuum plus matter, conserved energy-momentum tensor on the brane. It is usually separated in two parts [@Csaki], $$\label{decomposition of tau}
\tau_{\mu\nu} = \sigma g_{\mu\nu} + {\cal{T}}_{\mu\nu},$$ where $\sigma$ is the tension of the brane, which is interpreted as the vacuum energy density, and ${\cal{T}}_{\mu\nu}$ represents the energy-momentum tensor of [*ordinary*]{} matter in $4D$.
From (\[emt on the brane in terms of K\]) and (\[decomposition of tau\]) we get $$\label{K in terms of matter in the brane}
K_{\mu\nu} = - \frac{\epsilon k_{(5)}^2}{2} \left({\cal{T}}_{\mu\nu} - \frac{1}{3}g_{\mu\nu}({\cal{T}} + \sigma)\right).$$ Substituting this expression into (\[4D Einstein with T and K\]) we obtain [@Shiromizu] $$\label{EMT in brane theory}
^{(4)}G_{\mu\nu} = {\Lambda}_{(4)}g_{\mu\nu} + 8\pi G {\cal{T}}_{\mu\nu} - \epsilon k_{(5)}^4 \Pi_{\mu\nu} - \epsilon E_{\mu\nu},$$ where $$\label{definition of lambda}
\Lambda_{(4)} = - \epsilon \frac{k_{(5)}^4 \sigma^2}{12},$$ $$\label{effective gravitational coupling}
8 \pi G = - \epsilon \frac{k_{(5)}^4 \sigma}{6},$$ and $$\label{quadratic corrections}
\Pi_{\mu\nu} = \frac{1}{4} {\cal{T}}_{\mu\alpha}{\cal{T}}^{\alpha}_{\nu} - \frac{1}{12}{\cal{T}} {\cal{T}}_{\mu\nu} - \frac{1}{8}g_{\mu\nu}{\cal{T}}_{\alpha\beta}{\cal{T}}^{\alpha\beta} + \frac{1}{24}g_{\mu\nu}{\cal{T}}^2.$$ All these four-dimensional quantities have to be evaluated on ${\Sigma^{+}_{\psi_{b}}}$. They contain two important features; they give a working definition of the fundamental quantities $\Lambda_{(4)}$ and $G$ and contain higher-dimensional modifications to general relativity. Namely, local quadratic energy-momentum corrections via the tensor $\Pi_{\mu\nu}$, and the nonlocal effects from the free gravitational field in the bulk, transmitted by $E_{\mu\nu}$.
Matter in the brane. Gaussian coordinates
-----------------------------------------
In the braneworld literature the use of Gaussian coordinates in quite common. In these coordinates the function $F$ is given by (\[F in Gaussian coordinates, c2 = 0\]), which under the re-scaling $F \rightarrow \bar{F}^{a^2/(a + b)}$ becomes $\bar{F} = l(t) + \psi h(t)$.
If we locate the brane at $\psi = 0$, then the metric of the bulk is given by:
1. For $\psi > 0$ $$\label{bulk metric for psi positive}
dS^2_{(+)} = \frac{\left[\dot{l} + \psi \dot{h}\right]^2}{h^2}dt^2 - A \left[l(t) + \psi h(t)\right]^{p_{1}}dx^2 - B \left[l(t) + \psi h(t)\right]^{p_{2}}dy^2 - C \left[l(t) + \psi h(t)\right]^{p_{3}}dz^2 - d\psi^2.$$
2. For $\psi < 0$ $$\label{bulk metric for psi negative}
dS^2_{(-)} = \frac{\left[\dot{l} - \psi \dot{h}\right]^2}{h^2}dt^2 - A \left[l(t) - \psi h(t)\right]^{p_{1}}dx^2 - B \left[l(t) - \psi h(t)\right]^{p_{2}}dy^2 - C \left[l(t) - \psi h(t)\right]^{p_{3}}dz^2 - d\psi^2.$$
Using (\[extrinsic curvature\]) we calculate the non-vanishing components of $K_{\mu\nu} = {K_{\mu \nu}}_{|\Sigma^{+}_{\psi_{b}}}$. These are $$\label{extrinsic curvature in Gaussian coordinates}
K_{00} = \frac{\dot{l}\dot{h}}{h^2}, \;\;\;K_{11} = - \frac{A \alpha a l^{(p_{1} - 1)}h}{(a + b)},\;\;\;K_{22} = - \frac{B \beta a l^{(p_{2} - 1)}h}{(a + b)}, \;\;\;K_{33} = - \frac{C \gamma a l^{(p_{3} - 1)}h}{(a + b)}.$$ We assume that the matter in the brane satisfies the equation of state $$\label{equation of state for ordinary matter}
p = n \rho,$$ where $\rho = {\cal{T}}_{0}^{0}$, $p = (p_{x} + p_{y} + p_{z})/3$ and $p_{x} = - {\cal{T}}_{1}^{1}$, $p_{y} = - {\cal{T}}_{2}^{2}$, $p_{z} = - {\cal{T}}_{3}^{3}$. Using these expressions, from (\[emt on the brane in terms of K\]), with $\epsilon = - 1$, and (\[decomposition of tau\]) we obtain
$$\begin{aligned}
\label{ordinary matter quantities}
k_{(5)}^2 \sigma &=& - \frac{2}{(1 + n)}\left[\frac{\dot{h}}{\dot{l}} + \frac{(2 + 3n) a^2}{3(a + b)}\left(\frac{h}{l}\right)\right], \nonumber \\
k_{(5)}^2 \rho &=& \frac{2}{(1 + n)}\left[\frac{\dot{h}}{\dot{l}} - \frac{a^2}{3 (a + b)}\left(\frac{h}{l}\right)\right], \;\;\;n \neq -1, \;\;\;\dot{l} \neq 0.\end{aligned}$$
We notice that in cosmological applications the metric function $g_{00}$ is subjected to the condition [@Binetruy], [@strength] $$\label{condition on the metric}
{g_{00}}_{|_{brane}} = 1.$$ Thus $$\label{condition on h and l}
h(t) = s \; \dot{l}(t), \;\;\;s = \pm 1.$$ Therefore, we have two equations for the three unknown $\sigma, \rho$ and $l(t)$. Taking the covariant divergence of (\[decomposition of tau\]), it follows that to conserve both the total brane energy-momentum tensor $\tau_{\mu\nu}$ and the matter energy-momentum tensor ${\cal{T}}_{\mu\nu}$, we must have $\sigma = \sigma_{0} = $ constant. Then, using (\[condition on h and l\]) we integrate the first equation in (\[ordinary matter quantities\]) and obtain the scale factor as[^9]
$$\label{exponential expansion}
l(t) = \left[C_{1} e^{- s(n + 1)k_{(5)}^2 \sigma_{0} t/2} + C_{2}\right]^{\eta}, \;\;\; \eta \equiv \frac{3(a + b)}{(2 + 3n)a^2 + 3(a + b)},$$
where $C_{1}$ and $C_{2}$ are constants of integration. We note that $\eta$ is positive for arbitrary values of $\alpha$, $\beta$, $\gamma$ and $n > -1$. Therefore, if we choose $s = - 1$ and set $C_{2} = 0$, then the “origin" $l = 0$ is located at $t = - \infty$. Thus, from (\[ordinary matter quantities\]) we find $\rho \propto \sigma = \sigma_{0}$ for all $t$, regardless of the value of $n$ (but $n \neq - 1$). The resulting metric on the brane is de Sitter-like, with different rates of exponential expansion in every direction, similar to the models discussed in (\[barotropic case, 4D metric\]).
Non-Gaussian embeddings
-----------------------
The question may arise of whether the simplicity of the above scenario is not a consequence of the simplifying assumption of Gaussian coordinates. In order to investigate this question, we consider here the embedding that arises from the choice[^10]
$$\label{generalization of V dot}
\dot{V} \propto \frac{\dot{F} F^{q c/(a + b)}}{F'^2},$$
where $q$ is some constant. With this choice the metric in $5D$ can be written as
$$\label{metric for sigma constant without exponential expansion}
dS^2 = \frac{ F^{q c/(a + b)} \dot{F}^2}{F'^2} dt^2 - A F^{p_{1}}dx^2 - B F^{p_{2}}dy^2 - C F^{p_{3}}dz^2 - F^{q c/(a + b)} d\psi^2.$$
Now, using $(\partial \dot{V}/\partial \psi) = (\partial {V'}/\partial t)$ we find $$\label{F generalized}
F = \left[l(t) + \psi h(t)\right]^{(a + b)/(a + b - q c)}, \;\;\;\mbox{for}\;\;\; q \neq \frac{a + b}{c},$$ and $$\label{F for Generalization II}
F = l(t)e^{\psi h(t)},\;\;\; \mbox{for}\;\;\; q = \frac{a + b}{c},$$ where $l(t)$ and $h(t)$ are arbitrary functions of integration. Again, if we locate the brane at $\psi = 0$, then the metric in the ${\bf{Z}}_{2}$-symmetric bulk is obtained by replacing $\psi \rightarrow |\psi|$ in (\[F generalized\]) and (\[F for Generalization II\]). Following the steps used in Section $(4.2)$ we find $$\label{equation for sigma in the generalized model}
k_{(5)}^2 \sigma = - \frac{2 s }{n + 1}\left[\frac{\ddot{l}}{\dot{l}} + \frac{(2 + 3 n)a^2 + 3 q c}{3 (a + b - q c)}\;\left(\frac{\dot{l}}{l}\right)\right], \;\;\;\;q \neq \frac{a + b}{c}.$$ It is interesting to note that in the case where $q = (a + b)/c$, the equation for $\sigma$ can [*formally*]{} be obtained from (\[equation for sigma in the generalized model\]) by setting $q = 0$. Consequently, (\[F for Generalization II\]) yields models on the brane that are identical to those in Gaussian coordinates (\[ordinary matter quantities\]), although the metric in the $5D$ bulk is completely different in both cases.
The conclusion emanating from (\[equation for sigma in the generalized model\]) is that, within the context of the $5$-dimensional Kasner spacetimes under consideration, Gaussian and non-Gaussian embeddings generate the same physics on the brane. In particular, the assumption of constant $\sigma$, which is equivalent to a constant cosmological term $\Lambda_{(4)}$, obliges the universe to expand in a de Sitter [*anisotropic*]{} form regardless of (the choice of) the embedding. This is quite analogous to the cosmological “no-hair" theorem/conjecture of general relativity.
Vacuum solutions on the brane
-----------------------------
Since the extra dimension is spacelike, the solutions to the field equations are invariant under the transformation $(x, y, z) \leftrightarrow \psi$. However, the physics in $4D$ crucially depends on how we choose our ordinary $3D$ space.
In order to illustrate this, let us permute $\psi \leftrightarrow z$ in the solution given (\[metric for sigma constant without exponential expansion\]) and (\[F generalized\]). Also, to avoid misunderstanding we change $F(t, \psi) \rightarrow H(t, z)$. Using this notation, we find that the metric
$$\label{vaccum solutions on the brane}
dS^2 = \frac{ H^{q c/(a + b)} \dot{H}^2}{H_{z}^2} dt^2 - A H^{p_{1}}dx^2 - B H^{p_{2}}dy^2 - H^{q c/(a + b)} dz^2 \pm C H^{p_{3}}d\psi^2,$$
where $H_{z} \equiv \partial H/\partial z$ and $H = \left[l(t) + z h(t)\right]^{(a + b)/(a + b - q c)}$, is also a solution of the field equations $R_{AB} = 0$. Although (\[metric for sigma constant without exponential expansion\]) and (\[vaccum solutions on the brane\]) are diffeomorphic in $5D$, their interpretation in $4D$ is quite different. Specifically, unlike (\[metric for sigma constant without exponential expansion\]) in (\[vaccum solutions on the brane\]): (i) the extra dimension can be either spacelike or timelike, (ii) the spacetime slices $\Sigma_{\psi}$ are non-flat, and (iii) the metric of the spacetime is independent of $\psi$. As a consequence of the latter, the extrinsic curvature $K_{\mu \nu}$, defined by (\[extrinsic curvature\]), vanishes identically. Which in turn, by virtue of (\[emt on the brane in terms of K\]), implies $\tau_{\mu\nu} = 0$, i.e., the spacetime (the brane) is devoid of matter $({\cal{T}}_{\mu\nu} = 0)$ and $\Lambda_{(4)} = 0$.
Clearly, other $5D$ metrics with properties similar to (\[vaccum solutions on the brane\]) can be constructed from the solutions (\[the metric in Gaussian coordinates\])-(\[F in the synchronous coordinates\]) of section $2$ as well as from (\[F for perfect fluid\]) and (\[F, barotropic case\]) of section $3$. The conclusion here is that the spacetime part of the $5D$ Kasner-like metric (\[Kasner-like metric in terms of F-bar\]), after the transformation $\psi \leftrightarrow (x, y, z)$, can be interpreted as vacuum solutions in a braneworld ${\bf{Z}}_{2}-$symmetric scenario.
Static embeddings
==================
As we noted above, when a $5D$ metric is independent of $\psi$, the extra dimension can be either spacelike or timelike. This is a general feature of the $5D$ field equations [@XtraSym]. Therefore, after the transformation $\psi \leftrightarrow t$ the $5D$ metric still satisfies the field equations $R_{AB} = 0$. The interesting feature here is that after such a transformation the line element induced on $4D$ hypersurfaces $\Sigma_{\psi}$ is static, instead of dynamic as in sections $3$ and $4$.
A simple $5D$ line element that illustrates this feature, in a quite general way, can be obtained from the Kasner-like metric (\[Kasner-like metric in terms of F-bar\]) in the synchronous reference system. In fact, making the transformations ${\psi \leftrightarrow z}$, $F(t, \psi) {\rightarrow} H(t, z)$; and $\psi \leftrightarrow t$, $H(t, z) \rightarrow W(\psi, z)$, from (\[cosmological homogeneous solution in synchronous coordinates \]) and (\[F in the synchronous coordinates\]) we obtain $$\label{static solution}
dS^2 = C W^{p_{3}} dt^2 - A W^{p_{1}}dx^2 - B W^{p_{2}}dy^2 - \left(\frac{W_{z}}{W'}\right)^2 dz^2 + d\psi^2,$$ where (We recall the re-scalling of $F$ introduced at the end of section $2$.) $$\label{the function S}
W = M(z) + \psi N(z).$$ This is a solution of the $5D$ equations $R_{AB} = 0$ for any arbitrary functions $M(z)$ and $N(z)$. It explicitly depends on the extra dimension $\psi$, which now is timelike. At this point it is worthwhile to emphasize that in modern noncompactified $5D$ theories both, spacelike and timelike extra dimensions are physically admissible [@epsilon].
Once again the choice of the functions $M(z)$ and $N(z)$ depends on the version of $5D$ relativity we use to evaluate the properties of the matter content in $4D$. Bellow we illustrate this by considering the induced matter approach, used in section $3$, and the braneworld paradigm used in section $4$.
Static solutions with planar symmetry in conventional $4D$ general relativity
-----------------------------------------------------------------------------
It is not difficult to show that the components of the effective EMT, induced on spacetime hypersurfaces $\Sigma_{\psi}:\psi = \psi_{0} =$ constant, for the metric (\[static solution\]) satisfy algebraic relations similar, but not identical[^11], to those in (\[rel between the stresses\]) and (\[rel between T0 and the stresses\]), which are independent on the specific choice of $M$ and $N$. We omit them here and present the case where the effective matter quantities satisfy the barotropic linear equation of state (\[barotropic equation of state\]). In such a case we find $$\label{M for the general static solution}
M(z) = \bar{C} N(z)^{k} - \psi_{0}N(z), \;\;\;k \equiv - \frac{(\alpha^2 + \beta^2 + \gamma^2)[(3n + 1)(\alpha + \beta)] + 2 \gamma]}{(\alpha \beta + \alpha \gamma + \beta \gamma)[(3n + 1)(\alpha + \beta - \gamma) + 4 \gamma]},$$ where $\bar{C}$ is a constant of integration. Thus, in the $5D$ solution (\[static solution\]): $W = \bar{C} N^k + (\psi - \psi_{0})N$, which implies that the metric induced in $4D$ is independent of the choice of the hypersurface $\Sigma_{\psi}$. The effective energy density in $4D$ is given by $$\label{static rho effective}
8 \pi G \rho^{(eff)} = \frac{2 a^2 \gamma c \; N^{2(1 - k)}}{\bar{C}^2 [(3n + 1)(\alpha + \beta) + 2\gamma](a + b)^2 },$$ and the stresses are $$\label{the static stresses}
\frac{p_{x}^{(eff)}}{\rho^{(eff)}} = \frac{(3n - 1)\gamma - \alpha (3n + 1)}{2 \gamma}, \;\;\;\frac{p_{y}^{(eff)}}{\rho^{(eff)}} = \frac{(3n - 1)\gamma - \beta (3n + 1)}{2 \gamma}, \;\;\;\frac{p_{z}^{(eff)}}{\rho^{(eff)}} = \frac{(3n + 1)(\alpha + \beta) + 2\gamma}{2 \gamma}.$$ We note that for $k = 1$ these quantities are constants and $\rho^{(eff)} < 0$ for all values of $\alpha, \beta$ and $\gamma$. In what follows we assume $k \neq 1$.
Since ${g_{33}}_{|_{\Sigma_{\psi}}} = - {\bar{C}}^2 k^2 N^{2(k - 2)}(dN/dz)^2$ we can make the coordinate transformation $N^{(k - 2)} dN \rightarrow d\bar{z}$, i.e., $N \sim {\bar{z}}^{1/(k - 1)}$. In terms of this new coordinate the static metric in $5D$ is generated by (henceforth we omit the bar over $z$) $$\label{S generating static solutions }
W(t, \psi) = \bar{C}\left[z^{k/(k - 1)} + (\psi - \psi_{0}) z^{1/(k - 1)}\right]$$ The matter quantities induced in $4D$ decrease as $1/z^2$. Therefore, the above equations represent static “pancake-like" distributions where the matter is concentrated near the plane $z = 0$, while far from it $\rho \rightarrow 0$.
Except for the singularity at $z = 0$, the matter distribution presents “reasonable" physical properties. Indeed, for every value of $n $, the “physical" conditions $\rho^{(eff)} > 0$ and $\rho \geq |p_{x, y, z}|$ are satisfied in a wide range of parameters $\alpha, \beta$ and $\gamma$. As an illustration, in the case of axial symmetry with respect to $z$, for $n = 0$ these conditions hold in the range $- 2\beta/3 < \gamma < - \beta/2$ ($\alpha = \beta > 0$) or $-\beta/2 < \gamma < - 2\beta/3$ $(\alpha = \beta < 0)$. For $n = 1/3$, they hold if $- 2 \beta < \gamma \leq - \beta$ ($\alpha = \beta > 0$) or $- \beta \leq \gamma < - 2\beta$ ($\alpha = \beta < 0$). A similar analysis can be extended for other values of $n$.
A simpler solution can be obtained from the above expressions in the limiting case where $k = \infty$, which occurs for $(3n + 1)(\alpha + \beta - \gamma) + 4\gamma = 0$. In this case (\[S generating static solutions \]) simplifies to $W = \bar{C}\left[z + (\psi - \psi_{0})\right]$ and the matter quantities are obtained from (\[static rho effective\]), (\[the static stresses\]) by replacing $\gamma \rightarrow [(3n + 1)(\alpha + \beta)/3(n - 1)]$. In the case of axial symmetry, the line element becomes independent of the parameters $\alpha, \beta, \gamma$ and depends only on $n$. The effective matter quantities satisfy $\rho^{(eff)} > 0$ and $\rho \geq |p_{x, y, z}|$ for any $n$ in the range $- 1/3 \leq n < 1/3 $.
Static solutions on the brane
-----------------------------
We now proceed to use the braneworld technique for evaluating the matter quantities. If we locate the brane at $\psi = 0$, then the metric in the ${{\bf{Z}}_{2}}$-symmetric bulk is generated by $W = M(z) + |\psi| N(z)$. From (\[emt on the brane in terms of K\]), with $\epsilon = 1$, and (\[decomposition of tau\]) we obtain the components of ${\cal{T}}_{\mu\nu}$. Now the barotropic equation of state (\[equation of state for ordinary matter\]) yields a differential equation linking $M(z), N(z)$ and $\sigma$, which can be easily integrated for constant vacuum energy, $\sigma = \sigma_{0}$. Namely, we obtain
$$\label{Static solutions on the brane, N(z)}
N(z) = \frac{3 k_{(5)}^2\sigma_{0} (a + b)(1 + n) M(z)}{2(5 + 3n)[\gamma^2 + \gamma(\alpha + \beta)] + 4(2 + 3n)[\alpha^2 + \beta(\alpha + \beta)]} - E M(z)^{- \frac{a [(2 + 3n)(\alpha + \beta) + 3\gamma]}{(a + b)(2 + 3n)}},$$
where $E$ is a constant of integration. Using this expression we obtain $$\label{Static solutions on the brane, rho}
k_{(5)}^2 \rho = \frac{6 E a \gamma}{(a + b)(2 + 3n) M(z)^{\tilde{k}}} + \frac{2 k_{(5)}^2 \sigma_{0}[\alpha^2 + \beta^2 - \gamma^2 + \alpha \beta
- \gamma(\alpha + \beta)]}{\tilde{k}(a + b)(2 + 3n)},$$ where $$\label{Static solutions on the brane, definition of k tilde}
\tilde{k} = \frac{(5 + 3n)[\gamma^2 + \gamma(\alpha + \beta)] + 2(2 + 3n)[\alpha^2 + \beta(\alpha + \beta)]}{(\alpha^2 + \beta^2 + \gamma^2)(2 + 3n)}.$$ We note that $\rho = $ constant for $\gamma = 0$. Therefore, in what follows we assume $\gamma > 0$. Since $M(z)$ is an arbitrary function, without loss of generality we can choose it as[^12] $$\label{choice of M(z)}
M(z) \sim z^{2/\tilde{k}},$$ which is suggested by the decrease of the effective density discussed in section $5.1$. It is not difficult to see that $\rho$ is positive for a large number of values of $\alpha, \beta, \gamma$. The positivity of the first term is guaranteed by the constant of integration $E$. To illustrate the positivity of the second term, we once again consider the case with axial symmetry with respect to $z$. In this case we find $$\label{Static solutions on the brane, rho for axial symmetry}
\lim_{z \rightarrow \pm \infty}{\rho} = \frac{2\sigma_{0} (3\beta + \gamma)(\beta - \gamma)} {\gamma (5 + 3n)(2\beta + \gamma) + 6\beta^2(3n + 2)},$$ which is positive for any $\beta > \gamma$ and $n \geq - 2/3$.
The main conclusion from this section is that regardless of whether we use the braneworld paradigm or the induced matter approach, the basic picture in $4D$ is essentially the same. Namely that the $4D$ part of (\[static solution\]) represents static pancake-like distributions of matter.
Summary
=======
The vacuum Einstein field equations for the $5D$ FRW line element (\[Usual cosmological metric in 5D\]) allow complete integration in a number of cases. In particular for $\Phi = 1$, or $n = 1$, the $(t, \psi)$-component of the field equations provides a relation that leads to a set of first integrals [@Binetruy], [@JPdeL-isotropicCosm]. However, for the simplest anisotropic extension of (\[Usual cosmological metric in 5D\]), namely, the diagonal Bianchi type-I metric $$\label{Bianchi type I, conclusions}
dS^2 = n^2(t, \psi)dt^2 - \sum_{i = 1}^{3}b_{i}(t, \psi) \left(dx^{i}\right)^2 + \epsilon \Phi^2(t, \psi)d\psi^2,$$ this procedure does not work. (For a discussion, and a new point of view in the context of braneworld, see [@Antonio].)
Here we have pointed out that making the coordinate transformation $du \propto \left(n dt - \Phi d\psi\right)$, $dv \propto \left(n dt + \Phi d\psi\right)$, in (\[Bianchi type I, conclusions\]) with $\epsilon = - 1$, the field equations allow complete integration in several physical situations, viz., (\[new solution\]), (\[solution 2\]), (\[solution 3\]). The $4D$ interpretation of the $5D$ solutions requires the introduction of coordinates adapted to spacetime sections $\Sigma_{\psi}$. We introduced the $(t, \psi)$ coordinates by setting $u = F(t, \psi)$ and $v = V(t, \psi)$, and used a foliation of the $5D$ manifold such that $\Sigma_{\psi}$ is a hypersurface of the foliation that is orthogonal to the extra dimension with tangent ${\hat{n}}^{A} = \delta^{A}_{4}/\Phi$. From a mathematical point of view the functions $F$ and $V$ can be arbitrary, except for the fact that they have to satisfy (\[diagonal 5D metric\]). However, from a physical point of view, they are related to two important aspects of the construction of the spacetime, namely: (i) the choice of coordinates in $5D$, e.g., Gaussian normal coordinates adapted to $\Sigma_{\psi}$, and (ii) the formulation of physical conditions on the matter fields in $4D$, e.g., some an equation of state.
Our study shows that there is a great freedom for embedding a $4D$ spacetime in an anisotropic $5D$ cosmological model. Similar results but in a distinct context have been found in [@Antonio]. To simplify the algebraic expressions, but not the physics, in sections $3$, $4$ and $5$ we have devoted our attention to the study of $4D$ spacetimes embedded in the light-like Kasner cosmological metric, which is a simplified version of (\[light-like metric in t-psi coordinates\]).
We have seen that the simple one-variable line element (\[Lightlike Kasner solution\]) can accommodate a great variety of models in $4D$. Indeed, within the context of STM and braneworld theories, we have shown here that the Kasner-like metrics (\[Kasner-like metric in terms of F-bar\]) may be used or interpreted as embeddings for a large number of cosmological and static spacetimes in $4D$. Thus, apparently “different" astrophysical and cosmological scenarios in $4D$ might just be distinct versions of the same physics in $5D$ [@XtraSym].
This investigation can be extended, or generalized, in different ways. In particular, we have not fully examined the possible $4D$ interpretation of the self-similar homothetic solution (\[new solution\]). Neither, have we investigated the solutions (\[solution 2\]) and (\[solution 3\]). An important future development here is the question of how these solutions can be applied in the generalizations of Mixmaster or Belinskii-Khalatnikov-Lifschits oscillations, as well as other issues mentioned in section $1$, which appear in theories with one extra dimension.
#### Acknowledgments:
I wish to thank one of the anonymous referees for a careful reading of this manuscript as well as for helpful and constructive suggestions.
Appendix: Solving the field equations. Part II {#appendix-solving-the-field-equations.-part-ii .unnumbered}
==============================================
Here we show two more families of analytic solutions to the field equations (\[Rxx, Ryy, Rzz\]). With this aim we notice that these equations are greatly simplified if we introduce the function $$\label{definition of calV}
{\cal{V}} = e^{\lambda + \mu + \sigma},$$ in terms of which (\[Rxx, Ryy, Rzz\]) become $$\begin{aligned}
\label{Rxx, Ryy, Rzz in terms of calV}
4 \lambda_{u v} + \frac{\lambda_{u} {\cal{V}}_{v}}{{\cal{V}}} + \frac{\lambda_{v} {\cal{V}}_{u}}{{\cal{V}}} &=& 0, \nonumber \\
4 \mu_{u v} + \frac{\mu_{u} {\cal{V}}_{v}}{{\cal{V}}} + \frac{\mu_{v} {\cal{V}}_{u}}{{\cal{V}}}&=& 0, \nonumber \\
4 \sigma_{u v} + \frac{\sigma_{u} {\cal{V}}_{v}}{{\cal{V}}} + \frac{\sigma_{v} {\cal{V}}_{u}}{{\cal{V}}} &=& 0.\end{aligned}$$ Adding these equations and using (\[definition of calV\]) we obtain an equation for ${\cal{V}}$, namely, $$\label{equation for calV}
2 {\cal{V}}{\cal{V}}_{u v} - {\cal{V}}_{u} {\cal{V}}_{v} = 0,$$ whose general solution can be written as
$$\label{solution for calV}
{\cal{V}} = \left[\tilde{h}(u) + \tilde{g}(v)\right]^2,$$
where $\tilde{h}$ and $\tilde{g}$ are arbitrary functions of their arguments. Clearly, the self-similar solution discussed in section $2.1$ corresponds to the particular choice $$\label{the metric functions in terms of calV}
e^{\lambda} \propto {\cal{V}}^{\alpha/a}, \;\;\;e^{\mu} \propto {\cal{V}}^{\beta/ a}, \;\;\;e^{\sigma} \propto {\cal{V}}^{\gamma/ a},$$ which satisfies (\[Rxx, Ryy, Rzz in terms of calV\]) and (\[equation for calV\]) identically.
In what follows, as in section $2.1$, we introduce a new set of null coordinates $\tilde{u}$ and $\tilde{v}$ by the relations $\tilde{h}(u) = {\tilde{c}}_{1} \tilde{u}$ and $\tilde{g}(v) = {\tilde{c}}_{2} \tilde{v}$, where ${\tilde{c}}_{1}$ and ${\tilde{c}}_{2}$ are constants. In terms of these new coordinates ${\cal{V}} = \left({\tilde{c}}_{1} \tilde{u} + {\tilde{c}}_{2} \tilde{v}\right)^2$. Substituting this expression into the first of the equations (\[Rxx, Ryy, Rzz in terms of calV\]), and dropping the tilde characters, we obtain $$2 \left(c_{1} u + c_{2} v\right)\lambda_{u v} + c_{1} \lambda_{v} + c_{2} \lambda_{u} = 0.$$ A similar expression holds for $\mu$. The solutions bellow are obtained under the assumption that $e^{\lambda}$ and $e^{\mu}$ are separable functions of their arguments. In which case, the above equation implies that they are proportional to $e^{\pm \left(c_{1} u - c_{2} v\right)}$. The metric function $e^{\sigma} = {\cal{V}}\; e^{- \left(\lambda + \mu\right)}$ automatically satisfies the third equation in (\[Rxx, Ryy, Rzz in terms of calV\]) and is non-separable. Consequently, there are two different families of solutions corresponding to whether $e^{\lambda} \propto e^{- \mu}$ or $e^{\lambda} \propto e^{ \mu}$.
$\bullet$ In the case where $e^{\lambda} \propto e^{- \mu}$, the field equations $R_{uu} = 0$ and $R_{vv} = 0$ reduce to $$\label{separable 1}
2 {\cal{A}}_{u} - c_{1}\left(c_{1} u + c_{2} v\right){\cal{A}} = 0,\;\;\;\mbox{and}\;\;\; 2 {\cal{A}}_{v} - c_{2}\left(c_{1} u + c_{2} v\right){\cal{A}} = 0.$$ These equations completely determine the function ${\cal{A}}$ and assure the fulfillment of $R_{u v} = 0$. The final form of the solution is given by
$$\label{solution 2}
dS^2 = C_{0} e^{[\left(c_{1} u + c_{2}v\right)^2/4]} d u d v - C_{1} e^{\left(c_{1} u - c_{2} v\right)} d x^2 - C_{2} e^{- \left(c_{1} u - c_{2} v\right)} d y^2
- \left(C_{1} C_{2}\right)^{- 1}\left(c_{1} u + c_{2} v\right)^2 d z^2.$$
$\bullet$ Following the same steps as above we find that when $e^{\lambda} \propto e^{\mu}$ the solution is $$\label{solution 3}
dS^2 = C_{0} e^{[3 \left(c_{1} u + c_{2}v\right)^2/4]} e^{- 2\left(c_{1} u - c_{2} v\right)}d u d v -
C \; e^{\left(c_{1} u - c_{2} v\right)} \left(d x^2 + d y^2\right)
- C^{- 2} \left(c_{1} u + c_{2} v\right)^2 e^{- 2\left(c_{1} u - c_{2} v\right)} d z^2.$$ In the above line elements $\left(C, C_{0}, C_{1}, C_{2}\right)$ are constants of integration. We note that the resulting solutions are quite complicated even in the case where either $c_{1}$ or $c_{2}$ are set equal to zero. Although this is a great obstacle for the analytical interpretation of these metrics in $4D$, it allows us to appreciate the simplicity of the self-similar solutions discussed in the main text.
[99]{}
.
[^1]: E-Mail: [email protected], [email protected]
[^2]: Notation: $x^{\mu} = (x^0, x^1, x^2, x^3)$ are the coordinates in $4D$ and $\psi$ is the coordinate along the extra dimension. We use spacetime signature $(+, -, -, -)$, while $\epsilon = \pm 1$ allows for spacelike or timelike extra dimension, both of which are physically admissible for a detailed discussion see, e.g., [@epsilon].
[^3]: Without entering into technical details: extrapolating backwards in time towards the singularity, one finds an infinite number of alternating quasi-periodic Kasner-like epochs with different expansion rates [@MTW]-[@BKL].
[^4]: In the traditional interpretation of Sedov, Taub and Zel’dovich [@Sedov], self-similarity means that all dimensionless quantities in the theory can be expressed as functions only of a single similarity variable, which is some combination of the independent coordinates. In this way the field equations become a system of ordinary, instead of partial, differential equations.
[^5]: The proportionality coefficients $C_{0}$, $C_{1}$, $C_{2}$, $C_{3}$ can be set equal to unity without any loss of generality.
[^6]: We note that (\[diagonal 5D metric\]) remains invariant under the re-scaling $F \rightarrow \bar{F}^{a^2/(a + b)}$.
[^7]: We exclude $c = 0$ because it corresponds to empty space, i.e., $T_{\mu\nu} = 0$.
[^8]: In fact, the effective EMT defined by (\[4D Einstein with T and K\]) contains a contribution, given by $E_{\alpha\beta}$, which is the spacetime projection of the $5D$ Weyl tensor and connects the physics in $4D$ with the geometry in $5D$.
[^9]: From (\[effective gravitational coupling\]), with $\epsilon = - 1$, it follows that $\sigma$ must be positive in order to ensure $G > 0$.
[^10]: This is a simple combination between $\dot{V} \propto \dot{F}F^{[c(1 + 3n)/(a + b)]}$ for the anisotropic FRW models considered in Section $3.3$, and $\dot{V} \propto \dot{F}/F'^2$ for the Gaussian embedding discussed in Section $4.2$.
[^11]: For example (\[rel between the stresses\]) is now replaced by $- (\alpha + \gamma)T_{1}^{1} + (\beta + \gamma)T_{2}^{2} + (\beta - \alpha)T_{3}^{3} = 0$.
[^12]: We note that in the case under consideration $(\gamma \neq 0)$, $\tilde{k}$ never vanishes. In fact, for real parameters $\alpha$ and $\beta$ the quantity $[\alpha^2 + \beta(\alpha + \beta)]$ is always positive. On the other hand, $[\gamma^2 + \gamma(\alpha + \beta)] = 0$ requires $\alpha + \beta + \gamma = 0$, i.e., $a = 0$, which corresponds to Minkowski space in $5D$ and $4D$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'A. Brucalassi'
- 'L. Pasquini'
- 'R. Saglia'
- 'M.T. Ruiz'
- 'P. Bonifacio'
- 'I. Leão'
- 'B.L. Canto Martins'
- 'J.R. de Medeiros'
- 'L. R. Bedin'
- 'K. Biazzo'
- 'C. Melo'
- 'C. Lovis'
- 'S. Randich'
bibliography:
- 'master.bib'
date: 'Received / Accepted '
title: |
Search for giant planets in M67 III:\
excess of hot Jupiters in dense open clusters. [^1]
---
Introduction
============
Hot Jupiters (HJs) are defined as giant planets ($M_{p} > 0.3$[$M_{\mathrm{Jup}}$]{}) on short-period orbits (P < 10 days). They show an occurrance rate of $\sim$1.2% around Sun-like field stars [@Wright2012; @Mayor2011]. These close-in giant planets are highly unlikely to have formed in situ, and it is believed that they form beyond the snow line where solid ices are more abundant, allowing the planet cores to grow several times more massive than in the inner part of the proto-planetary disk before undergoing an inward migration. Of the mechanisms that are able to trigger migration, the two supported most often are dynamical interaction with the circumstellar disk and gravitational scattering caused by other planets [@Rasio96; @Lin97 **]. Other ideas include violent migration mechanism such as dynamical encounters with a third body (multi-body dynamical interaction). In particular, recent N-body simulations have shown that a planetary system inside a crowded birth-environment can be strongly destabilized by stellar encounters and dynamical interaction, which also favours the formation of HJs [@Davies2014; @Malmberg2011; @Shara2016]. Open clusters (OCs) hold great promise as laboratories in which properties of exoplanets and theories of planet formation and migration can be explored.
In Paper I [@Pasquini2012] we described a radial velocity (RV) survey to detect the signature of giant planets around a sample of main-sequence (MS) and giant stars in M67. The first three planets discovered were presented in @Brucalassi2014. One goal of this project is to investigate whether and how planet formation is influenced by the environment. Recent planet search surveys in OCs support that the statistics in OCs is compatible with the field [@Malavolta2016; @Brucalassi2014; @Meibom2013; @Quinn2014; @Quinn2012]. In this work we show that for M67 the frequency of HJs is even higher than in the field.
Observations and orbital solutions {#sec:Sample_Obs}
==================================
Of the 88 stars in the original M67 sample, 12 have been found to be binaries [@Pasquini2012]. Two additional binaries have recently been discovered (Brucalassi et al. 2016). The final sample therefore comprises 74 single stars (53 MS and turnoff stars and 21 giants) that are all high-probability members (from proper motion and radial velocity) of the cluster according to @Yadav2008 and @Sanders77.
**Parameters** YBP401 YBP1194 YBP1514
-------------------------------------------------- ---------------- --------------- ---------------
$\mathrm{\alpha}$ $(\mathrm{J2000})$ 08:51:19.05 08:51:00.81 08:51:00.77
$\mathrm{\delta}$ $(\mathrm{J2000})$ +11:40:15.80 +11:48:52.76 +11:53:11.51
Spec.type F9V G5V G5V
$m_{\mathrm{V}}$ $[\mathrm{mag}]$ 13.70 14.6 14.77
$B-V$ $[\mathrm{mag}]$ 0.607 0.626 0.680
$M\star$ \[$M_{\odot}$\] 1.14$\pm$0.02 1.01$\pm$0.02 0.96$\pm$0.01
${\ensuremath{\log g}}$ $[\mathrm{cgs}]$ 4.30$\pm$0.035 4.44$\pm$0.05 4.57$\pm$0.06
${\ensuremath{T_{\mathrm{eff}}}}$ $[\mathrm{K}]$ 6165$\pm$64 5780$\pm$27 5725$\pm$45
: Stellar parameters of the three M67 stars newly found to host planet candidates.[]{data-label="StarParam"}
The star YBP401 shows significant indications of a HJ companion and is analysed here in detail. We also present an update of the RV measurements for the stars YBP1194 and YBP1514, for which two other HJs were announced in our previous work [@Brucalassi2014].
Basic stellar parameters ($V$, $B-V$, ${\ensuremath{T_{\mathrm{eff}}}}$, ${\ensuremath{\log g}}$) with their uncertainties were taken from the literature. A distance modulus of 9.63$\pm$0.05 [@Pasquini2008] and a reddening of E(B-V)=0.041$\pm$0.004 [@Taylor2007] were assumed, stellar masses and radii were derived using the 4 Gyr theoretical isochrones from @Pietrinferni2004 and @Girardi2000. The parameters estimated from isochrone fitting agree within the errors with the values adopted from the literature. The main characteristics of the three host stars are listed in Table \[StarParam\].
{width="48.00000%"}
\[Fit\_YBP401\]
The RV measurements were carried out using the HARPS spectrograph [@Mayor03] at the ESO 3.6m telescope in high-efficiency mode (with R=90000 and a spectral range of 378-691 nm), with the SOPHIE spectrograph [@Bouchy06] at the OHP 1.93 m telescope in high-efficiency mode (with R=40000 and a range of 387-694 nm), with the HRS spectrograph [@Tull1998] at the Hobby Eberly Telescope (with R=60000 and a range of 407.6-787.5 nm), and with the HARPS-N spectrograph at the TNG on La Palma of the Canary Islands (spectral range of 383-693 nm and R=115000). Additional RV data points for giant stars have been observed between 2003 and 2005 [@Lovis2007] with the CORALIE spectrograph at the 1.2 m Euler Swiss telescope.
HARPS, SOPHIE, and HARPS-N are provided with a similar automatic pipeline. The spectra are extracted from the detector images and cross-correlated with a numerical G2-type mask. Radial velocities are derived by fitting each resulting cross-correlation function (CCF) with a Gaussian [@Baranne1996; @Pepe2002]. For the HRS, the radial velocities were computed using a series of dedicated routines based on IRAF and by cross-correlating the spectra with a G2 star template [@Cappetta2012]. We used nightly observations of the RV standard star HD32923 to correct all observations for each star to the zero point of HARPS [as explained in @Pasquini2012] and to take into account any instrument instability or systematic velocity shifts between runs. An additional correction was applied to the SOPHIE data to consider the low signal-to-noise ratio (S/N) of the observations [@Santerne2012 see Eq.(1)].
![Phased RV measurements and Keplerian best fit, best-fit residuals, and bisector variation for YBP1194. Same symbols as in Fig.\[Fit\_YBP401\], green dots: HRS measurements.](./Figures/FIT_YBP1194_2.jpg){width="48.00000%"}
\[FIT\_YBP1194\]
The RV measurements of our target stars were studied by computing the Lomb-Scargle periodogram [@Scargle1982; @Horne1986] and by using a Levenberg-Marquardt analysis [@Wright2009 RVLIN] to fit Keplerian orbits to the radial velocity data. The orbital solutions were independently checked using the Yorbit program (Segransan et al. 2011) and a simple Markov chain Monte Carlo (MCMC) analysis (see Table \[PlanetParamMCMC\]). We investigated the presence and variability of chromospheric active regions in these stars by measuring the variations of the core of the H$\alpha$ line with respect to the continuum, following a method similar to the one described in @Pasquini1991. The more sensitive Ca II H and K lines were not accessible because of the low S/N ratio of our observations. For each case we verified the correlation between the RVs and the bisector span of the CCF [calculated following @Queloz01] or with the full width at half maximum (FWHM) of the CCF.\
\
**YBP401.** According to @Yadav2008, this F9V MS star has a membership probability of 97% and a proper motion shorter than 6 mas/yr with respect to the average. @Vereshchagin2014 revised the membership list of @Yadav2008 and expressed doubts about the cluster membership for YBP401. However, the RV value considered for YBP401 in @Vereshchagin2014 has an uncertainty of $\sigma=130$[$\mathrm{m\, s^{-1}}$]{} and is not consistent with our measurements by more than 1$\sigma$. Recently, @Geller2015 confirmed YBP401 as a single cluster member.\
This target has been observed since January 2008: 19 RV points have been obtained with HARPS with a typical S/N of 15 (per pixel at 550 nm) and a mean measurement uncertainty of 15[$\mathrm{m\, s^{-1}}$]{} including calibration errors. Five additional RV measurements were obtained with SOPHIE and two with HARPS-N, with measurement uncertainties of 9.0[$\mathrm{m\, s^{-1}}$]{} and 11.0[$\mathrm{m\, s^{-1}}$]{} , respectively.
![Phased RV measurements and Keplerian best fit, best-fit residuals, and bisector variation for YBP1514. Same symbols as in Fig.\[FIT\_YBP1194\]. ](./Figures/FIT_YBP1514_2.jpg){width="48.00000%"}
\[FIT\_YBP1514\]
The final 26 RV measurements of YBP401 show a variability of $\sim$35[$\mathrm{m\, s^{-1}}$]{}and an average uncertainty of $\sim$14[$\mathrm{m\, s^{-1}}$]{} for the individual RV values. A clear peak is present in the periodogram (see Fig. \[Periodograms\]) at 4.08 days. A Keplerian orbit was adjusted to the RV data of YBP401 (see Fig. \[Fit\_YBP401\]), and the resulting orbital parameters for the planet candidate are reported in Tables \[PlanetParam\] and \[PlanetParamMCMC\]. We note that the non-zero eccentricity is consistent with e$=$0 within 2$\sigma$ and the other parameters change by less than 1$\sigma$ when fixing e$=$0. We included the eccentricity in the data analysis, which resulted in a better fit ($\chi_{red}^{2}\sim1$) and in reduced RV residuals. However, more precise observations are needed to constrain small non-zero eccentricities and to avoid overinterpreting the results (see discussions in Zakamska et al. 2011; Pont et al. 2011). The residuals have an rms amplitude of $\sim$13[$\mathrm{m\, s^{-1}}$]{} and the periodogram of the residuals does not show any clear periodicity when the main signal is removed (see Fig. \[Periodograms\]). Neither the bisector span nor the activity index present correlations with the RV variations (see Fig. \[BIS\_FWHM\_Ha\]); this excludes activity-induced variations of the shape or the spectral lines as the source of the RV measurements.\
\
**YBP1194 and YBP1514.** We have now collected 29 measurements for both YBP1194 and YBP1514, spanning seven years. The average RV uncertainty is $\sim$13.0[$\mathrm{m\, s^{-1}}$]{} for HARPS and SOPHIE, $\sim$26.0[$\mathrm{m\, s^{-1}}$]{} for HRS and $\sim$8.0[$\mathrm{m\, s^{-1}}$]{} for HARPS-N . Figures \[FIT\_YBP1194\] and \[FIT\_YBP1514\] show the phase-folded data points together with the best-fit solution and the residual over the time. The peak in the periodogram is more pronounced and the RV signal is better determined (see Fig. \[Periodograms\]) than in @Brucalassi2014. When the planet signature is removed, the rms of the residuals is $\sim$12.3[$\mathrm{m\, s^{-1}}$]{} for YBP1194 and $\sim$14.4[$\mathrm{m\, s^{-1}}$]{} for YBP1514.\
We note that the resulting updated orbital parameters are consistent within the errors with the previously published data (see Tables \[PlanetParam\] and \[PlanetParamMCMC\]).
Frequency of hot Jupiters in OCs
================================
The most striking result is that with the star YBP401 we have found three HJs around 66 MS and subgiant stars in M67 (53 stars if we only consider single stars). This gives a frequency of HJs of 4.5$^{+4.5}_{-2.5}$% and 5.6$^{+5.4}_{-2.6}$%. These results also agree with the HJs frequency (5.5$^{+5.5}_{-2.5}$%) obtained by a Monte Carlo analysis in our parallel work (Brucalassi et al. 2016 sub.). Our values are higher than those derived from the RV surveys around FGK stars [$1.2\%\pm0.38$ of HJs @Wright2012]. The comparison is even more striking when considering that the [*Kepler*]{} [^2] statistics of HJs is lower, around 0.4$\%$ [@Howard11]. However, the comparison between different samples and between simulations and observations is not trivial. The analysis of the [*Kepler*]{} and the RV surveys for instance use different selection criteria (radii vs. masses) and different intervals of orbital periods. @Dawson2013 showed that the discrepancy might be partially due to the different metallicity of the samples. Another effect to take into account is in the definition of the comparison samples. RV surveys are performed on pre-selected samples that have been corrected for the presence of binaries, while the [*Kepler*]{} statistics (and most of the simulations) refer to all FGK stars in the Cygnus field, without any previous selection for binaries. For M67 our survey sample was heavily pre-selected with the aim to eliminate all known and suspected binaries in advance. We therefore expect that when we compare our results on the whole sample (3/66 or 4.5$^{+4.5}_{-2.5}$%) with the [*Kepler*]{} statistics (0.4$\%$), an upper limit of the planet frequency will be provided, while the comparison of the frequency of the single-star sample (3/53 or 5.6$^{+5.4}_{-2.6}$% ) is expected to compare well with the 1.2$\%$ of the radial velocity surveys because they have gone through a similar selection process. Finally, based on a HJ occurrance rate of 1.2% like for field stars, one or two additional HJs may exist in M67 with a non-negligible probability of 5%.
For several years, the lack of detected planets in OCs was in contrast with the field results, but the recent discoveries [@Malavolta2016; @Brucalassi2014; @Meibom2013; @Quinn2014; @Quinn2012] have completely changed the situation.\
These results are difficult to reconcile with the early survey of @Paulson2004, who did not detect any HJs around 94 G-M stars of the Hyades cluster. Since the extrapolation from non-detection to non-existence of HJs critically depends on several assumptions such as stellar noise and real measurement errors, constraints on the allocated time, number of observations per star, sampling and planet mass, we cannot state at present whether the discrepancy is real until larger surveys are performed. If we were to add the results of the RV surveys in the OCs M67, Hyades, and Praesepe, we would determine a rate of 6 out of 240 HJs per surveyed stars (including some binaries), which is a high percentage when compared to 10 out of 836 HJ per surveyed stars in the field sample of @Wright2012. We can conclude that, contrary to early reports, the frequency of HJs discovered in the three OCs subject of recent RV surveys is higher than amongst the field stars. To explain the high frequency of HJs in M67, Hyades, and Praesepe, we may argue that the frequency of HJs depends on stellar metallicity, mass, or on dynamical history, and therefore environment. The dependence of planet frequency on stellar metallicity is complex: even if established very early [@Johnson2010; @Udry2007; @Fischer2005; @Santos2004], a real correlation seems to be present only for Jupiters around MS stars, while it does not hold for giant planets around evolved stars [@Pasquini2007] or for low-mass planets [@Mayor2011]. Both Hyades and Praesepe are metal rich [@Pace2008; @Ferreira2015], and this may explain the higher frequency of HJs in these clusters, but this is not the case of M67, which has a well-established solar metallicity and abundance pattern [@Randich2005]. The hypothesis that the high frequency of HJs in M67 or in OCs in general originates from the higher mass of the host star can be also excluded: the stars hosting HJs in M67 all have masses around one solar mass, which is very similar to the masses of the HJ hosts discovered in the field. A similar argument holds for the stars hosting HJs in Praesepe and in the Hyades.
Finally, environment is left as the most suitable option to explain the HJs excess. It has been suggested that dense birth-environments such as stellar clusters can have a significant effect on the planet formation process and the resulting orbital properties of single planets or planetary systems. Close stellar fly-by or binary companions can alter the structure of any planetary system and may also trigger subsequent planet-planet scattering over very long timescales [@Davies2014; @Malmberg2011]. This leads to the ejection of some planets, but it also seems to favour the conditions for the formation of HJs [@Shara2016]. As predicted by such mechanisms, M67 HJs show orbits with non-zero eccentricity, which is also true for the HJ found in the Hyades. The importance of the encounters is primarly determined by the local stellar density, the binary fraction, the collisional cross-section of the planetary system, and the timescale on which the planet is exposed to external perturbations. @Malmberg2011 produced simulations for a cluster of 700 stars and an initial half-mass radius of 0.38 parsecs, showing that a non-negligible number of stars spend long enough as a binary system and also that the majority of the stars is affected by at least one fly-by. M67 has more than 1400 stars at present, it is dominated by a high fraction of binaries [@Davenport2010] after loosing at least three quarters of its original stellar mass, and has suffered mass segregation. @Shara2016 have recently completed N-body simulations for a case similar to the one of M67, but with only a 10$\%$ of binaries, finding that HJs can be produced in 0.4$\%$ of cluster planetary systems when only considering initial fly-by encounters. This fraction is smaller than what we find in M67, and the influence of other migration mechanisms probably needs to be considered as well to explain our results. However, the same authors acknowledged that a higher fraction of binaries will strongly enhance the probability of HJ formation and therefore their frequency. Given that the binary fraction of M67 stars is currently very high [@Pasquini2012; @Mathieu1990] and that models show that it must also have been high at the origin [@Hurley2005], the high fraction of M67 HJs seems to qualitatively agree with the N-body simulations. The same simulations predict that after 5 Gyr the percentage of stars hosting HJs retained by the cluster is substantially higher than the percentage of stars not hosting HJs. Two factors can contribute to enhance HJ planet formation: the capability of producing HJs, and the capability of the cluster to retain stars hosting HJs. The interaction takes part well within the first Gyr of the cluster lifetime, so that the stars with HJs do not require to still have a stellar companion at the age of M67. @Geller2015 reported no evidence for nearby companions at the present epoch. Finally, considering that about one of ten HJs produces a transit, we suggest to carefully examine the [*Kepler*]{}/K2 observations [@Howell2014] for any transit.
LP acknowledges the Visiting Researcher program of the CNPq Brazilian Agency, at the Fed. Univ. of Rio Grande do Norte, Brazil. RPS thanks ESO DGDF, the HET project, the PNPS and PNP of INSU - CNRS for allocating the observations. MTR received support from PFB06 CATA (CONICYT).
[c]{} RV values Residuals BIS span FWHM\
\
\
\
------------ ------------------------------ ------------------------------ ------------------------------ ------------------- ------------
BJD RV $\sigma_{obs}$ BIS span H$_{\alpha}$ratio instrument
(-2450000) ([$\mathrm{km\, s^{-1}}$]{}) ([$\mathrm{km\, s^{-1}}$]{}) ([$\mathrm{km\, s^{-1}}$]{})
4491.3462 33.2165 0.0089 $-$0.0013333 0.03133 SOPHIE
4855.6494 33.1522 0.0110 $-$0.0241850 0.02735 HARPS
4859.5234 33.1978 0.0089 0.0150000 0.02843 SOPHIE
4861.8052 33.1559 0.0110 0.0573989 0.02435 HARPS
5190.8584 33.1793 0.0080 $-$0.0384314 0.02154 HARPS
5627.7314 33.2302 0.0270 0.0061686 0.02206 HARPS
5946.8267 33.1927 0.0180 0.0812266 0.02624 HARPS
5967.6030 33.1589 0.0170 $-$0.0273837 0.02377 HARPS
5978.5732 33.2331 0.0170 0.0187173 0.02635 HARPS
5985.5850 33.1612 0.0090 $-$0.0098333 0.02806 SOPHIE
6245.8511 33.1369 0.0150 0.0556374 0.02424 HARPS
6269.7925 33.1744 0.0170 0.0426396 0.02201 HARPS
6302.7832 33.1737 0.0220 0.0607550 0.02471 HARPS
6308.7578 33.1658 0.0110 0.0069258 0.02561 HARPS
6323.7534 33.1513 0.0170 0.0179823 0.02399 HARPS
6333.7314 33.2038 0.0140 0.0456952 0.02144 HARPS
6682.6787 33.1879 0.0090 0.0310414 0.02459 HARPS
6694.5210 33.1934 0.0124 0.0562494 - HARPS-N
6697.4912 33.2130 0.0092 0.0208414 - HARPS-N
6715.7002 33.1711 0.0110 0.0338823 0.02443 HARPS
6719.6357 33.1821 0.0140 0.0273266 0.02430 HARPS
6720.4028 33.1154 0.0145 $-$0.0233333 0.02147 SOPHIE
6721.5381 33.1395 0.0129 0.0215000 0.03277 SOPHIE
6977.8101 33.1514 0.0150 $-$0.0060115 0.02535 HARPS
6978.8369 33.1645 0.0160 0.0337306 0.02274 HARPS
6983.8188 33.2158 0.0120 0.0234168 0.02289 HARPS
------------ ------------------------------ ------------------------------ ------------------------------ ------------------- ------------
**Parameters** YBP401 YBP1194 YBP1514
----------------------------------------- ------------------------------ ---------------------------- ------------------------------
$P$ $[\mathrm{days}]$ 4.0873$^{+0.0003}_{-0.0002}$ 6.959$^{+0.001}_{-0.001}$ 5.1189$^{+0.0008}_{-0.0007}$
$T$ $[\mathrm{JD}]$ 2455974.23$^{+0.49}_{-0.49}$ 2455290.0$^{+0.4}_{-0.3}$ 245986.34$^{+0.28}_{-0.20}$
$e$ 0.141$^{+0.112}_{-0.113}$ 0.294$^{+0.077}_{-0.056}$ 0.332$^{+0.133}_{-0.127}$
$\omega$ $[\mathrm{deg}]$ -31.69$^{+43.0}_{-54.0}$ 99.14$^{+16.0}_{-16.0}$ -34.76$^{+17.35}_{-14.26}$
$\sqrt{e}\sin \omega$ -0.197$^{+0.152}_{-0.166}$ 0.535$^{+0.061}_{-0.064}$ -0.329$^{+0.134}_{-0.136}$
$\sqrt{e}\cos \omega$ 0.319$^{+0.104}_{-0.242}$ -0.086$^{+0.186}_{-0.171}$ 0.474$^{+0.092}_{-0.086}$
$K$ \[[$\mathrm{m\, s^{-1}}$]{}\] 48.911$^{+5.0}_{-6.0}$ 35.607$^{+2.2}_{-4.0}$ 50.97$^{+3.0}_{-3.0}$
$m\sin{i}$ \[[$M_{\mathrm{Jup}}$]{}\] 0.41$^{+0.06}_{-0.04}$ 0.32$^{+0.3}_{-0.2}$ 0.42$^{+0.03}_{-0.02}$
$\gamma$ \[[$\mathrm{km\, s^{-1}}$]{}\] 33.172$^{+0.003}_{-0.004}$ 34.185$^{+0.002}_{-0.002}$ 34.058$^{+0.003}_{-0.003}$
[^1]: Based on observations collected at the ESO 3.6m telescope (La Silla), at the 1.93m telescope of the Observatoire de Haute-Provence (OHP), at the Hobby Eberly Telescope (HET), at the Telescopio Nazionale Galileo (TNG, La Palma) and at the Euler Swiss Telescope.
[^2]: http://kepler.nasa.gov/
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Cyprien de Masson d’Autume[^1] Mihaela RoscaJack Rae Shakir Mohamed\
DeepMind\
`{cyprien,mihaelacr,jwrae,shakir}@google.com`
bibliography:
- 'main.bib'
title: Training Language GANs from Scratch
---
[^1]: Equal contribution.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper discusses asymptotic theory for penalized spline estimators in generalized additive models. The purpose of this paper is to establish the asymptotic bias and variance as well as the asymptotic normality of the penalized spline estimators proposed by Marx and Eilers (1998). Furthermore, the asymptotics for the penalized quasi likelihood fit in mixed models are also discussed.'
author:
- |
[Takuma Yoshida]{}$^{1}$ [and]{} [Kanta Naito]{}$^{1}$[^1]\
$^{1}$[*Graduate School of Science and Engineering*]{}\
[*Shimane University, Matsue 690-8504, Japan*]{}
title: Asymptotics for penalized splines in generalized additive models
---
[**Keywords**]{} Asymptotic normality, $B$-spline, Generalized additive model, Mixed model, Penalized spline.
Introduction
============
The generalized additive model(GAM) is a typical regression model, in which the relationship between the one-dimensional response $Y$ and the multidimensional explanatory $\vec{x}=(x_1,\cdots,x_D)$ is modeled by a link function $g(\cdot)$, as follows: $$g(E[Y|\vec{X}=x])=\eta(\vec{x})=\eta_1(x_1)+\cdots+\eta_D(x_D),$$ where each $\eta_j(j=1,\cdots,D)$ is a univariate regression function. If $Y$ has a Gaussian distribution, then $g$ is the identity function and, hence, $\eta(\vec{x})=E[Y|\vec{X}=\vec{x}]$. Additionally, the GAM can specify a distribution such as a Bernoulli, Poisson or Gamma distribution. In GAMs, the purpose is often to estimate $\eta$. The parametric and the nonparametric estimation techniques of $\eta$ have been established by several authors (see Hastie et al. (1990) and Wood (2006)). In this paper, $\eta$ is estimated via the penalized spline method. Penalized splines were introduced by O’Sullivan (1986) and Eilers and Marx (1996), and are recognized as an efficient technique for GAMs. Applications and theories of penalized splines in GAMs have been widely discussed, including by Aerts et al. (2002) and Ruppert et al. (2003).
To construct the estimator of $\eta_j$’s, a repetition update method, the so-called backfitting algorithm, is often used. However, when the response has a non-Gaussian distribution, such as a Bernoulli or Poisson distribution, the overall estimation procedure becomes complicated and its computation time grows, because the estimators are obtained by using a blend of backfitting and the Fisher-scoring algorithm. On the other hand, Marx and Eilers (1998) proposed a new penalized spline estimator without backfitting algorithms, which is denoted as the ridge corrected penalized spline estimator (RCPS). We will briefly describe the RCPS as we will focus on it later. The penalized spline estimator is obtained based on maximization of the penalized log-likelihood. However, it appears difficult to obtain the maximizer of the penalized log-likelihood $\ell$ as the Hessian of $\ell$ is not invertible. The RCPS is constructed based on the maximization of $\ell_\gamma$, which is $\ell$ plus an additional ridge penalty. Since the Hessian of $\ell_\gamma$ is invertible, the maximizer of $\ell_\gamma$ can be obtained via the Fisher-scoring algorithm. Thus, it is easy to construct the RCPS.
In univariate models($D=1$), Hall and Opsomer (2005), Claeskens et al. (2009), Kauermann et al. (2009) and Wang et al. (2011) researched the asymptotic properties of penalized spline estimators. Recently, Yoshida and Naito (2012) worked on the asymptotic distribution of penalized splines in an additive model. In contrast, Horowitz and Mammen (2004), Linton (2000) and Yu et al. (2008) studied the asymptotics for the kernel estimator in a GAM. However, the asymptotic results for penalized spline estimators in GAMs have not yet been sufficiently developed like they have been for their applications.
In this paper, the asymptotics for penalized splines in a GAM are discussed. Our main purpose is to establish the asymptotic normality of the RCPS. Kauermann et al. (2009) showed the asymptotic normality of the penalized spline estimator in generalized linear models(GLM). Hence, the results in this paper generalize the results of Kauermann et al. (2009). Furthermore, penalized spline smoothing can be linked to mixed models(see Lin and Zhang (1999) and Ruppert et al. (2003)). In generalized additive mixed models (GAMM), the penalized quasi likelihood method (PQL) is an efficient method for obtaining the estimator and predictor. We also show the asymptotic normality of the PQL fit.
This paper is organized as follows. In Section 2, the GAM is defined and the RCPS is constructed by the Fisher-scoring algorithm. Section 3 shows the asymptotic normality of the RCPS. Section 4 states the asymptotics for the PQL fits in a GAMM. In Section 5, the applications of the approximate confidence interval are addressed. Section 6 provides a numerical study to validate the asymptotic normality of the RCPS. Related discussions and issues for future research are addressed in Section 7, and proofs for theoretical results are given in the Appendix.
Penalized spline estimator in a GAM
===================================
Generalized additive spline model
---------------------------------
For the dataset $\{ (y_i,\vec{x}_{i})|i=1,\cdots,n\}$, consider an exponential family of the generalized additive model with a canonical link function $$\begin{aligned}
f(y_i|\vec{x}_i,\eta)=\exp \left( \frac{y_i\eta(\vec{x}_i)-c(\eta(\vec{x}_i))}{\phi } + h(y_i,\phi) \right),\quad i=1,\cdots,n, \label{model1}\end{aligned}$$ where $\vec{x}_i=(x_{i1},\cdots,x_{iD})$ is a $D$-variate explanatory variable, $\eta(\cdot)$ is an unknown natural parameter which has the additive formation $$\eta(\vec{x})=\eta_1(x_1)+\cdots+\eta_D(x_D),$$ where $\eta_j$’s is an unknown univariate function, $\phi$ is a dispersion parameter, and $c$ and $h$ are known functions. The canonical link function indicates $g^{-1}=c^\prime$, which leads to $E[Y_i|\vec{X}_i=\vec{x}_{i}]=g^{-1}(\eta(\vec{x}_i))=c^\prime(\eta(\vec{x}_i))$ and $V[Y_i |\vec{X}_i=\vec{x}_{i}]= \phi c^{\prime\prime}(\eta(\vec{x}_i ))$, where $c^\prime$ and $c^{\prime\prime}$ are the first and second derivatives of $c$. More general settings concerned with the link function were clarified by McCullagh and Nelder (1989). In this paper, for the natural parameter $\eta(\vec{x})$, we assume that $E[\eta_j(X_j)]=0 (j=1,\cdots,D)$ to ensure the identifiability of $\eta_j$. For simplicity, we hereafter ignore the role of the dispersion parameters in (\[model1\]) and set $\phi\equiv 1$, thus denoting $h(y,\phi)=h(y)$.
Our purpose is to estimate $\eta_j$ using nonparametric spline methods. We now prepare the $B$-spline model $$s(x)=\sum_{k=-p+1}^{K_{n}} B_k^{[p]}(x)b_{k,j},\ \ j=1,\cdots,D$$ as an approximation to $\eta_j(x)$, where $B_k^{[p]}(x)(k=-p+1,\cdots,K_{n})$ are the $p$th $B$-spline functions defined recursively as $$\begin{aligned}
B_k^{[0]}(x)&=&
\left\{
\begin{array}{cc}
1,& \kappa_{k-1}<x\leq \kappa_k,\\
0,& {\rm otherwise},
\end{array}
\right. \\
B_k^{[p]}(x)&=&\frac{x-\kappa_{k-1}}{\kappa_{k+p-1}-\kappa_{k-1}}B_k^{[p-1]}(x)+\frac{\kappa_{k+p}-x}{\kappa_{k+p}-\kappa_{k}}B_{k+1}^{[p-1]}(x),\end{aligned}$$ where $\kappa_k (k=-p+1,\cdots,K_n+p)$ are knots and $b_{k,j}$’s is an unknown parameter. Some fundamental properties of $B$-splines were detailed by de Boor (2001). We will write $B_k^{[p]}(x)=B_k(x)$ unless we specify the degree of $B$-splines, because we will mainly focus on the $p$th $B$-spline from now on. The suggested density of $Y_i$ is defined as $$\begin{aligned}
f(y_i|\vec{x}_i,\vec{b})=\exp \left( y_i Z(\vec{x}_i)^T\vec{b}-c(Z(\vec{x}_i)^T\vec{b}) + h(y_i) \right),\end{aligned}$$ where $Z(\vec{x})=(\vec{B}(x_1)^T\ \cdots\ \vec{B}(x_D)^T)^T$, $\vec{B}(x)=(B_{-p+1}(x)\ \cdots\ B_{K_n}(x))^T$, $\vec{b}=(\vec{b}_1^T\ \cdots\ \vec{b}_D^T)^T$ and $\vec{b}_j=(b_{-p+1,j}\ \cdots\ b_{K_n,j})^T$. Using the estimator $\hat{\vec{b}}_j=(\hat{b}_{-p+1,j}\ \cdots\ \hat{b}_{K_n,j})^T$ of $\vec{b}_j$, the estimator of $\eta_j(x_j)$ can be defined as $$\hat{\eta}_j(x_j)=\sum_{k=-p+1}^{K_n} B_{k}(x_j)\hat{b}_{k,j},\ \ j=1,\cdots,D.$$
The penalized spline estimator
------------------------------
To estimate $\vec{b}$, we prepare the log-likelihood $$\begin{aligned}
\frac{1}{n}\sum_{i=1}^n \log f(y_i|\vec{x}_i,\vec{b})=\frac{1}{n}\{\vec{y}^T (Z\vec{b})-\vec{1}^T c(Z\vec{b})\}+\frac{1}{n}\vec{1}^T h(\vec{y}),\end{aligned}$$ where $\vec{y}=(y_1\ \cdots\ y_n)^T$, $c(Z\vec{b})=(c(Z(\vec{x}_1)^T\vec{b})\ \cdots\ c(Z(\vec{x}_n)^T\vec{b}))^T$, $Z_k=(B_{-p+j}(x_{ik}))_{ij}$, $Z=[Z_1\ \cdots\ Z_D]$, and $h(\vec{y})=(h(y_1),\cdots,h(y_n))^T$. It is known that the spline estimator obtained by maximization of the log-likelihood tends to display ’wiggle behavior’. Hence, we consider using the penalized spline estimator to obtain a smooth curve. Define the penalized log-likelihood as follows $$\begin{aligned}
\ell(\vec{b},\lambda_n)
&=&
\frac{1}{n}\sum_{i=1}^n \log f(y_i|\vec{x}_i,\vec{b})-\sum_{j=1}^D \frac{\lambda_{jn}}{2n}\vec{b}_j^T \Delta_m^\prime \Delta_m \vec{b}_j\nonumber\\
&=&
\frac{1}{n}\{\vec{y}^T (Z\vec{b})-\vec{1}^T c(Z\vec{b})\}+\frac{1}{n}\vec{1}^T h(\vec{y})-\frac{1}{2n} \vec{b}^T Q_m(\lambda_n) \vec{b}, \label{pen1}\end{aligned}$$ where $\lambda_{jn}$ is the smoothing parameter $(j=1,\cdots,D)$, the $(K_n+p-m)\times (K_n+p)$th matrix $\Delta_m$ is the $m$th difference matrix, which is given by Marx and Eilers (1998) and $Q_m(\lambda_n)={\rm diag}[\lambda_{1n}\Delta_m^\prime \Delta_m\ \cdots\ \lambda_{Dn}\Delta_m^\prime \Delta_m]$. In general, the maximizer of (\[pen1\]) is obtained by the Fisher-scoring algorithm. As in the typical problem of spline methods in a GAM, however, the Hessian of $\ell(\vec{b},\lambda_n)$ is not invertible and so the Fisher-scoring method is not usable directly. To overcome this singularity problem, we can use backfitting algorithms (see Hastie and Tibshirani (1990)). However, the overall algorithm becomes complicated and the computation grows (see Section 1). These problems were discussed by Marx and Eilers (1998). We will next review the ridge corrected penalized spline estimator.
The ridge corrected penalized spline estimator
----------------------------------------------
Marx and Eilers (1998) proposed a nice estimation method for $\vec{b}$ without using a backfitting algorithm for penalized splines in the GAM context. They defined the ridge corrected penalized log-likelihood as $$\begin{aligned}
\ell(\vec{b},\lambda_n,\gamma_n)
=
\ell(\vec{b},\lambda_n)-\frac{\gamma_n}{2n}\vec{b}^T \vec{b}, \label{pen}\end{aligned}$$ where $\gamma_n>0$. Let $\hat{\vec{b}}=(\hat{\vec{b}}_1^T\ \cdots\ \hat{\vec{b}}_D^T)^T$ be the maximizer of (\[pen\]), which can be obtained directly via the Fisher-scoring method since the Hessian of $\ell(\vec{b},\lambda_n,\gamma_n)$ is invertible. The gradient $G(\vec{b},\lambda_n,\gamma_n)$ and Hessian $H(\vec{b},\lambda_n,\gamma_n)$ of $\ell(\vec{b},\lambda_n,\gamma_n)$ are obtained with $$\begin{aligned}
G(\vec{b},\lambda_n,\gamma_n)&=&\frac{\partial \ell(\vec{b},\lambda_n,\gamma_n)}{\partial \vec{b}}
=
\frac{1}{n}\{Z^T\vec{y}-Z^T c^\prime (Z\vec{b})\}-\frac{1}{n}Q_m(\lambda_n)\vec{b}-\frac{\gamma_n}{n}\vec{b},\\
H(\vec{b},\lambda_n,\gamma_n)&=&\frac{\partial^2 \ell(\vec{b},\lambda_n,\gamma_n)}{\partial \vec{b}\partial \vec{b}^T}
=
-\frac{1}{n}Z^T \operatorname{diag}[c^{\prime\prime} (Z\vec{b})]Z-\frac{1}{n}Q_m(\lambda_n)-\frac{\gamma_n}{n} I,\end{aligned}$$ where $c^{\prime}(Z\vec{b})$ and $c^{\prime\prime}(Z\vec{b})$ are $n$-vectors defined in the same manner as $c(Z\vec{b})$. The $k$-step iterated estimator $\vec{b}^{(k)}$ of $\vec{b}$ can be written as $$\begin{aligned}
{\small
\vec{b}^{(k)}=
(Z^T W^{(k-1)} Z+Q_m(\lambda_n)+\gamma_n I)^{-1}
Z^T W^{(k-1)}\left\{ Z\vec{b}^{(k-1)}+(W^{(k-1)})^{-1}\left\{\vec{y}-c^\prime (Z\vec{b}^{(k-1)})\right\}\right\},
}\end{aligned}$$ where $W^{(k-1)}={\rm diag}[c^{\prime\prime}(Z\vec{b}^{(k-1)})]$. As $k\rightarrow \infty$, $\vec{b}^{(k)}$ converges to $\hat{\vec{b}}$ if the initial $\vec{b}^{(0)}$ is appropriately chosen. The RCPS of $\eta_j(x_j)$ can be obtained as $\hat{\eta}_j(x_j)=\vec{B}(x_j)^T \hat{\vec{b}}_j$. In the next section, we discuss the asymptotic properties of $[\hat{\eta}_1(x_1)\ \cdots\ \hat{\eta}_D(x_D)]^T$.
Asymptotic theory
=================
Here, we list some assumptions regarding the asymptotics of the penalized spline estimator.\
[**Assumptions**]{}
1. The explanatory $\vec{X}=(X_1,\cdots,X_D)$ is distributed as $P(\vec{x})$ on $[0,1]^D$, where $[0,1]^D$ is the $D$-variate unit cube.
2. For $j=1,\cdots,D$, $\eta_j\in C^{p+1}$ and $c\in C^3$.
3. The knots for the $B$-spline basis are equidistantly located with $\kappa_k=k/K_n(k=-p+1,\cdots,K_n+p)$ and the number of knots satisfies $K_n=o(n^{1/2})$.
4. For the non-singularity of $H(\vec{b},\lambda_n,\gamma_n)$, $K_n$ is chosen such that $D(K_n+p)<n$.
5. The smoothing parameters $\lambda_{jn} (j=1,\cdots,D)$ are positive sequences such that $\lambda_{jn}^{-1}$ is larger than the maximum eigenvalue of $(Z_j^T Z_j)^{-1/2}\Delta_m^\prime \Delta_m(Z_j^T Z_j)^{-1/2}$.
6. Lastly, $\gamma_n=o(\lambda_nK_n^{-m})$, where $\lambda_n=\max_{j}\{\lambda_{jn}\}$.
For a random variable $U_n$, $E[U_n|\vec{X}_n]$ and $V[U_n|\vec{X}_n]$ denote the conditional expectation and variance of $U_n$ given $(\vec{X}_1,\cdots,\vec{X}_n)=(\vec{x}_1,\cdots,\vec{x}_n)$, respectively. Define the $(K_n+p)$th square matrix $G_{k}=(G_{k,ij})_{ij}$, where the $(i,j)$-th component is $$\begin{aligned}
G_{k,ij}=\int_{[0,1]^D} c^{\prime\prime}(\eta(\vec{x}))B_{-p+i}(x_k)B_{-p+j}(x_k)dP(\vec{x}).\end{aligned}$$ Using this, we get $\Gamma_j(\lambda_{jn})=(G_{j}+(\lambda_{jn}/n)\Delta_m^\prime \Delta_m)$. Let $$\begin{aligned}
\vec{b}_0=(\vec{b}_{10}^T \cdots\ \vec{b}_{D0}^T)^T=\underset{\vec{b}}{\operatorname*{argmin}}\left\{\frac{1}{n}\sum_{i=1}^nE\left[\left.\log \frac{f(Y_i|\vec{x}_i,\eta)}{f(Y_i|\vec{x}_i,\vec{b})}\right|\vec{X}_n\right]\right\} \label{b0}\end{aligned}$$ and let $\eta_{j0}(x_j)=\vec{B}(x_j)^T \vec{b}_{j0} (j=1,\cdots,D)$.
The asymptotic bias of $\hat{\eta}_j(x_j)$ can be written as $$E[\hat{\eta}_j(x_j)|\vec{X}_n]-\eta_j(x_j)=E[\hat{\eta}_j(x_j)|\vec{X}_n]-\eta_{j0}(x_j)+\eta_{j0}(x_j)-\eta_j(x_j).$$ In the following Proposition \[app\], the difference $\eta_{j0}(x_j)-\eta_j(x_j)$ is asymptotically evaluated. The asymptotics for $\hat{\eta}_j(x_j)-\eta_{j0}(x_j)=\vec{B}(x_j)^T (\hat{\vec{b}}_j-\vec{b}_{j0})$ can be shown in the following Theorem \[mv\] by using the Taylor expansion of $G(\hat{\vec{b}},\lambda_n,\gamma_n)$ around $\vec{b}_0$ (see Lemma \[ex\] in the Appendix), the properties of a partitioned matrix of $H(\vec{b},\lambda_n,\gamma_n)$ and its asymptotic results.
\[app\] Under the Assumptions, for $j=1,\cdots,D,$ $$\begin{aligned}
\eta_{j0}(x_j)-\eta_j(x_j)= b_{j,a}(x_j)+o(K_n^{-(p+1)}),\end{aligned}$$ where $$b_{j,a}(x)=-\frac{\eta^{(p+1)}_j(x)}{K_n^{p+1} (p+1)!}\sum_{k=1}^{K_n}I(\kappa_{k-1}\leq x<\kappa_k){\rm Br}_{p+1}\left(\frac{x-\kappa_{k-1}}{K_n^{-1}}\right),$$ $I(a<x<b)$ is the indicator function of an interval $(a,b)$ and ${\rm Br}_p(x)$ is the $p$th Bernoulli polynomial.
\[mv\] Under the Assumptions, for $j=1,\cdots,D,$ $$\begin{aligned}
E[\hat{\eta}_j(x_j)|\vec{X}_n]-\eta_j(x_j)
&=&b_{j,a}(x_j)+b_{j,\lambda}(x_j)+o_P(K_n^{-(p+1)})+o_P(\lambda_{jn}K_n^{1-m}n^{-1}),\\
V[\hat{\eta}_j(x_j)|\vec{X}_n]
&=&
\frac{1}{n}\vec{B}(x_j)^T\Gamma_j(\lambda_{jn})^{-1}\Gamma_j(0)\Gamma_j(\lambda_{jn})^{-1}\vec{B}(x_j)(1+o_P(1))\\
&=&O_P(K_n/n),\\
Cov(\hat{\eta}_i(x_i),\hat{\eta}_j(x_j))&=&O_P(n^{-1}),\end{aligned}$$ where $b_{j,a}(x_j)$ is given in Proposition \[app\], $$b_{j,\lambda}(x)=-\frac{\lambda_{jn}}{n}\vec{B}(x)^T \Gamma_j(\lambda_{jn})^{-1}\Delta_m^\prime \Delta_m\vec{b}_{j0}=O\left(\frac{\lambda_n K_n^{1-m}}{n}\right).$$
In Theorem \[mv\], the influence of $\gamma_n$ appears to be of only negligible order. In actuality, we can use very small $\gamma_n$ as long as $H(\vec{b},\lambda_n,\gamma_n)$ is nonsingular. For example, Marx and Eilers (1998) used $\gamma_n=10^{-6}$. Thus, it is understood that the influence of $\gamma_n$ is small theoretically and numerically. From Theorem \[mv\], the conditional Mean Squared Error(MSE) of $\hat{\eta}_j(x_j)$ can be obtained as follows.
Under the same assumption as Theorem \[mv\], it follows that $$\begin{aligned}
{\small
{\rm MSE}[\hat{\eta}_j(x_j)|\vec{X}_n]=E[\{\hat{\eta}_j(x_j)-\eta_j(x_j)\}^2|\vec{X}_n]=O_P\left(K_n^{-2(p+1)}+\lambda_{jn}^2K_n^{2(1-m)}n^{-2}\right)+O_P(K_nn^{-1}).
}\end{aligned}$$ Furthermore, the rate of convergence of the [MSE]{} of $\hat{\eta}_j(x_j)$ becomes $O(n^{-(2p+2)/(2p+3)})$ by taking $K_n=O(n^{1/(2p+3)})$, $\lambda_{jn}=O(n^{\nu}), \nu\leq (p+m+1)/(2p+3)$.
Compared with the kernel estimator, the asymptotic order of MSE of the RCPS is the same as that of the local $p$th polynomial estimator when $p$ is odd and the number of knots in the spline methods and the bandwidth $h_n$ in the kernel methods are connected by $K_n/h_n^{-1}=O(1)$(see Opsomer (2000)). Lyapunov’s condition of the central limit theorem yields the asymptotic normality of $[\hat{\eta}_1(x_1)\ \cdots\ \hat{\eta}_D(x_D)]^T$.
\[clt\] Suppose there exists $\delta\geq 2$ such that $E[|Y_i-c^\prime(\eta(\vec{x}_i))|^{2+\delta}|\vec{X}_i=\vec{x}_i]<\infty$. Furthermore, we assume $K_n=O(n^{1/(2p+3)})$ and $\lambda_{n}=O(n^\nu), \nu\leq (p+m+1)/(2p+3)$. Then, under the Assumptions, for any fixed point $\vec{x}=(x_1,\cdots,x_D)\in(0,1)^D$, as $n\rightarrow \infty$, $$\begin{aligned}
\sqrt{\frac{n}{K_n}}
\left[
\begin{array}{c}
\hat{\eta}_{1,\gamma}(x_1)-\eta_1(x_1)-{\rm Bias}_1(x_1)\\
\vdots\\
\hat{\eta}_{D,\gamma}(x_D)-\eta_D(x_D)-{\rm Bias}_D(x_D)
\end{array}
\right]
\stackrel{d}{\longrightarrow}N_D\left(
\vec{0}
,\Psi\right),\end{aligned}$$ where for $j=1,\cdots,D$, ${\rm Bias}_j(x_j)=b_{j,a}(x_j)+b_{j,\lambda}(x_j)$, and $\Psi=\operatorname{diag}[\psi_1(x_1)\ \cdots\ \psi_D(x_D)]$ with $$\psi_j(x_j)=\lim_{n\rightarrow \infty}\frac{1}{K_n}\vec{B}(x_j)^T\Gamma_j(\lambda_{jn})^{-1}\Gamma_j(0)\Gamma_j(\lambda_{jn})^{-1}\vec{B}(x_j),\ \ j=1,\cdots,D.$$
The proof of Theorem \[clt\] is almost the same as that of Theorem 2 of Yoshida and Naito (2012). The asymptotic order of the bias and variance of the RCPS in Theorem \[mv\] allows us to satisfy Lyapunov’s condition for $[\hat{\eta}_1(x_1)\ \cdots\ \hat{\eta}_D(x_D)]^T$.
We note that an approximate pointwise confidence interval of $\eta_j(x_j)$ can be constructed by using the asymptotic distributional result of $\hat{\eta}_j(x_j)$. However, the asymptotic bias and variance of $\hat{\eta}_j(x_j)$ contain unknown variables and, hence, these should be estimated. For example, we replace $\vec{b}_0$ and $G_j$ with $\hat{\vec{b}}$ and $n^{-1}Z_j^T \hat{W}Z_j$, respectively, where $\hat{W}=\operatorname{diag}[c^{\prime\prime}(Z\hat{\vec{b}})]$. Furthermore, as it is the pilot estimator of the $(p+1)$th derivative of $\eta_j$, we can utilize the $(p+1)$th derivative of the RCPS $\hat{\eta}_j$ with $(p+2)$ or higher degree splines. Thus, we can construct the estimator $\widehat{{\rm Bias}}_j(x_j)$ and $\hat{\psi}_j(x_j)$ of ${\rm Bias}_j(x_j)$ and $\psi_j(x_j)$, respectively. Consequently, we obtain an approximate confidence interval of $\eta_j(x_j)$ by the following Corollary.
\[inter\] Under the same assumption as Theorem \[clt\], a $100(1-\alpha)\%$ asymptotic confidence interval of $\eta_j(x_j)$ at any fixed point $x_j\in(0,1)$ is $$\left[\hat{\eta}_j(x_j)-\widehat{{\rm Bias}}_j(x_j)- z_{\alpha/2}\sqrt{\hat{\psi}_j(x_j)},\ \hat{\eta}_j(x_j)-\widehat{{\rm Bias}}_j(x_j)+ z_{\alpha/2}\sqrt{\hat{\psi}_j(x_j)}\right],$$ where $z_{\alpha/2}$ is the $(1-\alpha/2)$th normal percentile.
[**Remark 1**]{} We see from the proof of Theorem \[mv\] that the asymptotic form of $\hat{\eta}_j(x_j)$ can be written as $$\begin{aligned}
\hat{\eta}_j(x_j)-\eta_j(x_j)=\left\{\vec{B}(x_j)^T \Gamma_j(\lambda_{jn})^{-1}G_j(\vec{b}_0,\lambda_n,\gamma_n)+b_{j,a}(x_j)\right\}(1+o_P(1)) \label{asRCPS}\end{aligned}$$ under the same assumption as Theorem \[clt\], where $G_j(\vec{b}_0,\lambda_n,\gamma_n)$ is the $j$th $(K_n+p)$-subvector of $G(\vec{b}_0,\lambda_n,\gamma_n)$. From (2.8) of Kauermann et al. (2009), we see that $\hat{\eta}_j(x_j)$ and the penalized spline estimator based on the dataset $\{(y_i,x_{ij}):i=1,\cdots,n\}$ in GLM have the same asymptotic form. Thus, (\[asRCPS\]) indicates that the asymptotic results of the RCPS in the GAM include those in the GLM. Note that in GLM($D=1$), we do not need to use the ridge penalty because the Hessian of the penalized log-likelihood of $\vec{b}$ is strictly convex.
[**Remark 2**]{} Claeskens et al. (2009) showed the asymptotic bias and variance of the penalized spline estimator in a regression model with $D=1$. They studied the asymptotics for penalized splines in the following two asymptotic scenarios: (a) the value $K_q$ appeared in their paper, less than 1, and (b) $K_q\geq 1$. In our setting, Assumption 5 guarantees case (a) and so Theorem \[mv\] can be seen as the general version of Theorem 2 (a) of Claeskens et al. (2009) with respect to the model and dimension of covariates. If $\lambda_{jn}^{-1}$ is equal or smaller than the maximum eigenvalue of $(Z_j^TZ_j)^{-1}\Delta_m^T \Delta_m(Z_j^TZ_j)^{-1}$, the asymptotics for the penalized spline estimator in the GAM will be demonstrable, such as in Theorem 2 (b) of Claeskens et al. (2009).
[**Remark 3**]{} From Theorem \[clt\], it is understood that $[\hat{\eta}_1(x_1)\ \cdots\ \hat{\eta}_D(x_D)]^T$ are asymptotically mutually independent. Wand (1999) showed the asymptotic independence of the kernel estimator in additive models. Hence the penalized spline estimator and the kernel estimator have the same asymptotic property. The asymptotic independence of the joint distribution of $[\hat{\eta}_1(x_1)\ \cdots\ \hat{\eta}_D(x_D)]^T$ gives some justification for Corollary \[inter\], in which the approximate confidence interval is constructed based on the asymptotic result of the marginal distribution of $\hat{\eta}_j(x_j)$.
[**Remark 4**]{} Clearly, the penalized spline estimator can also be obtained via the backfitting algorithm. The asymptotic normality of the backfitting estimator can be shown, although it is not discussed in this paper. In additive models, Yoshida and Naito (2012) derived the asymptotic normality of the penalized spline estimator obtained by the backfitting algorithm.
[**Remark 5**]{} Theorems in this section have been shown for the RCPS with common $(p,K_n,m)$ in each covariate. When we construct $\hat{\eta}_j(x_j)$ using different $(p,K_n,m)$ in each $j$, the asymptotic normality of the RCPS can also be shown. In other words, for $\hat{\eta}_j(x_j)$ with $(p_j,K_{jn},m_j)$ which satisfy $(p_j,K_{jn},m_j)\not=(p_i,K_{in},m_i)$ ($j\not= i$), Theorems \[mv\] and \[clt\] hold.
The mixed model representation
==============================
In this section, we discuss the penalized spline estimator in relation to mixed models. We consider model (\[model1\]) again with $\eta_j(x_j)$ approximated by a $p$th truncated spline model: $$\sum_{i=0}^{p} \beta_{i,j} x^{i} +\sum_{k=1}^{K_n-1} u_{k,j}(x-\kappa_k)^{p}_+,$$ where $(x)_+=\max\{x,0\}$. We assume that the random vector $\vec{u}_j=(u_{1,j}\ \cdots\ u_{K_n-1,j})^T$ has density $\vec{u}_j \sim N(\vec{0},\sigma_j^2 I)$ with $\sigma_j^2<\infty$, and $\vec{u}_i$ and $\vec{u}_j$ are independent for $i\not= j$. Hence, $\vec{u}=[\vec{u}_1^T\ \cdots\ \vec{u}_D^T]^T$ distributes $N(\vec{0},\Sigma_u)$, where $\Sigma_u={\rm diag}[\sigma_1^2I\ \cdots\ \sigma_D^2I]$. Let $$\begin{aligned}
X_j=
\left[ \begin{array}{cccc}
1 & x_{1j} & \cdots & x^p_{1j} \\
\vdots & \vdots & \ddots & \vdots \\
1 & x_{nj} & \cdots & x^p_{nj}
\end{array} \right], \quad
R_j=
\left[ \begin{array}{ccc}
(x_{1j}-\kappa_{1})^p_+ & \cdots & (x_{1j}-\kappa_{K_n-1})^p_+ \\
\vdots & \ddots & \vdots \\
(x_{nj}-\kappa_{1})^p_+ & \cdots & (x_{nj}-\kappa_{K_n-1})^p_+
\end{array} \right],\end{aligned}$$ $S_j=[X_j\ R_j]$, $S=[S_1\ \cdots\ S_D]$, $\vec{\beta}_j=(\beta_{0,j}\ \cdots\ \beta_{p,j})^T$, $\vec{\theta}_j=[\vec{\beta}_j^T\ \vec{u}_j^T]^T$ and $\vec{\theta}=[\vec{\theta}_1^T\ \cdots\ \vec{\theta}_D^T]^T$. The suggested joint density of $(\vec{y},\vec{u})$ can be written as $$\begin{aligned}
f(\vec{y},\vec{u}:\vec{\beta})
&=&
f(\vec{y}|\vec{u}:\vec{\beta})f(\vec{u})\\
&=&
\exp\left[\vec{y}^T (S\vec{\theta})-\vec{1}^T c(S\vec{\theta})+\vec{1}^T h(\vec{y})\right]\frac{1}{\sqrt{(2\pi)^{D} |\Sigma_u|}}\exp\left[-\frac{1}{2}\vec{u}^T \Sigma_u^{-1}\vec{u}\right]\\
&=&
\exp\left[\vec{y}^T (S\vec{\theta})-\vec{1}^T c(S\vec{\theta})+\vec{1}^T h(\vec{y})\right]\frac{1}{\sqrt{(2\pi)^D |\Sigma_u|}}\exp\left[-\frac{1}{2}\vec{\theta}^T \Theta \vec{\theta}\right],\end{aligned}$$ where $\Theta={\rm diag}[\Theta_1\ \cdots\ \Theta_D]$ and $\Theta_j={\rm diag}[O_{p+1}\ \sigma_j^{-2}I]$. As a convenient method of obtaining the estimator $\hat{\vec{\beta}}$ of $\vec{\beta}$ and the predictor $\hat{\vec{u}}$ of $\vec{u}$, the PQL is often used. In the PQL context, for given $\sigma_1^2,\cdots,\sigma_D^2$, $(\hat{\vec{\beta}},\hat{\vec{u}})$ is defined as the maximizer of $$\begin{aligned}
\ell(\vec{\beta},\vec{u})=
\frac{1}{n}\log f(\vec{y},\vec{u}|\vec{\beta})=
\frac{1}{n}\{\vec{y}^T (S\vec{\theta})-\vec{1}^T c(S\vec{\theta})\}-\frac{1}{2n}\vec{\theta}^T \Theta\vec{\theta}+C(\Sigma_u),\end{aligned}$$ where $C(\Sigma_u)$ is not dependent on $\vec{\beta}$ and $\vec{u}$. Let $S(x)=(1\ x\ \cdots\ x^p\ (x-\kappa_1)_+^p\ \cdots\ (x-\kappa_{K_n-1})_+^p)^T$. Then, the PQL fit of $\eta_j(x_j)$ is defined as $\hat{\eta}_{j,P}(x_j)=S(x_j)^T \hat{\vec{\theta}}_j$, where $\hat{\vec{\theta}}_j=[\hat{\vec{\beta}}_j^T\ \hat{\vec{u}}_j^T]^T$.
We show the asymptotic distribution of $[\hat{\eta}_{1,P}(x_1)\ \cdots\ \hat{\eta}_{D,P}(x_D)]^T$. In order to achieve the asymptotic normality of the PQL fits, we consider the equivalence result between the $B$-spline model and the truncated spline model. By the fundamental property of the $B$-spline function, there exists a $(K_n+p)$th invertible matrix $L_j$ such that $Z_j=S_jL_j$. Then we obtain $Z=SL$ and $S\vec{\theta}=Z\vec{b}$ where $L={\rm diag}[L_1\ \cdots\ L_D]$ and $\vec{b}=L^{-1}\vec{\theta}$. Furthermore, $\ell(\vec{\beta},\vec{u})$ can be rewritten as $$\begin{aligned}
\ell(\vec{\beta},\vec{u})=\ell(\vec{b})
=
\frac{1}{n}\{\vec{y}^T (Z\vec{b})-\vec{1}^T c(Z\vec{b})\}-\frac{K_n^{2p}}{2n} \vec{b}^T Q_{p+1}(\Sigma_u) \vec{b}+C(\Sigma_u), \label{loglike}\end{aligned}$$ where $Q_{p+1}(\Sigma_u)={\rm diag}[\sigma_1^{-2}\Delta_{p+1}^T\Delta_{p+1}\ \cdots\ \sigma_D^{-2}\Delta_{p+1}^T\Delta_{p+1}]$. Here we have used the fact that $\vec{\theta}^T\Theta\vec{\theta}=\vec{b}^T L^T\Theta L\vec{b}=K_n^{2p}\vec{b}^T Q_{p+1}(\Sigma_u)\vec{b}$. Claeskens et al. (2009) clarified the equality $\vec{\theta}_j^T \Theta_j\vec{\theta}_j=K_n^{2p}\vec{b}_j^T \Delta_{p+1}^T\Delta_{p+1}\vec{b}_j$. By showing the asymptotic distribution of the maximizer $[\hat{\vec{b}}_{1,P}^T\ \cdots\ \hat{\vec{b}}_{D,P}^T]^T$ of $\ell(\vec{b})$, we obtain the asymptotic normality of $[\hat{\eta}_{1,P}(x_1)\ \cdots\ \hat{\eta}_{D,P}(x_D)]^T$, where $$\begin{aligned}
\hat{\eta}_{j,P}(x_j)=S(\vec{x})^T \hat{\vec{\theta}}_j=S(\vec{x})^TL_j L_j^{-1} \hat{\vec{\theta}}_j=Z(\vec{x})^T \hat{\vec{b}}_{j,P}.\end{aligned}$$
\[cltpql\] Suppose there exists $\delta\geq 2$ such that $E[|Y_i-c^{\prime}(\eta(\vec{x}_i))|^{2+\delta}|\vec{X}_i=\vec{x}_i]<\infty$ and $\eta_1,\cdots,\eta_D\in C^{p+1}$. Under $K_n=O(n^{1/(2p+3)})$ and $\sigma_j^{-2}=O(n^\nu),\nu<2/(2p+3)$, for any fixed point $\vec{x}\in(0,1)^D$, as $n\rightarrow \infty$, $$\begin{aligned}
\sqrt{\frac{n}{K_n}}
\left[
\begin{array}{c}
\hat{\eta}_{1,P}(x_1)-\eta_1(x_1)-{\rm Bias}_1(x_1)\\
\vdots\\
\hat{\eta}_{D,P}(x_D)-\eta_D(x_D)-{\rm Bias}_D(x_D)
\end{array}
\right]
\stackrel{d}{\longrightarrow}N_D\left(
\vec{0}
,\Psi_P\right),\end{aligned}$$ where ${\rm Bias}_j(x_j)=b_{j,a}(x_j)+b_{j,\sigma}(x_j)$, $b_{j,a}(x_j)$ is that given in Proposition 1, $$\begin{aligned}
b_{j,\sigma}(x_j)=-\frac{K_n^{2p}}{n\sigma_j^{2}}\vec{B}(x_j)^T G_j^{-1}\Delta_{p+1}^{T}\Delta_{p+1}\vec{b}_{j0}=O_P(K_n^{p}/(n\sigma_j^2)),\end{aligned}$$ $\Psi_P=\operatorname{diag}[\psi_{1,P}(x_1)\ \cdots\ \psi_{D,P}(x_D)]$ and $$\begin{aligned}
&&{\small \psi_{j,P}(x_j)}\\
&&{\small =\lim_{n\rightarrow \infty}\frac{1}{K_n}\vec{B}(x_j)^T (G_j+K_n^{2p}(n\sigma_j^2)^{-1}\Delta_{p+1}^T\Delta_{p+1})^{-1}G_j(G_j+K_n^{2p}(n\sigma_j^2)^{-1}\Delta_{p+1}^T\Delta_{p+1})^{-1}\vec{B}(x_j).
}\end{aligned}$$
[**Remark 6**]{} If we use $m=p+1$, the results of Theorem \[clt\] are asymptotically equivalent to those of Theorem \[cltpql\] by replacing $\lambda_{jn}$ with $K_n^{2p}\sigma_j^{-2}(j=1,\cdots,D)$.
[**Remark 7**]{} It should be noted that the maximum likelihood method or the restricted maximum likelihood method can be utilized for estimating $\sigma_j^2 (j=1,\cdots,D)$ by using pseudo data. These methods and estimation algorithm based on the Fisher-scoring algorithm are detailed by Breslow and Clayton (1993) and by Ruppert et al. (2003).
Applications
============
We apply the approximate confidence interval of each covariate $\eta_j(x_j)$ to real datasets. In all examples, $(p,m)=(3,2)$ is adopted. The number of knots and the smoothing parameters are chosen via generalized cross-validation. As a pilot estimator of $\eta_j^{(4)}(x_j)$ in ${\rm Bias}_j(x_j)$, we utilize the 4th derivative of the RCPS with a 5th degree $B$-spline model. To see the behavior of $\hat{\eta}_j(x_j)$, the partial residual plots $$\hat{\eta}_j(x_{ij})+\hat{W}^{-1}(y_i-c^\prime(\hat{\eta}(\vec{x}_i))),$$ for each $x_{ij}(j=1,\cdots,D)$ are displayed (see Cook and Dabrera (1998) and Landwehr et al. (1984)).
Kyphosis data
-------------
The kyphosis data set had 81 rows and 4 columns, representing data of children who have had corrective spinal surgery. This data is available in the software R (package ${\tt rpart}$). For this data, the logistic model $$\begin{aligned}
Y_i&\sim&{\rm Bernoulli}\left(\frac{\exp[\eta_1(x_{i1})+\cdots+\eta_3(x_{i3})]}{1+\exp[\eta_1(x_{i1})+\cdots+\eta_3(x_{i3})]}\right),\ \ i=1,\cdots,81\end{aligned}$$ is assumed, where $Y_i$ is a factor with levels absent(0) or present(1) indicating whether a kyphosis was present (1) after the operation, $x_{i1}$ is the age in months, $x_{i2}$ is the number of vertebrae involved and $x_{i3}$ is the number of the first (topmost) vertebra operated on. We construct the RCPS with $\gamma_n=10^{-6}$ and the approximate confidence intervals for each $\eta_j(x_j)$.
In Fig. \[kyphosis\], for $j=1,2,3$, the RCPS $\hat{\eta}_j(x_j)$, the 99$\%$ approximate pointwise confidence interval $$\begin{aligned}
\left[\hat{\eta}_j(x_j)-\widehat{{\rm Bias}}_j(x_j)- 2.58\sqrt{\hat{\psi}_j(x_j)},\ \hat{\eta}_j(x_j)-\widehat{{\rm Bias}}_j(x_j)+ 2.58\sqrt{\hat{\psi}_j(x_j)}\right],\end{aligned}$$ and the partial residual are all illustrated. For comparison, $\eta_j\pm 2\times$(standard error): $$\begin{aligned}
\left[\hat{\eta}_j(x_j)- 2\sqrt{\hat{\psi}_j(x_j)},\ \hat{\eta}_j(x_j)+ 2\sqrt{\hat{\psi}_j(x_j)}\right],\ j=1,2,3\end{aligned}$$ are also superimposed. In all covariates, smooth intervals are obtained. Marx and Eilers (1998) illustrated the RCPS and $\eta_j\pm 2\times$(standard error) for the same dataset in Fig.4 of their paper. Our results and theirs are similar. However, our interval is wiggles a bit because the asymptotic bias is corrected in each covariate.
![Plots of the kyphosis data with the RCPS (dashed), the 99$\%$ approximate confidence interval(solid), $\hat{\eta}_j\pm 2\times$(standard error) (dot-dashed line) and the partial residuals. The left, middle and right panels are for $\eta_1(x_1)$, $\eta_2(x_2)$ and $\eta_3(x_3)$, respectively. \[kyphosis\]](kyphosisf1.eps "fig:"){width="45mm" height="45mm"} ![Plots of the kyphosis data with the RCPS (dashed), the 99$\%$ approximate confidence interval(solid), $\hat{\eta}_j\pm 2\times$(standard error) (dot-dashed line) and the partial residuals. The left, middle and right panels are for $\eta_1(x_1)$, $\eta_2(x_2)$ and $\eta_3(x_3)$, respectively. \[kyphosis\]](kyphosisf2.eps "fig:"){width="45mm" height="45mm"} ![Plots of the kyphosis data with the RCPS (dashed), the 99$\%$ approximate confidence interval(solid), $\hat{\eta}_j\pm 2\times$(standard error) (dot-dashed line) and the partial residuals. The left, middle and right panels are for $\eta_1(x_1)$, $\eta_2(x_2)$ and $\eta_3(x_3)$, respectively. \[kyphosis\]](kyphosisf3.eps "fig:"){width="45mm" height="45mm"}
Air Pollution and Mortality data
--------------------------------
This data set contained air pollution and mortality data for the city of Milan, Italy, over 3652 consecutive days (i.e., 10 consecutive years: 1st January, 1980 to 30th December, 1989). The original data is available on the web site of Ruppert et al. (2003). The relationship between the number of deaths in a day and some explanatory variables is modeled as follows $$\begin{aligned}
Y_i \sim {\rm Poisson}[\exp(\eta_1(x_{i1})+\cdots+\eta_5(x_{i5}))],\ \ i=1,\cdots,102,\end{aligned}$$ where $Y_i$ is the total number of deaths in a day, $x_{i1}$ is the number of days since 31st December, 1979, $x_{i2}$ the mean daily temperature in degrees celcius, $x_{i3}$ is the relative humidity, $x_{i4}$ is a measure of sulfur dioxide levels (SO2) in ambient air and $x_{i5}$ is the total amount of suspended particles (TSP) in ambient air. All of these have been measured on public holidays within the 3652 days, giving a sample size of $n=102$. We constructed the RCPS of $\eta_j(x_j)$ and the 99$\%$ approximate confidence intervals. In Fig. \[air\], the RCPS, the $99\%$ approximate confidence intervals, $\hat{\eta}_j\pm2\times$(standard error) and the partial residual for each $x_j$ are illustrated. We see that the effect of $\widehat{{\rm Bias}}_j(x_j)$ is somewhat large for all covariates.
![Plots of air pollution and mortality data with the RCPS, the 99$\%$ approximate confidence interval, $\hat{\eta}_j\pm 2\times$(standard error) and the partial residuals. \[air\]](airf1.eps "fig:"){width="45mm" height="40mm"} ![Plots of air pollution and mortality data with the RCPS, the 99$\%$ approximate confidence interval, $\hat{\eta}_j\pm 2\times$(standard error) and the partial residuals. \[air\]](airf2.eps "fig:"){width="45mm" height="40mm"} ![Plots of air pollution and mortality data with the RCPS, the 99$\%$ approximate confidence interval, $\hat{\eta}_j\pm 2\times$(standard error) and the partial residuals. \[air\]](airf3.eps "fig:"){width="45mm" height="40mm"}\
![Plots of air pollution and mortality data with the RCPS, the 99$\%$ approximate confidence interval, $\hat{\eta}_j\pm 2\times$(standard error) and the partial residuals. \[air\]](airf4.eps "fig:"){width="45mm" height="40mm"} ![Plots of air pollution and mortality data with the RCPS, the 99$\%$ approximate confidence interval, $\hat{\eta}_j\pm 2\times$(standard error) and the partial residuals. \[air\]](airf5.eps "fig:"){width="45mm" height="40mm"}
Simulation
==========
In this section, we validate Theorem \[clt\] numerically by simulation. The true natural parameter utilized in the simulation is defined as $\eta(\vec{x})=\eta_1(x_1)+\eta_2(x_2)+\eta_3(x_3)$, where $\eta_1(x_1)=\sin(2\pi x_1)$, $\eta_2(x_2)=2\cos(2\pi x_2)$ and $\eta_3(x_3)=\sin^2((\pi/2) x_3)$. The design points $(x_{i1},x_{i2},x_{i3})$ are created by $$\begin{aligned}
\left[
\begin{array}{c}
x_{i1}\\
x_{i2}\\
x_{i3}
\end{array}
\right]
=
\left[
\begin{array}{ccc}
(1+\rho+\rho^2)^{-1}&0&0\\
0&(1+2\rho)^{-1}&0\\
0&0&(1+\rho+\rho^2)^{-1}
\end{array}
\right]
\left[
\begin{array}{ccc}
1&\rho&\rho^2\\
\rho&1&\rho\\
\rho^2&\rho&1
\end{array}
\right]
\left[
\begin{array}{c}
z_{i1}\\
z_{i2}\\
z_{i3}
\end{array}
\right],\end{aligned}$$ where $z_{ij}(i=1,\cdots,n,j=1,2,3)$ are generated independently from $U(0,1)$, the uniform distribution on $[0,1]$. We prepared two types of the design, with (i) $\rho=0$ and (ii) $\rho=0.2$. Then, the true functions are corrected to satisfy $E[\eta_j(X_j)]=0$ in each (i) and (ii). The response $Y_i$ is generated from $$\begin{aligned}
Y_i\sim{\rm Bernoulli}\left(\frac{\exp[\eta_1(x_{i1})+\eta_2(x_{i2})+\eta_3(x_{i3})]}{1+\exp[\eta_1(x_{i1})+\eta_2(x_{i2})+\eta_3(x_{i3})]}\right),\ \ i=1,\cdots,n. \label{ber1}\end{aligned}$$
Our purpose is to compare the density of $N(0,1)$ and the kernel density estimate of the simulated $U_j (j=1,2,3)$, as well as the density of $N_2(\vec{0},I_2)$ and the kernel density estimate of the simulated $[U_1,U_2]^T$, $[U_1,U_3]^T$ and $[U_2,U_3]^T$ to validate Theorem \[clt\], where $$\begin{aligned}
\left[
\begin{array}{c}
U_1\\
U_2\\
U_3
\end{array}
\right]
=
\sqrt{\frac{n}{K_n}}
\left[
\begin{array}{c}
\displaystyle\frac{\hat{\eta}_1(x_1)-\eta_1(x_1)-\widehat{{\rm Bias}}_1(x_1)}{\hat{\psi}_1(x_1)}\\
\displaystyle\frac{\hat{\eta}_2(x_2)-\eta_2(x_2)-\widehat{{\rm Bias}}_2(x_2)}{\hat{\psi}_2(x_2)}\\
\displaystyle\frac{\hat{\eta}_3(x_3)-\eta_3(x_3)-\widehat{{\rm Bias}}_3(x_3)}{\hat{\psi}_3(x_3)}
\end{array}
\right]. \label{2norm}\end{aligned}$$ Here, $$\begin{aligned}
\hat{\psi}_j(x_j)=\frac{1}{K_n}\vec{B}(x_j)^T\hat{\Gamma}_j(\lambda_{jn})^{-1}\hat{\Gamma}_j(0)\hat{\Gamma}_j(\lambda_{jn})^{-1}\vec{B}(x_j),\end{aligned}$$ $\hat{\Gamma}_j(\lambda_{jn})=n^{-1}(Z_j^T\hat{W}Z_j+\lambda_{jn}\Delta_m^\prime \Delta_m)$. For $j=1,2,3$, $\widehat{{\rm Bias}}_j(x_j)$ is constructed using the same method as that in the previous section. The bandwidth discussed by Sheather and Jones (1991) is utilized for kernel density estimates. The simulation algorithm described as follows:
1. For $j=1,2,3$ and $i=1,\cdots,n$, generate $x_{ij}$ from (i) or (ii).
2. Generate the data $\{(y_{i},\vec{x}_{i})|i=1,\cdots,n\}$ from (\[ber1\]).
3. Calculate $\hat{\eta}_j(x_j) (j=1,2,3)$ at a fixed point $(x_1,x_2,x_3)=(0.5,0.5,0.5)$.
4. Calculate the values of (\[2norm\]).
5. Iterate from Step 2 to Step 4, 10000 times.
6. Draw the kernel density estimate of $U_1, U_2$ and $U_3$ and compare with the density of $N(0,1)$.
7. Draw the kernel density estimate of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$, and compare with the density of $N_2(\vec{0},I_2)$.
To construct $\hat{\eta}_j(x_j) (j=1,2,3)$, we utilize the cubic $B$-spline ($p=3$) and the second difference matrix ($m=2$). Furthermore $K_n=2\lceil n^{2/5}\rceil$, $\lambda_{1n}=0.1\sqrt{n/K_n}$, $\lambda_{2n}=0.01\sqrt{n/K_n}$ and $\lambda_{3n}=\sqrt{n/K_n}$ are used. The ridge parameter is chosen as $\gamma_n=10^{-4}$. The sample sizes are set at $n=100$ and $n=1000$.
In Fig. \[log.simu\], the density estimate of (\[2norm\]) with (i), and the densities of the normal distribution are shown. As the sample size increases, the asymptotic normality of the RCPS in Theorem \[clt\] can be observed numerically. We see from (1,1), (2,2) and (3,3) panels that the density estimate becomes close to 0 when $n=1000$. When $n=100$, a large correlation between $U_i$ and $U_j$ can be observed. However, as $n$ increases, the correlation becomes small. The results with the correlated design (ii) are drawn in Fig. \[log.simu.cor\]. The density estimate of $U_2$ appears to be far from $N(0,1)$, even when $n=1000$. However, we also find that $[U_i\ U_j]^T$ tends to become close to $N_2(\vec{0},I_2)$ as $n\rightarrow \infty$. We have confirmed that the density estimate with $Y_i\sim {\rm Poisson(\eta(\vec{x}_i))}$ tends to become close to the normal distribution as $n$ increases, though this is not shown in this paper. However, the speed of convergence of the density estimate with the Poisson model was somewhat slower than with the Bernoulli model.
![The density estimate of $U_i$, $[U_i\ U_j]^T$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and an uncorrelated design. For $i=1,2,3$, the $(i,i)$ panels are the density estimates of $U_i$ for $n=100$(dot-dashed) and $n=1000$(dashed), and the density of $N(0,1)$(solid). The (2,1), (3,1) and (3,2) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=100$ and the density of $N(\vec{0},I_2)$(solid). The (1,2), (1,3) and (2,3) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=1000$ and the density of $N(\vec{0},I_2)$(solid). In each panel, the contour lines of $N(\vec{0},I_2)$ are the same as that of the density estimate. \[log.simu\]](log1density.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]^T$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and an uncorrelated design. For $i=1,2,3$, the $(i,i)$ panels are the density estimates of $U_i$ for $n=100$(dot-dashed) and $n=1000$(dashed), and the density of $N(0,1)$(solid). The (2,1), (3,1) and (3,2) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=100$ and the density of $N(\vec{0},I_2)$(solid). The (1,2), (1,3) and (2,3) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=1000$ and the density of $N(\vec{0},I_2)$(solid). In each panel, the contour lines of $N(\vec{0},I_2)$ are the same as that of the density estimate. \[log.simu\]](log-1-1000.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]^T$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and an uncorrelated design. For $i=1,2,3$, the $(i,i)$ panels are the density estimates of $U_i$ for $n=100$(dot-dashed) and $n=1000$(dashed), and the density of $N(0,1)$(solid). The (2,1), (3,1) and (3,2) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=100$ and the density of $N(\vec{0},I_2)$(solid). The (1,2), (1,3) and (2,3) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=1000$ and the density of $N(\vec{0},I_2)$(solid). In each panel, the contour lines of $N(\vec{0},I_2)$ are the same as that of the density estimate. \[log.simu\]](log-2-1000.eps "fig:"){width="45mm" height="45mm"}\
![The density estimate of $U_i$, $[U_i\ U_j]^T$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and an uncorrelated design. For $i=1,2,3$, the $(i,i)$ panels are the density estimates of $U_i$ for $n=100$(dot-dashed) and $n=1000$(dashed), and the density of $N(0,1)$(solid). The (2,1), (3,1) and (3,2) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=100$ and the density of $N(\vec{0},I_2)$(solid). The (1,2), (1,3) and (2,3) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=1000$ and the density of $N(\vec{0},I_2)$(solid). In each panel, the contour lines of $N(\vec{0},I_2)$ are the same as that of the density estimate. \[log.simu\]](log-1-100.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]^T$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and an uncorrelated design. For $i=1,2,3$, the $(i,i)$ panels are the density estimates of $U_i$ for $n=100$(dot-dashed) and $n=1000$(dashed), and the density of $N(0,1)$(solid). The (2,1), (3,1) and (3,2) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=100$ and the density of $N(\vec{0},I_2)$(solid). The (1,2), (1,3) and (2,3) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=1000$ and the density of $N(\vec{0},I_2)$(solid). In each panel, the contour lines of $N(\vec{0},I_2)$ are the same as that of the density estimate. \[log.simu\]](log2density.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]^T$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and an uncorrelated design. For $i=1,2,3$, the $(i,i)$ panels are the density estimates of $U_i$ for $n=100$(dot-dashed) and $n=1000$(dashed), and the density of $N(0,1)$(solid). The (2,1), (3,1) and (3,2) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=100$ and the density of $N(\vec{0},I_2)$(solid). The (1,2), (1,3) and (2,3) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=1000$ and the density of $N(\vec{0},I_2)$(solid). In each panel, the contour lines of $N(\vec{0},I_2)$ are the same as that of the density estimate. \[log.simu\]](log-3-1000.eps "fig:"){width="45mm" height="45mm"}\
![The density estimate of $U_i$, $[U_i\ U_j]^T$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and an uncorrelated design. For $i=1,2,3$, the $(i,i)$ panels are the density estimates of $U_i$ for $n=100$(dot-dashed) and $n=1000$(dashed), and the density of $N(0,1)$(solid). The (2,1), (3,1) and (3,2) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=100$ and the density of $N(\vec{0},I_2)$(solid). The (1,2), (1,3) and (2,3) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=1000$ and the density of $N(\vec{0},I_2)$(solid). In each panel, the contour lines of $N(\vec{0},I_2)$ are the same as that of the density estimate. \[log.simu\]](log-2-100.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]^T$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and an uncorrelated design. For $i=1,2,3$, the $(i,i)$ panels are the density estimates of $U_i$ for $n=100$(dot-dashed) and $n=1000$(dashed), and the density of $N(0,1)$(solid). The (2,1), (3,1) and (3,2) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=100$ and the density of $N(\vec{0},I_2)$(solid). The (1,2), (1,3) and (2,3) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=1000$ and the density of $N(\vec{0},I_2)$(solid). In each panel, the contour lines of $N(\vec{0},I_2)$ are the same as that of the density estimate. \[log.simu\]](log-3-100.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]^T$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and an uncorrelated design. For $i=1,2,3$, the $(i,i)$ panels are the density estimates of $U_i$ for $n=100$(dot-dashed) and $n=1000$(dashed), and the density of $N(0,1)$(solid). The (2,1), (3,1) and (3,2) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=100$ and the density of $N(\vec{0},I_2)$(solid). The (1,2), (1,3) and (2,3) panels are the density estimates of $[U_1\ U_2]^T$, $[U_1\ U_3]^T$ and $[U_2\ U_3]^T$(dashed) for $n=1000$ and the density of $N(\vec{0},I_2)$(solid). In each panel, the contour lines of $N(\vec{0},I_2)$ are the same as that of the density estimate. \[log.simu\]](log3density.eps "fig:"){width="45mm" height="45mm"}\
![The density estimate of $U_i$, $[U_i\ U_j]$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and the correlated design. The description of each panel is the same as in Fig. \[log.simu\]. \[log.simu.cor\]](log1densitycor.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and the correlated design. The description of each panel is the same as in Fig. \[log.simu\]. \[log.simu.cor\]](log-1-1000cor.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and the correlated design. The description of each panel is the same as in Fig. \[log.simu\]. \[log.simu.cor\]](log-2-1000cor.eps "fig:"){width="45mm" height="45mm"}\
![The density estimate of $U_i$, $[U_i\ U_j]$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and the correlated design. The description of each panel is the same as in Fig. \[log.simu\]. \[log.simu.cor\]](log-1-100cor.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and the correlated design. The description of each panel is the same as in Fig. \[log.simu\]. \[log.simu.cor\]](log2densitycor.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and the correlated design. The description of each panel is the same as in Fig. \[log.simu\]. \[log.simu.cor\]](log-3-1000cor.eps "fig:"){width="45mm" height="45mm"}\
![The density estimate of $U_i$, $[U_i\ U_j]$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and the correlated design. The description of each panel is the same as in Fig. \[log.simu\]. \[log.simu.cor\]](log-2-100cor.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and the correlated design. The description of each panel is the same as in Fig. \[log.simu\]. \[log.simu.cor\]](log-3-100cor.eps "fig:"){width="45mm" height="45mm"} ![The density estimate of $U_i$, $[U_i\ U_j]$ and the density of $N(0,1)$ and $N(\vec{0},I_2)$ with the Bernoulli model and the correlated design. The description of each panel is the same as in Fig. \[log.simu\]. \[log.simu.cor\]](log3densitycor.eps "fig:"){width="45mm" height="45mm"}\
Discussion
==========
This paper showed the asymptotic normality of the penalized spline estimator in the GAM. The results of this paper generalize Theorem 1 of Kauermann et al. (2009) and Theorem 2 of Yoshida and Naito (2012). The main tools used to prove our Theorems were the spline approximation theories and the asymptotic properties of the band matrices. By applying their properties, the asymptotics for penalized splines in other models can be investigated for further study.
In spline smoothing, the determination of smoothing parameters is very important. Many researchers have addressed this problem by using grid search methods, such as Mallow’s $C_p$, cross-validation and generalized cross-validation. Since the computation time of a grid search is dramatically increased when $D>1$, more direct methods would be a useful area of study. It may be possible to discuss the selection of smoothing parameters based on the asymptotic properties in this paper.
In recent years, the so-called high-dimensional additive models characterized by $^^ ^^ n<D$" have been studied by many authors such as Meier et al. (2009), Huang et al. (2010) and Fan et al. (2011). These previous works are based on unpenalized $B$-spline estimators. Although it is beyond the scope of this paper, the asymptotics for penalized splines in high dimensional additive models would be interesting to explore.
Appendix {#appendix .unnumbered}
========
For a matrix $X_n=(X_{ij,n})_{ij}$, if $\displaystyle\max_{i,j}\{n^{\alpha}|X_{ij,n}|\}=O_{P}(1)(o_{P}(1))$, then it is written as $X_n=O_{P}(n^{-\alpha}\vec{1}\vec{1}^T)(o_{P}(n^{-\alpha}\vec{1}\vec{1}^T))$. When $X_n$ is vector, it is written as $X_n=O_{P}(n^{-\alpha}\vec{1})$. We define $W_0=\operatorname{diag}[c^{\prime\prime}(Z\vec{b}_0)]$, $G_{j,n}=n^{-1}Z_j^T W_0Z_j$, $G_{i,j,n}=n^{-1}Z_i^T W_0Z_j$ and $G_{j,i,n}=G_{i,j,n}^T$. In the sequel, we use $H_{j,n}=G_{jn}+(\lambda_{jn}/n)\Delta_m^\prime \Delta_m+(\gamma_n/n) I (i,j=1,\cdots,D,i\not=j)$.
We need 3 additional Lemmas as follows.
\[G1\] $G_{j,n}$, $G_{i,j,n}$ and $H_{j,n}$ satisfy $G_{j,n}=O_P(K_n^{-1}\vec{1}\vec{1}^T)$, $G_{i,j,n}=O_P(K_n^{-2}\vec{1}\vec{1}^{T})$ and $H_{j,n}^{-1}=O_P(K_n\vec{1}\vec{1}^T)$. Let $A=(a_{ij})_{ij}$ be $(K_n+p)\times (K_n+p)$ matrix. Assume that as $K_n\rightarrow \infty$, $A=O_P(K_n^\alpha\vec{1}\vec{1}^T)$. Then, under the Assumptions, $G_{jn}A=O_P(K_n^{\alpha-1}\vec{1}\vec{1}^T)$, $G_{i,j,n}A=O_P(K_n^{\alpha-2}\vec{1}\vec{1}^{T})$ and $H_{j,n}^{-1}A=O_P(K_n^{1+\alpha}\vec{1}\vec{1}^T)$.
\[G2\] Let $A_{D,n}$ be $\{D(K_n+p)\}\times \{D(K_n+p)\}$ matrix. Assume that as $K_n\rightarrow \infty$, $A_{D,n}=O_P(K_n^{\alpha}\vec{1}\vec{1}^T)$. Then, under the Assumptions, $A_{D,n}H(\vec{b}_0,\lambda_n,\gamma_n)^{-1}=O_P(K_n^{\alpha+1}\vec{1}\vec{1}^T)$
\[G3\] Under the assumption, for $j=1,\cdots,D$, $\Delta_m\vec{b}_{j0}=O(K_n^{-m}\vec{1})$.
Lemma \[G1\] can be proven by the properties of the integral of $B$-spline basis and the inverse of band matrices detailed in Claeskens et al. (2009) and Yoshida and Naito (2012). Then, Assumption 5 of this paper indicates that the case $K_q<1$ of Claeskens et al. (2009). The proof of Lemma \[G2\] is addressed in Yoshida and Naito (2012) by induction for $D$. Lemma \[G3\] can be shown from the derivative property of $B$-spline model: $s^{(m)}(x)=\vec{B}^{[p-m]}(x)^T K_n^m\Delta_m\vec{b}$. The above equality and Proposition 1 yield $\vec{B}^{[p-m]}(x)^T K_n^m\Delta_m\vec{b}_{j0}=\eta_j^{(m)}(x)(1+o(1))$. Since the asymptotic order of $\vec{B}^{[p-m]}(x)^T K_n^m\Delta_m\vec{b}_{j0}$ and each component of $K_n^m\Delta_m\vec{b}_{j0}$ are the same as $O(1)$, Lemma \[G3\] holds. The details are clarified in Section 2 of Claeskens et al. (2009).
\[ex\] Under the Assumptions, $$\begin{aligned}
\hat{\vec{b}}-\vec{b}_0=-H(\vec{b}_0,\lambda_n,\gamma_n)^{-1}G(\vec{b}_0,\lambda_n,\gamma_n)+o_P\left(\left\{\left(\frac{\lambda_nK_n^{1-m}}{n}\right)^2+\frac{K_n}{n}\right\}\vec{1}\right).\end{aligned}$$
proof of Lemma \[ex\] {#proof-of-lemma-ex .unnumbered}
---------------------
We use the Taylor expansion of $G(\hat{\vec{b}},\lambda_n,\gamma_n)$ around $\vec{b}_0$, giving $$\begin{aligned}
0&=&G(\hat{\vec{b}},\lambda_n,\gamma_n)\\
&=&G(\vec{b}_0,\lambda_n,\gamma_n)+H(\vec{b}_0,\lambda_n,\gamma_n)(\hat{\vec{b}}-\vec{b}_0)\\
&&+ \{H(\vec{b}_0+\Omega(\hat{\vec{b}}-\vec{b}_0),\lambda_n,\gamma_n)-H(\vec{b}_0,\lambda_n,\gamma_n)\}(\hat{\vec{b}}-\vec{b}_0),\end{aligned}$$ where $\Omega={\rm diag}[\omega_{1}\ \cdots\ \omega_{D(K_n+p)}]$ and $\omega_i\in (0,1)$. Therefore, $\hat{\vec{b}}-\vec{b}_0$ can be written as $$\begin{aligned}
\hat{\vec{b}}-\vec{b}_0
&=&
\left\{-H(\vec{b}_0,\lambda_n,\gamma_n)\right\}^{-1}G(\vec{b}_0,\lambda_n,\gamma_n)\nonumber\\
&&-
H(\vec{b}_0,\lambda_n,\gamma_n)^{-1}\{H(\vec{b}_0+\Omega(\hat{\vec{b}}-\vec{b}_0),\lambda_n,\gamma_n)-H(\vec{b}_0,\lambda_n,\gamma_n)\}(\hat{\vec{b}}-\vec{b}_0)\nonumber\\
&=&
\left\{-H(\vec{b}_0,\lambda_n,\gamma_n)\right\}^{-1}G(\vec{b}_0,\lambda_n,\gamma_n)\nonumber\\
&&+
H(\vec{b}_0,\lambda_n,\gamma_n)^{-1}\left(\frac{1}{n}Z^T\left\{W(\vec{b}_0+\Omega(\hat{\vec{b}}-\vec{b}_0))-W(\vec{b}_0) \right\}Z\right)(\hat{\vec{b}}-\vec{b}_0)
,\label{ex21}\end{aligned}$$ where $W(\vec{b})=\operatorname{diag}[c^{\prime\prime}(Z(\vec{x}_i)^T\vec{b})]$. Furthermore, for $i=1,\cdots,n$, the Taylor expansion yields $$\begin{aligned}
c^{\prime\prime}(Z(\vec{x}_i)^T\{\vec{b}_0+\Omega(\hat{\vec{b}}-\vec{b}_0)\})
&=&c^{\prime\prime}(Z(\vec{x}_i)^T\vec{b}_0)\\
&&+
c^{(3)}[Z(\vec{x}_i)^T\vec{b}_0+\theta_iZ(\vec{x}_i)^T\Omega(\hat{\vec{b}}-\vec{b}_0)]Z(\vec{x}_i)^T\Omega(\hat{\vec{b}}-\vec{b}_0),\end{aligned}$$ where $\theta_i\in(0,1)$. Hence from Proposition 1, we obtain $$\begin{aligned}
W(\vec{b}_0+\Omega(\hat{\vec{b}}-\vec{b}_0))-W(\vec{b}_0)
&=&\operatorname{diag}[c^{(3)}(Z(\vec{x}_i)^T\vec{b}_0+\theta_iZ(\vec{x}_i)^T\Omega(\hat{\vec{b}}-\vec{b}_0))Z(\vec{x}_i)^T\Omega(\hat{\vec{b}}-\vec{b}_0)]\\
&=&\operatorname{diag}[c^{(3)}(\eta(\vec{x}_i))Z(\vec{x}_i)^T\Omega(\hat{\vec{b}}-\vec{b}_0)](1+o(1))\\
&\equiv& R(\hat{\vec{b}}).\end{aligned}$$
For simplicity, we rewrite $G=G(\vec{b}_0,\lambda_n,\gamma_n)$ and $H=-H(\vec{b}_0,\lambda_n,\gamma_n)$. Then (\[ex21\]) can be written as $$\begin{aligned}
\hat{\vec{b}}-\vec{b}_0=H^{-1}G+H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)(\hat{\vec{b}}-\vec{b}_0). \label{ex22}\end{aligned}$$ We now prove $$\begin{aligned}
H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)(\hat{\vec{b}}-\vec{b}_0)
=
o_P\left(\left\{\left(\frac{\lambda_nK_n^{1-m}}{n}\right)^2+\frac{K_n}{n}\right\}\vec{1}\right). \label{purp1}\end{aligned}$$ From (\[ex22\]), the left hand side of (\[purp1\]) can be evaluated as $$\begin{aligned}
&&H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)(\hat{\vec{b}}-\vec{b}_0)\\
&&=
H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)H^{-1}G+\left\{H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)\right\}^2(\hat{\vec{b}}-\vec{b}_0).\end{aligned}$$ First we show the asymptotic order of $R(\hat{\vec{b}})$. The $i$th component of $R(\hat{\vec{b}})$ can be written by (\[ex22\]) as $$\begin{aligned}
&&c^{(3)}(\eta(\vec{x}_i))Z(\vec{x}_i)^T\Omega(\hat{\vec{b}}-\vec{b}_0)(1+o(1))\\
&&= c^{(3)}(\eta(\vec{x}_i))Z(\vec{x}_i)^T\Omega\left\{H^{-1}G+H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)(\hat{\vec{b}}-\vec{b}_0)\right\}(1+o(1)).\end{aligned}$$ By calculating the expectation and the square root of variance of each component of $G$, we obtain with Lemma \[G3\] $$G=O_P\left(\left\{\frac{\lambda_n}{nK_n^{m}}+\frac{1}{\sqrt{nK_n}}\right\}\vec{1}\right).$$ Therefore Lemma \[G2\] yields that for $\vec{z}\in[0,1]^D$, $$\begin{aligned}
Z(\vec{z})^T\Omega H^{-1}G=o_P\left(\left\{\frac{\lambda_nK_n^{1-m}}{n}+\sqrt{\frac{K_n}{n}}\right\}\right).\end{aligned}$$ Since $c^{(3)}(\eta(\vec{x}))$ is bounded near $\eta(\vec{x})$ for $\vec{x}\in[0,1]^D$, we have with tedious but easy calculation that $$\begin{aligned}
&&\left|c^{(3)}(\eta(\vec{x}_i))Z(\vec{x}_i)^T\Omega\left\{H^{-1}G+H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)(\hat{\vec{b}}-\vec{b}_0)\right\}\right|(1+o(1))\nonumber\\
&&\leq \sup_{\vec{z}\in[0,1]^D}\left|c^{(3)}(\eta(\vec{z}))Z(\vec{z})^T\Omega\left\{H^{-1}G+H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)(\hat{\vec{b}}-\vec{b}_0)\right\}\right|(1+o(1))\nonumber\\
&&=o_P\left(\frac{\lambda_nK_n^{1-m}}{n}+\sqrt{\frac{K_n}{n}}\right).\label{Rcal}\end{aligned}$$ Then Lemmas \[G1\] and \[G2\] and (\[Rcal\]) yield $$\begin{aligned}
H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)H^{-1}G
&=&
H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)O_P\left(\left\{\frac{\lambda_nK_n^{1-m}}{n}+\sqrt{\frac{K_n}{n}}\right\}\right)\\
&=&
H^{-1}o_P\left(\frac{1}{K_n}\left\{\frac{\lambda_nK_n^{1-m}}{n}+\sqrt{\frac{K_n}{n}}\right\}^2\right)\\
&=&
o_P\left(\left\{\frac{\lambda_nK_n^{1-m}}{n}+\sqrt{\frac{K_n}{n}}\right\}^2\vec{1}\right).\end{aligned}$$ Further we get with simple calculation $$\begin{aligned}
\left\{H^{-1}\left(\frac{1}{n}Z^T R(\hat{\vec{b}})Z\right)\right\}^2(\hat{\vec{b}}-\vec{b}_0)=o_P\left(\left\{\frac{\lambda_nK_n^{1-m}}{n}+\sqrt{\frac{K_n}{n}}\right\}^2\vec{1}\right).\end{aligned}$$ This implies (\[purp1\]) and completes the proof. $\Box$
proof of Proposition \[app\] {#proof-of-proposition-app .unnumbered}
----------------------------
Barrow and Smith (1978) showed that for $j=1,\cdots,D$, there exists $\vec{b}_j^*\in\mathbb{R}^{K_n+p}$ such that $$\begin{aligned}
\sup_{z\in(0,1)}\left|\eta_j(z)+b_{j,a}(z)-\vec{B}(z)^T\vec{b}_j^*\right|
=o(K_n^{-(p+1)}).\end{aligned}$$ Let $\eta_j^*(z)=\vec{B}(z)^T\vec{b}_j^*$, $\eta^*(\vec{x})=\sum_{j=1}^D \eta^*_j(x_j)$, $\eta_0(\vec{x})=\sum_{j=1}^D \eta_{j0}(x_j)$ and $b_a(\vec{x})=\sum_{j=1}^D b_{j,a}(x_j)$. We now prove that $$\begin{aligned}
\vec{b}_{j0}-\vec{b}_j^*=o(K_n^{-(p+1)}\vec{1}),\ \ j=1,\cdots,D. \label{pur}\end{aligned}$$ Since the asymptotic order of $\eta_{j0}(x_j)-\eta_j^*(x_j)$ and that of $\vec{b}_0-\vec{b}^*$ are the same, if (\[pur\]) is satisfied, we obtain for any $x_j\in(0,1)$, $|\eta_{j0}(x_j)-\eta_j^*(x_j)|=o(K_n^{-(p+1)})$ hence Proposition 1 holds.
From the definition of $\vec{b}_0$, we have $$\begin{aligned}
\frac{1}{n}\sum_{i=1}^n E\left[\left.\log \frac{f(Y_i|\vec{x}_i,\eta)}{f(Y_i|\vec{x}_i,\vec{b}_0)}\right|\vec{X}_n\right]
\leq
\frac{1}{n}\sum_{i=1}^n E\left[\left.\log \frac{f(Y_i|\vec{x}_i,\eta)}{f(Y_i|\vec{x}_i,\vec{b}^*)}\right|\vec{X}_n\right],\label{ineq}\end{aligned}$$ where $\vec{b}^*=((\vec{b}^*_1)^T\ \cdots\ (\vec{b}^*_D)^T)^T$. The Taylor expansion to $c(\eta^*(\vec{x}_i))$ around $\eta(\vec{x}_i)$ yields $$\begin{aligned}
&&\frac{1}{n}\sum_{i=1}^n E\left[\left.\log \frac{f(Y_i|\vec{x}_i,\eta)}{f(Y_i|\vec{x}_i,\vec{b}^*)}\right|\vec{X}_n\right]\\
&&=
\frac{1}{n}\sum_{i=1}^n\left[c^\prime(\eta(\vec{x}_i))\{\eta(\vec{x}_i)-\eta^*(\vec{x}_i)\}-\{c(\eta(\vec{x}_i))-c (\eta^*(\vec{x}_i))\}\right]\nonumber\\
&&=
\frac{1}{2n}\sum_{i=1}^n\left[\{\eta(\vec{x}_i)-\eta^*(\vec{x}_i)\}^2c^{\prime\prime} (\eta(\vec{x}_i))(1+o(1))\right]\nonumber\\
&&=
\frac{1}{2n}\sum_{i=1}^n\left[b_{a}(\vec{x}_i)^2c^{\prime\prime} (\eta(\vec{x}_i)(1+o(1))\right]\\
&&= O(K_n^{-2(p+1)}).\end{aligned}$$ Therefore we obtain $|\eta(\vec{x}_i)-\eta_{0}(\vec{x}_i)|= o(1)$, by which $$\begin{aligned}
\frac{1}{n}\sum_{i=1}^n E\left[\left.\log \frac{f(Y_i|\vec{x}_i,\eta)}{f(Y_i|\vec{x}_i,\vec{b}_0)}\right|\vec{X}_n\right]
&=&\frac{1}{n}\sum_{i=1}^n\left[\{\eta(\vec{x}_i)-\eta_{0}(\vec{x}_i)\}^2c^{\prime\prime} (\eta(\vec{x}_i))(1+o(1))\right]\nonumber\\
&=&
\frac{1}{n}(\vec{\eta}-Z\vec{b}_0)^T W (\vec{\eta}-Z\vec{b}_0), \label{b0def}\end{aligned}$$ where $W=\operatorname{diag}[c^{\prime\prime}(\eta(\vec{x}_i))(1+o(1))]$ and $\vec{\eta}=(\eta(\vec{x}_1)\ \cdots\ \eta(\vec{x}_n))^T$. It is easy to show that $\vec{b}_0$ satisfies $$\begin{aligned}
\frac{1}{n}Z^T W Z \vec{b}_0=\frac{1}{n}Z^T W \vec{\eta}\end{aligned}$$ since $\vec{b}_0$ is the minimizer of (\[b0def\]). Further, from the definition of $\vec{b}^*$, we have $$\begin{aligned}
\vec{\eta}=Z\vec{b}^* -\vec{B}_a +o(K_n^{-(p+1)}\vec{1}),\end{aligned}$$ where $\vec{B}_a=(b_a(\vec{x}_1)\ \cdots\ b_a(\vec{x}_n))^T$. Hence, we obtain $$\begin{aligned}
\frac{1}{n}Z^T W Z (\vec{b}_0-\vec{b}^*)=-\frac{1}{n}Z^T W \{\vec{B}_a+o(K_n^{-(p+1)}\vec{1})\}. \label{dist2}\end{aligned}$$ By noting $Z=[Z_1\ \cdots\ Z_D]$, the $k$th component of first $(K_n+p)$ block of $n^{-1}Z^T W\vec{B}_a$ can be calculated as $$\begin{aligned}
\left(n^{-1}Z_1^T W\vec{B}_a\right)_k
&=&
\frac{1}{n}\sum_{i=1}^n c^{\prime\prime}(\eta(\vec{x}_i))B_{-p+k}(x_{i1})b_a(\vec{x}_{i})(1+o(1))\nonumber\\
&=&
\sum_{j=1}^D \left[\frac{1}{n}\sum_{i=1}^n c^{\prime\prime}(\eta(\vec{x}_i))B_{-p+k}(x_{i1})b_{j,a}(x_{ij})\right](1+o(1))\nonumber\\
&=&
\sum_{j=1}^D \int_{[0,1]^D} c^{\prime\prime}(\eta(\vec{x}))B_{-p+k}(x_{1})b_{j,a}
(x_{j})dP(\vec{x})(1+o(1))\nonumber\\
&=&o(K_n^{-(p+2)}).\label{brsp}\end{aligned}$$ Here the last equality in (\[brsp\]) can be obtained by mimicking the proof of Lemma 10 of Agarwal and Studden (1980). Similarly since the row sum of $n^{-1}Z^T W$ has an order $O(K_n^{-1})$, we get (\[dist2\]) as $$\begin{aligned}
\frac{1}{n}Z^T W Z (\vec{b}_0-\vec{b}^*)=-\frac{1}{n}Z^T W \{\vec{B}_a+o(K_n^{-(p+1)}\vec{1})\}=o(K_n^{-(p+2)}\vec{1}).\end{aligned}$$ From Lemma \[G1\], we have $n^{-1}Z^T W Z=O(K_n^{-1}\vec{1}\vec{1}^T)$. Therefore $$\vec{b}_0-\vec{b}^*=o(K_n^{-(p+1)}\vec{1})$$ and (\[pur\]) can be proven. $\Box$
To complete the proof of Theorem 1, first we will obtain (A) the asymptotic form of $\hat{\vec{b}}_D-\vec{b}_{D0}$. And then (B) we will derive the asymptotic form of $E[\hat{\vec{b}}_D-\vec{b}_{D0}|\vec{X}_n]$ and $V[\hat{\vec{b}}_D|\vec{X}_n]$. Similar argument will be applied to $\hat{\vec{b}}_j-\vec{b}_{j0} (j=1,\cdots,D-1).$
proof of Theorem \[mv\] {#proof-of-theorem-mv .unnumbered}
-----------------------
First we aim to show (A). From Lemma \[ex\], we obtain $$\begin{aligned}
\hat{\vec{b}}-\vec{b}_0
&=&
\left\{-H(\vec{b}_0,\lambda_n,\gamma_n)\right\}^{-1}G(\vec{b}_0,\lambda_n,\gamma_n)+R_n(\hat{\vec{b}})\\
&=&
\left(\frac{1}{n}Z^T W_0 Z+ \frac{1}{n}Q_m(\lambda_n)+\frac{\gamma_n}{n} I\right)^{-1}G(\vec{b}_0,\lambda_n,\gamma_n)+ R_n(\hat{\vec{b}}), \end{aligned}$$ where $$R_n(\hat{\vec{b}})=o_P\left(\left\{\left(\frac{\lambda_nK_n^{1-m}}{n}\right)^2+\frac{K_n}{n}\right\}\vec{1}\right).$$ Let $M_D=n^{-1}(Z^T W_0 Z+Q_m(\lambda_n)+\gamma_n I)$ and let $\Lambda_{j,\gamma}=n^{-1}(Z_j^T W_{0} Z_j+\lambda_{jn}\Delta_m^\prime \Delta_m +\gamma_n I)$. Then, $M_{D}$ can be written by using $M_{D-1}$ as $$\begin{aligned}
M_{D}
=
\left[
\begin{array}{cccc}
\Lambda_{1,\gamma}&G_{1,2,n}&\cdots&G_{1,D,n}\\
G_{2,1,n}&\Lambda_{2,\gamma}&\cdots&G_{2,D,n}\\
\vdots&\vdots &\ddots&\vdots\\
G_{D,1,n}&\cdots&\cdots&\Lambda_{D,\gamma}
\end{array}
\right]
=
\left[
\begin{array}{cc}
M_{D-1}&R^T\\
R&\Lambda_{D,\gamma}
\end{array}
\right],\label{Mmat1}\end{aligned}$$ where $R=[G_{D,1,n}\ \cdots\ G_{D,D-1,n}]$. From the result of partitioned matrix (see, Horn and Johnson (1985)), we have $$\begin{aligned}
M_D^{-1}=\left[
\begin{array}{cc}
M^{-1}_{D-1}+M^{-1}_{D-1}R^T V^{-1}RM^{-1}_{D-1}&-M^{-1}_{D-1}R^T V^{-1}\\
-V^{-1}RM^{-1}_{D-1}&V^{-1}
\end{array}
\right],\end{aligned}$$ where $V=\Lambda_{D,\gamma}-RM^{-1}_{D-1}R^T$. Let $G_{(-D)}(\vec{b}_0,\lambda_n,\gamma_n)$ and $G_D(\vec{b}_0,\lambda_n,\gamma_n)$ be the first $(D-1)(K_n+p)$th subvector and last $(K_n+p)$th subvector of $G(\vec{b}_0,\lambda_n,\gamma_n)$. Then, $$\begin{aligned}
\hat{\vec{b}}-\vec{b}_0&\equiv&\left[
\begin{array}{cc}
\hat{\vec{b}}_{(-D)}-\vec{b}_{(-D)0}\\
\hat{\vec{b}}_D-\vec{b}_{D0}
\end{array}
\right]\\
&=&
\left[
\begin{array}{cc}
M^{-1}_{D-1}+M^{-1}_{D-1}R^T V^{-1}RM^{-1}_{D-1}&-M^{-1}_{D-1}R^T V^{-1}\\
-V^{-1}RM^{-1}_{D-1}&V^{-1}
\end{array}
\right]
\left[
\begin{array}{c}
G_{(-D)}(\vec{b}_0,\lambda_n,\gamma_n)\\
G_D(\vec{b}_0,\lambda_n,\gamma_n)
\end{array}
\right]\\
&&\quad + R_n(\hat{\vec{b}}),\end{aligned}$$ from which we have $$\begin{aligned}
\hat{\vec{b}}_D-\vec{b}_{0,D}
&=&
V^{-1}G_D(\vec{b}_0,\lambda_n,\gamma_n)-V^{-1}RM^{-1}_{D-1}G_{(-D)}(\vec{b}_0,\lambda_n,\gamma_n)+R_n\\
&=&
\Lambda_{D,\gamma}^{-1}G_D(\vec{b}_0,\lambda_n,\gamma_n) + v_n(\vec{b}_0)+R_{D,n}(\hat{\vec{b}}),\end{aligned}$$ where $$\begin{aligned}
v_n(\vec{b}_0)
&=&
-V^{-1}RM^{-1}_{D-1}G_{(-D)}(\vec{b}_0,\lambda_n,\gamma_n)+\left\{V^{-1}-\Lambda_{D,\gamma}^{-1}\right\}G_D(\vec{b}_0,\lambda_n,\gamma_n)\end{aligned}$$ and $R_{D,n}(\hat{\vec{b}})$ is last $(K_n+p)$th subvector of $R_n(\hat{\vec{b}})$.
In following, we shall start to show (B). The expectation of $\hat{\vec{b}}_D-\vec{b}_{0,D}$ can be written as $$\begin{aligned}
E[\hat{\vec{b}}_D-\vec{b}_{0,D}|\vec{X}_n]
&=&
\Lambda_{D,\gamma}^{-1}E[G_D(\vec{b}_0,\lambda_n,\gamma_n)|\vec{X}_n] + E[v_n(\vec{b}_0)|\vec{X}_n]+E[R_{D,n}(\hat{\vec{b}})|\vec{X}_n].\end{aligned}$$ First, $E[R_{D,n}(\hat{\vec{b}})|\vec{X}_n]=o_P(\lambda_nK_n^{1-m}n^{-1})$ is satisfied. In the sequel, because $$\begin{aligned}
E[G(\vec{b}_0,0,0)|\vec{X}_n]
&=&
\frac{1}{n}\sum_{i=1}^n E\left[\left.\frac{\partial}{\partial \vec{b}} \log f(Y_i|x_i,\vec{b}_0)\right|\vec{X}_n\right]=0,\end{aligned}$$ we have with Lemma \[G3\] $$\begin{aligned}
E[G(\vec{b}_0,\lambda_n,\gamma_n)|\vec{X}_n]&=&E[G(\vec{b}_0,0,0)|\vec{X}_n]-\frac{1}{n}Q_m(\lambda_n)\vec{b}_0-\frac{\gamma_n}{n} \vec{b}_0\\
&=&
O_P\left(\frac{\lambda_nK_n^{-m}}{n}\vec{1}\right).\end{aligned}$$ From Lemmas \[G1\] and \[G2\], on the other hand, we obtain $$\begin{aligned}
V^{-1}-\Lambda_{D,\gamma,n}^{-1}
&=&
\Lambda_{D,\gamma}^{-1}(I-RM^{-1}_{D-1}R^T \Lambda_{D,\gamma}^{-1})^{-1}-\Lambda_{D,\gamma}^{-1}\\
&=&
\Lambda_{D,\gamma}^{-1}RM^{-1}_{D-1}R^T \Lambda_{D,\gamma}^{-1}(I-RM^{-1}_{D-1}R^T \Lambda_{D,\gamma}^{-1})^{-1}\\
&=&
O_P\left(K_n\frac{1}{K_n^2}K_n\frac{1}{K_n^2}K_n\vec{1}\vec{1}^T\right)\\
&=&
O_P\left(K_n^{-1}\vec{1}\vec{1}^T\right).\end{aligned}$$ Therefore we have with straightforward calculation $$\begin{aligned}
&&E[v_n(\vec{b}_0)|\vec{X}_n]\\
&&=
-V^{-1}RM^{-1}_{D-1}E[G_{(-D)}(\vec{b}_0,\lambda_n,\gamma_n)]+\left\{V^{-1}-\Lambda_{D,\gamma,n}^{-1}\right\}E[G_D(\vec{b}_0,\lambda_n,\gamma_n)]\\
&&=
O_P\left(K_n\frac{1}{K_n^2}K_n\frac{\lambda_nK_n^{-m}}{n}\vec{1}\right)
+
O_P\left(\frac{\lambda_nK_n^{-(m+1)}}{n}\vec{1}\right)\\
&&=
O_P\left(\frac{\lambda_nK_n^{-m}}{n}\vec{1}\right).\end{aligned}$$
Above calculations are combined into $$\begin{aligned}
E[\hat{\vec{b}}_D-\vec{b}_{D0}|\vec{X}_n]
&=&
-\frac{\lambda_{Dn}}{n}\Lambda_{D,\gamma}^{-1}\Delta_m^\prime \Delta_m\vec{b}_{D0}-\frac{\gamma_n}{n}\Lambda_{D,\gamma,n}^{-1}\vec{b}_{D0}+o_P\left(\frac{\lambda_nK_n^{1-m}}{n}\vec{1}\right)\\
&=&
-\frac{\lambda_{Dn}}{n}\Gamma_D(\lambda_{Dn})^{-1}\Delta_m^\prime \Delta_m\vec{b}_{D0}+o_P\left(\frac{\lambda_nK_n^{1-m}}{n}\vec{1}\right).\end{aligned}$$ Here we have used the fact $\Lambda_{D,\gamma}^{-1}=\Gamma_D(\lambda_{Dn})^{-1}(I+o_P(\vec{1}\vec{1}^T))$ and $(\gamma_n/n)\Gamma_D(\lambda_{Dn})^{-1}\vec{b}_{D0}=o(\lambda_nK_n^{1-m}n^{-1}\vec{1})$. Hence, we finally obtain $$\begin{aligned}
E[\hat{\eta}_{D,\gamma}(x_D)-\eta_{D,0}(x)|\vec{X}_n]
&=&
E[\vec{B}(x_D)^T (\hat{\vec{b}}_D-\vec{b}_{D0})|\vec{X}_n]\\
&=&
b_{D,\lambda}(x_D)+o_P(\lambda_nK_n^{1-m}n^{-1}).\end{aligned}$$ This implies that the first assertion of Theorem \[mv\]. The variance of $\hat{\eta}_{D,\gamma}(x_D)=\vec{B}(x_D)^T\hat{\vec{b}}_D$ can be written as $$\begin{aligned}
V[\hat{\eta}_{D,\gamma}(x_D)|\vec{X}_n]
&=&
\vec{B}(x_D)^T\Lambda_{D,\gamma}^{-1}V[G_D(\vec{b}_0,\lambda_n,\gamma_n)|\vec{X}_n]\Lambda_{D,\gamma}^{-1}\vec{B}(x_D)(1+o_P(1)).\end{aligned}$$ since it is easy to find that the conditional variance of $v_n(\vec{b}_0)$ can be shown to be $o_P(K_n/n)$. By noting $$\begin{aligned}
V[G_D(\vec{b}_0,\lambda_n,\gamma_n)|\vec{X}_n]
=
\frac{1}{n^2}Z_D^T V[\vec{y}|\vec{X}_n]Z_D=\frac{1}{n^2}Z_D^T W Z_D=\frac{1}{n}G_{D} +o_P((K_n/n)^{-1}\vec{1}\vec{1}^T),\end{aligned}$$ we have the second assertion as $$\begin{aligned}
&&V[\hat{\eta}_{D,\gamma}(x_D)|\vec{X}_n]\\
&&=
\frac{1}{n}\vec{B}(x_D)^T\Lambda_{D,\gamma}^{-1}G_{D} \Lambda_{D,\gamma}^{-1}\vec{B}(x_D)(1+o_P(1))\\
&&=
\frac{1}{n}\vec{B}(x_D)^T\Gamma_D(\lambda_{Dn})^{-1} \Gamma_D(0) \Gamma_D(\lambda_{Dn})^{-1}\vec{B}(x_D)(1+o_P(1)).\end{aligned}$$ Also it is easily confirmed by straightforward calculation with Lemma \[G1\] that for $j\not= k$, $$\begin{aligned}
Cov(\hat{\eta}_j(x_j),\hat{\eta}_{k}(x_k))
=
\frac{1}{n}\vec{B}(x_j)^T\Lambda_{j,\gamma}^{-1}\left(\frac{1}{n}Z_j^T W Z_k\right)\Lambda_{k,\gamma}^{-1}\vec{B}(x_k)(1+o_P(1))
=O_P(n^{-1}),\end{aligned}$$ this completely the proof. $\Box$
proof of Theorem \[cltpql\] {#proof-of-theorem-cltpql .unnumbered}
---------------------------
Let $\hat{\vec{b}}_P=[\hat{\vec{b}}_{1,P}^T\ \cdots\ \hat{\vec{b}}_{D,P}^T]^T$ and let $\tilde{\vec{b}}=(\tilde{\vec{b}}_1^T\ \cdots\ \tilde{\vec{b}}_D^T)^T$ be the maximizer of $$\begin{aligned}
G(\vec{b},\Sigma_u,\tilde{\gamma}_n)=\frac{1}{n}\{\vec{y}^T (Z\vec{b})-\vec{1}^T c(Z\vec{b})\}+\frac{1}{n}\vec{1}^T h(\vec{y})-\frac{1}{2n} \vec{b}^T Q_{p+1}(\Sigma_u) \vec{b}-\frac{\tilde{\gamma}_n}{2n}\vec{b}^T \vec{b}\end{aligned}$$ with respect to $(\vec{b}_1^T\ \cdots\ \vec{b}_D^T)^T$, where $\tilde{\gamma}_n=o(K_n^{p-1}/\sigma_j^2)$ and let $\tilde{\eta}_j(x_j)=\vec{B}(x_j)^T\tilde{\vec{b}}_j (j=1,\cdots,D)$. Then the asymptotic normality of $[\tilde{\eta}_1(x_1)\ \cdots\ \tilde{\eta}_D(x_D)]^T$ can be obtained by the same manner to the proof of Theorem \[clt\] with $m=p+1$. Similar to Lemma \[ex\], $\hat{\vec{b}}_P$ can be written as $$\begin{aligned}
\hat{\vec{b}}_P-\tilde{\vec{b}}&=&(Z^T\tilde{W}Z+Q_{p+1}(\Sigma_u)+\tilde{\gamma}_nI)^{-1}G(\hat{\vec{b}}_P,\Sigma_u,\tilde{\gamma}_n)+r_n\\
&=&
\tilde{\gamma}_n(Z^T\hat{W}Z+Q_{p+1}(\Sigma_u)+\tilde{\gamma}_nI)^{-1}\hat{\vec{b}}_P+r_n\end{aligned}$$ where $\hat{W}=\operatorname{diag}[c^{\prime\prime}(Z\hat{\vec{b}}_P)]$ and $r_n=o_P(\sqrt{K_n/n}\vec{1})$ is the remainder. Then Lemma \[G2\] yields $\hat{\vec{b}}_P-\tilde{\vec{b}}=O_P(\tilde{\gamma}_nK_nn^{-1}\vec{1})=o_P(\sqrt{K_n/n}\vec{1})$, by which $\hat{\eta}_{j,P}(x_j)-\tilde{\eta}_j(x_j)=o_P(\sqrt{K_n/n}) (j=1,\cdots,D)$. This leads to Theorem \[cltpql\]. $\Box$
(2002). Some theory for penalized spline generalized additive models. . [**103**]{}, 455–470.
(1980). Asymptotic integrated mean square error using least squares and bias minimizing splines. , 1307–25.
(1978). Asymptotic properties of best $L_2[0,1]$ approximation by splines with variable knots. , 293–304.
(1993). Approximate inference in generalized linear mixed models. , 9–25.
(2009). Asymptotic properties of penalized spline estimators. , 529–44.
(1998). Partial residual plots in generalized linear models. , 730–739.
(2001).
(1996). Flexible smoothing with [*B*]{}-splines and penalties. , 89-121.(with Discussion)
(2011). Nonparametric independence screening in sparse ultra-high-dimensional additive models. , 544–557.
(2005). Theory for penalized spline regression. , 105–18.
(1990). . London Chapman & Hall.
(2009). . Springer-Verlag.
(1985). . Cambridge University Press.
(2004). Nonparametric estimation of an additive model with a link function. ,2412–2443.
(2010). Variable selection in nonparametric additive models. , 2282–2313.
(2009). Some asymptotic results on generalized penalized spline smoothing. B [**71**]{}, 487–503.
(1984). Graphical methods for assessing logistic regression models. . [**79**]{}, 61–71.
(1999). Inference in generalized additive mixed models by using smoothing splines. B [**61**]{}, 381–400.
(2000). Efficient estimation of generalized additive nonparametric regression models. . [**16**]{} 502–523.
(1989). Chapman & Hall, London.
(1998). Direct generalized additive modeling with penalized likelihood. & [*Data Anal.*]{} [**28**]{}, 193–209.
(2009). High-dimensional additive modeling. . 3779–3821.
(2000). Asymptotic properties of Backfitting estimators. , 166–79.
(1986). A statistical perspective on ill-posed inverse problems. , 505–27.(with discussion)
(2003). .
(1991). A reliable data-based bandwidth selection method for kernel density estimation. 53, 683–690.
(1999). A central limit theorem for local polynomial backfitting estimators. , 57–65.
(2011). On the asymptotics of penalized spline smoothing. , 1–17.
(2006). . CRC: Chapman & Hall.
(2012). Asymptotics for penalized additive $B$-spline regression. , in press.
(2008). Smooth backfitting in generalized additive models. ,228–260.
[^1]: E-mail: [[email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the influence of a DC bias voltage $V$ on quantum interference corrections to the measured differential conductance in metallic mesoscopic wires and rings. The amplitude of both universal conductance fluctuations (UCF) and Aharonov-Bohm effect (ABE) is enhanced several times for voltages larger than the Thouless energy. The enhancement persists even in the presence of inelastic electron-electron scattering up to $V\sim 1$mV. For larger voltages electron-phonon collisions lead to the amplitude decaying as a power law for the UCF and exponentially for the ABE. We obtain good agreement of the experimental data with a model which takes into account the decrease of the electron phase-coherence length due to electron-electron and electron-phonon scattering.'
author:
- 'C. Terrier, D. Babić[^1], C. Strunk[^2], T. Nussbaumer and C. Schönenberger'
title: 'The Amplitude of Non-Equilibrium Quantum Interference in Metallic Mesoscopic Systems'
---
If the size of a conductor is of order of the electron phase-coherence length $L_\varphi (T)$ the wave character of electrons leads to experimentally observable quantum interference contributions to the conductance $G$. These are the aperiodic and periodic fluctuations $\delta G$ of the conductance around its average value. In the former case, called universal conductance fluctuations, the interference pattern is formed as a superposition of contributions from a continuous range of possible interference paths. Averaged over either impurity configuration, magnetic field or energy, the root-mean-square (rms) conductance fluctuation $\delta G_{rms}$ is of the order $\sim e^2/h$ [@leealt]. The periodic fluctuations, known as the Aharonov-Bohm effect [@ab], are observed if the interference is imposed by geometry of a sample, most commonly in the form of a loop. If the loop is threaded by a magnetic flux $\phi$, $\delta G (\phi)$ exhibits periodic oscillations with a period $h/e$, and $\delta G_{rms}$ is again $\sim e^2/h$. Both the UCF [@skocpol] and the ABE [@umbach] are suppressed by ensemble averaging if independent phase-coherent units are connected in series. The above behaviour is characteristic of the linear-response regime, [*i.e.*]{} of $eV \ll k_B T$ or $eV \ll E_c$, where $E_c$ is the coherence energy (Thouless energy) determined by the size of the conductor. In non-equilibrium ($eV \gg E_c, k_B T$) the fluctuations are expected to be remarkably different, as predicted theoretically by Larkin and Khmel’nitskiĭ (LK) [@lk]. If inelastic processes can be neglected the rms fluctuation $\delta g_{rms}$ of the [*differential conductance*]{} $g$ increases with $V$ according to $\delta g_{rms} \sim (e^2/h) \sqrt{V/V_c}$, where $V_c=E_c/e$. This, at first sight surprising result, can be understood as follows. At $V \gg V_c$ the relevant energy range for the transport subdivides into $N=V/V_c$ uncorrelated energy intervals, each contributing to the fluctuations of the current by an amount $\sim (e^2/h) V_c$. Incoherent superposition of these contributions leads to $\delta g_{rms}$ being $N^{1/2}$ times larger than $e^2/h$. Inelastic scattering at large voltages destroys quantum interference. The enhancement of $\delta g_{rms}(V)$ is thereby suppressed and $\delta g_{rms}$ eventually decreases with increasing voltage [@lk]. Experimental studies of both the UCF [@webb1; @kaplan; @ralph; @schafer] and ABE [@webb1; @ct; @haus] under non-equilibrium conditions have been done, but did not result in a satisfactory understanding of $\delta g (V)$. Moreover, [@ralph] reported on a voltage-independent $\delta g_{rms}$, [@webb1] on a decrease of $\delta g_{rms}(V)$ for $V > V_c$, whereas in [@schafer; @ct; @haus] it was found that $\delta g_{rms}(V)$ shows a non-monotonic behaviour. We also note that explanations [@ct; @haus] of the non-monotonic behaviour of $\delta g(V)$, observed in the present work as well, were at a qualitative level.
In this paper we report on the non-equilibrium UCF and ABE in diffusive gold samples measured over a wide voltage range of $V \gg k_B T / e, V_c$ (strongly $V$-dominated energy averaging) that covers both the low-voltage $(V/V_c)^{1/2}$ enhancement of $\delta g_{rms}$ and its suppression at large voltages. Emphasis is put on the decay of $\delta g_{rms}^{UCF}(V)$ and $\delta g_{rms}^{ABE}(V)$. It is shown that $\delta g_{rms}(V)$ decays as a power law for the UCF and exponentially for the ABE. A quantitative comparison between simple models and the experiments allows to extract the voltage-dependent phase-coherence length $L_{\varphi}(V)$ and to discuss the nature of inelastic scattering processes out of equilibrium.
The samples were produced by electron-beam lithography and evaporation of 99.99 % pure gold. The substrate was silicon covered with 400 nm of SiO$_2$. Three 20 nm thick samples of different planar geometries were prepared [@samples]: (1) for the ABE measurements a ring of average diameter 1 $\mu$m and line width 0.09 $\mu$m (sample S$_{ABE}$), with a resistance of 25.8 $\Omega$ at 0.3 K; (2) for the UCF measurements a $L_w = 1.5$ $\mu$m long and 0.13 $\mu$m wide wire (sample S$_{UCF}$), with a resistance of 18.4 $\Omega$ at 0.3 K; (3) for the weak-localisation (WL) measurements a 98 $\mu$m long and 0.17 $\mu$m wide wire (sample S$_{WL}$). Sample S$_{WL} $ was made relatively long in order to suppress the UCF. All samples were produced under identical conditions (the same source of gold and the same evaporation parameters). The diffusion constants $D$ are consequently very similar: 116 cm$^2$/s for samples S$_{UCF}$ and S$_{WL}$, and 110 cm$^2$/s for sample S$_{ABE}$. The measurements were carried out in a $^3$He cryostat with cryogenic rf-filtering, using a low-frequency ($37$Hz) lock-in technique to measure the differential conductance. Typical voltage resolution was $\sim 0.3$ nV. Non-equilibrium UCF and ABE were measured at $T=300$ mK by superimposing a comparatively large DC bias voltage $V_{DC}$ to the small excitation voltage $V_{AC}$. The following hierarchy of energies was always maintained: $eV_{DC} \gg k_B T \geq eV_{AC} > E_c$. Ensemble averaging was achieved by measuring the UCF and ABE over a magnetic-field range of $\sim 2$T, largely exceeding the correlation field. Typical sweep rates were 0.1 mT/s. Fig.1a displays raw data of a differential-conductance measurement on sample S$_{ABE}$, taken at $V_{DC}=0.5$ mV. Both types of the fluctuations are present. In Fig.1b we show the Fourier transform of the same data, exhibiting a well-defined peak at the position corresponding to $\phi=h/e$. From the magnetoconductance traces we have extracted $\delta g_{rms}^{ABE}$ and $\delta g_{rms}^{UCF}$ for sample S$_{ABE}$, as shown in Fig.2 (discussed in more detail later), and $\delta g_{rms}^{UCF}$ for sample S$_{UCF}$. The values of $\delta g_{rms}^{UCF}$ have been determined as standard deviations of the whole differential-conductance traces. The amplitudes $\delta g_{rms}^{ABE}$ have been calculated by averaging the periodic part of the fluctuations over $\sim 500$ periods, after the coarse background has been removed numerically.
The WL was measured on sample S$_{WL}$ in the linear-response regime ($V_{DC}=0$) and in the temperature range 0.3 - 10 K. By fitting a one-dimensional WL expression [@altshulerwl] to the low-field magnetoconductance data we obtained the linear-response $L_\varphi$ as a function of temperature, as shown in the inset to Fig.3. Similarly to published work [@mohanty; @gougam], $L_\varphi (T)$ saturates at low temperatures (below 1 K), which has been attributed to impurity mediated inelastic electron-electron scattering [@glazman]. At high temperatures $L_\varphi$ follows a $L_\varphi \propto T^{-q}$ dependence with $q \approx 1.2$ that suggests dephasing by electron-phonon interaction. At $T= 300$ mK, $L_\varphi = 3$ $\mu$m and $\tau_\varphi = \sqrt{L_\varphi^2 / D} = 0.77$ ns, implying that in the linear-response regime the diffusion of electrons is coherent throughout the samples $S_{UCF}$ and $S_{ABE}$. $L_\varphi$ has to be compared further with the effective length of the sample $L_{c0}$ which depends on the particular geometry and the coherence phenomenon investigated. For the UCF in sample S$_{UCF}$, $L_{c0} = L_w = 1.5$ $\mu$m. In the case of ABE in sample S$_{ABE}$, $L_{c0} = C_r \approx 3.14$ $\mu$m, where $C_r$ is the circumference of the ring. Finally, for the UCF in sample S$_{ABE}$, $L_{c0}$ is calculated as the distances ring - voltage contacts plus $C_r /2$, which gives $\approx 2.2$ $\mu$m. Thus, $L_\varphi$(300 mK) is always bigger or very similar to $L_{c0}$, so that the linear-response coherence energy is set by the time of electron diffusion through the sample according to $E_{c0} = eV_{c0} = \hbar D / L_{c0}^2$.
With increasing temperature or applied voltage the coherence length decreases, and once it becomes shorter than $L_{c0}$ the interference contributions are suppressed. The effect of temperature was demonstrated by Milliken [*et al.*]{} [@milliken]. Their method was to determine $L_\varphi (T)$ from the WL and to use this to describe the decay of [*equilibrium*]{} $\delta G$, which turned out to be a power law for the UCF and exponential for the ABE. Our approach is complementary: we keep the bath temperature constant and investigate the voltage dependence of the [*non-equilibrium*]{} $\delta g$. While Milliken [*et al.*]{} observed a monotonically-decreasing $\delta G (T)$, in our case $\delta g (V)$ is a non-monotonic function, as shown in Fig.2. For small $V_{DC}$ (but still much larger than $V_{c0}$), $\delta g_{rms}$ is enhanced by $V_{DC}$. At higher voltages $\delta g_{rms}$ decreases with $V_{DC}$, [*faster for the*]{} ABE [*than for the*]{} UCF, which is in qualitative agreement with the results of Milliken [*et al.*]{}
Let us first turn to the details of non-equilibrium UCF. For convenience we denote the phase-coherence length at a finite bias voltage by $L_{\varphi}(V)$, while $L_{\varphi}(T)$ referes to the linear-response value. If $V_{DC} \gg V_{c0}, k_B T/e$ and $L_{\varphi} (V_{DC}) > L_{c0}$, the LK enhancement of $\delta g_{rms}$ holds and one expects that $\delta g_{rms} \propto \sqrt{V_{DC}/V_{c0}}$. Once $L_{\varphi} (V_{DC})$ becomes substantially smaller than $L_{c0}$ inelastic processes start to be important for most of the electrons, and the sample effectively subdivides into $L_{c0}/L_ {\varphi}$ uncorrelated phase-coherent sections in series. The LK enhancement in a single section is still valid if the voltage drop $V^{sec}_{DC} = V_{DC} (L_{\varphi}/L_{c0})$ over a section exceeds the corresponding coherence voltage $V_c^{sec} = V_{c0} (L_{c0}/L_{\varphi})^2$. Following the LK arguments, $\delta g_{rms}$ corresponding to a single section is proportional to $(V^{sec}_{DC} / V_c^{sec})^{1/2} = [ (V_{DC}/V_{c0}) (L_{\varphi}/L_{c0})^3]^{1/2}$. Incoherent addition of these contributions of sections in series leads to the [*differential conductance*]{} being suppressed by a factor $(L_{\varphi}/L_{c0})^{1/2}$, and hence $$%
\delta g_{rms}^{UCF} = A_{UCF} \frac{e^2}{h} \sqrt{ \frac{V_{DC}}{V_{c0}} }
\left( \frac{L_{\varphi} (V_{DC})}{ L_{c0}} \right) ^2 \; ,
%$$ where $A_{UCF}$ is a constant prefactor. This expression is different from that resulting from the linear-response ensemble averaging, and valid if $V_{DC} > V_{c0}$ and $L_{\varphi} < L_{c0}$ [@remark]. It enables an unambiguous determination of $L_{\varphi} (V_{DC})$. Since the form of $L_{\varphi} (V_{DC})$ is not known [*a priori*]{}, we cannot determine $A_{UCF}$ and $L_{\varphi} (V_{DC})$ simultaneously from a direct fit to the data. To avoid this problem we fix $A_{UCF}$ by setting $L_\varphi= L_{c0}$ for the first measured point which satisfies $V_{DC} \gg V_{c0}$, and then simply extract $L_{\varphi} (V_{DC})$ for all the other points from the measured data. For the two sets of UCF data, [*i.e.*]{} for samples S$_{UCF}$ and S$_{ABE}$, we have obtained excellent agreement in $L_{\varphi}(V_{DC})$, as we show in Fig.3 by open (S$_{ABE}$) and full (S$_{UCF}$) squares. Normalised to the corresponding values of $L_{c0}$ both results collapse onto a single curve. The phase-coherence length deduced from our non-equilibrium experiment $L_{\varphi}(V)$ has a qualitatively similar functional form as $L_{\varphi}(T)$ obtained from the WL measurements (apart from the saturation of $L_\varphi (T)$, since saturation of $L_\varphi (V)$ is not expected for $eV \gg k_B T$ [@glazman]). Two regimes with different power-law dependences are discernible. In the region of increasing $\delta g_{rms}(V_{DC})$ there is a power law $L_{\varphi}(V_{DC}) \propto V_{DC}^{-s}$, where $s = 0.19 \pm 0.02$. For decreasing $\delta g_{rms} (V_{DC})$, [*i.e.*]{} where $L_{\varphi}(V_{DC})$ is considerably smaller than $L_{c0}$, the dependence changes to $L_{\varphi} \propto V_{DC}^{-p}$ with $p = 0.57 \pm 0.03$. Inserting thus found $L_{\varphi}$ back into Eq.1 results in the dotted line in Fig.2.
We now discuss the non-equilibrium ABE. Here the LK theory cannot be directly applied to extract $\delta g_{rms}$. Subdivision of a ring into phase-coherent sections makes no sense in this case, as only those electrons which stay coherent over the whole length $C_r$ contribute to the interference. However, since the same physics governs the UCF and ABE, $\delta g_{rms}^{ABE}$ is expected to be $\propto \sqrt{V_{DC}/V_{c0}}$ in the regime where inelastic scattering is absent. This is indeed the case, as shown in Fig.2. The form of the suppression of $\delta g_{rms}^{ABE}$ can be inferred from a theoretical analysis of DiVincenzo and Kane [@dk]. They have shown that inelastic processes influence the ABE in two ways. First, the fluctuations are suppressed as $\exp( - \beta C_r / L_{\varphi})$, where $\beta$ is of order unity in the range $1 < C_r / L_{\varphi} < 3$ and smaller in the case of a true exponential decay occurring in the asymptotic limit $C_r / L_{\varphi} \rightarrow \infty$. Second, the proper form of $E_c$ in the presence of inelastic processes is given by $E_c^{in} = \hbar D / C_r^{2-\alpha} L_{\varphi}^\alpha = E_{c0} (C_r/L_{\varphi})^\alpha$ with $\alpha \approx 1.3$. This renormalisation of the coherence energy is a consequence of the statistics for those electrons that diffuse around a ring without being scattered inelastically [@dk]. Roughly, the probability distribution for an electron to contribute to the ABE has a maximum at a value of the electron traversal time $\sim D /C_r L_{\varphi}$ [@dk]. Combining the above arguments we can write for $L_{\varphi} < C_r$: $$%
\delta g_{rms}^{ABE} = A_{ABE} \frac{e^2}{h} \sqrt{ \frac{V_{DC}}{V_{c0}} }
\left( \frac{L_{\varphi}(V_{DC})}{C_r} \right) ^{0.65}
e^ { - \beta C_r / L_{\varphi}(V_{DC}) } \; .
%$$ Good agreement of the ABE data with both Eq.2 and the UCF data is obtained for $\beta \approx 1.1$, as we show in Fig.2 by the solid curve and in Fig.3 by crosses.
Below we discuss how the observed power law dependences in $L_\varphi (V_{DC})$ can be linked to microscopic scattering processes. From Figs.2,3 we can distinguish two regimes (denoted by [I]{} and [II]{} in Fig.3): [(I)]{} $\delta g_{rms} (V_{DC})$ increases with $V_{DC}$ and $L_{\varphi}(V_{DC}) \propto V_{DC}^{-0.19}$, and [(II)]{}, $\delta g_{rms}(V_{DC})$ decreases with $V_{DC}$ while $L_{\varphi}(V_{DC}) \propto V_{DC}^{-0.57}$.
The phase-coherence length is determined by the dephasing rate $\tau_{\varphi}^{-1}$ via $L_{\varphi}=\sqrt{D\tau_{\varphi}}$. $\tau_{\varphi}^{-1}$ can be expressed as an average of the inelastic scattering rate $\tau_{in}^{-1}(\epsilon)$, ($\epsilon$ is the energy exchanged in the interaction) of electrons over the accessible energy range given by the width of the electron distribution function $f$ [@gougam]. If the electron-electron scattering is dominant, the scattering rate is energy dependent as described by the kernel function $K(\epsilon)=\kappa_{\eta}\epsilon^{-\eta}$. The energy interval $E$ accessible for the scattering is set by $k_BT$ in equilibrium and $eV_{DC}$ in non-equilibrium. At high energies, where single scattering events determine $\tau_\varphi$, a simple argument [@gougam] leads to $\tau_{\varphi,ee}^{-1}(E) \propto
E^{1/\eta}$ [@saturation]. Two choices for $\eta$ are currently under debate: $\eta=3/2$ for the disorder enhanced Coulomb interaction [@gougam] and $\eta=2$ for magnetic impurity mediated interaction [@glazman; @pierre]. This results in $L_{\varphi,ee}^{-1}(V_{DC}) \propto
V_{DC}^{1/3}$ for $\eta = 3/2$, and $L_{\varphi,ee}^{-1}(V_{DC}) \propto V_{DC}^{1/4}$ for $\eta = 2$. Hence the kernel exponent $\eta=2$ appears more consistent with our experimental observation of $L_\varphi^{-1} \propto
V_{DC}^{0.19}$. This is supported by the observed saturation of the equilibrium $\tau_\varphi$ for temperatures below .
We emphasise that - independently of the precise interaction mechanism - the rather weak decrease of $L_{\varphi,ee}$ with increasing $V$ is insufficient to suppress the LK enhancement of $\delta g_{rms}$. This is seen as follows. We recall that the condition for the appearance of the LK enhancement in a phase-coherent section of the wire is $V_{DC}^{sec} > V_c^{sec}$, which can be written as $V_{DC}/V_{c0} > (L_{c0}/L_{\varphi})^3$. Since $L_{\varphi,ee}
\propto V_{DC}^{-1/4}$, the right-hand side of the above inequality increases more slowly with $V_{DC}$ than the left-hand side, and the condition for the enhancement is maintained over the whole voltage range of increasing $\delta g_{rms} (V_{DC})$.
While the regime [I]{} is determined by electron-electron scattering and a non-equilibrium electron distribution function $f$, the stronger decay of $L_{\varphi}(V)$ in the regime [II]{} is caused by electron-phonon scattering. It can be very well described by equilibrium properties if we assume local thermal equilibrium with an elevated electron temperature $T_{el} (V_{DC})$. Electron-phonon scattering leads to $\tau_{\varphi,ep} \propto T_e^{-m}$, where $m$ equals 3 for very clean samples, 4 for strongly disordered samples, and 2 - 3 for samples of intermediate degree of disorder [@schmid]. Since $T_e$ is related to the applied voltage by $T_e \propto V^{2/(2+m)}$ [@anderson], one obtains $\tau_{\varphi,ep} \propto V^{-2m/(2+m)}$ and $L_{\varphi,ep} \propto V^{-m/(2+m)}$. Our result $L_{\varphi}(V_{DC}) \propto V_{DC}^{-0.57 \pm 0.03}$ gives $m = 2.6 \pm 0.3$, which agrees well with the value $2.5$ obtained in a noise measurement on similar samples [@henny3]. Moreover, our WL result ($L_{\varphi,ep}(T)\propto T^{-m/2}$) is $m = 2.4$, in excellent agreement with the above values. We also note that even though $T_e$ rises up to $\sim 10 - 15$ K at high $V_{DC}$ the energy averaging given by the LK approach remains essentially unaffected because $T_e$ increases more slowly than the voltage itself. Although our phenomenological analysis leads to a consistent interpretation of our data, further theoretical work is needed to fully understand the dephasing in the case of strong non-equilibrium.
In conclusion, we have measured the universal conductance fluctuations and Aharonov-Bohm effect in mesoscopic gold samples under highly non-equilibrium conditions of large applied bias voltages. The rms fluctuation $\delta g_{rms}$ of the differential conductance initially increases with voltage $\propto \sqrt{V/V_c}$, which demonstrates the validity of the theoretical prediction by Larkin and Khmel’nitskiĭ. This increase is followed by a decay of $\delta g_{rms}$ at higher voltages, where inelastic scattering becomes substantial. The amplitude decays as a power law for the universal conductance fluctuations and exponentially in the case of the Aharonov-Bohm effect. The decrease of the phase-coherence length with increasing voltage is in good agreement with the inferred $L_{\varphi}(V)$ dependences for electron-electron and electron-phonon scattering. In particular, the electron-electron collisions are not sufficient to suppress the enhancement mechanism. This work was supported by the Swiss National Science Foundation.
[99]{}
.
.
.
.
\[\]; .
.
.
.
.
.
.
The results for these samples are representative of the data obtained by measurements carried out on altogether eleven samples. We did not find any significant sample dependence of the results.
\[\].
.
.
.
.
Strictly speaking, the conditions are $V_{DC}\gg V_{c0}$ and $L_{\varphi}\ll L_{c0}$.
.
The saturation of $\tau_\varphi (T)$ at very low energies is not covered by the simple argument of Ref. 15.
in
in .
PhD Thesis, University of Basel, 1998.
[^1]: Present address: Department of Physics, Faculty of Science, University of Zagreb, Croatia.
[^2]: Present address: Institute for Experimental and Applied Physics, University of Regensburg, Germany.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Pair wave functions (PWF) which are eigenfunctions of the reduced density $2$-matrix are considered for a homogeneous Bose liquid. With the Bogoliubov principle of the correlation weakening it is proved that the distribution of the “dissociated" pair states over momenta is exactly the product of the single-particle distribution functions. Thus, the “dissociated" pair states are naturally classified as condensate-condensate, condensate-supracondensate and supracondensate-supracondensate ones provided the Bose-Einstein condensate exists. The condensate-condensate as well as condensate-supracondensate PWF are expressed in terms of the averages of products of the creation and destruction Bose operators. This leads to the simple interpretation of the anomalous averages as the “scattering parts" of the condensate-condensate and condensate-supracondensate PWF. It is shown that in contrast to the Fermi liquid, the appearance of the anomalous averages for the Bose liquid does not necessarily mean that there exist bound states of pairs of particles. The PWF in the Hartree-Fock-Bogoliubov (HFB) approach are found. Given the density of the condensate is not zero, there are no bound pair states in the HFB scheme. The expansion of the pair correlation function in the set of PWF is very useful in order to take into account both short-range and long-range spatial correlations. Applications (possible and already realized) of the formalism developed are discussed.'
address: 'Bogoliubov Laboratory of Theoretical Physics, Joint Institute for Nuclear Research, 141980, Dubna, Moscow region, Russia'
author:
- 'A. Yu. Cherny [@email]'
date: 'February 8, 2000'
title: '**Pair Wave Functions in a Bose Liquid**'
---
\#1\#2 \#1 \#1[\_[\#1]{}]{} \#1[\_[\#1]{}\^]{} =-3mm
[2]{}
Introduction {#1}
============
It is well-known that for the homogeneous Fermi liquid the appearance of anomalous averages $\langle{\hat\psi}^{\dagger}(x_{1})
{\hat\psi}^{\dagger}(x_{2})\rangle$ implies that there exists a macroscopic number of bound pairs of particles with zero momentum (here ${\hat\psi}^{\dagger}(x)$ and ${\hat\psi}(x)$ are the field operators). In the case of the pair condensate like this, the properties of the system change radically. In particular, this leads to the superfluidity of the Fermi liquid (superconductivity in an electron system, superfluidity in $^{3}$He).
For a homogeneous Bose liquid, the superfluidity is usually associated with the macroscopic number of particles themselves in the zero-momentum state (the Bose-Einstein condensate). In the Bogoliubov theory for a weakly imperfect Bose gas, repulsive interaction between particles depletes the condensate, while makes it thermodynamically stable. In this case the anomalous averages $\langle{\hat\vartheta}^{\dagger}(\r{1})
{\hat\vartheta}^{\dagger}(\r{2})\rangle$ play a crucial role also, where ${\hat\vartheta}^{\dagger}({\bf r})$ is the field operator creating a supracondensate particle at a point ${\bf r}$, see Eq. (\[thetakrr\]).
Parallels between the Bogoliubov transformations for boson [@bog47] and fermion [@bog58] systems gave birth to pair theories being special cases of the HFB theory (see, e.g. Ref. [@kobe] and references therein). In the framework of the HFB approach, the gap in the single-particle excitation energy appears [@hohmartin] that may be interpreted, by analogy with the theory of superconductivity, as an evidence for the bound states of pairs of particles. Below we show that in the presence of a single-particle condensate, there are no bound pair states of bosons in the HFB theory. Thus, it has been verified in terms of PWF that the gap is unphysical, which has been stressed in many papers [@hohmartin; @hugpines].
The concept of wave functions for a group of $m$ particles in medium can be rigorously introduced with eigenfunctions of the $m$-th-order reduced density matrix, or the $m$-matrix. Indeed, the system of $m$ particles is a subsystem of that of $N$ particles. So, its state is not pure even in the situation when the system as a whole has a wave function. In general a subsystem is specified by the density matrix (see, e.g. Ref. [@Landau]). The $m$-matrix is of use when we have a noncoherent superposition of the $m$-particle wave functions. Löwdin and Shull used wave functions like these for one and two particles to describe electron states in atoms, and called them the “natural orbitals" and “natural geminals", respectively (see Ref. [@coleman]). The concept of PWF for fermions has been used by Shafroth and his coworkers in papers on the theory of superconductivity (for a review, see Ref. [@blatt]). In what follows, we shall use the term PWF for the wave functions of particle pairs, using the notation of Bogoliubov who made the most clear presentation of the concept of PWF for fermions [@bogquasi].
This paper concerns PWF for the Bose liquid below the temperature of the condensation, $T_{c}$. Analysis is carried out in general form as far as it is possible, with a special emphasis on the bound pair states and that part of PWF which is responsible for the correlations of particles at short distances. Consideration of these correlations is of special interest when investigating concrete boson systems for which the potential of two-particle interaction $\Phi(r)$ is, as a rule, strongly singular, or, in other words, of the hard-core type (i.e. potential goes to infinity like $1/r^{m}$ ($m>3$) as $r\rightarrow 0$, and, consequently, there is no Fourier transform of it). For example, the well-known Lennard-Jones potential is of this form. In this case standard approximations based on the weak-coupling pertrubation theory lead to divergencies. For example, the statistical average of the interaction energy per particle $$\varepsilon_{int}=\frac{1}{N}\Bigl\langle\frac{1}{2}\sum_{i\not=j} \Phi(|{\bf
r}_{i}-{\bf r}_{j}|)\Bigr\rangle=\frac{n}{2}\int\d^{3}r\,\Phi(r)g(r)
\label{uaver}$$ (here $g(r)$ is the pair distribution function, $n$ is the density of the particles) is infinite in the simplest Hartree approximation corresponding to $g(r)=1$. A reason for the divergency is obvious as the short-range spatial correlations of two particles are not properly taken into consideration. In order to escape the mentioned trouble a real potential is usually replaced by various effective potentials for which the Fourier transforms exist (for the Bose case see the $t$-matrix approach [@brueck], the pseudopotential with positive scattering length, or the hard-sphere model, [@lee] and the method by Beliaev [@bel]). This replacement implies that the effective potential contains all “information" on the short-range correlations of particles. A range of validity of this replacement for the many-boson systems is discussed in the textbook [@popov] and, for an inhomogeneous gas, in the recent paper [@burnett] (see the discussion at the end of Sec. \[3a\]). In the lowest order in density $n$ of the Bose gas such effective-interaction procedures are reduced to the Bogoliubov model with the effective potential \[for example, in the pseudopotential method the Fourier transform of the “bare" potential is replaced by $\Phi(k)=4\pi a/m$, where $a$ is the scattering length obtained from the two-body Schrödinger equation\]. However, this approach is found to be thermodynamically inconsistent [@CSPRE], which can manifest itself in various unphysical results. For example, one can obtain no depletion of the Bose condensate within the pseudopotential approach, if one tries to apply a self-consistent treatment [@huang]. Besides, we are not able to obtain the correct value of the interaction energy (\[uaver\]) by direct calculations using $g(r)$ [@CSEPJB]. Thus, it looks interesting and promising to realize an alternative variational scheme that is self-consistent from the very beginning, as the short-range correlations are properly taken into account for all the terms in the expansion of $g(r)$ (\[gr\]) in the set of PWF. Thus, PWF are “regular channels" which can be used to cancel the divergencies in a self-consistent manner (see discussion in Secs. \[3a\] and \[4\]). The first results derived in Refs. [@CSPRE; @CSJPS; @CSPRE2] demonstrate that the formalism of the PWF can actually be employed as a basis for the self-consistent variational treatment for a Bose gas with the strongly singular interaction potential.
General properties of the eigenfunctions and eigenvalues of the correlation function (\[f2\]) (differing from the $2$-matrix only by a norm) are interesting in itself when an eigenvalue becomes a macroscopic quantity (off-diagonal long-range order [@yang]). With the Bogoliubov principle of correlation weakening we shall show that eigenvalues of the correlation function (\[f2\]) which belong to continuous spectrum (“dissociated" pair states) are expressed as the product of the single-particle momentum distributions. Thus, the “dissociated" pair states are naturally classified as condensate-condensate, condensate-supracondensate and supracondensate-supracondensate ones provided the Bose-Einstein condensate exists. The condensate-condensate as well as condensate-supracondensate PWF are expressed in terms of the averages of products of the creation and destruction Bose operators. This leads to the simple interpretation of the anomalous averages as the “scattering parts" of the condensate-condensate and condensate-supracondensate PWF (see Sec. \[3\]).
In the next section, the PWF are introduced for boson systems by analogy with Bogoliubov’s paper [@bogquasi], and the continuous spectrum of the correlation function (\[f2\]) is calculated. In Sec. \[3\] we consider the general structure of the $2$-matrix below $T_{c}$, the PWF in an explicit form and their physical interpretation. In Sec. \[3a\] the PWF are evaluated in HFB approach as well as in the Bogoliubov model, and some applications of the developed formalism are considered. Main results are summarized in Sec. \[4\].
The concept of particle pair states for bosons {#2}
==============================================
Let us consider a homogeneous system of $N$ bosons, and, for simplicity, let the spin of the particles be equal to zero. The Hamiltonian is assumed to be invariant with respect to translations and the transformation of particle momentum ${\bf p}_{i}\rightarrow -{\bf p}_{i}$. The latter implies that the Hamiltonian does not change when the canonical transformation ${\hat a}_{{\bf
k}}\to {\hat a}_{-{\bf k}},\ {\hat a}^{\dagger}_{{\bf k}}\to {\hat
a}^{\dagger}_{-{\bf k}}$ is performed. In particular, this condition is satisfied for pairwise interactions with a potential $\Phi({\bf r})=
\Phi(r)$. A state of the system is determined by a density matrix $$\begin{aligned}
\rho(\rp{1},\ldots,\rp{N};\r{1},\ldots,\r{N})=\sum_{n}
&&w_{n}\Psi_{n}^{*}(\r{1},\ldots,\r{N})
\nonumber \\
&&\times\Psi_{n}(\rp{1},\ldots,\rp{N}),
\label{rho}\end{aligned}$$ where $\sum_{n}w_{n}=1$, $w_{n}\geq 0$, and $\Psi_{n}$ are orthonormal system of functions being symmetric with respect to any permutations of particles.
The $2$-matrix is defined as $$\begin{aligned}
\rho_{2}&&(\rp{1},\rp{2};\r{1},\r{2})=\int_{V}\d^{3}r_{3}\cdots \d^{3}r_{N}
\nonumber \\
&&\times\rho(\rp{1},\rp{2},\r{3},\ldots,\r{N};\r{1},\r{2},\r{3},\ldots,
\r{N}),
\label{rho2def}\end{aligned}$$ and can be expressed in terms of the field operators ${\hat\psi}(\r{})$ and ${\hat\psi}^{\dagger}(\r{})$ (see, e.g. Ref. [@boglec]) $$\rho_{2}(\rp{1},\rp{2};\r{1},\r{2})=\frac{1}{N(N-1)}
\langle{\hat\psi}^{\dagger}(\r{1}){\hat\psi}^{\dagger}(\r{2})
{\hat\psi}(\rp{2}){\hat\psi}(\rp1{})\rangle
\label{rho2vtor}$$ Here $\langle\cdots\rangle=\mbox{Tr}\,(\cdots\hat\rho)$ stands for the statistical average over the state $\hat\rho$. The density matrix $\rho_{2}$ is normalized, namely $$\mbox{Tr}\,\rho_{2}=\int_{V}\d^{3}r_{1}\d^{3}r_{2}\;
\rho_{2}(\r{1},\r{2};\r{1},\r{2})=1,$$ therefore, any matrix element (\[rho2vtor\]) has asymptotic behaviour as $1/V^{2}$, in the thermodynamic limit $V\rightarrow\infty$, $N/V= n = \mbox{const}$. Therefore, it is more convenient to work with the pair correlation function $$F_{2}(\r{1},\r{2};\rp{1},\rp{2})=
\langle{\hat\psi}^{\dagger}(\r{1}){\hat\psi}^{\dagger}(\r{2})
{\hat\psi}(\rp{2}){\hat\psi}(\rp1{})\rangle.
\label{f2}$$
The boundary conditions for $F_2$ [@Note3a] follow from the principle of the correlation weakening at macroscopical separations [@bogquasi]: $$\begin{aligned}
\langle {\hat\psi}^{\dagger}({\bf r}_1) {\hat\psi}^{\dagger}({\bf r}_2)
&&{\hat\psi}({\bf r}_2^{\prime}){\hat\psi} ({\bf r}_1^{\prime})\rangle\to
\nonumber \\
&&\langle{\hat\psi}^{\dagger}({\bf r}_1) {\hat\psi}^{\dagger}({\bf r}_2)\rangle
\langle{\hat\psi} ({\bf r}_2^{\prime}){\hat\psi} ({\bf r}_1^{\prime})\rangle
\label{corr1}\end{aligned}$$ when $${\bf r}_1 - {\bf r}_2 = \mbox{const},\
{\bf r}_1^{\prime} - {\bf r}_2^{\prime} = \mbox{const},\
|{\bf r}_1^{\prime} - {\bf r}_1| \to \infty ;
\label{limit1}$$ $$\begin{aligned}
\langle {\hat\psi}^{\dagger}({\bf r}_1) {\hat\psi}^{\dagger}({\bf r}_2)
&&{\hat\psi}({\bf r}_2^{\prime}){\hat\psi} ({\bf r}_1^{\prime})\rangle\to
\nonumber \\
&&\langle {\hat\psi}^{\dagger}({\bf r}_1) {\hat\psi}({\bf r}_1^{\prime})\rangle
\langle{\hat\psi}^{\dagger} ({\bf r}_2){\hat\psi} ({\bf r}_2^{\prime})\rangle
\label{corr2}\end{aligned}$$ when $${\bf r}_1 - {\bf r}_1^{\prime} = \mbox{const},\
{\bf r}_2 - {\bf r}_2^{\prime} = \mbox{const},\
|{\bf r}_1 - {\bf r}_2| \to \infty.
\label{limit2}$$ It should be stressed that limits (\[corr1\]) and (\[corr2\]) are valid either the Bose-Einstein condensation takes place or not. In the first case the anomalous averages in the expression (\[corr1\]) are not zero [@note7a].
As the expression (\[f2\]) is a Hermitian kernel, we can expand it in the orthonormal set of its eigenprojectors (in the Hilbert space): $$F_{2}(\r{1},\r{2};\rp{1},\rp{2})=
\sum_{\nu}N_{\nu}\Psi_{\nu}^{*}(\r{1},\r{2})\Psi_{\nu}(\rp{1},\rp{2})
\label{f2psinu}$$ where $$\int_{V}\d^{3}r_{1}\d^{3}r_{2}\;{|\Psi_{\nu}
(\r{1},\r{2})|}^{2}=1.
\label{psinorm}$$ Eigenfunctions $\Psi_{\nu}(\r{1},\r{2})$, which at the same time are eigenfunctions of $2$-matrix (\[rho2vtor\]), are called the wave functions of pairs of particles, or PWF. From (\[f2\]), (\[f2psinu\]) and (\[psinorm\]) we get $$\int_{V}\d^{3}r_{1}\d^{3}r_{2}\;F_{2}(\r{1},\r{2};\r{1},\r{2})=
\langle {\widehat N}^{2}-{\widehat N}\rangle$$ $$=N(N-1)=\sum_{\nu}N_{\nu},$$ i.e. the sum of all $N_{\nu}$ is the total number of pairs. Therefore, the non-negative quantities $N_{\nu}$ can be interpreted as the mean number of the pairs in the state $\nu$, any pair being doubly taken. The ratio $w_{\nu}=N_{\nu}/[N(N-1)]$ is the probability of observing a particle pair in the pure state with the wave function $\Psi_{\nu}({\bf r}_1, {\bf r}_2)$. Here, as one might expect, $\sum_{\nu} w_{\nu}=1$.
From the definition (\[f2\]) it follows that $$F_{2}(\r{1},\r{2};\rp{1},\rp{2})=
F_{2}(\r{2},\r{1};\rp{1},\rp{2})=
F_{2}(\r{1},\r{2};\rp{2},\rp{1})$$ hence, $\Psi_{\nu}(\r{1},\r{2})=\Psi_{\nu}(\r{2},\r{1})$, i.e. the PWF for bosons are symmetric with respect to permutations.
For equilibrium states, the momentum of centre of mass of a pair of particles ${\bf q}$ is a quantum number corresponding to PWF, so the index $\nu$ can be represented as $\nu=(\omega,{\bf q})$, where $\omega$ stands for other indices. PWF can be written in the following form $$\Psi_{\nu}(\r{1},\r{2})=\psi_{\omega,{\bf q}}(\r{1}-\r{2})
\frac{\exp [i {\bf q}(\r{1}+\r{2})/2]}{\sqrt{V}}.
\label{psinurazd}$$ Let us demonstrate this. Consider an equilibrium state described by the Gibbs canonical ensemble. As the Hamiltonian is invariant under translations, the total momentum ${\bf P}$ is conserved. So, we have $$\Psi_{n}(\r{1}+{\bf a},\ldots,\r{N}+{\bf a})=
\Psi_{n}(\r{1},\ldots,\r{N})\exp[i{\bf P}{\bf a}],
\label{psip1}$$ Without loss of generality we put ${\bf P}=0$, and introduce new variables $${\bf R}=(\r{1}+\r{2})/2,\quad {\bf r}=\r{1}-\r{2},
\label{Rr}$$ and similarly for $\rp{1}$ and $\rp{2}$. Rewrite (\[rho\]) with the help of (\[psip1\]) setting ${\bf a}={\bf R}$ and ${\bf a}={\bf R}^{\prime}$, respectively. Substituting rewritten Eq. (\[rho\]) into (\[rho2def\]) and performing the Fourier transformations with respect to $\rp{3}-{\bf
R}^{\prime},\ldots,\rp{N}-{\bf R}^{\prime},\r{3}-{\bf R},\ldots,\r{N}-{\bf
R}$ we obtain: $$\rho_{2}(\rp{1},\rp{2},\r{1},\r{2})=\frac{1}{V}\sum_{{\bf q}}
e^{i{\bf q}({\bf R}^{\prime}-{\bf R})}
\rho_{{\bf q}}(\rp{},\r{}),
\label{rhoq}$$ where $\rho_{{\bf q}}$ is a non-negative Hermitian kernel. Expanding it in the set of its eigenprojectors $\psi_{\omega,{\bf q}}({\bf
r}')\psi^{*}_{\omega,{\bf q}}(\r{})$ we arrive at Eqs. (\[f2psinu\]) and (\[psinurazd\]). Analogous proof can be given for the Gibbs grand canonical ensemble.
Then the expression (\[f2psinu\]) can be written in the form $$\begin{aligned}
F_{2}(\r{1},\r{2};\rp{1},\rp{2})&=&\sum_{\omega,{\bf q}}
\frac{N_{\omega,{\bf q}}}{V}
\psi^{*}_{\omega,{\bf q}}(\r{1}-\r{2})\psi_{\omega,{\bf q}}(\rp{1}-\rp{2})
\nonumber \\
&&\times\exp\left[i\frac{{\bf q}}{2}(\rp{1}+\rp{2}-\r{1}-\r{2})\right].
\label{f2omq}\end{aligned}$$ For $\psi_{\omega,{\bf q}}(\r{})$, which can be interpreted as the wave function of a pair of particles in the centre-of-mass system, from (\[psinorm\]) and (\[psinurazd\]) we get the following normalization condition: $$\int_{V}\d^{3}r\;|\psi_{\omega,{\bf q}}(\r{})|^{2}=1.
\label{psiomqn}$$ As the wave function $\psi_{\omega,{\bf q}}(\r{})$ is an EF of the kernel $\rho_{{\bf q}}(\rp{},\r{})$ (see Eq. (\[rhoq\])), it can belong either to a discrete or a continuous spectrum, so we have two possible situations [@note3]:\
(1) $\psi_{\omega,{\bf q}}(\r{})\rightarrow 0$ at $r\rightarrow
\infty$ (bound state of a pair of particles).\
(2) $\psi_{\omega,{\bf q}}(\r{})\rightarrow \sqrt{2}\cos ({\bf p}
{\bf r})$ at $r\rightarrow \infty$ (“dissociated" states of a pair of particles corresponding to scattering with relative momentum ${\bf p}$, the factor $\sqrt{2}$ stands for the appropriate normalization when $p\not=0$).
In the first case, $\omega=i$ is a discrete index numbering bound pair states. We denote $\psi_{\omega,{\bf q}}(\r{})=\varphi_{{\bf q}, i} (\r{})$ and keep the normalization (\[psiomqn\]): $$\int_{V}\d^{3}r\;|\varphi_{{\bf q},i}(\r{})|^{2}=1.
\label{svnorm}$$ In the second case [@note4], $\omega={\bf p}$ and $$\psi_{\omega,{\bf q}}(\r{})=\frac{\varphi_{{\bf p},{\bf q}}
(\r{})}{\sqrt{V}}.$$ Due to Eq. (\[psiomqn\]) we obtain the normalization $$\frac{1}{V}\int_{V}\d^{3}r\;|\varphi_{{\bf p},{\bf q}}
(\r{})|^{2}=1.
\label{freenorm}$$ With the variables (\[Rr\]) the expression (\[f2omq\]) can be rewritten in the following form:
------------------------------------------------------------------------
------------------------------------------------------------------------
$$\begin{aligned}
F_{2}(\r{1},\r{2};\rp{1},\rp{2})=F_{2}(\r{},{\bf R};
\rp{},{\bf R}^{\prime})
&=&\sum_{{\bf q},i}\frac{N_{q,i}}{V}
\varphi^{*}_{{\bf q},i}(\r{})\varphi_{{\bf q},i}(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})]
\nonumber \\
&&+\sum_{{\bf p},{\bf q}}\frac{N_{{\bf p},{\bf q}}}{V^{2}}
\varphi^{*}_{{\bf p},{\bf q}}(\r{})\varphi_{{\bf p},{\bf q}}
(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})].
\label{f2quasi}\end{aligned}$$ In the limit $V\rightarrow \infty$ we should replace the quasidiscrete momentum sums by the corresponding integrals: $$\begin{aligned}
F_{2}(\r{},{\bf R};\rp{},{\bf R}^{\prime})
&=&\sum_{i}\int \d^{3}q\;w_{i}(q)
\varphi^{*}_{{\bf q},i}(\r{})\varphi_{{\bf q},i}(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})]
\nonumber \\
&&+\int \d^{3}p\,\d^{3}q\;w({\bf p},{\bf q})
\varphi^{*}_{{\bf p},{\bf q}}(\r{})\varphi_{{\bf p},{\bf q}}
(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})].
\label{f2asymp}\end{aligned}$$
------------------------------------------------------------------------
------------------------------------------------------------------------
[2]{} Thus, one can see from Eqs. (\[f2quasi\]), (\[f2asymp\]) that $Vw_{i}(q)\d^{3}q$ is the number of bound pairs in state $i$ with momentum ${\bf q}$ in an infinitesimally small momentum volume $\d^{3}q$, and $V^{2}w({\bf p},{\bf q})\d^{3}p\,\d^{3}q$ is the number of “dissociated", or scattered, pairs with relative momentum ${\bf p}$ and centre-of-mass momentum ${\bf q}$ in infinitesimally small momentum volumes $\d^{3}p$ and $\d^{3}q$.
In the centre-of-mass system the replacement ${\bf p} \to - {\bf p}$ corresponds to the permutation of particles. So, the following symmetric relations take place: $$w({\bf p},{\bf q})=w(-{\bf p},{\bf q}),
\label{wpqsimm}$$ $$\varphi_{{\bf p},{\bf q}}({\bf r})=\varphi_{{\bf p},{\bf q}}(-{\bf r})
=\varphi_{-{\bf p},{\bf q}}({\bf r}).
\label{phisymm}$$
Let us now prove, with the principle of the correlation weakening, that the distribution of the particle pairs over the “scattering" states $w({\bf
p},{\bf q})$ is expressed in terms of the single-particle momentum distribution $w_{sing}({\bf k}) = n_{0} \delta({\bf k}) + n(k)/(2\pi)^{3}$, where $\delta({\bf k})$ stands for the $3$-dimensional $\delta$-function, $n(k)=\langle {\hat a}_{\bf k}^{\dagger} {\hat a}_{\bf k}\rangle$ is the single-particle occupation number. Suppose that the temperature is larger than the temperature of the Bose-Einstein condensation, so, $n_{0}=0$ and $w_{sing}(k)$ is a regular function, i.e. it does not contain $\delta$-function after the thermodynamic limit. On the one hand, in the limiting situation of (\[limit2\]) we have the relation (\[corr2\]), which can be written as
------------------------------------------------------------------------
------------------------------------------------------------------------
$$\begin{aligned}
F_{2}(\r{1},\r{2};\rp{1},\rp{2})
&\to&\int\frac{\d^3 p_1}{(2\pi)^3} n(p_1) \exp[i{\bf p}_1
({\bf r}_1^{\prime}-{\bf r}_1)]
\int\frac{\d^3 p_2}{(2\pi)^3} n(p_2) \exp[i{\bf p}_2
({\bf r}_2^{\prime}-{\bf r}_2)] \nonumber \\
&=&\int \d^3q\,\d^3p \frac{n({\bf q}/2+{\bf p})\,
n({\bf q}/2-{\bf p})}{(2\pi)^6}
\exp[i{\bf p}({\bf r}^{\prime}-{\bf r})]
\exp[i{\bf q}({\bf R}^{\prime}-{\bf R})],
\label{F1F1}\end{aligned}$$ where, passing to the last equality, we introduced the new variables ${\bf
q}={\bf p}_1+{\bf p}_2$ and ${\bf p}=({\bf p}_1- {\bf p}_2)/2$ and used notations (\[Rr\]). On the other hand, when (\[limit2\]) is true, we have $$r=|{\bf r}_2-{\bf r}_1| \to \infty,\
r^{\prime}=|{\bf r}_2^{\prime}-{\bf r}_1^{\prime}|\to \infty,
|{\bf r}+{\bf r}^{\prime}| \to \infty,\ {\bf R}^{\prime}-{\bf R}
=\mbox{const},\ {\bf r}^{\prime}-{\bf r}=\mbox{const}.$$
[2]{} Hence, in this limit we obtain $$\varphi_{{\bf q},i}(\r{})\rightarrow 0,\quad
\varphi_{{\bf p},{\bf q}}(\r{})\rightarrow \sqrt{2}\cos ({\bf p}{\bf r}).$$ Therefore, it follows from (\[f2asymp\]) that in the limiting case $(\ref{limit2})$ we have $$\begin{aligned}
F_{2}(\r{1},\r{2};\rp{1},\rp{2}) \to\int
&&\d^3 q\,\d^3 p\; w({\bf p},{\bf q})
2\cos({\bf p}{\bf r})\cos({\bf p}{\bf r}')
\nonumber \\
&&\times\exp[i{\bf q}({\bf R}^{\prime}-{\bf R})].
\label{F2lim}\end{aligned}$$ Further, the Riemann’s theorem [@Note6] used while integrating over ${\bf
p}$ and relation (\[wpqsimm\]) allow us to rewrite (\[F2lim\]) as $$\begin{aligned}
F_{2}(\r{1},\r{2};\rp{1},\rp{2}) \to \int
&&\d^3q\,\d^3p\;w({\bf p},{\bf q})
\exp\left[i{\bf p}({\bf r}^{\prime}-{\bf r})\right]
\nonumber \\
&&\times\exp\left[i{\bf q}({\bf R}^{\prime}-{\bf R})\right].
\label{F2lim2a}\end{aligned}$$ The right-hand side of Eq. (\[F1F1\]) is equal to that of (\[F2lim2a\]) at all the values of space variables $\widetilde{\bf r}={\bf r}'-{\bf r}$ and $\widetilde{\bf R}={\bf R}'-{\bf R}$. Hence, we derive the following equality: $$w({\bf p},{\bf q})=\frac{n({\bf q}/2+{\bf p})
n({\bf q}/2-{\bf p})}{(2\pi)^6}.
\label{wpqsms}$$
Below the temperature of the Bose-Einstein condensation a macroscopic number of particles $N_{0}$ occupies the zero-momentum state: $n_{0}=\langle {\hat
a}_{0}^{\dagger}{\hat a}_{0}\rangle/V =N_{0}/V =\mbox{const}$, when $V\to\infty$, and we have in the thermodynamic limit $$\langle {\hat\psi}^{\dagger}({\bf r}_1) {\hat\psi}({\bf r}_{1}^{\prime})\rangle
=\int \d^{3}p_{1} \Bigl(n_{0}\delta({\bf p}_{1})+\frac{n(p_{1})}{(2\pi)^{3}}
\Bigr)$$ $$\times\exp\left[i{\bf p}_{1}({\bf r}_{1}^{\prime}-{\bf r}_{1})\right].$$ By analogy with the above proof one can readily be convinced that the distribution $w({\bf p},{\bf q})= w(\frac{{\bf p}_{1}-{\bf p}_{2}}{2},{\bf
p}_{1}+{\bf p}_{2})$ contains the following parts: $$\begin{aligned}
1)\quad && n^{2}_{0}\delta({\bf p}_{1})\delta({\bf p}_{2}),
\label{wpqsing1}\\
2)\quad && n_{0}\delta({\bf p}_{1})\frac{n(p_{2})}{(2\pi)^{3}}
+n_{0}\delta({\bf p}_{2})\frac{n(p_{1})}{(2\pi)^{3}},
\label{wpqsing2}\\
3)\quad && \frac{n(p_{1})}{(2\pi)^{3}} \frac{n(p_{2})}{(2\pi)^{3}}.
\label{wpqsing3}\end{aligned}$$ The expression (\[wpqsing1\]) corresponds to condensate pair states, or condensate-condensate pairs (both particle of a couple are in the condensate); the terms (\[wpqsing2\]), to condensate-supracondensate pairs; and the last term (\[wpqsing3\]), to supracondensate-supracondensate ones. Thus, owing to the expressions (\[wpqsing1\])-(\[wpqsing3\]) we obtain the natural classification of the “dissociated" pair states.
We note that the total number of bound pairs $N_{b} \simeq V\sum_{i}\int
\d^{3}q\, w_{i}(q)$ is asymptotically proportional to $V^{1}$,\
while the total number of “dissociated" pairs $N_{d} \simeq V^{2} \int
\d^{3} p\,\d^{3}q\; w({\bf p},{\bf q})=N^{2}$ is asymptotically proportional to $V^{2}$. From an intuitive point of view it is obvious, as a given particle in a liquid can form bound states with $M$ particles ($M=\mbox{const}$ as $V\rightarrow \infty$), while it forms “dissociated" states with the other $N-1-M$ particles. So, we arrive at the following asymptotic equation: $$N_{b}+N_{d}=N(N-1)\simeq N_{d}\simeq N^{2}.$$
Diagonal elements of $F_{2}(\r{1},\r{2};\rp{1},\rp{2})$ depend on squared absolute values of the PWF and determine the behaviour of the static pair distribution function $g(r)$, which can be directly observed in scattering experiments: $$g(r)=\frac{V^{2}}{N(N-1)}F_{2}(\r{},{\bf R};\r{},{\bf R})$$ $$=\frac{V^{2}}{N(N-1)}\sum_{\omega,{\bf q}}\frac{N_{\omega,{\bf q}}}{V}
|\psi_{\omega,{\bf q}}({\bf r})|^{2}.$$ After the thermodynamic limit this expression takes a form $$\begin{aligned}
g(r)= \frac{1}{n^{2}}\biggl(
&&\sum_{i}\int \d^{3}q\;w_{i}(q)
|\varphi_{{\bf q},i}(\r{})|^{2}
\nonumber \\
&&+\int \d^{3}p\,\d^{3}q\;w({\bf p},{\bf q})
|\varphi_{{\bf p},{\bf q}}(\r{})|^{2}
\biggr).
\label{gr}\end{aligned}$$ One can easily be convinced that $g(r)\rightarrow 1$ as $r\rightarrow
\infty$. The behaviour of $g(r)$ for $r\rightarrow 0$ is determined by the PWF at small distances. Because of repulsion, the probability of finding three particles in a very small volume is much lower than that of finding two particles in the same volume. Therefore, the short-range behaviour of two particles is determined by the ordinary two-body Schrödinger equation with the pairwise interaction $\Phi(r)$. Thus, the PWF for short distances between two particles is proportional to the usual wave function $\varphi_{\bf p}^{(0)}(\r{})$ of the two-body problem [@kimball]: $$\varphi_{{\bf p},{\bf q}}(\r{})\to C\varphi_{\bf p}^{(0)}(\r{}),
\label{bcond1}$$ when $r\to 0$. This implies that for a singular potential the PWF goes to zero for $r\rightarrow 0$, and it follows that $g(r)\rightarrow 0$ as $r\rightarrow 0$.
Structure of pair correlation function and pair wave functions for Bose liquid with condensate {#3}
==============================================================================================
In order to examine the behaviour of the pair correlation function (\[f2\]) with a condensate we will employ the procedure proposed first by Bogoliubov [@bog47] and lately justified with the help of the principle of correlation weakening [@bogquasi]: in the calculation of averages [@note1] of any number of the field operators we can substitute the $C$-number $\sqrt{N_{0}}$ for the operators ${\hat a}_{0}$ and ${\hat
a}_{0}^{\dagger}$ involved in ${\hat\psi}(\r{})$ and ${\hat\psi}^{\dagger}(\r{})$. Here $N_{0}$ is the number of particles in the condensate. It should be stressed that, after Bogoliubov, this substitution gives [*asymptotically exact values*]{} of the averages in the limit $V\rightarrow \infty$ and [*does not imply any approximations*]{}. Indeed, we have for any average that involves the operator ${\hat a}_{0}/\sqrt{V}$: $$\lim_{V\to\infty}\langle \frac{{\hat a}_{0}}{\sqrt{V}}\cdots{\hat\psi}({\bf r}')
\cdots\rangle$$ $$=\lim_{V\to\infty}\frac{1}{V}\int_{V}\d^{3}r\langle {\hat\psi}({\bf r})
\cdots{\hat\psi}({\bf r}')\cdots\rangle$$ $$=\lim_{V\to\infty}\frac{1}{V}\int_{V}\!\d^{3}r\langle{\hat\psi}({\bf r})\rangle
\langle\cdots{\hat\psi}({\bf r}')\cdots\rangle\!=\!
\sqrt{n_{0}}\langle\cdots{\hat\psi}({\bf r}')\cdots\rangle.$$ Here we use the definition $\langle{\hat\psi}({\bf r})\rangle=
\langle{\hat\psi}^{\dagger}({\bf r})\rangle=\sqrt{n_{0}}$ ($n_{0}=N_{0}/V$ is the density of the condensate) and the fact that the main contribution to the value of the integral comes from infinitely large $r$. Thus, we can put $$\begin{aligned}
{\hat\psi}(\r{})&=&\sqrt{n_{0}}+{\hat\vartheta}(\r{}),\nonumber \\
{\hat\psi}^{\dagger}(\r{})&=&\sqrt{n_{0}}+{\hat\vartheta}^{\dagger}(\r{}),
\label{psikrr}\end{aligned}$$ where $$\begin{aligned}
{\hat\vartheta}(\r{})&=&\frac{1}{\sqrt{V}}\sum_{{\bf p}\not=0}{\hat a}_{{\bf p}}
e^{i{\bf p}{\bf r}},\nonumber \\
{\hat\vartheta}^{\dagger}(\r{})&=&\frac{1}{\sqrt{V}}\sum_{{\bf p}\not=0}
{\hat a}^{\dagger}_{{\bf p}}e^{-i{\bf p}{\bf r}}
\label{thetakrr}\end{aligned}$$ are supracondensate field operators. Substituting the expression (\[psikrr\]) into (\[f2\]) and using the relation $\langle{\hat\vartheta}({\bf r})\rangle=\langle{\hat\vartheta}^{\dagger }
({\bf r})\rangle=0$ we obtain
------------------------------------------------------------------------
------------------------------------------------------------------------
$$\begin{aligned}
F_{2}(\r{1},\r{2};\rp{1},\rp{2})=F^{(1)}_{2}(\r{1},\r{2};\rp{1},
\rp{2})+F^{(2)}_{2}(\r{1},\r{2};\rp{1},\rp{2})
+F^{(3)}_{2}(\r{1},\r{2};\rp{1},\rp{2})
+F^{(4)}_{2}(\r{1},\r{2};\rp{1},\rp{2}),
\label{f2tot}\end{aligned}$$ where we introduce the notations $$\begin{aligned}
F^{(1)}_{2}(\r{1},\r{2};\rp{1},\rp{2})&=&n_{0}^{2}+
n_{0}\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}^{\dagger}(\r{2})
\rangle
+n_{0}\langle{\hat\vartheta}(\rp{2}){\hat\vartheta}(\rp{1})\rangle,
\label{f21}\\
F^{(2)}_{2}(\r{1},\r{2};\rp{1},\rp{2})&=&n_{0}\Bigl(
\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}(\rp{1})\rangle+
\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}(\rp{2})\rangle
+\langle{\hat\vartheta}^{\dagger}(\r{2}){\hat\vartheta}(\rp{2})\rangle
+\langle{\hat\vartheta}^{\dagger}(\r{2}){\hat\vartheta}(\rp{1})\rangle
\Bigr),
\label{f22}\\
F^{(3)}_{2}(\r{1},\r{2};\rp{1},\rp{2})&=&\sqrt{n_{0}}\Bigl(
\langle{\hat\vartheta}^{\dagger}(\r{2}){\hat\vartheta}(\rp{2}){\hat\vartheta}(\rp{1})\rangle
+\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}(\rp{2}){\hat\vartheta}(\rp{1})\rangle
\nonumber \\
&&\phantom{\sqrt{n_{0}}\Bigl(}+\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}^{\dagger}(\r{2}){\hat\vartheta}(\rp{2})
\rangle
+\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}^{\dagger}(\r{2})
{\hat\vartheta}(\rp{1})
\rangle\Bigr),
\label{f23}\\
F^{(4)}_{2}(\r{1},\r{2};\rp{1},\rp{2})&=&
\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}^{\dagger}(\r{2})
{\hat\vartheta}(\rp{2}){\hat\vartheta}(\rp{1})\rangle.\label{f24}\end{aligned}$$ The expressions (\[f21\])-(\[f24\]) can be expanded in systems of projectors (in the Hilbert space). For (\[f21\])-(\[f23\]) the corresponding projectors can be expressed in an explicit form in terms of averages of the creation and destruction Bose operators ${\hat a}_{{\bf p}}$ and ${\hat a}^{\dagger}_{{\bf p}}$. For (\[f21\]) we have $$F^{(1)}_{2}(\r{1},\r{2};\rp{1},\rp{2})=n_{0}^{2}\varphi^{*}
(\r{1}-\r{2})
\varphi(\rp{1}-\rp{2})
-\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}^{\dagger}(\r{2})\rangle
\langle{\hat\vartheta}(\rp{2}){\hat\vartheta}(\rp{1})\rangle,$$ where $$\begin{aligned}
\varphi(r)&=&1+\psi(r),\label{phir}\\
\psi(r)&=&\psi(\r{1}-\r{2})=\psi^{*}(\r{1}-\r{2})=
\langle{\hat\vartheta}(\r{1}){\hat\vartheta}(\r{2})\rangle/n_{0}
=\frac{1}{(2\pi)^{3}}\int \d^{3}k\;\psi(k)e^{i{\bf k}{\bf r}},
\label{psir}\\
\psi(k)&=&\langle {\hat a}_{{\bf k}}{\hat a}_{-{\bf k}}\rangle/n_{0}.
\label{psik}\end{aligned}$$ In a similar manner we get $$F_{2}^{(2)}(\r{},{\bf R};\rp{},{\bf R}^{\prime})=
\int \d^{3}p\,\d^{3}q\;w^{(1)}({\bf p},{\bf q})
\sqrt{2}\cos({\bf p}\r{})\sqrt{2}\cos({\bf p}\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})].
\label{f22red}$$ Here $$w^{(1)}({\bf p},{\bf q})=2n_{0}\delta({\bf q}/2-{\bf p})
\frac{n({\bf q}/2+{\bf p})}{(2\pi)^{3}}
=2n_{0}\delta({\bf q}/2-{\bf p})\frac{n(q)}{(2\pi)^{3}}
\label{w1pq}$$ is the momentum distribution for condensate-supracondensate pairs corresponding to the expression (\[wpqsing2\]).
Performing the Fourier transformations, we reduce $F_{2}^{(3)}$ to the following form $$\begin{aligned}
F_{2}^{(3)}(\r{},{\bf R};\rp{},{\bf R}^{\prime})=
\int \d^{3}p\,\d^{3}q\;w^{(1)}({\bf p},{\bf q})
\left(\psi_{{\bf p}}(\rp{})\sqrt{2}\cos({\bf p}\r{})+
\psi^{*}_{{\bf p}}(\r{})\sqrt{2}\cos({\bf p}\rp{})\right)
\times\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})].
\label{f23red}\end{aligned}$$ where $$\begin{aligned}
\psi_{{\bf p}}(\r{})&&=
\frac{1}{(2\pi)^{3}}\int \d^{3}k\;\psi_{{\bf p}}({\bf k})
e^{i{\bf k}{\bf r}},
\label{psipr}\\
\psi_{{\bf p}}({\bf k})&&=
\psi_{{\bf p}}(-{\bf k})=\psi_{-{\bf p}}({\bf k})
=\sqrt{N_{0}}\frac{1}{\sqrt{2}n_{0}}
\frac{\langle {\hat a}^{\dagger}_{2{\bf p}}{\hat a}_{{\bf p}+{\bf k}}
{\hat a}_{{\bf p}-{\bf k}}\rangle}{n(2p)}.
\label{psipk}\end{aligned}$$ The second equation in (\[psipk\]) is a consequence of the invariance of averages with respect to the transformation ${\hat a}_{{\bf k}}\to {\hat
a}_{-{\bf k}},\ {\hat a}^{\dagger}_{{\bf k}}\to {\hat a}^{\dagger}_{-{\bf
k}}$, as the Hamiltonian does not change when the transformation is performed. The factor $\sqrt{N_{0}}$ compensates the decrease of the average value $\langle {\hat a}^{\dagger}_{2{\bf p}}{\hat a}_{{\bf p}+{\bf k}} {\hat
a}_{{\bf p}-{\bf k}}\rangle\sim V^{-1/2}$, so, the quantity $\psi_{{\bf
p}}({\bf k})$ is of order $1$ in the limit $V\rightarrow \infty$.
From (\[psipr\]) and (\[psipk\]) we obtain the symmetry properties $$\psi_{{\bf p}}({\bf r})=\psi_{{\bf p}}(-{\bf r})=
\psi_{-{\bf p}}({\bf r}).$$ Summing the expressions (\[f22red\]) and (\[f23red\]) we get $$F_{2}^{(2)}(\r{},{\bf R};\rp{},{\bf R}^{\prime})+
F_{2}^{(3)}(\r{},{\bf R};\rp{},{\bf R}^{\prime})
=\int \d^{3}p\,\d^{3}q\;w^{(1)}({\bf p},{\bf q})
\varphi^{*}_{{\bf p}}(\r{})\varphi_{{\bf p}}(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})]$$ $$-\int \d^{3}q\;2n_{0}\frac{n(q)}{(2\pi)^{3}}
\psi^{*}_{{\bf q}/2}(\r{})\psi_{{\bf q}/2}(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})],$$ where $$\varphi_{{\bf p}}(\r{})=\varphi_{-{\bf p}}(\r{})=
\varphi_{{\bf p}}(-\r{})=\sqrt{2}\cos({\bf p}\r{})+
\psi_{{\bf p}}(\r{}),
\label{phipr}$$ in accordance with (\[phisymm\]).
Now it is not difficult to express (\[f2tot\]) in a form similar to (\[f2asymp\]) $$\begin{aligned}
F_{2}(\r{},{\bf R};\rp{},{\bf R}^{\prime})=
n_{0}^{2}\varphi^{*}(r)\varphi(r')
+\int\!\d^{3}p\,\d^{3}q\;2n_{0}\delta({\bf q}/2-{\bf p})\frac{n(q)}{(2\pi)^{3}}
\varphi^{*}_{{\bf p}}(\r{})\varphi_{{\bf p}}(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})]
+{\widetilde F}_{2}(\r{},{\bf R};\rp{},{\bf R}^{\prime}),
\label{f2final}\end{aligned}$$ where $$\begin{aligned}
{\widetilde F}_{2}(\r{},{\bf R};\rp{},{\bf R}^{\prime})=
{\widetilde F}_{2}(\r{1},\r{2};\rp{1},\rp{2})
&=&F^{(4)}_{2}(\r{1},\r{2};\rp{1},\rp{2})-
\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}^{\dagger}(\r{2})\rangle
\langle{\hat\vartheta}(\rp{2}){\hat\vartheta}(\rp{1})\rangle\nonumber \\
&&-\int \d^{3}q\;2n_{0}\frac{n(q)}{(2\pi)^{3}}
\psi^{*}_{{\bf q}/2}(\r{})\psi_{{\bf q}/2}(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})].
\label{f2tild}\end{aligned}$$
------------------------------------------------------------------------
------------------------------------------------------------------------
[2]{} As it is proved in Sec. \[2\], the eigenvalues of $F_{2}(\r{},{\bf
R};\rp{},{\bf R}^{\prime})$ are given by Eqs. (\[wpqsing1\])-(\[wpqsing3\]). So, one can conclude that the projectors $\varphi^{*}(r)\varphi(r')$ and $\varphi^{*}_{{\bf q}/2}(\r{})\exp[-i{\bf
q}{\bf R}] \varphi_{{\bf q}/2}(\rp{}) \exp[i{\bf q}{\bf R}^{\prime}]$ are nothing else but the eigenprojectors for the pair correlation function, and, hence, $\varphi(r)$ and $\varphi_{{\bf q}/2}({\bf r})\exp[i{\bf q}{\bf R}]$ are the PWF for the Bose system below $T_{c}$. The first term in (\[f2final\]) corresponds to pairs in the condensate; the second, to condensate-supracondensate pairs; the last one, ${\widetilde F}_{2}$, to pairs in the supracondensate, and, maybe, to bound pairs. The PWF $\varphi(r)$ and $\varphi_{\bf p}({\bf r})$ given by expressions (\[phir\])-(\[psik\]) and (\[psipr\])-(\[phipr\]) respectively are normalized in accordance with (\[freenorm\]). The condensate PWF differs essentially from $1$ at short distances and goes to $1$ when distances between particles in a pair increase. The situation is similar to the scattering of two bare particles with zero momenta at infinity. The anomalous averages $\langle{\hat\vartheta}^{\dagger}(\r{1})
{\hat\vartheta}^{\dagger}(\r{2})\rangle$ in PWF (\[phir\]) have essentially different physical interpretation in comparison with the fermion systems: they determine the “scattering" part of the condensate PWF, while for fermion systems anomalous averages $\langle{\hat\psi}^{\dagger}(\r{1})
{\hat\psi}^{\dagger}(\r{2})\rangle$ themselves are PWF for the bound state of pairs of particles with zero momentum [@bogquasi; @CSPRB].
In the condensate-supracondensate PWF (\[phipr\]) we drop the index corresponding to the centre-of-mass momentum as ${\bf q}=\pm 2{\bf p}$. This is a consequence of the $\delta$-function being present in the pair momentum distribution (\[w1pq\]). For the same reason in the condensate-condensate PWF (\[phir\]) the indices are omitted at all.
The functions $\psi(r)$ and $\psi_{{\bf p}}({\bf r})$ defined by Eqs. (\[psir\]) and (\[psipr\]) are responsible for particle short-distance correlations, therefore, owing to (\[bcond1\]) for strongly singular potentials we have $$\psi(r=0)=-1,\quad\psi_{{\bf p}}({\bf r}=0)=-\sqrt{2},
\label{bcond}$$ (see the discussion at the end of Sec. \[2\]).
It should be emphasized that the pair correlation function for supracondensate particles is ${\widetilde F}_{2}$ rather than $F_{2}^{(4)}$ defined by Eq. (\[f24\]). Since the kernel (\[f2final\]) is non-negative, the kernel (\[f2tild\]) is also non-negative. Consequently, the last two terms in the expression for $\widetilde {F}_{2}$ are reduced to zero by the terms in $F_{2}^{(4)}$. These terms look like those corresponding to bound states of pairs. If the kernel $F_{2}^{(4)}$ written in form (\[f2asymp\]) contains other terms associated with bound states, then the latter are also contained in ${\widetilde F}_{2}$ and, consequently, in the final expression (\[f2final\]). The distribution over supracondensate-supracondensate pair states is given by the expression (\[wpqsing3\]), and thus, we arrive at the final formula for ${\widetilde F}_{2}$: $$\begin{aligned}
{\widetilde F}_{2}&&(\r{},{\bf R};\rp{},{\bf R}^{\prime})
\nonumber \\
&&=\sum_{i}\int \d^{3}q\;w_{i}(q)
\varphi^{*}_{{\bf q},i}(\r{})\varphi_{{\bf q},i}(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})]\nonumber \\
&&\phantom{=}+\int \d^{3}p\,\d^{3}q\;
\frac{n({\bf q}/2+{\bf p})}{(2\pi)^{3}}
\frac{n({\bf q}/2-{\bf p})}{(2\pi)^{3}}
\nonumber \\
&&\phantom{=+\int}\times\varphi^{*}_{{\bf p},{\bf q}}(\r{})\varphi_{{\bf p},{\bf q}}
(\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})],
\label{f2tildfin}\end{aligned}$$ where the symmetry properties (\[phisymm\]) are fulfilled. Thus, in order to answer the question about real bound states of pairs of particles, one must carry out a detailed analysis taking into account all groups of the terms (\[f21\])-(\[f24\]). We emphasize that we do not specify the size of the bound pairs \[the characteristic range of the PWF $\varphi_{{\bf
q},i}(\r{})$\]. If it is much more than the mean distance between particles, then, following Bogoliubov [@bogquasi], one may call these pairs “quasi-molecules" [@note8]. If the radius of the bound particle couples is of the order of the mean distance between particles or, even, less then one may speak about ordinary molecules. Note that “quasi-molecules" are not stable for a Bose system within the mean-field variational treatment [@nozieres] which is equivalent to the HFB approach. It is a difficult question whether it is the case beyond the mean-field approximation, so, [*a priori*]{} we cannot ignore existence of “quasi-molecules". The bound pair states can totally be a result of the collective effects, as in the case of Fermi systems [@CSPRB]. We note that such situation can even be realized in the case of three particles [@baum].
It is not difficult to see with the help of the Riemann theorem [@Note6] that the boundary condition (\[corr1\]) is satisfied for (\[f2final\]), (\[f2tildfin\]) in the limit (\[limit1\]).
The expressions (\[f21\])-(\[f24\]) can be interpreted in the conventional way, in terms of scattering processes for pairs of particles $({\bf p}_{1},{\bf p}_{2})\rightarrow({\bf p}^{\prime}_{1},{\bf
p}^{\prime}_{2})$, where ${\bf p}_{1}$, ${\bf p}_{2}$, ${\bf p}^{\prime}_{1}$ and ${\bf p}^{\prime}_{2}$ are the particle momenta. For this purpose one should calculate the averages in the momentum representation, i.e. make Fourier transformation. This procedure corresponds to expansion of the PWF in plane waves. Thus, the first term in (\[f21\]) is related to ${\bf
p}_{1}={\bf p}_{2}={\bf p}^{\prime}_{1}={\bf p}^{\prime}_{2}=0$; the second one, to ${\bf p}_{1},\ {\bf p}_{2}\not=0$ and ${\bf p}^{\prime}_{1}={\bf
p}^{\prime}_{2}=0$; and so on. The terms in (\[f23\]) correspond to processes when one of the four momenta is zero. These processes become dominant in the vicinity of $T_{c}$ [@note7], when $N_{0}/N\ll 1$. In this case $\left(\frac{N_{0}}{N}\right)^{2}\ll\frac{N_{0}}{N}\ll
\sqrt{\frac{N_{0}}{N}}$, so one can expect that $F^{(3)}_{2}$ is important in the expression (\[f2tot\]). Taking into account that the thermodynamic potential is tightly connected with the correlation function, one can assume that the terms in (\[f23\]) are important to obtain correct thermodynamics at those temperatures. Recall that these are precisely the terms being responsible for the short-range correlations for the condensate-supracondensate pairs.
Pair wave functions in Hartree-Fock-Bogoliubov and Bogoliubov models {#3a}
====================================================================
Let us consider the HFB approach from the point of view of PWF. In the HFB method (see the detailed discussion in Ref. [@hohmartin]), a Hamiltonian is approximated by a quadratic form of the Bose operators ${\hat\alpha}_{{\bf
p}}^{\dagger}$ and ${\hat\alpha}_{{\bf p}}$ connected with initial operators ${\hat a}_{{\bf p}}^{\dagger}$ and ${\hat a}_{{\bf p}}$ by the canonical Bogoliubov transformations. Consequently, all averages can be calculated with the Wick-Bloch-De Dominicis theorem [@bloch]. Thus, we have
------------------------------------------------------------------------
------------------------------------------------------------------------
$$\begin{aligned}
F^{(3)}_{2}(\r{1},\r{2};\rp{1},\rp{2})&=&0,
\label{f32hfb}\\
F^{(4)}_{2}(\r{1},\r{2};\rp{1},\rp{2})&=&
\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}(\rp{1})\rangle
\langle{\hat\vartheta}^{\dagger}(\r{2}){\hat\vartheta}(\rp{2})\rangle
+\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}(\rp{2})\rangle
\langle{\hat\vartheta}^{\dagger}(\r{2}){\hat\vartheta}(\rp{1})\rangle
+\langle{\hat\vartheta}^{\dagger}(\r{1}){\hat\vartheta}^{\dagger}(\r{2})\rangle
\langle{\hat\vartheta}(\rp{2}){\hat\vartheta}(\rp{1})\rangle.
\label{f24hfb}\end{aligned}$$ The correlation function (\[f2tot\]) can easily be rewritten in the following form $$\begin{aligned}
F_{2}(\r{},{\bf R};\rp{},{\bf R}^{\prime})=
n_{0}^{2}\varphi^{*}(r)\varphi(r')
+\int &&\d^{3}p\,\d^{3}q\;
\Bigl[2n_{0}\delta({\bf q}/2-{\bf p})\frac{n(q)}{(2\pi)^{3}}
+\frac{n({\bf q}/2+{\bf p})}{(2\pi)^{3}}
\frac{n({\bf q}/2-{\bf p})}{(2\pi)^{3}}\Bigr]\nonumber \\
&&\times\sqrt{2}\cos({\bf p}\r{})\sqrt{2}\cos({\bf p}\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})],
\label{f2hfb}\end{aligned}$$ where $\varphi(r)$ is related to the averages via (\[phir\]). In so doing, $n(p)$ and $\psi(k)$ are expressed in terms of Hamiltonian parameters in a self-consistent manner. From (\[f2hfb\]) one can see that there are no bound states of particles in the framework of the HFB approach. We note that the expression (\[f2hfb\]) corresponds to the ideal Bose gas when $\varphi(r)=1$ and $n(p)$ is the Bose-Einstein distribution with a zero chemical potential. One can see that in this case all PWF are symmetrized plane waves, as it is to be expected.
In the HFB method, only the condensate PWF differs from the plane wave. Since PWF are intimately related to the excitation spectrum, one can assume that it is a rough approximation, which has led to the unphysical gap in the single-particle excitation energy. As Eq. (\[f32hfb\]) is fulfilled for all temperatures, then the HFB model cannot correctly describe the thermodynamic behaviour of the system in the vicinity of $T_{c}$ ($T<T_{c}$). It should be pointed out that the expression (\[f2hfb\]) is valid provided $n_{0}\not=0$. If $n_{0}=0$, but $\langle {\hat\vartheta}^{\dagger}(\r{1})
{\hat\vartheta}^{\dagger}(\r{2}) \rangle \not=0$, as in the theory of Valatin and Butler [@val] (see the careful analysis by Nozières and Saint James [@nozieres]), then $F_{2}^{(1)}=F_{2}^{(2)}=F_{2}^{(3)}=0$ and $F_{2}^{(4)}$ is determined by (\[f24hfb\]). In this case the anomalous averages really correspond to the bound pairs of particles with $q=0$: $$F_{2}(\r{},{\bf R};\rp{},{\bf R}^{\prime})=
\rho_{0}\varphi^{*}(r)\varphi(r')
+\int \d^{3}p\,\d^{3}q\;
\frac{n({\bf q}/2+{\bf p})}{(2\pi)^{3}}
\frac{n({\bf q}/2-{\bf p})}{(2\pi)^{3}}
\sqrt{2}\cos({\bf p}\r{})\sqrt{2}\cos({\bf p}\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})],$$
------------------------------------------------------------------------
------------------------------------------------------------------------
[2]{} where $\rho_{0}$ is the density of the bound pairs $$\rho_{0}=\int \d^{3}r_{1}\;
|\langle{\hat\vartheta}(\r{1}){\hat\vartheta}(\r{2})\rangle|^{2},$$ and $\varphi(\r{1}-\r{2})= \langle {\hat\vartheta}(\r{1}) {\hat\vartheta}(\r{2})
\rangle /\sqrt{\rho_{0}}$ is PWF normalized in accordance with (\[svnorm\]) [@note10].
The Bogoliubov theory of weakly imperfect Bose gas is a particular case of the HFB theory when the condensate depletion is small: $(n-n_{0})/n\ll 1$. It can be obtained directly from (\[f2tot\]) if we neglect terms of the third and forth orders in ${\hat\vartheta}$ and ${\hat\vartheta}^{\dagger}$ [@note2] $$\begin{aligned}
F_{2}&&(\r{},{\bf R};\rp{},{\bf R}^{\prime})
\nonumber \\
=&&n_{0}^{2}\varphi^{*}(r)\varphi(r')
+\int \d^{3}p\,\d^{3}q\;
2n_{0}\delta({\bf q}/2-{\bf p})\frac{n(q)}{(2\pi)^{3}}\nonumber \\
&&\times\sqrt{2}\cos({\bf p}\r{})\sqrt{2}\cos({\bf p}\rp{})
\exp[i{{\bf q}}({\bf R}^{\prime }-{\bf R})].
\label{f2bog}\end{aligned}$$ Comparing the exact expressions (\[f2final\]) and (\[f2tildfin\]) with Eq. (\[f2bog\]) one can reveal that the Bogoliubov model involves the following approximations. First, the supracondensate-supracondensate part is omitted, which is completely justified as the condensate depletion is small. Second, the condensate-supracondensate PWF are symmetrized plane waves \[$\psi_{{\bf p}}(\r{})=0$ in Eq. (\[phipr\])\], hence, the boundary conditions (\[bcond\]) for a hard-core potential cannot be satisfied by any choice of possible parameters. In particular, the change of the “bare" potential by the effective one leaves this property of the model unaltered. The approximation $\varphi_{\bf p}(\r{})=\sqrt{2} \cos({\bf p}{\bf r})$ implies that the Bogoliubov model is nothing else but the ideal gas approximation for the condensate-supracondensate PWF, while the scattering part of the condensate-condensate PWF (\[psir\]) is not zero but small and can be evaluated within the Born approximation for the in-medium PWF [@CSPRE; @CSJPS; @CSPL]. Therefore, the Bogoliubov model does not properly take into account the short-range correlations of bosons. So, in the framework of this model we obviously obtain the divergency when evaluating the mean energy for the “bare" hard-core potential due to the infinite contribution of the condensate-supracondensate “channel" involved in the representation (\[gr\]) to the mean interaction energy per particle (\[uaver\]). The hard-sphere model can be treated as the Bogoliubov model with the replacement $\Phi(r)\to\delta({\bf r})4\pi a/m$. Even after this substitution we face another divergency. The point is that the Bogoliubov model involves, in the implicit form, the next-to-Born term for the scattering amplitude: $a_{1}=-b$, $b=m/[4\pi(2\pi)^3]
\int\d^3k\,\Phi^{2}(k)/(2T_k)$, where $T_{k}=k^{2}/(2m)$. The hard-sphere replacement implies that $\Phi(k)=4\pi a/m=\mbox{\rm const}$, which leads to the divergency associated with the parameter $b$, and, hence, to divergent the mean energy. A cure for this difficulty is to replace the Bogoliubov expression (\[f2bog\]) by the exact formula (\[f2final\]) in which the term $\widetilde{F}_{2}$ can be neglected due to a small condensate depletion: $$\begin{aligned}
&&F_{2}(\r{},{\bf R};\rp{},{\bf R}^{\prime})=
n_0^2\varphi^*(r)\varphi(r')
\nonumber \\
&&+2n_0\int\!\frac{\d^3q}{(2\pi)^3}n(q)\varphi_{{\bf q}/2}^*({\bf r})
\varphi_{{\bf q}/2}({\bf r}')
\exp[i{\bf q}({\bf R}'-{\bf R})].
\label{f2ans}\end{aligned}$$ In so doing the PWF $\varphi(r)$ and $\varphi_{\bf p}({\bf r})$ should be determined in a self-consistent manner which provides the boundary conditions (\[bcond1\]) \[and, hence, (\[bcond\])\]. Along this line with the help of the variational procedure we can derive the following non-linear equation: $$U(k)=\Phi(k)
-\frac{1}{2}\int\frac{\d^3q}{(2\pi)^3}\frac{\Phi(|{\bf k}
-{\bf q}|)U(q)}{\sqrt{\widetilde{T}^2_q+
2n\widetilde{T}_q U(q)}}
\label{Uk}$$ valid at zero temperature and small densities [@CSPRE; @CSJPS]. Here for the function $\widetilde{T}_{q}$ we can put, within a good accuracy, $\widetilde{T}_{q}\simeq T_{q}=q^{2}/(2m)$, $\Phi(k)$ is the Fourier transform of the potential $\Phi(r)$, and $U(k)=\int\d^{3}r\,\varphi(r)\Phi(r) \exp(-i{\bf k}{\bf r})$ can be treated as a scattering amplitude in a medium. In the limit $n\to0$ Eq. (\[Uk\]) is reduced to the ordinary Lippmann-Schwinger equation for the scattering amplitude which corresponds to the zero scattering momentum. Note that Eq. (\[Uk\]) looks like an equation for the many-body $t$-matrix $\Gamma({\bf
k},{\bf k}',{\bf q};\omega)$ at $\omega=0$ (see, e.g. the recent review [@shigrif], Sec. 4). However, there exist essential differences. First, in general, the equation for the $t$-matrix is frequency dependent in contrast to Eq. (\[Uk\]). Second, $\Gamma({\bf k},{\bf k}',{\bf
q};\omega=0)$ can not be associated with in-medium scattering amplitude $U_{{\bf k}',{\bf q}}({\bf k})= \int\d^{3}r\,\varphi_{{\bf k}',{\bf q}}({\bf
r}) \Phi(r) \exp(-i{\bf k}{\bf r})$. In particular, the condensate-supracondensate PWF (\[psipr\])-(\[phipr\]) is characterized by the index ${\bf k}'=\pm{\bf q}/2$ and expressed via “triple" averages (\[psipk\]) which are completely neglected in the many-body $t$-matrix approximation as well as in the Bogoliubov, HFB and Popov ones [@shigrif]. Thus, one cannot associate the PWF $\varphi_{{\bf
k}',{\bf q}}({\bf r})$ with the function $\chi_{{\bf k}',{\bf
q};\omega=0}({\bf r})$ whose Fourier transform is usually defined as $\Gamma({\bf k},{\bf k}',{\bf q};\omega)=1/(2\pi)^{3}
\int\d^{3}p\,\Phi(p)\chi_{{\bf k}',{\bf q};\omega}({\bf k}-{\bf p})$ \[see Ref. [@shigrif], Eq. (A.23)\]. The latter is often called “effective wave function in a medium". Third, Eq. (\[Uk\]) is obtained by means of [*variational*]{} method [*beyond*]{} a mean-field approximation [@CSPRE; @CSJPS], while an equation for $t$-matrix is usually derived by summing a certain set of the diagrams. For this reason the relation $\widetilde{T}_{q}\simeq T_{q}=q^{2}/(2m)$ is an approximation for Eq. (\[Uk\]) while the $t$-matrix formalism deals exactly with $T_{q}$. Thus, the relation between the $t$-matrix and PWF formalism is rather subtle and beyond the scope of this paper. We only note that our definition of the PWF as the eigenfunctions of the 2-matrix is consistent with that of quantum mechanics. Namely, we operate with the functions which determine the average energy of the interaction per particle (\[uaver\]) by means of the representations (\[gr\]) in accordance with the formalism of quantum statistical mechanics. While the physical sence of the functions $\chi_{{\bf
k}',{\bf q};\omega=0}({\bf r})$ is rather unclear.
At sufficiently large momentum $k$ the main contribution in the integral in the right-hand side of Eq. (\[Uk\]) comes from the large momenta $q$. As $\lim_{q\to\infty}U(q)/\widetilde{T}_{q}=0$, for $k\to\infty$ Eq. (\[Uk\]) is also reduced to the two-body Lippmann-Schwinger equation, which leads to the boundary condition (\[bcond1\]). However, it follows from Eq. (\[Uk\]) that at finite density $n$ $\psi(k)\propto 1/k$ at small $k$, and, hence, $\psi(r) \propto 1/r^{2}$ at $r\to\infty$, in contrast to the two-body problem which implies $\psi(r)\propto 1/r$. Thus, peculiar overscreening takes place for the condensate-condensate PWF. Details concerning Eq. (\[Uk\]) can be found in Refs. [@CSPRE; @CSJPS; @CSPRE2].
With the help of the formalism of the PWF one can obtain the exact relationship between the chemical potential and the PWF (\[phir\]) and (\[phipr\]) in the presence of the Bose-Einstein condensate. This relationship is of special interest, for it allows us to see in an explicit form how the renormalization of an interatomic interaction takes place as applied to the Gross-Pitaevskii equation, at least in the homogeneous case. Let us consider at first an inhomogeneous Bose system with the Hamiltonian
------------------------------------------------------------------------
------------------------------------------------------------------------
$$\widehat H=\int\d^{3}r\,{\hat\psi}^{\dagger}({\bf r},t)
\Bigl(-\frac{\nabla^{2}}{2m}+V_{ext}({\bf r})\Bigr){\hat\psi}({\bf r},t)+
\int\d^{3}r\d^{3}r'\,\Phi(|{\bf r}-{\bf r}'|)
{\hat\psi}^{\dagger}({\bf r},t){\hat\psi}^{\dagger}({\bf r}',t)
{\hat\psi}({\bf r}',t){\hat\psi}({\bf r},t),
\label{ham}$$ where the potential $V_{ext}({\bf r})$ corresponds to the external forces, and $\Phi(r)$ is the “bare" interaction potential. Note that the Bose field operators used in Secs. \[2\] and \[3\] are related to the zero time: $\psi({\bf r})=\psi({\bf r},t=0)$, $\psi^{\dagger}({\bf r}) =
\psi^{\dagger}({\bf r},t=0)$. Using the Heisenberg equation of motion with the Hamiltonian (\[ham\]) for the Bose field operator ${\hat\psi}({\bf r},t)$ and taking average values we get $$i\frac{\partial \phi({\bf r},t)}{\partial t}=
\Bigl(-\frac{\nabla^{2}}{2m}+V_{ext}({\bf r})\Bigr)\phi({\bf r},t)+
\int\d^{3}r'\,\Phi(|{\bf r}-{\bf r}'|)
\langle{\hat\psi}^{\dagger}({\bf r}',t){\hat\psi}({\bf r}',t)
{\hat\psi}({\bf r},t)\rangle,
\label{psi0}$$ where $\phi({\bf r},t)=\langle{\hat\psi}({\bf r},t)\rangle$ \[and, respectively, $\phi^{*}({\bf r},t)=\langle{\hat\psi}^{\dagger}({\bf
r},t)\rangle$\]. On the other hand, one can make use of the Bogoliubov’s procedure of “extracting" the $c$-number parts of the Bose field operators in the Hamiltonian (\[ham\]) (see discussion in Sec. \[3\]): $$\begin{aligned}
{\hat\psi}(\r{},t)&=&\phi({\bf r},t)+{\hat\vartheta}(\r{},t),
\nonumber \\
{\hat\psi}^{\dagger}(\r{},t)&=&\phi^{*}({\bf r},t)+
{\hat\vartheta}^{\dagger}(\r{},t).
\label{psikrr1}\end{aligned}$$ Thus, in the stationary equilibrium case the grand canonical potential $\Omega$ depends on the $c$-number complex parameter $\phi({\bf r})$ that can be considered as the order parameter. So, the equilibrium value of the grand canonical potential $\Omega=\Omega(T,V,\mu;\{\phi({\bf r})\},\{\phi^{*}({\bf
r})\})$ corresponds to a minimum with respect to the order parameter (the parameter $\mu$ stands for the chemical potential). Using the well-known expression for an infinitesimal change of the potential $\delta\Omega=\langle\delta(\widehat H-\mu \widehat N)\rangle$ one can obtain from the condition $\delta\Omega/\delta\phi({\bf r})=
\delta\Omega/\delta\phi^{*}({\bf r})=0$: $$\mu\phi({\bf r})=
\Bigl(-\frac{\nabla^{2}}{2m}+V_{ext}({\bf r})\Bigr)\phi({\bf r})+
\int\d^{3}r'\,\Phi(|{\bf r}-{\bf r}'|)
\langle{\hat\psi}^{\dagger}({\bf r}'){\hat\psi}({\bf r}')
{\hat\psi}({\bf r})\rangle.
\label{statpsi0}$$ It is not difficult to see that Eq. (\[statpsi0\]) is nothing but the stationary form of Eq. (\[psi0\]) corresponding to the solution $\phi({\bf
r},t)=\phi({\bf r})\exp(-i\mu t)$ because in this case we have for the time-dependent “triple" averages $$\langle{\hat\psi}^{\dagger}({\bf r}_{2},t){\hat\psi}({\bf r}'_{2},t)
{\hat\psi}({\bf r}'_{1},t)\rangle=\exp(-i\mu t)
\langle{\hat\psi}^{\dagger}({\bf r}_{2}){\hat\psi}({\bf r}'_{2})
{\hat\psi}({\bf r}'_{1})\rangle.
\label{timeevol}$$ Indeed, the principle of the correlation weakening requires that at $|{\bf r}_{1}|\to\infty$ $$\langle{\hat\psi}^{\dagger}({\bf r}_{1}){\hat\psi}^{\dagger}({\bf r}_{2})
{\hat\psi}({\bf r}'_{2}){\hat\psi}({\bf r}'_{1})\rangle=
\langle{\hat\psi}^{\dagger}({\bf r}_{1},t){\hat\psi}^{\dagger}({\bf r}_{2},t)
{\hat\psi}({\bf r}'_{2},t){\hat\psi}({\bf r}'_{1},t)\rangle\to$$ $$\langle{\hat\psi}^{\dagger}({\bf r}_{1},t)\rangle
\langle{\hat\psi}^{\dagger}({\bf r}_{2},t)
{\hat\psi}({\bf r}'_{2},t){\hat\psi}({\bf r}'_{1},t)\rangle=
\langle{\hat\psi}^{\dagger}({\bf r}_{1})\rangle
\langle{\hat\psi}^{\dagger}({\bf r}_{2})
{\hat\psi}({\bf r}'_{2}){\hat\psi}({\bf r}'_{1})\rangle.$$
[2]{} The relation (\[timeevol\]) follows from the last equation as $\langle{\hat\psi}^{\dagger}({\bf r}_{1})\rangle= \phi^{*}({\bf r}_{1})$ and $\langle{\hat\psi}^{\dagger}({\bf r}_{1},t)\rangle= \phi^{*}({\bf
r}_{1})\exp(i\mu t)$. The stationary Gross-Pitaevskii equation can be obtained from the exact equation (\[statpsi0\]) with the help of the replacement: $$\begin{aligned}
\Phi(r)&\to&\frac{4\pi a}{m}\delta({\bf r}),
\label{replPhi} \\
\langle{\hat\psi}^{\dagger}({\bf r}'){\hat\psi}({\bf r}')
{\hat\psi}({\bf r})\rangle&\to&
\phi^{*}({\bf r}')\phi({\bf r}')
\phi({\bf r}),
\label{replpsi}\end{aligned}$$ where $a$ is the scattering length which can be defined with the help of the “bare" scattering amplitude: $4\pi a/m=U^{(0)}(0)$ \[$U^{(0)}(k)$ is the solution of Eq. (\[Uk\]) at $n=0$ and $\widetilde{T}_{q} =
T_{q}=q^{2}/(2m)$\]. However, within that approach it is not clear how the renormalization (\[replPhi\]) takes place. The most transperent way of understanding this replacement in the homogeneous case is based on the PWF formalism. In this situation $V_{ext}({\bf r})=0$, $\phi=\sqrt{n_{0}}\exp(i\chi)$, and we can put the phase $\chi$ equal to zero. So, for the homogeneous system the exact equation (\[statpsi0\]) reads $$\mu=\frac{1}{\sqrt{n_0}}\int\d^3r'\,\Phi(|{\bf r}-{\bf r}'|)
\langle{\hat\psi}^{\dagger}({\bf r}'){\hat\psi}({\bf r}')
{\hat\psi}({\bf r})\rangle.
\label{mu}$$ This equation has been obtained in the momentum representation in the Bogoliubov’s paper [@bogquasi] (see also Ref. [@string]). Using the specific expressions for the scattering parts of the condensate-condensate and supracondensate-condensate PWF given by Eqs. (\[psik\]) and (\[psipk\]), one can represent Eq. (\[mu\]) in the following form: $$\mu=n_0 U(0)+
\sqrt{2}\int\frac{\d^3q}{(2\pi)^3}n(q)
U_{{\bf q}/2}({\bf q}/2),
\label{mu1}$$ here $U_{{\bf p}}({\bf k})= \int\d^3r\,\varphi_{\bf p}({\bf r}) \Phi(r)
\exp(-i{\bf k}{\bf r})$. Equation (\[mu1\]) is the exact representation for the chemical potential via the condensate-condensate and condensate-supracondensate PWF and the occupation numbers $n(k)$. It can be used in order to obtain the density expansion for $\mu$ at $n\to 0$. With the help of the expression for $n(k)$ established within the variational scheme mentioned above [@CSPRE; @CSJPS], $$n(k)=\frac{1}{2}\Biggl(\frac{\widetilde{T}_k+n U(k)}
{\sqrt{\widetilde{T}_k^2+2n\widetilde{T}_k U(k)}}
-1\Biggr),
\label{nk}$$ one can derive for the condensate density $$\begin{aligned}
n_0&=&n\left(1-\int \frac{\d^{3}k}{(2\pi)^{3}} \frac{n(k)}{n}\right)
\nonumber \\
&=&n\left(1-\frac{8}{3\sqrt{\pi}}\sqrt{na^3} + \cdots\right),
\label{depletion}\end{aligned}$$ where the substitution ${\bf k}={\bf k}'\sqrt{2mn}$ in the integral is employed. Besides, at $n\to 0$ from Eq. (\[Uk\]) one can obtain the density correction for the in-medium scattering amplitude [@CSEPJB; @CSJPS; @CSPRE2] $$U(0) = U^{(0)}(0)\Bigl(1 + \frac{8}{\sqrt{\pi}}\sqrt{na^3}+ \cdots \Bigr).
\label{correction}$$ Using Eqs. (\[depletion\]) and (\[correction\]) and making the “scaling" substitution ${\bf q}={\bf q}'\sqrt{2mn}$ in the integral in Eq. (\[mu1\]), we arrive at the familiar density expansion [@brueck; @lee; @bel] $$\mu=nU^{(0)}(0)\left(1+\frac{32}{3\sqrt{\pi}}
\sqrt{n a^3}+\cdots\right).
\label{mudens}$$ In so doing, the explicit formula for the condensate-supracondensate PWF $\varphi_{\bf p}({\bf r})$ are needless, but only the relation $\lim_{{\bf
p}\to 0}\varphi_{\bf p}({\bf r})=\sqrt{2}\varphi(r)$ is of use. The factor $\sqrt{2}$ appears as due to the normalization condition (\[freenorm\]).
Now it is not difficult to see that in the homogeneous case the Gross-Pitaevskii equation corresponds to the simplest Hartree approximation with the renormalized potential (\[replPhi\]): $$\mu\simeq n_0 U^{(0)}(0).
\label{mupit}$$ The range of validity of this approximation is determined by the next-to-leading terms in Eqs. (\[depletion\])-(\[mudens\]) and readily expressed as $\sqrt{na^{3}}\ll 1$. In the paper of Prokakis et al. [@burnett] the criterion for validity of the Gross-Pitaevskii equation is formulated as $$n\frac{4\pi\hbar^{2}a}{m}\ll \hbar\omega,
\label{critburn}$$ where the frequency $\omega$ characterizes the harmonic trap potential. In the homogeneous case ($\omega=0$) it cannot be fulfilled, while the Gross-Pitaevskii approximation (\[replPhi\]), (\[replpsi\]) is valid, if the condition $\sqrt{na^{3}}\ll 1$ is realized. Thus, the criterion (\[critburn\]) appears to be refined.
Summary {#4}
=======
The reduced density matrix of the second order (\[rho2vtor\]) is a fundamental characteristic of a many-particle system, its eigenfunctions, PWF, being the pure states of two particles, selected in an arbitrary way, and its eigenvalues being the probabilities of finding a couple in those states. Thus, the properties of the $2$-matrix are of interest in itself and from the point of view of various applications as well.
With the Bogoliubov principle of correlation weakening we express the momentum distribution of “dissociated" pair states in terms of the single-particle occupation numbers \[see the expressions (\[wpqsing1\])-(\[wpqsing3\])\]. This allows us to represent the pair correlation function (\[f2\]) in the form (\[f2final\]), (\[f2tildfin\]), where the condensate-condensate and condensate-supracondensate PWF are explicitly given by Eqs. (\[phir\])-(\[psik\]) and (\[psipr\])-(\[phipr\]), respectively. It should be stressed that this representation is exact and do not implies any mean-field approximation as well as various model assumptions since it is solely based on the Bogoliubov’s principle of the correlation weakening. The transparent physical interpretation of the anomalous averages (\[psir\]) and (\[psipr\]) as the “scattering parts" of the corresponding PWF is proposed. The boundary conditions (\[bcond\]) are obtained in the case of hard-core, or strongly singular, potentials. The condensate-supracondensate PWF is determined by the “triple" average (\[psipk\]). The importance of these averages is emphasized in Ref. [@burnett] for the effects of condensate–exited-state interactions in the evolution of condensate mean field, which are in accordance with our results. In the textbook [@griffin] the crucial role of the contribution of the “triple" averages is discussed in the context of “dielectric formalism" approach. Those averages are not zero in so called Beliaev-Popov approximation discussed recently in [@shigrif].
Thus, the anomalous averages $\langle{\hat\vartheta}(\r{1}){\hat\vartheta}(\r{2})\rangle$ are responsible for the spatial correlations of two particles in the condensate, and do not imply that bound states of pairs of particles exist; the terms (\[f23\]) are found to be responsible for the spatial correlations of condensate-supracondensate pairs of particles and can be expected to be important in thermodynamics near the critical temperature $T_{c}$. In the HFB approach, the PWF are symmetrized plane waves except for the condensate pairs provided the condensate density is not equal to zero.
The formalism of PWF is useful to construct various approximations which take into account both short-range and long-range spatial correlations of particles in a self-consistent manner. This is due to the fact that all the pair wave functions in the expansion (\[f2final\]), (\[f2tildfin\]) are “regular channels" from the point of view of its behaviour at small $r$. Along this line one can obtain the system of non-linear integro-differential equations for PWF, which in the weak-coupling approximation leads to the Bogoliubov model of the weakly interacting Bose gas [@CSPL]. For a dilute Bose gas the non-linear integral equation (\[Uk\]) that takes into account both short- and long-range spatial correlations can be obtained with the help of variational method beyond a mean-field approach [@CSEPJB; @CSJPS; @CSPRE2]. The exact relationship between PWF and the chemical potential (\[mu1\]) allows us to obtain the density expansion for the chemical potential (\[mudens\]) in the strong-coupling regime (Bose gas is dilute, but the potential is strongly singular as the Lennard-Jones one). In this regime one can derive the expansion for the mean energy per particle in powers of the boson density $n$, the famous results [@lee; @bel] being reproduced without any divergencies which are inherent in the pseudopotential approximation [@CSEPJB; @CSJPS; @CSPRE2]. Thus, this expansion can be readily obtained from Eq. (\[mudens\]) with the thermodynamic relation $\mu=\partial (\varepsilon(n) n)/\partial n$: $$\varepsilon=\frac{1}{n}\int\limits_{0}^{n}\d n'\,\mu(n')
=\frac{2\pi\hbar^2 a n}{m}
\Bigl(1+\frac{128}{15\sqrt{\pi}}\sqrt{na^3}+\cdots\Bigr).$$ Moreover, in the framework of our approach we can obtain by [*direct*]{} calculation the interaction energy per particle (\[uaver\]) (as well as the kinetic one) [@CSEPJB]. However, this is impossible within the effective-potential approach of Refs. [@brueck; @lee; @bel], which obviously indicates the incorrect behaviour of the pair distribution function $g(r)$ (see the detailed discussions in Refs. [@CSEPJB; @CSPRE2]).
It is worth noting that the concept of PWF is helpful not only in the case of the Bose systems but also in the situation of the Fermi ones. Indeed, employing this formalism one can easily prove that below the temperature of the superconducting phase transition there always exist the bound states of fermion pairs beyond the pair condensate [@CSPRB].
The author would like to thank A. A. Shanenko for fruitful discussions.
E-mail: [email protected]
N. N. Bogoliubov, J. of Phys. (USSR) [**11**]{}, 23 (1947), reprinted in D. Pines, ed., [*The many-body problem*]{} (W.A. Benjamin, New York, 1961).
N. N. Bogoliubov, Nuovo Cimento [**7**]{}, 794 (1958); J. G. Valatin, Nuovo Cimento [**7**]{}, 843 (1958).
D. H. Kobe, Ann. Phys. (N.Y.) [**47**]{}, 15 (1968).
P. C. Hohenberg and P. C. Martin, Ann. Phys. (N.Y.) [**34**]{}, 291 (1965).
N. M. Hugenholtz and D. Pines, Phys. Rev. [**116**]{}, 489 (1959); A. Miller, D. Pines and P. Nozières, Phys. Rev. [**127**]{}, 1452 (1962); J. Gavoret and P. Nozières, Ann. Phys. (N.Y.) [**28**]{}, 349 (1964).
L. D. Landau and E. M. Lifshitz, [*Course of Theoretical Physics*]{}, Vol. 3, [*Quantum Mechanics – Non-relativistic Theory*]{} (Pergamon, New York, 1977), Sec. 14.
A. J. Coleman, Rev. Mod. Phys. [**35**]{}, 668 (1963).
J. M. Blatt, [*Theory of superconductivity*]{}, (Academic, New York, 1964).
N. N. Bogoliubov, [*Quasi-averages*]{}, preprint D-781, JINR, Dubna (1961) \[English transl. N. N. Bogoliubov, [*Lectures on Quantum Statistics*]{}, Vol. 2 (Gordon and Breach, New York, 1970) p. 1\].
K. A. Brueckner and K. Sawada, Phys. Rev. [**106**]{}, 1117 (1957); Phys. Rev. [**106**]{}, 1128 (1957); N. M. Hugenholtz and D. Pines, Phys. Rev. [**116**]{}, 489 (1959).
T. D. Lee, K. Huang and C. N. Yang, Phys. Rev. [**106**]{}, 1135 (1957).
S. T. Beliaev, Zh. Eksp. Teor. Fiz. (USSR) [**34**]{}, 433 (1958) \[Sov. Phys. JETP [**7**]{}, 299 (1958)\].
V. N. Popov, [*Functional Integrals in Quantum Field Theory and Statistical Physics*]{} (Reidel, Dordrecht, 1983).
N. P. Prokakis, K. Burnett, and H. T. C. Stoof, Phys. Rev. A [**57**]{}, 1230 (1998).
A. Yu. Cherny, A. A. Shanenko, Phys. Rev. E [**60**]{}, R5 (1999).
K. Huang and P. Tommasini, J. Res. Natl. Inst. Stand. Technol. [**101**]{}, 435 (1996).
A. Yu. Cherny, A. A. Shanenko, submitted to Eur. Phys. J. B.
A. Yu. Cherny, A. A. Shanenko, J. Phys. Studies (Lviv) [**3**]{}, No. 3, 272 (1999), the Issue to memory of N. N. Bogoliubov; E-print cond-mat/9904217.
A. Yu. Cherny, A. A. Shanenko, E-print cond-mat/9909199, submitted to Phys. Rev. E.
C. N. Yang, Rev. Mod. Phys. [**34**]{}, 694 (1962).
N. N. Bogoliubov, [*Lectures on Quantum Statistics*]{}, Vol. 1 (Gordon and Breach, New York, 1967) p. 39.
After the thermodynamic limit.
In the normal state, above $T_{c}$, it is natural to propose that the asymptotic of the pair correlation function in limit (\[limit2\]) is determined by the expression in the right hand side of Eq. (\[corr2\]) plus the “exchange" term $\langle {\hat\psi}^{\dagger}({\bf r}_1)
{\hat\psi}({\bf r}_2^{\prime})\rangle \langle{\hat\psi}^{\dagger} ({\bf
r}_2){\hat\psi} ({\bf r}_1^{\prime})\rangle$. This term certainly goes to zero in limit (\[limit2\]) since we have $\langle{\hat\psi}^{\dagger} ({\bf
r}_1){\hat\psi} ({\bf r}_2)\rangle\to \langle{\hat\psi}^{\dagger} ({\bf
r}_1)\rangle \langle{\hat\psi} ({\bf r}_2)\rangle=0$ when $|{\bf r}_1-{\bf
r}_2|\to\infty$ in accordance with the principle of the correlation weakening.
We mean the behaviour of these functions after the limit $V\rightarrow \infty$.
If the spin of any particle is not zero and there is no magnetic ordering, then the index can be represented in the form $\omega=({\bf p},S,m_{S})$, where $S$ is a complete spin of a pair, and $m_{S}$ is its $z$-projection. The following symmetry properties take place: $$\varphi_{{\bf p},{\bf q},S,m_{S}}({\bf r},\sigma_{1},\sigma_{2})=
\varphi_{{\bf p},{\bf q},S,m_{S}}(-{\bf r},\sigma_{2},\sigma_{1})$$ $$=\varphi_{-{\bf p},{\bf q},S,m_{S}}({\bf r},\sigma_{2},\sigma_{1}),$$ where $\sigma_{1}$ and $\sigma_{2}$ are single-particle $z$-projections of the spin (compare with Eq. (\[phisymm\])). The same is valid for bound states of pairs.
The Riemann’s theorem asserts that for any regular function $f({\bf q})$ (i.e. $f({\bf q})$ does not contain $\delta$-function) we have $\lim_{r \to \infty}\int \d^3 q\, f({\bf q}) \exp(i{\bf q}{\bf r})=0,$ provided the integral $\int \d^3 q\, f({\bf q}) \exp(i{\bf q}{\bf r})$ exists.
The similar idea, as applied to the electron gas, was proposed in the paper: J. C. Kimball, Phys. Rev. A [**7**]{}, 1648 (1973).
Bogoliubov introduced the notion “quasi-averages" for systems with spontaneously broken symmetry. So, by the notion “averages" we mean “quasi-averages".
For fermions the situation like this takes place for the bound pairs in the BCS-model.
P. Nozières and D. Saint James, J. Phys. (Paris) [**43**]{}, 1133 (1982).
A. Yu. Cherny, A. A. Shanenko, Phys. Rev. B [**60**]{}, 1276 (1999).
B. Baumgartner, J. Phys. A: Math. Gen. [**30**]{}, L741 (1997).
Of course, except for the “supracondensate" processes with ${\bf p}^{\prime}_{1},\ {\bf p}^{\prime}_{2},
\ {\bf p}_{1},\ {\bf p}_{2}\not=0$.
C. Bloch and C. De Dominicis, Nucl. Phys. [**7**]{}, 459 (1958).
J. G. Valatin and D. Butler, Nuovo Cimento [**10**]{}, 37 (1958).
As it is shown by P. Nozières and D. Saint James [@nozieres], bound pair states can only exist for a system of bosons with an internal spin structure when the two-body interaction $\Phi$ depends on the spins of the particles. We do not consider the spin of the particles for the sake of simplicity, though we can easily generalize the concept of the PWF for this case [@bogquasi; @note4; @CSPRB].
In the Bogoliubov theory, $F^{(3)}_{2}=F^{(4)}_{2}=0$. We keep the term $\langle {\hat\vartheta}^{\dagger}(\r{1})
{\hat\vartheta}^{\dagger}(\r{2}) \rangle \langle {\hat\vartheta}(\rp{2})
{\hat\vartheta}(\rp{1}) \rangle$ from $F^{(4)}_{2}$ for retaining the property of non-negative definition of the kernel (\[f2bog\]).
A. Yu. Cherny, A. A. Shanenko, Phys. Lett. A [**250**]{}, 170 (1998).
H. Shi and A. Griffin, Phys. Rep. [**304**]{}, 1 (1998).
S. Stringary, J. Low Temp. Phys. [**84**]{}, 279 (1991), we notice the differences in notations.
A. Griffin, [*Excitations in a Bose-Condensed Liquid*]{} (Cambridge University Press, New York, 1993).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report on an optical imaging design based on common-path off-axis digital holography, using a multiplexed volume Bragg grating. In the reported method, a reference optical wave is made by deflection and spatial filtering through a volume Bragg grating. This design has several advantages including simplicity, stability and robustness against misalignment.'
author:
- Leo Puyo
- 'Jean-Pierre Huignard'
- Michael Atlan
title: ' Off-axis digital holography with multiplexed volume Bragg gratings '
---
Introduction
============
The success of the Zernike phase-contrast microscope [@Zernike1942] has led to a number of subsequent phase-imaging methods, including Nomarski differential interference contrast microscopy [@Lang1968] and Hoffman modulation contrast microscopy [@HoffmanGross1975]. Those phase-contrast methods merge brightness and phase information in the recorded image. Later on, quantitative phase shifts of an object wave beating against a reference wave were measured by phase-contrast digital holography [@CucheBevilacqua1999]. In digital holography, this phase is derived from the measurement of spatiotemporal variations in intensity of the interference pattern, digitized by a sensor array, and digital image rendering by wave propagation computation. Since then, holographic phase microscopy has emerged as a powerful imaging technique for probing quantitatively refractive index changes in transparent samples [@MarquetRappaz2005]. Optical phase microscopy techniques based on interferometric recordings and computational image rendering have become widely available since then [@MarquetDepeursingeMagistretti2014; @McLeodOzcan2016; @MaRajshekhar2016], and have opened the way to tomographic phase microscopy [@Wolf1969; @CharriereMarian2006; @HaeberleBelkebir2010; @CotteToy2013; @JinZhou2017].\
On-axis interferometry, the original Gabor configuration for hologram recording [@Gabor1948], has the advantage of being easilly realized experimentally, but lacks of sensitivity [@BhardwajlPanigrahi2016] and might prevent accurate phase reconstructions [@Qu2010; @Jericho2012]. In contrast, in off-axis recording configuration [@LeithUpatnieks1962], the average propagation directions of the reference and the object optical waves, interfering in the sensor plane, are slightly tilted. The spatial spectrum of the hologram is shifted by a quantity which scales as the average tilt angle. This tilt permits the separation and discrimination of self-beating and cross-beating interferometric contributions of the object and the reference optical fields; spurious interferometric contributions can be filtered-off [@Cuche2000], and phase imaging can be performed straightforwardly [@CucheBevilacqua1999]. In addition, the optical pathes of light impinging on the sensor can have either both probed the sample in common-path configurations or taken different pathways in separate arms configurations. Common-path interferometry configurations, for which both waves travel the same pathways (quasi-)through the sample [@Weijuan2009; @Grishin2016; @SerabynLiewerLindensmith2016; @RostykusMoser2017] prevent phase noise from pathlength fluctuations of the reference beam decorrelated from the object beam, mode hops, and increase the overall stability of the interferogram. A significant improvement of off-axis interferometry with short coherence radiation was achieved by the use of diffraction gratings to tilt the coherence plane with respect to the direction of propagation of the reference wave, so that cross-interference patterns cover all the detector array [@Monemhaghdoust2011; @ChoiYang2012; @DuboisYourassowsky2012; @Witte2012; @MonemahghdoustMonfort2013; @SlabyKolman2013; @GuoWangHu2017].\
We propose here a new simple off-axis digital holographic microscopic scheme inspired by 1- point-diffraction interferometry approaches [@Smartt1975; @MedeckiTejnilGoldberg1996; @GaoYaoMin2011], and their refinements [@IndebetouwKlysubun1999; @PopescuIkedaDasariFeld2006; @WangMillet2011; @Shaked2012], in which a reference optical wave is formed by spatial filtering of a replica of the object wave. 2- off-axis implementations of phase-contrast digital holographic microscopy [@CucheBevilacqua1999; @GirshovitzShaked2013; @LuLiuLau2014; @Grishin2016; @WallaceRider2015], 3- lateral shearing interferometry [@Primot1993; @BonMaucort2009], 4- the angular filtering properties of volume Bragg gratings [@Kogelnik1969; @Ludman1981], recorded in the volume of doped glasses with photothermorefractive processes [@EfimovGlebov1999], and 5- an implementation of volume Bragg gratings for obtaining digital holograms in a self-reference configuration in a conventional microscope [@MonemhaghdoustMontfort2014]. In the reported design, a reference optical wave, suitable for off-axis holographic interferometry is made by deflection through a volume Bragg grating.
Experimental setup
==================
Optical configuration
---------------------
In the experimental setup sketched in Fig. \[fig\_Setup\], a microscopic object is illuminated in transmission by a laser (Crystalaser DL660-100, wavelength $\lambda\sim$660 nm, coherence length 0.3 mm). The transmitted object field can be magnified by a microscope objective or merely occulted partially by a pupil, before passing through a diffractive optical element, and impinges on the sensor array of a camera (Ximea MQ042MG-CM, array size: 2048$\times$2048 pixels, pitch: $5.5 \, \mu \rm m$). In the absence of microscope objective, the angular bandwidth $(\Delta \theta_0, \Delta \vartheta_0)$ of the object field $E_{00}$ impinging on the sensor array is limited by the aperture stop of the system, which can either be a pupil introduced between the object and the sensor, or the sensor itself. In the former case, $\Delta \theta_0 \approx {a_x}/{l}$, and $\Delta \vartheta_0 \approx {a_y}/{l}$, where $a_x$ and $a_y$ are the aperture dimensions in the $x$ and $y$ directions, and $l$ is the object-to-pupil distance. In the latter case, $\Delta \theta_0 \approx {D_x}/{L}$, and $\Delta \vartheta_0 \approx {D_y}/{L}$, where $D_x$ and $D_y$ are the sensor dimensions in the $x$ and $y$ directions, and $L$ is the object-to-sensor distance. If a microscope objective is present, the angular bandwidth of the object field is limited by the numerical aperture ${\rm NA}$ of the objective : $\Delta \theta_0 = \Delta \vartheta_0 = 2 \arcsin({\rm NA})$. A usual method to engineer a reference optical wave suitable for off-axis holographic interferometry from the object wave itself consists in reducing the spatial frequency content of the latter by low-pass filtering of the transverse projections of the transmitted angular wavenumbers $(k_x,k_y)$ with a pinhole [@Smartt1975; @IndebetouwKlysubun1999; @IndebetouwKlysubun2001; @KlysubunIndebetouw2001], or by defocus by curved mirrors [@HongKim2013], a spatial light modulator [@RosenBrooker2007], or lenses [@NaikPedrini2014; @MuhammadNguyen2016; @SerabynLiewerLindensmith2016].
Angular filtering by a multiplexed Bragg grating
------------------------------------------------
In our novel approach, a reference optical wave is made by deflection at the Bragg angle through a multiplexed grating. For this purpose, a transmission volume Bragg grating created by double interferometric exposure in a photothermorefractive glass was realized by OptiGrate [@EfimovGlebov1999; @CiapurinGlebov2006]. It displays a periodic change in the refractive index $n$ of the form $$\label{eq_RefractiveIndex}
n(x,y) = n_0 + n_1 \sin \left( \frac{2 \pi x}{\Lambda_x} \right) + n_2 \sin \left( \frac{2 \pi y}{\Lambda_y} \right)$$ where $n_0 = 1.498$ is the average refractive index of the glass, and $n_1 = n_2 \sim 1.4\times 10^{-5}$ is its modulation depth in each transverse direction, $x$, and $y$. It acts as a superposition of two perpendicular phase gratings, sketched in Fig. \[fig\_FilteringConfiguration\]. The thickness of the multiplexed grating is $d = 8.9 \, \rm mm$. The grating periods are $\Lambda_x = 18.9 \, \mu \rm m$, and $\Lambda_y = 18.3 \, \mu \rm m$. The relative diffraction efficiency (for plane waves, when the Bragg condition is fulfilled) of the horizontal and the vertical grating are $|E_{01}|^2/(|E_{00}|^2+|E_{01}|^2) = 0.56$, and $|E_{10}|^2/(|E_{00}|^2+|E_{10}|^2) = 0.55$.\
The Bragg condition defines the angle of deflection of a given input beam. For an unslanted grating, in the configuration sketched in Fig. \[fig\_FilteringConfiguration\], we have $$\label{eq_BraggAngle}
2 \sin( \theta_{\rm B} ) = \frac{\lambda}{\Lambda_x} \quad {\rm and} \quad 2 \sin( \vartheta_{\rm B} ) = \frac{\lambda}{\Lambda_y}$$ where $\Lambda_x$ and $\Lambda_y$ are the grating periods. The spatial filtering properties of transmission volume Bragg gratings constrain the degree of collimation of the input beam allowed to be deflected by twice the Bragg angles. The angular efficiency, defined by the grating angular diffraction efficiency [@Goodman2005] sets the angular support $\Delta\theta_1$ and $\Delta\vartheta_1$ of the deflected fields in $x$ and $y$ respectively, given by $$\label{eq_AngularEfficiency}
\Delta \theta_1 \approx n_0 \frac{\Lambda_x}{d} \quad {\rm and} \quad \Delta \vartheta_1 \approx n_0 \frac{\Lambda_y}{d}$$ where $n_0$ is the average refractive index of the material, and $d$ is the thickness of the grating [@Kogelnik1969; @Ludman1981]. When both Bragg conditions are fulfilled for the multiplexed Bragg grating (Fig. \[fig\_FilteringConfiguration\](c)), the input angles of the object beam are $\theta_{\rm B}$ and $\vartheta_{\rm B}$, the deviation angles of the first diffraction orders are predicted by Bragg’s law (Eq. \[eq\_BraggAngle\]), and only low spatial frequencies $k_x \in [-k\Delta\theta_1/2,k\Delta\theta_1/2]$ and $k_y \in [-k\Delta\vartheta_1/2,k\Delta\vartheta_1/2]$ are allowed to pass in the deflected beam, yet the spatial frequency content of the non-deflected beam is unaffected by transmission through the grating. The angular representation of the corresponding optical fields, $E_{00}$ and $E_{11}$ respectively, are reported in Fig \[fig\_DiffractionOrders\]. The shift and angular bandwidth of the fields $E_{10}$ and $E_{01}$ deflected by the first and second regular gratings of Fig. \[fig\_FilteringConfiguration\] are also reported in Fig \[fig\_DiffractionOrders\]. The spatial filtering property is the key to the realization of the tilted reference wave in common-path transmission interferometric configuration. With the chosen grating thickness of $d \simeq 9 {\,\rm mm}$, this angular filter of acceptance $\Delta\theta_1=\Delta\vartheta_1\simeq 3.3 \,{\rm mrad}$ is equivalent to the angular selectivity $\sim D/f \simeq 3.3 \,{\rm mrad}$ of a pinhole of diameter $D = 330 \,\mu {\rm m}$, set in the focal plane of a converging lens of focal length $f = 10 \,{\rm cm}$ in a point-diffraction interferometer [@Smartt1975; @MedeckiTejnilGoldberg1996; @GaoYaoMin2011]. The choice of grating periods $\Lambda_x$ and $\Lambda_y$ is dictated by the Nyquist–Shannon sampling theorem : the deflection angles of the propagation directions of $E_{11}$ and $E_{00}$ have to satisfy $2\theta_{\rm B} \in [-\theta_{\rm S}/2,\theta_{\rm S}/2]$, and $2\vartheta_{\rm B} \in [-\vartheta_{\rm S}/2,\vartheta_{\rm S}/2]$ where $$\label{eq_AntennaAcceptanceAngle}
\theta_{\rm S} \approx \frac{\lambda}{d_x} \quad {\rm and} \quad \vartheta_{\rm S} \approx \frac{\lambda}{d_y}$$ are the angular acceptances of the coherent detection on an array detector [@Siegman1966], with pixels spaced by $d_{x}$ and $d_{y}$ along $x$ and $y$, respectively. The pixel pitches of the camera used in the experiments are $d_{x} = d_{y} = 5.5 \, \mu \rm m$. Eq. \[eq\_BraggAngle\], Eq. \[eq\_AngularEfficiency\], and Eq \[eq\_AntennaAcceptanceAngle\] are valid for small input and output angles, in the paraxial approximation [@Goodman2005]. In order to achieve typical off-axis detunings between the object and the reference waves, deflection angles have to be about one half of the Nyquist angles $\theta_{\rm S}/2$ and $\vartheta_{\rm S}/2$ typically, and hence satisfy the relations $2\theta_{\rm B} \sim \theta_{\rm S}/4$, and $2\vartheta_{\rm B} \sim \vartheta_{\rm S}/4$, which impose grating periods of $\Lambda_x \sim 4 d_x$ and $\Lambda_y \sim 4 d_y$.
Experimental validation of angular filtering and deflection by a multiplexed Bragg grating
------------------------------------------------------------------------------------------
To illustrate the spatial filtering properties of the Bragg orders, we placed a converging lens between the grating and the sensor array to form the image of the resolution target in the detection plane. In Fig. \[fig\_ExperimentalBraggOrdersLaser658nmImgFullTarget\](a), the Bragg condition is not met for any beam direction, only the non-deflected beam (field $E_{00}$) is transmitted. The angular bandwidth of the object field $E_{00}$ is completely transmitted through the grating, and spatial frequencies are bounded by $(k\Delta\theta_0, k\Delta\vartheta_0)$, as sketched in Fig. \[fig\_DiffractionOrders\]. In Fig. \[fig\_ExperimentalBraggOrdersLaser658nmImgFullTarget\](b), the Bragg condition is met only for $\theta$ tilt ($x$ direction); the non-deflected beam (field $E_{00}$) and the beam deflected by $2\theta_{\rm B}$ (field $E_{10}$) are transmitted. The angular bandwidth of the deflected field $E_{10}$ is reduced. Its spatial frequencies are bounded by $(k\Delta\theta_1, k\Delta\vartheta_0)$, as sketched in Fig. \[fig\_DiffractionOrders\]. Hence, horizontal bars of the deflected image are no longer visible. In Fig. \[fig\_ExperimentalBraggOrdersLaser658nmImgFullTarget\](c), the Bragg condition is met only for $\vartheta$ tilt ($y$ direction); the non-deflected beam (field $E_{00}$) and the beam deflected by $2\vartheta_{\rm B}$ (field $E_{01}$) are transmitted. The angular bandwidth of the deflected field $E_{01}$ is reduced. Its spatial frequencies are bounded by $(k\Delta\theta_0, k\Delta\vartheta_1)$, as sketched in Fig. \[fig\_DiffractionOrders\]. Hence, vertical bars of the deflected image are no longer visible. When the Bragg condition is fulfilled for both directions, the non-deflected beam $E_{00}$, the deflected and spatially-filtered beams $E_{10}$ and $E_{01}$ are transmitted, as reported in Fig. \[fig\_ExperimentalBraggOrdersLaser658nmImgFullTarget\](d). A fourth beam, deflected both in $\theta$ and $\vartheta$ by $2\theta_{\rm B}$ and $2\vartheta_{\rm B}$ respectively is transmitted (field $E_{11}$). The angular bandwidth of the deflected field $E_{11}$ is reduced. Its spatial frequencies are bounded by $(k\Delta\theta_1, k\Delta\vartheta_1)$, as sketched in Fig. \[fig\_DiffractionOrders\]. Hence, both horizontal and vertical bars of the deflected image are no longer visible.
Off-axis hologram reconstruction
================================
Fresnel transformation
----------------------
Holograms were reconstructed by Fresnel transformation [@Goodman1967; @SchnarsJuptner1994; @Schnars2002; @KimYuMann2006; @PicartLeval2008; @VerrierAtlan2011; @PiedrahitaQuintero2015; @Latychevskaia2015]; in practice, we used the software Holovibes [@Holovibes] for real-time hologram rendering, which performed image rendering of complex-valued holograms $H(x,y,t)$ from the stream of rescaled interferograms $I(x,y,t)$ $$\begin{aligned}
\label{eq_FresnelTransform}
\nonumber H(x,y,t) = \frac{i}{\lambda z}\exp \left( -ikz \right) \iint I(x',y',t)\\
\times \exp \left[\frac{-i \pi}{\lambda z} \left((x-x')^2 + (y-y')^2\right) \right] {\rm d}x' {\rm d}y'\end{aligned}$$ The parameter $z$ is the reconstruction distance for which the image of the target appears on the magnitude of the hologram $|H|$. The interferometric contributions in the interferogram plane that create fringe sets (Fig. \[fig\_InterferogramAndHologramLaser658nm\](a), and Fig. \[fig\_InterferogramAndHologramLaser658nm\](b)) are separated in the Fourier reciprocal plane (Fig. \[fig\_InterferogramAndHologramLaser658nm\](c)) and the hologram reconstruction plane (Fig. \[fig\_InterferogramAndHologramLaser658nm\](e)).
Amplitude imaging
-----------------
In order to record holographic interferograms, we removed the lens between the grating and the sensor array, and we increased the aperture stop of the system to widen the lateral extension of the deflected fields, and make the four contributions interfere in an overlapping region covering most of the sensor area, as shown in the experimental interferogram displayed in Fig. \[fig\_InterferogramAndHologramLaser658nm\](a). A zoomed view in the region of interest depicted by the box, showing the juxtaposition of horizontal and vertical fringes, is reported in Fig. \[fig\_InterferogramAndHologramLaser658nm\](b). When the Bragg conditions are fulfilled for both directions, angular spectra of $E_{10}$, $E_{01}$, and $E_{11}$ are filtered by the Bragg grating angular selectivity, as sketched in Fig. \[fig\_DiffractionOrders\]. The total transmitted field is the sum of four components forming an interferogram on the sensor array $I = |E_{\rm t}|^2$, where $$\label{eq_SumOfFields}
E_{\rm t} = E_{00} + E_{10} + E_{01} + E_{11}$$ Each set of fringes corresponds to the interference between couples of field components, except self-beating contributions. The reconstructed hologram $H$ by linear Fresnel transformation (Eq. \[eq\_FresnelTransform\]) shifts those contributions according to this fringe structure. The spatial Fourier transform ${\cal F} \{ I \}$ of the interferogram $I$ displayed in Fig. \[fig\_InterferogramAndHologramLaser658nm\](a) takes the form $$\label{eq_SpatialConvolutionFields}
{\cal F} \{ I \} = E_{\rm t} * E_{\rm t}^{*}$$ where $*$ is the spatial convolution product. For the sake of notation simplicity, the fields $E$ either refer to distributions in the sensor plane, or at the exit face of the volume Bragg grating, or their reciprocal planes. The magnitude $|{\cal F} \{ I \}|$ is displayed in logarithmic scale in Fig. \[fig\_InterferogramAndHologramLaser658nm\](c), on which one can see nine diffraction locations of the 16 interferometric terms. A sketch of the 16 interference terms locations in a reciprocal plane of the interferogram is reported in Fig. \[fig\_InterferometricContributions\].\
Among the 16 interferometric terms of ${\cal F} \{ I \}$, the term $E_{00} * E_{11}^{*}$ highlighted in Fig. \[fig\_InterferogramAndHologramLaser658nm\](e), and displayed in Fig. \[fig\_InterferogramAndHologramLaser658nm\](d), and its complex conjugate $E_{00}^{*} * E_{11}$ are shifted in opposite corners of the reconstructed hologram, and used as off-axis holograms. These contributions are the result of the interference between the transmitted field $E_{00}$, and the field of reduced angular support in both directions $(x,y)$ $E_{11}$ which acts as a flattened reference wave. In the remaining corners, the beating contributions $E_{10} * E_{01}^{*}$ and $E_{10}^{*} * E_{01}$ are present. These terms are the result of the beat between transmitted and partly filtered fields in the horizontal and vertical directions. In the four sides, the beating contributions $E_{10} * E_{01}^{*}$ and $E_{10}^{*} * E_{01}$ are present. These terms are the result of the beat between transmitted and partly filtered fields in the horizontal and vertical directions.
Phase imaging
-------------
In order to obtain phase images, holograms were processed to remove artifacts generated by the off-axis configuration [@TrujilloCastaneda2016] and the phase curvature [@KreuzerJericho2001] of the reference wave. The interferometric order $E_{00} * E_{11}^{*}$ was selected and re-centered in the Fourier space while the rest of the Fourier components were cropped out [@CucheBevilacqua1999; @TrujilloCastaneda2016]. These operations allowed for the removal of signal corresponding to other interferometric contribution and also of the off-axis phase tilt. A Fresnel transformation (Eq. \[eq\_FresnelTransform\]) was carried out onto the interferograms acquired in the presence \[Fig. \[fig\_PhaseAndAmplitudeHologramRendering\](a)\] and absence \[Fig. \[fig\_PhaseAndAmplitudeHologramRendering\](b)\] of target. The phase image reported in Fig. \[fig\_PhaseAndAmplitudeHologramRendering\](g) is the difference between the phase of the reconstructed holograms acquired with \[Fig. \[fig\_PhaseAndAmplitudeHologramRendering\](e)\] and without \[Fig. \[fig\_PhaseAndAmplitudeHologramRendering\](f)\] target. The accuracy of the phase measurement might be hindered by a not perfectly flat reference wave in the reported results.
Conclusions
===========
In conclusion, the reported off-axis common-path holographic interferometer design performs spatial filtering with a multiplexed volume Bragg grating, which makes it suited to the detection of optical absorption and index variations. The described interferometer is lensless, and may be adapted to microscopic imaging. Lateral resolution, robustness against vibration and aberration, spatiotemporal coherence requirements of the radiation and accuracy of phase imaging might be further investigated. The reported results made use of one thick grating; alternatively, the use of two thinner multiplexed gratings could result in the same deflection and filtering properties [@OttSeGall2015].
Funding {#funding .unnumbered}
=======
This work was supported by the Investments for the Future program (LabEx WIFI: ANR-10-LABX-24, ANR-10-IDEX- 0001-02 PSL\*), and European Research Council (ERC Synergy HELMHOLTZ, $\#$610110).
[10]{}
Frits Zernike. Phase contrast, a new method for the microscopic observation of transparent objects. , 9(7):686–698, 1942.
Walter Lang. . Oberkochen, Carl Zeiss, 1982.
Robert Hoffman and Leo Gross. Modulation contrast microscope. , 14(5):1169–1176, 1975.
Etienne Cuche, Frédéric Bevilacqua, and Christian Depeursinge. Digital holography for quantitative phase-contrast imaging. , 24(5):291–293, 1999.
P. Marquet, B. [Rappaz]{}, P. J. [Magistretti]{}, E. [Cuche]{}, Y. [Emery]{}, T. [Colomb]{}, and C. [Depeursinge]{}. Digital holographic microscopy: a noninvasive contrast imaging technique allowing quantitative visualization of living cells with subwavelength axial accuracy. , 30:468–470, March 2005.
Pierre Marquet, Christian Depeursinge, and Pierre J Magistretti. Review of quantitative phase-digital holographic microscopy: promising novel imaging technique to resolve neuronal network activity and identify cellular biomarkers of psychiatric disorders. , 1(2):020901–020901, 2014.
Euan McLeod and Aydogan Ozcan. Unconventional methods of imaging: computational microscopy and compact implementations. , 79(7):076001, 2016.
Lihong Ma, Gannavarpu Rajshekhar, Ru Wang, Basanta Bhaduri, Shamira Sridharan, Mustafa Mir, Arindam Chakraborty, Rajashekar Iyer, Supriya Prasanth, Larry Millet, et al. Phase correlation imaging of unlabeled cell dynamics. , 6, 2016.
Emil Wolf. Three-dimensional structure determination of semi-transparent objects from holographic data. , 1(4):153–156, 1969.
F. Charri[è]{}re, A. Marian, F. Montfort, J. Kuehn, T. Colomb, E. Cuche, P. Marquet, and C. Depeursinge. Cell refractive index tomography by digital holographic microscopy. , 31(2):178–180, 2006.
Olivier Haeberl[é]{}, Kamal Belkebir, H Giovaninni, and Anne Sentenac. Tomographic diffractive microscopy: basics, techniques and perspectives. , 57(9):686–699, 2010.
Yann Cotte, Fatih Toy, Pascal Jourdain, Nicolas Pavillon, Daniel Boss, Pierre Magistretti, Pierre Marquet, and Christian Depeursinge. Marker-free phase nanoscopy. , 7(2):113–117, 2013.
Di Jin, Renjie Zhou, Zahid Yaqoob, and Peter TC So. Tomographic phase microscopy: principles and applications in bioimaging. , 34(5):B64–B77, 2017.
D. Gabor. A new microscopic principle. , 161:777–778, 1948.
Chanchal bhardwaj and Pradipta Panigrahi. Comparison between digital inline and off-axis holography system for particle field characterization. , page W3A.8, 2016.
Weijuan Qu, Oi Choo Chee, Yingjie Yu, and Anand Asundi. Recording and reconstruction of digital gabor hologram. , 121(23):2179–2184, 2010.
MH Jericho, HJ Kreuzer, M Kanka, and R Riesenberg. Quantitative phase and refractive index measurements with point-source digital in-line holographic microscopy. , 51(10):1503–1515, 2012.
E.N. Leith and J. Upatnieks. Reconstructed wavefronts and communication theory. , 52(10):1123–1128, 1962.
Etienne Cuche, Pierre Marquet, and Christian Depeursinge. Spatial filtering for zero-order and twin-image elimination in digital off-axis holography. , 39(23):4070, 2000.
Qu Weijuan, Yu Yingjie, Chee Oi Choo, and Anand Asundi. Digital holographic microscopy with physical phase compensation. , 34(8):1276–1278, 2009.
Oleg V Grishin, Ivan V Fedosov, and Valery V Tuchin. Simple technique of fourier-transform holographic microscope with compensation of phase aberration. In [*Saratov Fall Meeting 2015*]{}, pages 99171W–99171W. International Society for Optics and Photonics, 2016.
Eugene Serabyn, Kurt Liewer, Chris Lindensmith, Kent Wallace, and Jay Nadeau. Compact, lensless digital holographic microscope for remote microbiology. , 24(25):28540–28548, 2016.
Manon Rostykus and Christophe Moser. Compact lensless off-axis transmission digital holographic microscope. , 25(14):16652–16659, Jul 2017.
Zahra Monemhaghdoust, Fr[é]{}d[é]{}ric Montfort, Yves Emery, Christian Depeursinge, and Christophe Moser. Dual wavelength full field imaging in low coherence digital holographic microscopy. , 19(24):24005–24022, 2011.
Youngwoon Choi, Taeseok Daniel Yang, Kyoung Jin Lee, and Wonshik Choi. Full-field and single-shot quantitative phase microscopy using dynamic speckle illumination. , 36(13):2465–2467, 2011.
F Dubois and C Yourassowsky. Full off-axis red-green-blue digital holographic microscope with led illumination. , 37(12):2190–2192, 2012.
Stefan Witte, Andrius Plau[ş]{}ka, Margreet C Ridder, Laura van Berge, Huibert D Mansvelder, and Marie Louise Groot. Short-coherence off-axis holographic phase microscopy of live cell dynamics. , 3(9):2184–2189, 2012.
Zahra Monemhaghdoust, Frederic Montfort, Etienne Cuche, Yves Emery, Christian Depeursinge, and Christophe Moser. Full field vertical scanning in short coherence digital holographic microscope. , 21(10):12643–12650, 2013.
Tom[á]{}[š]{} Slab[y]{}, Pavel Kolman, Zbyn[ě]{}k Dost[á]{}l, Martin Anto[š]{}, Martin Lo[š]{}t’[á]{}k, and Radim Chmel[í]{}k. Off-axis setup taking full advantage of incoherent illumination in coherence-controlled holographic microscope. , 21(12):14747–14762, 2013.
Rongli Guo, Fan Wang, Xiaoying Hu, and Wenqian Yang. Off-axis low coherence digital holographic interferometry for quantitative phase imaging with an led. , 2017.
RN Smartt and WH Steel. Theory and application of point-diffraction interferometers. , 14(S1):351, 1975.
H Medecki, E Tejnil, KA Goldberg, and J Bokor. Phase-shifting point diffraction interferometer. , 21(19):1526–1528, 1996.
Peng Gao, Baoli Yao, Junwei Min, Rongli Guo, Juanjuan Zheng, Tong Ye, Irina Harder, Vanusch Nercissian, and Klaus Mantel. Parallel two-step phase-shifting point-diffraction interferometry for microscopy based on a pair of cube beamsplitters. , 19(3):1930–1935, 2011.
Guy Indebetouw and Prapong Klysubun. Space-time digital holography: A three-dimensional microscopic imaging scheme with an arbitrary degree of spatial coherence. , 75(14):2017–2019, 1999.
Gabriel Popescu, Takahiro Ikeda, Ramachandra R Dasari, and Michael S Feld. Diffraction phase microscopy for quantifying cell structure and dynamics. , 31(6):775–777, 2006.
Zhuo Wang, Larry Millet, Mustafa Mir, Huafeng Ding, Sakulsuk Unarunotai, John Rogers, Martha U Gillette, and Gabriel Popescu. Spatial light interference microscopy (slim). , 19(2):1016–1026, 2011.
Natan T Shaked. Quantitative phase microscopy of biological samples using a portable interferometer. , 37(11):2016–2018, 2012.
Pinhas Girshovitz and Natan T Shaked. Compact and portable low-coherence interferometer with off-axis geometry for quantitative phase microscopy and nanoscopy. , 21(5):5701–5714, 2013.
Yujie Lu, Yunhui Liu, and Tak Kit Lau. Simple, portable, and low-cost microscope based on off-axis digital holography using two spherical waves. , 39(15):4549–4552, 2014.
J Kent Wallace, Stephanie Rider, Eugene Serabyn, Jonas K[ü]{}hn, Kurt Liewer, Jody Deming, Gordon Showalter, Chris Lindensmith, and Jay Nadeau. Robust, compact implementation of an off-axis digital holographic microscope. , 23(13):17367–17378, 2015.
Jerome Primot. Three-wave lateral shearing interferometer. , 32(31):6242–6249, 1993.
Pierre Bon, Guillaume Maucort, Benoit Wattellier, and Serge Monneret. Quadriwave lateral shearing interferometry for quantitative phase microscopy of living cells. , 17(15):13080–13094, 2009.
Herwig Kogelnik. Coupled wave theory for thick hologram gratings. , 48(9):2909–2947, 1969.
Jacques E Ludman. Approximate bandwidth and diffraction efficiency in thick holograms. , 50(3):244–246, 1982.
Oleg M Efimov, Leonid B Glebov, Larissa N Glebova, Kathleen C Richardson, and Vadim I Smirnov. High-efficiency bragg gratings in photothermorefractive glass. , 38(4):619–627, 1999.
Zahra Monemhaghdoust, Fr[é]{}d[é]{}ric Montfort, Yves Emery, Christian Depeursinge, and Christophe Moser. Off-axis digital holographic camera for quantitative phase microscopy. , 5(6):1721–1730, 2014.
Joseph W Goodman. . Roberts and Company Publishers, 2005.
G. Indebetouw and P. Klysubun. Spatiotemporal digital microholography. , 18:319–325, February 2001.
Prapong Klysubun and Guy Indebetouw. A posteriori processing of spatiotemporal digital microholograms. , 18(2):326–331, 2001.
Jisoo Hong and Myung K Kim. Single-shot self-interference incoherent digital holography using off-axis configuration. , 38(23):5196–5199, 2013.
Joseph Rosen and Gary Brooker. Digital spatially incoherent fresnel holography. , 32(8):912–914, 2007.
Dinesh N Naik, Giancarlo Pedrini, Mitsuo Takeda, and Wolfgang Osten. Spectrally resolved incoherent holography: 3d spatial and spectral imaging using a mach–zehnder radial-shearing interferometer. , 39(7):1857–1860, 2014.
Dilband Muhammad, Cuong M Nguyen, Jihoon Lee, and Hyuk-Sang Kwon. Spatially incoherent off-axis fourier holography without using spatial light modulator (slm). , 24(19):22097–22103, 2016.
Igor V Ciapurin, Leonid B Glebov, and Vadim I Smirnov. Modeling of phase volume diffractive gratings, part 1: transmitting sinusoidal uniform gratings. , 45(1):015802–015802, 2006.
A.E. Siegman. The antenna properties of optical heterodyne receivers. , 5(10):1588, 1966.
J. W. Goodman and R. W. Lawrence. digital formation of images from electronically detected holograms. , 11(3):77–79, 1967.
U. Schnars and W. Juptner. Direct recording of holograms by a ccd target and numerical reconstruction. , 33:179–181, 1994.
U. Schnars and W. P. O. Juptner. Digital recording and numerical reconstruction of holograms. , 13:R85–R101, 2002.
Myung K Kim, Lingfeng Yu, and Christopher J Mann. Interference techniques in digital holography. , 8(7):S518–S523, 2006.
Pascal Picart and Julien Leval. General theoretical formulation of image formation in digital fresnel holography. , 25(7):1744–1761, 2008.
Nicolas Verrier and Michael Atlan. Off-axis digital hologram reconstruction: some practical considerations. , 50(34):H136–H146, Dec 2011.
Pablo Piedrahita-Quintero, Raul Casta[ñ]{}eda, and Jorge Garcia-Sucerquia. Numerical wave propagation in imagej. , 54(21):6410–6415, 2015.
Tatiana Latychevskaia and Hans-Werner Fink. Practical algorithms for simulation and reconstruction of digital in-line holograms. , 54(9):2424–2434, 2015.
http://holovibes.com.
Carlos Trujillo, Ra[ú]{}l Casta[ñ]{}eda, Pablo Piedrahita-Quintero, and Jorge Garcia-Sucerquia. Automatic full compensation of quantitative phase imaging in off-axis digital holographic microscopy. , 55(36):10299–10306, 2016.
HJ Kreuzer, MJ Jericho, IA Meinertzhagen, and W. Xu. Digital in-line holography with photons and electrons. , 13(47):10729–10741, 2001.
Daniel Ott, Marc SeGall, Ivan Divliansky, George Venus, and Leonid Glebov. High-contrast filtering by multipass diffraction between paired volume bragg gratings. , 54(31):9065–9070, 2015.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Subject matching performance in iris biometrics is contingent upon fast, high-quality iris segmentation. In many cases, iris biometrics acquisition equipment takes a number of images in sequence and combines the segmentation and matching results for each image to strengthen the result. To date, segmentation has occurred in 2D, operating on each image individually. But such methodologies, while powerful, do not take advantage of potential gains in performance afforded by treating sequential images as volumetric data. As a first step in this direction, we apply the Flexible Learning-Free Reconstructoin of Neural Volumes (FLoRIN) framework, an open source segmentation and reconstruction framework originally designed for neural microscopy volumes, to volumetric segmentation of iris videos. Further, we introduce a novel dataset of near-infrared iris videos, in which each subject’s pupil rapidly changes size due to visible-light stimuli, as a test bed for FLoRIN. We compare the matching performance for iris masks generated by FLoRIN, deep-learning-based (SegNet), and Daugman’s (OSIRIS) iris segmentation approaches. We show that by incorporating volumetric information, FLoRIN achieves a factor of 3.6 to an order of magnitude increase in throughput with only a minor drop in subject matching performance. We also demonstrate that FLoRIN-based iris segmentation maintains this speedup on low-resource hardware, making it suitable for embedded biometrics systems.'
author:
- |
Jeffery Kinnison\
University of Notre Dame\
Notre Dame, Indiana, USA\
[[email protected]]{}
- |
Mateusz Trokielewicz\
Research and Academic Computer Network\
Warsaw, Poland\
[[email protected]]{}\
- |
Camila Carballo\
University of Notre Dame\
Notre Dame, Indiana, USA\
[[email protected]]{}\
- |
Adam Czajka\
University of Notre Dame\
Notre Dame, Indiana, USA\
[[email protected]]{}\
- |
Walter Scheirer\
University of Notre Dame\
Notre Dame, Indiana, USA\
[[email protected]]{}\
bibliography:
- 'florin.bib'
title: 'Learning-Free Iris Segmentation Revisited: A First Step Toward Fast Volumetric Operation Over Video Samples'
---
Introduction and Application Context {#introduction}
====================================
Iris segmentation is perhaps the most important step in the iris recognition pipeline. Given the complex, rich iris texture used to extract identifying features for subject matching, segmentations must correspond closely to the actual iris. Lack of precise segmentation results in misalignment of tiny iris features and, in consequence, a high probability of false matches or non-matches. In the 25 years following Daugman’s seminal work on iris recognition [@Daugman_TPAMI_1993], a wide variety of approaches to iris segmentation have achieved high matching accuracy. Despite the wealth of methodologies, iris segmentation that simultaneously operates in real-time and affords a high degree of accuracy and generalization still seems to be just beyond our grasp.
The obvious tradeoff between speed and accuracy has led us to introduce a new paradigm for iris segmentation in this paper by processing iris image sequences as volumetric data. Since iris recognition cameras usually capture iris videos internally, and apply various quality checks to pick a single iris image to be used in feature extraction, we may assume that such videos are available in current commercial deployments. As a first, exploratory, step in this direction, we propose to apply the recently introduced Flexible Learning-Free Reconstruction of Neural Volumes (FLoRIN) pipeline [@shahbazi2018flexible], which was originally designed for processing volumetric image stacks of neural microscopy data. FLoRIN allows us to use information about the iris location not only in 2D space but also along the temporal axis, for example across multiple frames of an iris video.
We show that when FLoRIN is used, a minimal and acceptable drop in the overall iris recognition accuracy is compensated for by iris segmentation processing times that are a factor of $3.56$ faster than deep-learning-based solutions. When compared to other learning-free state-of-the-art iris segmentation techniques, FLoRIN-based processing is approximately an order of magnitude faster. This performance is maintained on low-resource hardware, making FLoRIN a good candidate for embedded iris matching solutions. Moreover, the FLoRIN-based iris segmentation does not need training, which is required for deep-learning-based methods, instead requiring tuning two threshold values over the interval $[0, 1]$. Following publication, we will make both the FLoRIN-based iris segmentation software and the new database of iris videos publicly available.
Summarizing, the novel contributions of this paper are:
1. The application of the FLoRIN framework for volumetric iris segmentation, along with the source code of the resulting FLoRIN-based volumetric and learning-free iris segmentation tool (presented in Sec. \[florin\]).
2. A novel database of iris near-infrared videos, presenting various pupil sizes as a result of visible light stimuli (presented in Sec. \[pupil-dynamics\]).
3. Evaluations of FLoRIN-based, deep learning-based, and Daugman’s segmentation approaches on a novel benchmark in terms of the matching accuracy and speed (described in Sec. \[results\]).
4. An evaluation of FLoRIN-based and Daugman’s segmentation implementations for embedded subject matching on a Raspberry Pi (presented in Sec. \[results\])
Related Work
============
As the first step of iris recognition, iris segmentation has been approached in a variety of ways. Early and, for many years, dominant methods relied upon the circular structure of both the pupil and the iris, as proposed by Daugman [@Daugman_TPAMI_1993]. In addition to circular approximations of iris boundaries, those methods detected eyelids, eyelashes, specular reflections, and excluded these occlusions from feature matching.
Subsequent approaches departed from circular approximations, and in many cases were inspired by Daugman’s suggestion to use a Fourier series to provide a more complex model of the iris boundary [@Daugman_TSMCB_2007]. Arvacheh and Tizhoosh [@arvacheh2006segmentation] proposed using active contour models to detect the pupillary and limbic boundaries of the iris, taking into consideration the fact that neither the pupil nor the iris are perfectly circular, but are near-circular. They also introduced an algorithm that iteratively detects and excludes eyelid areas that cover the iris. In another approach, Shah and Ross [@shah2009iris] proposed a method for extracting the iris using geodesic active contours to locate the boundaries of the iris in a given image. They focused on finding the boundaries of the iris, taking into consideration the fact that the iris is not necessarily circular due to the presence of these occlusions.
Other methods approached iris segmentation from a localization perspective. Luengo *et al.* [@luengo2009robust] proposed using mathematical morphology to extract the outer boundaries of both the pupil and the iris. They dealt with occlusions by simply removing portions of the segmentation which likely contained parts of the eyelid, eyelashes, reflections, etc. In another approach, He *et al.* [@he2009toward] extracted the iris center from the image and from that detected the iris boundaries. The novelty of their algorithm was the use of a histogram filter to detect irregularities in the shape of the eyelids.
More recently, iris segmentation has moved away from these classical approaches toward machine learning oriented methods. Li *et al.* [@li2010robust] proposed a method which uses k-means clustering to detect the outer boundary of the iris. In a similar fashion, Sahmoud and Abuhaiba [@sahmoud2013efficient] proposed using k-means clustering to determine the region of the iris.
Recent advances in deep learning-based segmentation have resulted in many applications of convolutional neural network architectures to iris segmentation. Jalilian and Uhl [@Jalilian_DLinBiometrics_2017] proposed to use several types of convolutional encoder-decoder networks and reported better performance for deep learning-based approaches when compared to conventional algorithms such as Daugman’s approach [@osiris], WAHET [@Uhl_ICB_2012], CAHT [@Rathgeb_IB_2013] and IFPP [@Uhl_IAR_2012]. Arsalan proposed to use a modified VGG-Face network to segment visible-light [@Arsalan_Symmetry_2017] and near-infrared [@Arsalan_Sensors_2018] iris images. Other deep learning-based iris segmentation techniques include using a re-trained U-Net [@Lozej_IWOBI_2018], semi-parallel Fully Convolutional Deep Neural Networks [@bazrafkan2018end], Generative Adversarial Networks [@Bezerra_SIBGRAPI_2018] and Mask R-CNN [@Ahmad_ArXiv_2018]. SegNet [@SegNet2016] was successfully re-trained by Trokielewicz [@TrokielewiczIWBF2018] to segment very challenging post-mortem iris images.
The ideal segmentation method would be fast, generalizable across subjects and sensors, and explainable to enable error detection and correction. Learning-based solutions are promising, however long training times and bias toward training datasets limit the performance of standard machine learning and deep learning models in the wild. Moreover, deep learning models are black boxes wherein errors are difficult to explain. By contrast, the FLoRIN pipeline used in this work is a deterministic set of operations that requires only setting two threshold values. Moreover, as will be shown, it operates on sequential image data more quickly than standard methods, enabling real-time iris segmentation. Last but not least, the number of open-sourced iris segmentation tools is currently very limited, despite a large number of papers proposing various solutions. FLoRIN-based iris segmentation software is offered with this paper.
The FLoRIN Framework {#florin}
====================
{width="\linewidth"}
The FLoRIN framework is a multi-stage pipeline with flexible image processing steps in each stage. FLoRIN was originally developed to meet the challenges of segmenting volumes of neural microscopy, enabling automatic discovery of non-standard structures like cells, vasculature, and myelinated axons. By incorporating volumetric data into the segmentation process through the N-Dimensional Neighborhood Thresholding (NDNT, Section \[florin-ndnt\]) algorithm [@shahbazi2018flexible], FLoRIN was able to boost the signal of features of interest without requiring any machine learning. In this way, FLoRIN achieved state of the art results across a number of imaging modalities in neural microscopy.
Conventionally, the FLoRIN pipeline consists of three major stages:
1. **Segmentation**. The images are loaded into FLoRIN and optionally processed to improve contrast, for example by histogram equalization, Weiner filters, etc. The images are then thresholded in 2D or 3D using NDNT [@shahbazi2018flexible] parameterized by a threshold value in range \[0, 1\] and a pixel neighborhood size. The neighborhood size can specify a neighborhood in two or three dimensions, the latter enabling the incorporation of volumetric data into the thresholding. The binarized NDNT output is then passed to the next stage.
2. **Identification**. The Segmentation stage output is processed to remove holes and artifacts, and morphological operations may be applied in 2D or 3D to refine the segmentation. Connected components are discovered and grouped by geometric and grayscale properties. This form of weak classification enables labeling different classes of structure; in neural microscopy this groups connected components corresponding to neurons, axons, etc. In iris segmentation, we use this stage to identify the pupil for separation from the iris pixels.
3. **Reconstruction**. The grouped connected components are labeled and output into image files for analysis. Properties of the connected components are saved to file, from which statistics about the discovered structures may be computed. In the case of iris segmentation, the Reconstruction stage outputs binary masks corresponding to the discovered iris pixels.
FLoRIN can operate on 2D images or 3D volumes, including time series or video data, at all stages. In particular, 3D segmentation and identification allows FLoRIN to account for volumetric data. For iris segmentation, we apply the 3D FLoRIN pipeline to videos of irises, wherein the volumetric dimension is time. This allows for the segmentation of the iris and pupil as features spanning multiple images, increasing the probability of segmenting the features of interest. In this work, we focused on adapting the Segmentation and Identification stages of FLoRIN, tailoring the pipeline to process iris images as shown in Figure \[fig:pipeline\]. Algorithm \[alg:pipeline\] provides a full description of the pipeline.
Create an empty array $seg$.\
N-Dimensional Neighborhood Thresholding {#florin-ndnt}
---------------------------------------
Central to FLoRIN is the NDNT algorithm, which is applied during the Segmentation stage to binarize images and volumes. NDNT is based on the thresholding method introduced by Bradley and Roth [@bradley2007adaptive], which uses the integral image to compute neighborhood statistics around each pixel and threshold based on the neighborhood average pixel intensity. This method was extended by Shahbazi *et al.* [@shahbazi2018flexible] to operate in *n* dimensions, allowing for the method to account for volumetric and channel-level information.
NDNT was shown to outperform standard thresholding methods when operating on neural microscopy data, generalizing to multiple imaging modalities and features. Moreover, NDNT operates solely on the pixel neighborhood histograms, requiring tuning a single thresholding parameter in range $[0, 1]$. Compared to learning-based methods, NDNT bypasses long training times and is explainable based on the input images and neighborhood size. In the case of iris segmentation, it can account for sequential images of the same iris, strengthening the signal for segmentation.
Baseline Iris Segmentation {#baselines}
==========================
In this section we describe the two baseline iris segmentation methods — SegNet [@SegNet2016] and OSIRIS [@osiris] — that we compare against, as well as the subject matching pipeline used to evaluate all segmentation methods.
SegNet DCNN Model {#baselines-segnet}
-----------------
To evaluate the accuracy of the proposed iris-adapted FLoRIN pipeline in comparison with a state-of-the-art segmentation method, we deploy a SegNet model [@SegNet2016] trained on several standard iris biometrics datasets their corresponding ground truth masks. SegNet is a fully convolutional encoder-decoder architecture, with the encoder stage being a modified VGG-16 model with removed fully connected layers, and the decoder stage comprising convolutional and upsampling layers corresponding to the max pooling layers of the encoding stage. SegNet has achieved state of the art performance on semantic segmentation tasks, including a recent successful use in segmentation of challenging post-mortem iris images [@TrokielewiczIWBF2018].
Training Procedure and Data
---------------------------
To tune the original SegNet model to segment iris images, we performed a re-training procedure with images drawn from several publicly available databases, including the Biosec baseline corpus [@biosec] (1200 images), the BATH database[^1] [@bath] (148 images), the ND0405 database[^2] (1283 images), the UBIRIS database [@ubiris] (1923 images), and the CASIA-V4-Iris-Interval database[^3] (2639 images). The ground truth masks came from the IRISSEG-CC dataset by Halmstad University [@halmstadGT] (for the Biosec set), from the IRISSEG-EP dataset by the University of Salzburg [@salzburgGT] (CASIA-V4-Iris-Interval, Notre Dame 0405, and UBIRIS). For the BATH subset, we created the binary masks ourselves. The model was trained in MATLAB for 120 epochs with a batch size of 4, using an SGDM optimizer with a momentum of 0.9, learning rate of 0.001, and L2 regularization of 0.0005. The training examples were shuffled after each epoch. In total, training took approximately one day on an NVIDIA GTX 1070 GPU.
{width="\textwidth"}
OSIRIS Segmentation and Matching {#baselines-osiris}
--------------------------------
In addition to the SegNet model, we also evaluate the original, unmodified OSIRIS segmentation. OSIRIS is an open-source academic matcher, which implements the principles of original Daugman concept, including iris segmentation using the circular Hough transform and subsequent refinement with active contours, iris image normalization, and Gabor-based filtering, which yields binary codes that allow for efficient matching. In this paper, we use OSIRIS in three different modes:
1. [**baseline conventional:**]{} stock OSIRIS for both segmentation, matching, and encoding
2. [**baseline DCNN:**]{} SegNet masks and OSIRIS encoding and matching
3. [**proposed:**]{} FLoRIN masks and OSIRIS encoding and matching
Each mode allows us to obtain ROC (Receiver Operating Characteristic) curves, which we use to denote the recognition accuracy that each set of iris masks enables.
Benchmark Iris Video Dataset {#pupil-dynamics}
============================
To enable processing volumetric data with FLoRIN, we used a newly collected database of iris videos, acquired in near-infrared, which is released as a part of this work. To include a wide spectrum of pupil dilation, we followed the acquisition protocol defined by Czajka [@Czajka_TIFS_2015], and collected 30-second videos, at a rate of 25 frames per second, of an iris stimulated by visible light. That is, during the first 15 seconds, 375 iris images are taken in darkness (yet small pupil oscillations, called [*hippus*]{} are still present). During the next 5 seconds, 125 iris images are taken right after the visible light is switched on. This forces the pupil’s rapid constriction and acquisition of sample under varying pupil size. The last 250 images are taken right after the visible light is switched off, making the pupil to dilate and providing additional samples with varying size of the iris texture. Hence, each video comprises 750 iris images. All samples are compliant with ISO/IEC 19794-6 standard. Fig. \[fig:dataSample\] presents selected frames taken from an example video, along with OSIRIS, SegNet and FLoRIN segmentations.
42 subjects participated in the data acquisition. We collected from one to four videos for each subject at the 25 FPS rate. Except for 17 videos, all sequences comprise 750 frames, and the total number of iris images considered in this work is 117,717. Researchers interested in obtaining a copy of this dataset are requested to follow the instructions provided at <http://...>[^4].
Experiments {#results}
===========
A key question we sought to answer in our experiments is: how close to top performing iris segmentation performance can one get with a learning-free method that is optimized for speed? To answer this question, we processed the collected iris video data with the proposed FLoRIN pipeline (Section \[florin\]. We then compared FLoRIN matching performance with segmentations generated by SegNet and OSIRIS (Section \[baselines-segnet\]), both of which are limited to 2D processing. All code used in these evaluations will be released upon publication.
For the assessment of the matching accuracy, we normalize the segmentation outputs produced by the SegNet and FLoRIN algorithms to fit them into the OSIRIS recognition pipeline (Section \[baselines-osiris\]. Circular Hough Transform (CHT) is employed to approximate the iris and pupil boundaries in the obtained binary masks. OSIRIS to normalizes the iris images and corresponding masks onto a dimensionless polar-coordinate rectangle, which is then used in the Gabor filtering and encoding stage.
Genuine matching performance was computed by first selecting every tenth frame of each video, then comparing the selected frames. Impostor matching performance was computed by first selecting every twentieth frame of the first video for each eye, then comparing the selected frames representing different eyes. Approximately 100k genuine comparisons and approximately 200k impostor comparisons were conducted. Comparison pairs were identical for all tested methods.
Our analysis shows that FLoRIN dramatically increases the speed of segmentation while incurring only minor penalties to subject matching performance. For each video, FLoRIN processed batches of 5 frames at a time. All throughput analyses were computed on a system with 4 Intel Xeon E5-2650 v4 processors, 24GB of RAM, and an NVIDIA GTX 1080ti GPU, using versions of the videos downsampled to $320 \times 240$ pixels per frame.
[**Method**]{} [**AUC**]{} [**EER**]{} [**Mean FPS**]{}
------------------------------------ ------------- ------------- ----------------------
[FLoRIN [@shahbazi2018flexible]]{} 0.962 0.074 **37.38 $\pm$ 2.58**
[SegNet [@SegNet2016]]{} 0.992 **0.016** $10.50 \pm 1.14$
[OSIRIS [@osiris]]{} **0.996** 0.017 $\:\,3.24 \pm 0.28$
: Matching Performance and Segmentation Throughput. The mean frames per second (FPS) is computed as the average per-video FPS over all 159 videos.[]{data-label="tab:perf"}
FLoRIN offers a substantial increase in throughput over SegNet and OSIRIS. As shown in Table \[tab:perf\], FLoRIN is a factor of 3.56 faster than SegNet and an order of magnitude faster than OSIRIS processing the same videos. This throughput increase is a direct result of processing videos in 3D: a batch of frames is segmented in parallel using surrounding frames to boost the signal, then all frames in the batch are post-processed simultaneously to improve the segmentation. This scheme introduces data parallelism across the time dimension of the videos which is unavailable to SegNet, OSIRIS, and other 2D segmentation methods.
![Receiver operating characteristic of the False Match Rate (FMR) versus the inverse False Non-Match Rate (FNMR) for FLoRIN, SegNet, and OSIRIS on the pupil dynamics dataset.[]{data-label="fig:roc"}](icb_2019_florin_roc.png){width="\columnwidth"}
The increased speed of FLoRIN comes with minor subject matching performance degradation. The receiver operating characteristic (ROC) curve of the inverse False Negative Match Rate vs the False Match Rate of subject matching for each method is shown in Figure \[fig:roc\]. FLoRIN has an area under the curve (AUC) of 0.96 and an equal error rate (EER) of 0.07 (Table \[tab:perf\]), while SegNet and OSIRIS both boast an AUC of 0.99 and an EER of 0.02. Qualitatively, we find that the FLoRIN segmentations tend to under-segment the pupil, leading to a greater proportion of non-iris pixels included in the normalized masks. This issue will be addressed with additional quality control operations as FLoRIN continues to develop.
[0.19]{} {width="\textwidth"}
[0.19]{} {width="\textwidth"}
[0.19]{} {width="\textwidth"}
[0.19]{} {width="\textwidth"}
[0.19]{} {width="\textwidth"}
Throughout this experiment, FLoRIN thresholding parameters were manually selected on a per-video basis, however in many cases the threshold values transferred between videos of the same subject. Parameter values were selected by a sweep over the real-valued domain $[0, 1]$ applied to a subset of each video. The results were then manually examined to determine the optimal threshold values. By reusing intermediate data computed by the NDNT algorithm that is invariant to threshold value, we were able to quickly segment these subsets across a large number of threshold values. We evaluated this scheme across subsets of 5 frames of each video with threshold values spaced at $10^{-2}$ and found that the parameter sweep could be completed in approximately $2$s for a given window size.
Recommendations for Threshold Parameters {#results-tvals}
----------------------------------------
While manually tuning the iris and pupil threshold values for each of the videos, we discovered a number of trends in the selection of threshold values and neighborhood windows sizes. We break these recommendations down by feature to enable others to apply FLoRIN to new data.
**Iris Parameters.** As the iris spans a large percentage of each frame in a video, we found that a larger neighborhood was better able to capture the iris pixels. In every video, we used a window of size $2\times 256\times 256$ around each pixel, which accounts for the previous and subsequent frames. Given the size of the window, we recognize that this is similar to a global thresholding method with the inclusion of volumetric data. Threshold values for the iris tended to the domain $[0.2, 0.5]$, with a sensitivity of $0.01$.
**Pupil Parameters.** Pupil segmentation was more sensitive to neighborhood size, relative to the shape of the image histogram. In the case of a histogram tending toward the right, a large $2\times 256\times 256$ neighborhood with threshold values in domain $[0.7, 0.95]$ sufficed to capture the pupil. In extreme cases with a tall peak of low-intensity values, for example when the subject wore eyeliner, we found that $2\times 64\times 64$ or $2\times 32\times 32$ neighborhoods were better suited to isolate the pupil. The domain of threshold values differs for smaller neighborhood sizes, typically between $[0.1\mbox{--}0.3]$, with a sensitivity of $0.01$.
In both cases, the threshold values could be narrowed to a subset of $[0, 1]$. These findings indicate that the optimal threshold value is a function of the image histogram and the window size. This function could be used to develop an automatic parameterization scheme, either through machine learning or another optimization scheme.
Performance on Low-Resource Hardware {#results-pi}
------------------------------------
Given the large improvement in running time performance of FLoRIN over SegNet and OSIRIS, Table \[tab:perf\], we hypothesized that FLoRIN would be suitable for low-resource hardware and embedded systems. We conducted a second timing experiment on a Raspberry Pi 3 with 4 cores and 1GB of RAM. We processed the first 20 frames of each video on the Pi with FLoRIN and OSIRIS using the same setup used to process the full videos.
With the limited resources available to the Pi, FLoRIN processed the video subset with a mean throughput of $4.50 \pm 0.36$ frames per second per video. This was an order of magnitude greater than OSIRIS, which had a throughput of $0.38 \pm 0.03$ frames per second per video. The throughput afforded by FLoRIN on the Pi indicates that FLoRIN is fast enough to enable segmentation on embedded systems, e.g. commercial iris recognition hardware.
Discussion {#conclusion}
==========
Deep learning has, in many ways, revolutionized the field of biometrics. However, it is not the only way to approach problems like iris segmentation. Legitimate criticisms of deep learning exist in the form of long training times, the need for large amounts of hand-labeled training data, and the complexity of optimizing various hyperparameters. And frustratingly, even when all of these problems have been addressed for a single dataset or operational scenario, the move to a different setting forces one to start all over again. Generalization remains an open problem within the field of machine learning at large.
Our turn back to learning-free methods is a direct response to this dilemma. Elaborate training regimes leaning on massive data collection and annotation efforts can be dispensed with in favor of immediate inference operation. A move to a new dataset or acquisition setting may be as simple as making a few adjustments to a minimal set of free parameters before deployment. Several decades worth of work on image processing and computer vision should not be ignored — the literature is filled with older techniques that can be updated for today’s problems to achieve remarkable performance gains and avoid the generalization dilemma. The FLoRIN approach is just one example of this.
Given the observed speed of iris segmentation with FLoRIN, we propose that the pipeline used in this work can be deployed as a semi-automated annotator for generating ground truth. Using FLoRIN, a new dataset may be processed rapidly, and the output masks may be proofread by human annotators to create pixel-level ground truth labels. This scheme will enable rapid release and dissemination of new databases for use across the biometrics community.
This study represents the initial application of FLoRIN to iris segmentation, and as such required manually tuning the parameters of the NDNT algorithm. We are in the process of evaluating a number of methods for automatic parameterization, including grid search across threshold values using Reverse Classification Accuracy [@valindria2017reverse] as a guide and regression models to map images to threshold values. Such automation will reduce the need for a human in the loop. Based on the results of Section \[results-pi\], we believe an automatically-parameterized FLoRIN will be ideal for embedded iris biometrics systems.
[^1]: <http://www.bath.ac.uk/elec-eng/research/sipg/irisweb/>
[^2]: \[cvrl\]<https://cvrl.nd.edu/projects/data/>
[^3]: <http://www.cbsr.ia.ac.cn/english/IrisDatabase.asp>
[^4]: link removed temporarily to make this submission anonymous
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
------------------------------------------------------------------------
We study the thermodynamics of the near horizon of near extremal Kerr geometry ($near-NHEK$) within an $AdS_2/CFT_1$ correspondence. We do this by shifting the horizon by a general finite mass, which does not alter the geometry and the resulting solution is still diffeomorphic to $NHEK$, however it allows for a Robertson Wilczek two dimensional Kaluza-Klein reduction and the introduction of a finite regulator on the $AdS_2$ boundary. The resulting asymptotic symmetry group of the two dimensional Kaluza-Klein reduction leads to a non-vanishing quantum conformal field theory ($CFT$) on the respective $AdS_2$ boundary. The $s$-wave contribution of the energy-momentum-tensor of the $CFT$, together with the asymptotic symmetries, generate a Virasoro algebra with calculable center and non-vanishing lowest Virasoro eigen-mode. The central charge and lowest eigen-mode reproduce the $near-NHEK$ Bekenstein-Hawking entropy via the statistical Cardy Formula and our derived central charge agrees with the standard Kerr/$CFT$ Correspondence. We also compute the Hawking temperature of the shifted $near-NHEK$ by analyzing quantum holomorphic fluxes of the Robinson and Wilczek two dimensional analogue fields.\
------------------------------------------------------------------------
\
Keywords: Black Hole Thermodynamics; Black-Hole/CFT Duality; Quantum Gravity.\
PACS numbers: 11.25.Hf, 04.60.-m, 04.70.-s
author:
- |
Ananda Guneratne$^{a}$[^1], Leo Rodriguez$^{a}$[^2], Sujeev Wickramasekara$^{a}$[^3]\
and Tuna Yildirim$^{b}$[^4]\
*$^{a}$Department of Physics*\
*Grinnell College*\
*Grinnell, IA 50112*\
*$^{b}$Department of Physics and Astronomy*\
*The University of Iowa*\
*Iowa City, IA 52242*
bibliography:
- 'cftgr.bib'
title: '**Near-Extremal Kerr $AdS_2\times S^2$ Solution and Black-Hole/Near-Horizion-CFT Duality**'
---
------------------------------------------------------------------------
Introduction {#sec:intro}
============
Black hole thermodynamic quantities [@hawk2; @hawk3; @beken], $$\begin{aligned}
\label{eq:htaen}
\begin{cases}
T_H=\frac{\hbar\kappa}{2\pi}&\text{Hawking Temperature}\\
S_{BH}=\frac{A}{4\hbar G}&\text{Bekenstein-Hawking Entropy}
\end{cases},\end{aligned}$$ provided an ample testing bed for most current competing theories of quantum gravity. It is widely believed that any viable ultraviolet completion of general relativity should reproduce some variant of [eq:htaen]{}, perhaps modulo some real finite parameter that would need to be fixed by experiment. To date there is a plethora of different approaches for arriving at [eq:htaen]{}, with string theories and loop quantum gravity the predominant competitors and no clear consensus of which approach should be preferred over the other.
Black Hole Temperature and Effective Action {#sec:bht}
-------------------------------------------
Since Hawking’s original analysis of the density of quantum states in terms of Bogolyubov coefficients, analysis via effective actions and their associated energy momentum tensors of the semiclassical matter fields has been explored in various settings for arriving at $T_H$[@mukwipf; @balfab2; @balfab; @cadtr; @qpz; @Camblong:2004ec; @Camblong:2004ye; @Yuan:2011gq; @LLRphd]. Of particular interest is the realization by Robinson and Wilczek (RW) that anomalous two dimensional chiral theories in the near horizon of black holes are rendered unitary by requiring the black hole to radiate at temperature $T_H$ [@robwill; @isowill; @msoda; @gango; @Jin; @Jinwu; @chen; @chen2; @pwu; @nampark; @setare; @petro; @rabin3; @rabin; @rabin2; @rabin4; @Banerjee:2008sn; @Wu:2011im; @Akhmedova:2008au; @Zampeli:2012tv; @srv]. This procedure requires a dimensional reduction yielding two dimensional analogues for various types of four dimensional black holes (RW2DA), beyond the basic Schwarzschild case, coupled to two dimensional matter fields.
More recently Rodriguez and Yildirim showed by analyzing quantum conformal matter in the background of certain RW2DA black holes that the resulting energy momentum tensor (EMT) is holomorphic in the horizon limit or at the asymptotic infinity boundary [@ry; @Button:2010kg]. Furthermore, this resulting holomorphic EMT is dominated by one component equaling the four dimensional Hawking flux of temperature $T_H$ weighted by a central charge $c$. A plausible interpretation of this result is that in the near horizon regime a four or higher dimensional spacetime metric exhibits a Kaluza-Klein reduction into two dimensional fields ${g\indices}{^{(2)}_\mu_\nu}$, ${\mathcal}{A}_\mu$ and $\Phi$. In this interpretation ${\mathcal}{A}_\mu$ and $\Phi$ are of gravitational origin, yet mathematically behave as two dimensional $U(1)$ gauge and conformal scalar fields. Thus a quantum field theoretic study of $\Phi$ in two dimensions with respect to ${g\indices}{^{(2)}_\mu_\nu}$ and ${\mathcal}{A}_\mu$ will have quantum gravitational implications in the near horizon of four dimensional black holes. Depending on the asymptotic symmetry group of certain RW2DA, this may suggest a sort of $AdS_2/CFT_1$ relationship for computing four dimensional black hole temperature.
Holographic Black Hole Entropy {#sec:bhent}
------------------------------
The fact that quantum gravity may be dual to a $CFT$ [@Maldacena:1997re] has spawned a surge, led by Strominger [@strom2; @kerrcft], Carlip [@Carlip:2011ax; @carlip; @carlip3; @carlip2] and Park [@Park:1999tj; @Park:2001zn; @kkp], in applying $CFT$ techniques to compute Bekenstein-Hawking Entropy of various black holes [@ry; @Button:2010kg; @cadss; @silva; @bgk; @Barnes:2009zn; @Astefanesei:2009sh; @Banados:2011sd; @Majhi:2011ws; @Majhi:2012tf]. The most notable such example is by far the Kerr/$CFT$ correspondence and its extensions [@kerrcft; @SheikhJabbaria:2011gc; @deBoer:2011zt; @Yavartanoo2012410; @springerlink:10.1140/epjc/s10052-012-1911-7; @rasmussen:2010xd; @rasmussen:2010sa; @Chen:2010yu; @Chen:2010bh; @Li:2010ch; @Castro:2010fd; @Krishnan:2010pv; @kerrcftstring; @kerrcftsugra; @kerrcftind; @daCunha:2010jj; @Wu:2009di; @Huang:2010yg; @Compere:2012jk], where the general idea is that the asymptotic symmetry group (ASG), preserving certain metric boundary or fall off conditions, is generated by a Virasoro algebra with calculable central extension: $$\begin{aligned}
\label{eq:vir}
\left[{\mathcal{Q}}_m,{\mathcal{Q}}_n\right]=(m-n){\mathcal{Q}}_{m+n}+\frac{c}{12}m\left(m^2-1\right)\delta_{m+n,0}.\end{aligned}$$ The Bekenstein-Hawking entropy is then obtained from Cardy’s Formula [@cardy2; @cardy1] in terms of $c$ and the proper normalized lowest eigen-mode via: $$\begin{aligned}
\label{eq:cf}
S=2\pi\sqrt{\frac{c\cdot{\mathcal{Q}}_0}{6}}.\end{aligned}$$ In the Kerr/$CFT$ case a thermal Cardy Formula is employed $$\begin{aligned}
\label{eq:tcf}
S_{BH}=\frac{\pi^2}{3}c_LT_L,\end{aligned}$$ which depends on the left Frolov-Thorne vacuum temperature $T_L$ for generic Kerr geometry [@frolovthorne]. This is in part due to the vanishing surface gravity of the extremal Kerr geometry, which is usually employed in regulating the quantum charges of [eq:vir]{}, thus leading to a finite ${\mathcal{Q}}_0^{L,R}$. However, a finite zero mode may be inferred by the identification: $$\begin{aligned}
\label{eq:fttdef}
\frac{\partial S_{CFT}}{\partial {\mathcal{Q}}_0}=\frac{\partial S_{BH}}{\partial {\mathcal{Q}}_0}=\frac{1}{T}\Rightarrow {\mathcal{Q}}_0=\frac{\pi^2}{6}c T^2.\end{aligned}$$ Requiring that the ASG contains a proper $SL(2,\mathbb{R})$ subgroup in the above equation, yields the general value $T=\frac{1}{2\pi}$, which is the case for Kerr/$CFT$, where $T_L=\frac{1}{2\pi}$, $T_R=0$, and $c_L=c_R=c=12J$ and the extremal Bekenstein-Hawking entropy is recovered via the thermal Cardy Formula: $$\begin{aligned}
\label{eq:tcf2}
S_{BH}=\frac{\pi^2}{3}c(T_L+T_R)=2\pi J.\end{aligned}$$ From the definition [eq:fttdef]{} we see that $T$ in general should be unitless. An interesting choice is to employ the Hawking temperature scaled by the finite time regulator $1/\kappa$ giving the general result $T=\frac{1}{2\pi}$ and also extends smoothly to extremality[^5].
The general Kerr entropy may be obtained by analyzing the Frolov-Thorne vacuum in the near extremal case, where $$\begin{aligned}
\label{eq:nefttemp}
T_L=\frac{(GM)^2}{2\pi J}~\text{and}~T_R=\frac{\sqrt{(GM)^4-(GJ)^2}}{2\pi J}.\end{aligned}$$ Using these values and $c=12J$ in [eq:tcf2]{} yields the standard area law $$\begin{aligned}
\label{eq:mal}
S_{BH}=2\pi\left(GM^2+\sqrt{G^2M^4-J^2}\right).\end{aligned}$$ Yet, the above result requires combining quantities derived separately at extremality and near-extremality. It is also not obvious that the combination of temperatures in [eq:nefttemp]{}, $T=T_L+T_R$, yields the value $\frac{1}{2\pi}$ except in the extremal limit. However recasting the values of $c,~T_L,~T_R$ into more general variables we see that: $$\begin{aligned}
\label{eq:nexctlr}
c=\frac{3A}{2\pi G},~T_{L}=\frac{4(GM)^2}{A}~\text{and}~T_{R}=\frac{4\sqrt{(GM)^4-(GJ)^2}}{A}\end{aligned}$$ substituting these values into [eq:tcf]{} yields $S_{BH}=\frac{A}{4G}$ and assuming they smoothly extend back to non-extremality we have: $$\begin{aligned}
T=T_{L}+T_{R}=\frac{4(GM)^2}{A}+\frac{4\sqrt{(GM)^4-(GJ)^2}}{A}=\frac{1}{2\pi},\end{aligned}$$ which would provide a more wholesome computation of the near-extremal Kerr black hole entropy. This is precisely the aim of this note, to construct a near horizon $CFT$ dual for the $near-NHEK$ spacetime and compute its corresponding entropy within a statistical Cardy formula [eq:cf]{}, without mixing result derived separately at extremality and near-extremality. We will do this within an $AdS_2/CFT_1$ correspondence by performing a RW two dimensional reduction of the $near-NHEK$ geometry in a specific finite mass gauge following similar constructions as found in [@Button:2010kg; @Castro:2009jf].
In [@Button:2010kg] Button, Rodriguez, Whiting and Yildirim showed that the RW2DA of non-extremal Kerr-Newman-$AdS$ is asymptotically $AdS_2$, for a specific choice of metric and gauge field fall off conditions, with effective near horizon functional derived via the RW dimensional reduction procedure. Evaluating the resulting effective EMT on the $AdS_2$ boundary and computing its response to a total symmetry transformation yields a central charge: $$\begin{aligned}
\label{eq:centerp1}
\frac{c}{24\pi}=\frac{r_+^2+a^2}{4\pi G\Xi}\Rightarrow c=\frac{3A}{2\pi G},\end{aligned}$$ which is in agreement with the Kerr/$CFT$ correspondence within the appropriate limits, i.e. $$\begin{aligned}
\lim_{{\ell}\to\infty,~Q\to0,~a\to GM}c=12J.\end{aligned}$$ Computation of the asymptotic symmetry algebra yields the normalized zero mode $$\begin{aligned}
{\mathcal{Q}}_0=\frac{A}{16\pi G},\end{aligned}$$ which together with the central charge and [eq:cf]{} yields the Bekenstein-Hawking Entropy of the KNAdS black hole. It is interesting to note that the lowest Virasoro eigen-mode satisfies $$\begin{aligned}
\mathcal{Q}_0=GM_{irr}^2\end{aligned}$$ where $M_{irr}^2$ is the irreducible mass of the KN$AdS$ black hole and agrees with [eq:nexctlr]{}. This suggests a possible addition to the $AdS/CFT$ dictionary, that the eigen-value of the $CFT$’s Hamiltonian is proportional to the irreducible mass of its black hole dual.
An additional distinct example of reducing near horizon dynamics to two dimensional physics was introduced by Castro and Larsen [@Castro:2009jf] by recasting the $NHEK$ [@Bardeen:1999px] into two dimensional fields ${g\indices}{^{(2)}_\mu_\nu}$, $\psi$ and ${\mathcal}{A}_\mu$ $$\begin{aligned}
\label{eq:clnhek}
ds^2=\frac{1+\cos^2{\theta}}{2}\left[ds_{(2)}^2+e^{-2\psi}{\ell}^2d\theta^2\right]+e^{-2\psi}{\ell}^2\frac{2\sin^2{\theta}}{1+\cos^2{\theta}}\left(d\phi+{\mathcal}{A}\right)^2\end{aligned}$$ and evaluating the Einstein-Hilbert action over [eq:clnhek]{} and integrating out angular degrees of freedom. A careful study of the asymptotic boundary currents of the resulting functional, within a well defined variational principle, yields a one dimensional quadratic two form with transformtion Law: $$\begin{aligned}
\label{eq:clemt}
\delta_{\epsilon+\Lambda}T_{tt}=&T_{tt}\xi'(t)+\xi(t)T'_{tt}+\frac{12J}{12}{\ell}\xi'''(t)+{\mathcal}{O}\left(e^{-\rho/{\ell}}\right)\end{aligned}$$ in Gauss normal coordinates. This closely resembles the transformation law for the energy momentum tensor of a conformal field with center $c=12J$, in agreement with the Kerr/$CFT$ correspondence [@kerrcft]. However, a computation of the asymptotic symmetry algebra within this work was not performed.
We will proceed in Section \[sec:nensea\], by reviewing the relevant $near-NHEK$ geometry and introduce the finite mass gauge by solving the vacuum Einstein field equations for generic four dimensional Einstein-Hilbert Theory with a specific initial ansatz. In Section \[sec:qftnhnek\] we perform the RW dimensional reduction to the geometry of interest and derive a two dimensional near horizon Liouville-like $CFT$ in terms of the resulting RW2DA fields. Analysis of the asymptotic symmetries of this theory leads to black hole thermodynamics of the near-extremal Kerr throat. Next, in Section \[sec:2da\], we derive a normalized $AdS_2$ effective action via off shell analysis of the Einstein-Hilbert action directly within the finite mass gauge. The resulting boundary counter term contribution to the effective action yields a precise one dimensional quadratic two form with calculable central charge, reproducing the results of the previous section. Finally in Section \[sec:concom\] we close with a discussion of our results and hint towards future work.
Geometry {#sec:nensea}
========
We are interested in studying four dimensional vacuum solutions which in the near horizon have the form: $$\begin{aligned}
\label{eq:knadsnhm}
ds^2=K_1\left(\theta\right){g\indices}{^{(2)}_\mu_\nu}dx^\mu dx^\nu+K_2\left(\theta\right)e^{-2\varphi}d\theta^2+K_3\left(\theta\right)e^{-2\varphi}\left[d\phi+\mathcal{A}\right]^2,\end{aligned}$$ The above two dimensional field splitting provides a robust platform for constructing $CFT$ duals for relevant classical spacetimes with non vanishing surface gravity (near-extremal), by analyzing the ASG of the Kaluza-Klein fields ${g\indices}{^{(2)}_\mu_\nu}$, ${\mathcal}{A}$ and $\varphi$. The $NHEK$ is a four dimensional vacuum solution, which is derived by taking the extremal near horizon limit: $$\begin{aligned}
\label{eq:nhekcoor}
r=GM+\lambda U,~t'=\frac{t}{\lambda},~\phi'=\phi+\frac{t}{2GM\lambda},~\lambda\to\infty\end{aligned}$$ of the generic Kerr metric: $$\begin{aligned}
\label{eq:kerrm}
{\begin{split}}ds^2_{Kerr}=&-\frac{\Sigma\Delta}{\left(r^2+a^2\right)^2-\Delta a^2\sin^2\theta}dt'^2+\Sigma\left[\frac{dr^2}{\Delta}+d\theta^2\right]\\
&+\frac{\left(\left(r^2+a^2\right)^2-\Delta a^2\sin^2\theta\right)\sin^2\theta}{\Sigma}\left[d\phi'+\frac{2rGMa}{\left(r^2+a^2\right)^2-\Delta a^2\sin^2\theta}dt'\right]^2,
{\end{split}}\end{aligned}$$ where $$\begin{aligned}
\label{eq:kerrdef}
{\begin{split}}\Sigma=&r^2+a^2\cos^2\theta,\\
\Delta=&\left(r-r_+\right)\left(r-r_-\right),\\
r_\pm=&GM\pm\sqrt{(GM)^2-a^2},\\
a=&\frac{J}{M},
{\end{split}}\end{aligned}$$ which yields: $$\begin{aligned}
\label{eq:nhek}
ds^2_{NHEK}=\frac{1+\cos^2{\theta}}{2}\left[-\frac{U^2}{{\ell}^2}dt^2+\frac{{\ell}^2}{U^2}dU^2+{\ell}^2d\theta^2\right]+{\ell}^2\frac{2\sin^2{\theta}}{1+\cos^2{\theta}}\left(d\phi+\frac{U}{{\ell}^2}dt\right)^2.\end{aligned}$$ The above metric is of the form [eq:knadsnhm]{}, and may be tuned to near-extremality via the finite temperature gauge yielding: $$\begin{aligned}
\label{eq:nnhek}
ds^2_{near-NHEK}=\frac{1+\cos^2{\theta}}{2}\left[-\frac{U^2-\epsilon^2}{{\ell}^2}dt^2+\frac{{\ell}^2}{U^2-\epsilon^2}dU^2+{\ell}^2d\theta^2\right]+{\ell}^2\frac{2\sin^2{\theta}}{1+\cos^2{\theta}}\left(d\phi+\frac{U}{{\ell}^2}dt\right)^2,\end{aligned}$$ where $\epsilon=\frac{1}{2\lambda}\left(r_+-r_-\right)$ is a finite excitation above extremality and ${\ell}^2=2G^2M^2$.
Finite Mass Gauge {#sec:nhgeo}
-----------------
To aid in our $CFT$ dual construction we want to endow the $near-NHEK$ geometry with a finite $ADM$ mass and angular momentum parameter $M$ and $a$. We will do this by solving the Einstein Field Equations directly for specific boundary conditions, as opposed to implementing a parameter tuning process in the transformations of [eq:nhekcoor]{}. Starting with [eq:knadsnhm]{} as our ansatz, we impose the following symmetry conditions:
- The horizon is topologically $S^2$, i.e. $$\begin{aligned}
\label{eq:hs2c}
K_2K_3=\sin^2\theta~\text{and}~K_1=K_2.\end{aligned}$$
- The metric is maximally isometric in $t$ and $\phi$, which implies $$\begin{aligned}
\label{eq:misotphi}
{g\indices}{^{(2)}_{\mu\nu}}={g\indices}{^{(2)}_{\mu\nu}}(r),~\varphi=\varphi(r)~\text{and}~{\mathcal{A}}={\mathcal{A}}_\mu(r)dx^\mu.\end{aligned}$$
- The metric is axially symmetric in four dimensions and spherically symmetric in two: $$\begin{aligned}
\label{eq:axisphs}
ds^2={g\indices}{^{(2)}_{\mu\nu}}dx^\mu dx^\nu=-f(r)dt^2+\frac{dr^2}{f(r)}~\text{and}~{\mathcal{A}}={\mathcal{A}}_t(r)dt.\end{aligned}$$
Implementing these conditions within the four dimensional vacuum Einstein field equations we obtain the relatively simple set of coupled differential equations: $$\begin{aligned}
\label{eq:refeq}
\begin{cases}
\frac{\varphi '(r) K_2'(\theta )}{K_2(\theta )}=0\\
~\\
K_2(\theta ) \left(-2 e^{-2 \varphi(r)} f'(r) \varphi'(r)+f(r)\left(4e^{-2 \varphi(r)}\varphi'(r)^2-2e^{-2\varphi(r)}\varphi''(r)\right)-2\right)+\\
K_2''(\theta )-3\cot(\theta)K_2'(\theta)=0\\
~\\
\sin (\theta ) \left(e^{-4 \varphi (r)} \sin ^2(\theta ) {\mathcal{A}}_t'(r)^2+K_2'(\theta)^2\right)+\sin (\theta ) K_2(\theta )^2 \left(-2 e^{-2 \varphi (r)} f'(r) \varphi'(r)\right.+\\
\left.f(r)\left(4 e^{-2 \varphi (r)} \varphi '(r)^2-2 e^{-2 \varphi (r)} \varphi''(r)\right)-2\right)-K_2(\theta ) \left(\sin (\theta ) K_2''(\theta )+\cos (\theta) K_2'(\theta )\right)=0\\
~\\
e^{-6 \varphi (a)} \sin ^2(\theta ) {\mathcal{A}}_t'(r)^2-e^{-4 \varphi (r)} K_2(\theta )^2 f''(r)-e^{-2
\varphi (r)} \left(K_2(\theta )^2 \left(2 f(r) \left(4 e^{-2 \varphi (r)} \varphi '(r)^2-\right.\right.\right.\\
\left.\left.\left.2 e^{-2
\varphi (r)} \varphi ''(r)\right)-2 e^{-2 \varphi (r)} f'(r) \varphi '(r)\right)-K_2'(\theta
)^2+K_2(\theta ) \left(K_2''(\theta )+\cot (\theta ) K_2'(\theta
)\right)\right)+\\
4 f(r) e^{-4 \varphi (r)} K_2(\theta )^2 \varphi '(r)^2=0
\end{cases}\end{aligned}$$ where prime denotes derivation with respect to the argument of the given function. A general solution to the above field equations is given by: $$\begin{aligned}
\label{eq:gbhs}
\begin{cases}
\varphi(r)=&\varphi\\
K_2(\theta)=&\frac{C_1 (1-\cos (\theta ))^{3/2} \sqrt{\cos ^2(\theta )-1}}{\sqrt{\cos (\theta )+1}}-\frac{C_2 \sqrt{1-\cos (\theta )} \cos (\theta )
\sqrt{\cos ^2(\theta )-1}}{(\cos (\theta )-1) \sqrt{\cos (\theta )+1}}\\
{\mathcal{A}}_t(r)=&\frac{r \sqrt{C_2^2-4 C_1 C_2}}{e^{-2\varphi}}+C_3\\
f(r)=&\frac{r^2}{e^{-2\varphi}}+r C_5+C_4
\end{cases}\end{aligned}$$ where $C_i$’s are integration constants and the full line element is given in [eq:knadsnhm]{} and reads: $$\begin{aligned}
\label{eq:knadsnhmgs}
{\begin{split}}ds^2=&\left(\frac{C_1 (1-\cos (\theta ))^{3/2} \sqrt{\cos ^2(\theta )-1}}{\sqrt{\cos (\theta )+1}}-\frac{C_2 \sqrt{1-\cos (\theta )} \cos (\theta )\sqrt{\cos ^2(\theta )-1}}{(\cos (\theta )-1) \sqrt{\cos (\theta )+1}}\right)\cdot\\
&\left[-\left(\frac{r^2}{e^{-2\varphi}}+r C_5+C_4\right)dt^2+\frac{1}{\frac{r^2}{e^{-2\varphi}}+r C_5+C_4}dr^2+e^{-2\varphi}d\theta^2\right]+\\
&\frac{\sin^2(\theta)}{\frac{C_1 (1-\cos (\theta ))^{3/2} \sqrt{\cos ^2(\theta )-1}}{\sqrt{\cos (\theta )+1}}-\frac{C_2 \sqrt{1-\cos (\theta )} \cos (\theta )
\sqrt{\cos ^2(\theta )-1}}{(\cos (\theta )-1) \sqrt{\cos (\theta )+1}}}\cdot\\
&e^{-2\varphi}\left[d\phi+\left(\frac{r \sqrt{C_2^2-4 C_1 C_2}}{e^{-2\varphi}}+C_3\right)dt\right]^2
{\end{split}}\end{aligned}$$
Next, we impose the final set of boundary conditions to fix the integration constants in the solution above:
\[con:kerr\]\[Finite Mass Gauge\]
- The inner and outer horizons are located at: $$\begin{aligned}
\label{eq:csh}
f(r_\pm)=0~\text{and}~r_\pm=GM\pm\sqrt{(GM)^2-a^2}.\end{aligned}$$
- The ADM mass is non zero and equal to the parameter $M$:
- The total angular momentum is given by $J=\frac{a^2+r^2_+}{2G}$.
Applying these conditions yields the final solution to [eq:refeq]{} $$\begin{aligned}
\label{eq:kbhs}
\begin{cases}
e^{-2\varphi}=&r_+^2+a^2\\
K_2(\theta)=&\frac{1+\cos^2\theta}{2}\\
{\mathcal{A}}_t(r)=&\frac{r-2GM}{r_+^2+a^2}\\
f(r)=&\frac{r^2-2rGM+a^2}{r_+^2+a^2}
\end{cases},\end{aligned}$$ with line element $$\begin{aligned}
\label{eq:knadsnhmkerr}
{\begin{split}}ds^2=&\frac{1+\cos^2\theta}{2}\left[-\frac{r^2-2rGM+a^2}{r_+^2+a^2}dt^2+\frac{r_+^2+a^2}{r^2-2rGM+a^2}dr^2+\left(r_+^2+a^2\right)d\theta^2\right]+\\
&\frac{2\sin^2\theta}{1+\cos^2\theta}\left(r_+^2+a^2\right)\left[d\phi+\left(\frac{r-2GM}{r_+^2+a^2}\right)dt\right]^2.
{\end{split}}\end{aligned}$$ The above line element clearly exhibits global $AdS_2\times S^2$ topology and is diffeomorphic to [eq:nhek]{} and [eq:nnhek]{} as we will demonstrate shortly. However, the above line element is written in Boyer-Lindquist type coordinates and exhibits two physical parameters $M$ and $a$, which will be useful for tuning purposes in our CFT construction.
Surface Gravity {#sec:NHnEKsg}
---------------
To compute the surface gravity of the gauged $near-NHEK$ ($gnNHEK$) we will exploit the fact that it is maximally isometric with respect to the coordinates $t$ and $\phi$ and define the general Killing vector: $$\begin{aligned}
\label{eq:gkvec}
\xi=\xi_{(t)}+\Omega_H\xi_{(\phi)},\end{aligned}$$ where $\Omega_H$ is the minimum of the function $$\begin{aligned}
\label{}
\frac{d\phi}{dt}=-\frac{g_{t\phi}}{g_{\phi\phi}}\pm\sqrt{\left(\frac{g_{t\phi}}{g_{\phi\phi}}\right)^2-\frac{g_{tt}}{g_{\phi\phi}}}\end{aligned}$$ evaluated on the horizon $r_+$. Given this above Killing vector and making use of the geodetic equation to rearrange Frobenius’ theorem $$\begin{aligned}
\label{eq:fbt}
\nabla_{[\alpha}\xi_{\mu}\xi_{\nu]}=0\end{aligned}$$ for hypersurface orthogonal congruences of null generators, we obtain: $$\begin{aligned}
\label{eq:fbtsg}
\kappa^2=-\frac12\left.\nabla^\mu\xi^\nu\nabla_\mu\xi_\nu\right\vert_{r_+}.\end{aligned}$$ Evaluating this over the connection of [eq:knadsnhmkerr]{} yields: $$\begin{aligned}
\label{eq:sgnhnek}
\kappa_{gnNHEK}=\frac12f'(r_+)=\frac{r_+-GM}{r_+^2+a^2},\end{aligned}$$ which leaves us with a finite non zero value, but may be tuned to zero for the case when $a\to GM$ i.e. when we approach extremality. We will also note the horizon are of the $gnNHEK$ black hole, which comes from evaluating the integral $$\begin{aligned}
A=\int d^2x\sqrt{\gamma}=4\pi\left(r_+^2+a^2\right),\end{aligned}$$ where $\gamma_{ab}$ are the metric degrees of freedom left over in [eq:knadsnhmkerr]{} after setting $dr=dt=0.$
$NHEK$-Map {#sec:NHEKl}
----------
As mentioned above, the solution [eq:knadsnhmkerr]{} is a $near-NHEK$ solution which can be mapped into [eq:nhek]{} via a similar coordinate relationship between $U$ and $r$ as in [@Bardeen:1999px]. Starting with [eq:knadsnhmkerr]{} and applying the coordinate redefinitions: $$\begin{aligned}
\label{eq:nhekcoor24}
r=GM+\lambda U,~t=\frac{\tilde t}{\lambda},~\phi=\tilde \phi+\frac{\tilde t}{2GM\lambda},\end{aligned}$$ and in stead of taking the limit as $\lambda\to0$ we set $a\to GM$ resulting in the $NHEK$ solution: $$\begin{aligned}
\label{eq:nhekfus}
ds^2=\frac{1+\cos^2{\theta}}{2}\left[-\frac{U^2}{{\ell}^2}d\tilde t^2+\frac{{\ell}^2}{U^2}dU^2+{\ell}^2d\theta^2\right]+{\ell}^2\frac{2\sin^2{\theta}}{1+\cos^2{\theta}}\left(d\tilde \phi+\frac{U}{{\ell}^2}d\tilde t\right)^2.\end{aligned}$$ In fact, as discussed in [@Amsel:2009ev; @Maldacena:1998uz; @Bredberg:2009pv], there always exists a coordinate mapping between a $near-NHEK$ and a $NHEK$ type solution as evident in the form of their Kretschmann invariants: $$\begin{aligned}
R_{\alpha\mu\beta\nu}R^{\alpha\mu\beta\nu}=\frac{1536 \sin ^2\theta (-52 \cos (2 \theta )+\cos (4 \theta )-45)}{(\cos (2 \theta )+3)^6 {\ell}^4}
\begin{cases}
{\ell}^2=2G^2M^2& NHEK\\
{\ell}^2=2G^2M^2& near-NHEK\\
{\ell}^2=r^2_++a^2& gnNHEK
\end{cases}.\end{aligned}$$ Notice, that for the choice $a=GM$ all invariants exhibit ${\ell}^2=2G^2M^2$, i.e. the extremal limit of [eq:knadsnhmkerr]{}.
Quantum Fields in $gnNHEK$ Spacetime {#sec:qftnhnek}
====================================
We will now study the resulting near horizon matter theory via the RW dimensional reduction procedure. Our goal, for this section, will be to apply our previous techniques from [@ry; @Button:2010kg] to study the resulting thermodynamics of the $gnNHEK$ within an $AdS_2/CFT_1$ formalism.
RW Dimensional Reduction {#sec:RWdr}
------------------------
In our initial ansatz leading to the $gnNHEK$ solution we assumed a specific decomposition of our four dimensional spacetime into two dimensional black hole and matter fields. However, we have not shown that these fields are the correct RW2DA useful in a holographic study of the quantum spacetime in the near horizon regime.
Let us consider a single free scalar field in the background of [eq:knadsnhmkerr]{} with functional: $$\begin{aligned}
\label{eq:freescalar4}
{\begin{split}}S_{free}=&\frac12\int d^4x\sqrt{-g}g^{\mu\nu}\partial_{\mu}\varphi\partial_\nu\varphi\\
=&-\frac12\int d^4x\,\varphi\left[\partial_\mu\left(\sqrt{-g}g^{\mu\nu}\partial_\nu\right)\right]\varphi\\
=&-\frac12\int d^4x\,\varphi\left[\partial_t\left( -\sin{\theta}\left(a^2+r_+^2\right)\frac{r_+^2+a^2}{r^2-2rGM+a^2}\partial_t\right)+\right.\\
&\partial_r\left( \sin{\theta}\left(a^2+r_+^2\right)\frac{r^2-2rGM+a^2}{r_+^2+a^2}\partial_r\right)+\partial_\theta\left( \sin{\theta}\partial_\theta\right)+\\
&\partial_\phi\left( \left\{-\sin{\theta}\left(a^2+r_+^2\right)\left(\frac{r-2GM}{r_+^2+a^2}\right)^2\frac{r_+^2+a^2}{r^2-2rGM+a^2}+\frac{(\cos{(2 \theta )}+3)^2}{16\sin\theta}\right\}\partial_\phi\right)+\\
&\left.2\partial_t\left( \sin{\theta}\left(a^2+r_+^2\right)\frac{r-2GM}{r_+^2+a^2}\frac{r_+^2+a^2}{r^2-2rGM+a^2}\partial_\phi\right)\right]\varphi.
{\end{split}}\end{aligned}$$ The above functional is reduced to a two dimensional theory by expanding the four dimensional scalar field in terms of spherical harmonics $$\begin{aligned}
\label{eq:sphdecom}
\varphi(t,r,\theta,\phi)=\sum_{lm}\varphi_{lm}(r,t)Y\indices{_l^m}(\theta,\phi),\end{aligned}$$ where $\varphi_{lm}$ is a complex interacting two dimensional scalar field, and integrating out angular degrees of freedom. Transforming to tortoise coordinates defined as: $$\begin{aligned}
\label{eq:ttcoor}
\frac{dr^{*}}{dr}=\frac{1}{f(r)}\end{aligned}$$ and considering the region very close to $r_+$ we find the two dimensional action is much reduced since all interaction, mixing and potential terms ($\sim l(l+1)\ldots$) are weighted by a factor of $f(r(r*))\sim e^{2\kappa r^{*}}$, which vanishes exponentially fast as $r\to r_+$. This leaves us with an infinite collection of massless charged scalar fields in the very near horizon region, with $U(1)$ gauge charge $m$ and remnant functional: $$\begin{aligned}
\label{eq:nhapw}
S=-\frac{r_+^2+a^2}{2}\int d^2x\;\varphi^{*}_{lm}\left[-\frac{1}{f(r)}\left(\partial_t-im{\mathcal{A}}_t\right)^2+\partial_rf(r)\partial_r\right]\varphi_{lm}.\end{aligned}$$ Thus, we arrive at the RW2DA for the $ngNHEK$ solution given by: $$\begin{aligned}
\label{eq:2drwamet}
{g\indices}{^{(2)}_\mu_\nu}=\left(\begin{array}{cc}-f(r) & 0 \\0 & \frac{1}{f(r)}\end{array}\right)\end{aligned}$$ and $U(1)$ gauge field $$\begin{aligned}
\label{eq:rw2dgf}
{\mathcal{A}}={\mathcal{A}}_tdt.\end{aligned}$$
Given the initial ansatz [eq:knadsnhm]{}, it is not surprising that the only relevant physical fields in the region $r\sim r_+$ are the RW2DAs above and makes the holographic statement that we may learn much about the quantum nature of spacetime in the near horizon regime via the semiclassical analysis of ${g\indices}{^{(2)}_\mu_\nu}$, ${\mathcal{A}}$ and $\varphi_{lm}$.
Effective Gravitational Action and Asymptotic Symmetries {#sec:asefa}
--------------------------------------------------------
We would like to interpret [eq:nhapw]{} as a useful action for gravity in the near horizon of classical four dimensional spacetime. This can be done by only considering the $s-wave$ contribution and making a field redefinition rendering the scalar field unitless [@Yale:2010tn; @Chung:2010xy; @Chung:2010xz; @solodukhin:1998tc]. The $s-wave$ approximation is sensible in this scenario, since we will interpret $\varphi_{lm}$ as a component of the gravitational field and hence it should be real and unitless. Most of the interesting gravitational dynamics seem to be contained in this region or approximation [@strom1][^6]. Motivated by these arguments we make the field redefinition $$\begin{aligned}
\label{eq:sfrd}
\varphi_{00}=\sqrt{\frac{6}{G}}\psi,\end{aligned}$$ where $\psi$ is now unitless and the $\sqrt{6}$ was chosen to recover the Einstein coupling $\frac{1}{16\pi G}$ in the quantum gravitational effective action of [eq:nhapw]{} within the $s$-wave approximation. Applying the field redefinition [eq:sfrd]{} to [eq:nhapw]{} yields: $$\begin{aligned}
\label{eq:tdtrw2}
S^{(2)}[\psi,g]=\frac{3(r_+^2+a^2)}{G}\int d^2x\sqrt{-g^{(2)}}\psi\left[D_{\mu}\left(\sqrt{-g^{(2)}}{g\indices}{_{(2)}^{\mu\nu}}D_{\nu}\right)\right]\psi,\end{aligned}$$ where $D_\mu$ is the gauge covariant derivative. In addition to redressing the scalar field, our choice of field redefinition has also rendered the effective coupling unitless, hinting towards a finite quantum theory. The effective action of this quantum theory, which may be extracted via zeta-function regularization of the functional determinant in [eq:tdtrw2]{}, is given by the sum of two functionals [@isowill; @Leutwyler:1984nd]: $$\begin{aligned}
\label{eq:nhpcft}
\Gamma=&\Gamma_{grav}+\Gamma_{U(1)},\end{aligned}$$ where $$\begin{aligned}
{\begin{split}}\Gamma_{grav}=&\frac{(r_+^2+a^2)}{16\pi G\Xi}\int d^2x\sqrt{-g^{(2)}}R^{(2)}\frac{1}{\square_{g^{(2)}}}R^{(2)}~\mbox{and}\\
\Gamma_{U(1)}=&\frac{3 e^2 (r_+^2+a^2)}{\pi G\Xi}\int {\mathcal}{F}\frac{1}{\square_{g^{(2)}}}{\mathcal}{F}.
{\end{split}}\end{aligned}$$ Next, we introduce the auxiliary scalars $\Phi$ and $B$ satisfying: $$\begin{aligned}
\label{eq:afeqm}
\square_{g^{(2)}} \Phi=R~\mbox{and}~\square_{g^{(2)}} B=\epsilon^{\mu\nu}\partial_\mu {\mathcal{A}}_\nu,\end{aligned}$$ which transforms the functional [eq:nhpcft]{} into a Liouville $CFT$ of the form: $$\begin{aligned}
\label{eq:nhlcft}
{\begin{split}}S_{NHCFT}=&\frac{(r_+^2+a^2)}{16\pi G\Xi}\int d^2x\sqrt{-g^{(2)}}\left\{-\Phi\square_{g^{(2)}}\Phi+2\Phi R^{(2)}\right\}\\
&+\frac{3 e^2 (r_+^2+a^2)}{\pi G\Xi}\int d^2x\sqrt{-g^{(2)}}\left\{-B\square_{g^{(2)}}B\right.\\
&+\left.2B \left(\frac{\epsilon^{\mu\nu}}{\sqrt{-g^{(2)}}}\right)\partial_\mu A_\nu\right\}
{\end{split}}\end{aligned}$$
Now, we turn our attention to computing the ASG of [eq:tdtrw2]{}. The behavior of the RW2DA fields at large $r$ is defined by $$\begin{aligned}
\label{eq:as2dwa}
{g\indices}{^{(0)}_\mu_\nu}=&
\left(
\begin{array}{cc}
-\frac{r^2}{{\ell}^2}+\frac{2 r G M}{{\ell}^2}-\frac{a^2}{{\ell}^2}+\mathcal{O}\left(\left(\frac{1}{r}\right)^3\right)& 0 \\
0 & \frac{{\ell}^2}{r^2}+\mathcal{O}\left(\left(\frac{1}{r}\right)^3\right)
\end{array}
\right),\\
\label{eq:asgf}
\mathcal{A}\indices{^{(0)}_t}=&\frac{r}{{\ell}^2}-\frac{2 G M}{{\ell}^2}+\mathcal{O}\left(\left(\frac{1}{r}\right)^3\right),\end{aligned}$$ which yield an asymptotically $AdS_2$ configuration with Ricci Scalar, $R=-\frac{2}{l^2}+O\left(\left(\frac{1}{r}\right)^1\right)$, where ${\ell}^2=r_+^2+a^2$. In addition, we impose the following metric and gauge field fall-off conditions: $$\begin{aligned}
\label{eq:mbc}
\delta g_{\mu\nu}=
\left(
\begin{array}{cc}
\mathcal{O}\left(\left(\frac{1}{r}\right)^3\right)&
\mathcal{O}\left(\left(\frac{1}{r}\right)^0\right) \\
\mathcal{O}\left(\left(\frac{1}{r}\right)^0\right) &
\mathcal{O}\left(r\right)
\end{array}
\right)~\mbox{and}~\delta \mathcal{A}=\mathcal{O}\left(\left(\frac{1}{r}\right)^0\right),\end{aligned}$$ which imply the following set of asymptotic metric preserving diffeomorphisms: $$\begin{aligned}
\label{eq:dpr}
\xi_n=\xi_1(r)\frac{e^{i n \kappa\left(t\pm r^*\right)}}{\kappa}\partial_t+\xi_2(r)\frac{e^{i n \kappa\left(t\pm r^*\right)}}{\kappa}\partial_r,\end{aligned}$$ where $r^*$ is the tortoise coordinate defined by $dr^*=\frac{1}{f(r)}dr$, $$\begin{aligned}
\xi_1=C r e^{i n \kappa r^*},~\xi_2=\frac{i r C (r-G M)}{\kappa n \left(r^2-2 r G M+a^2\right)},\end{aligned}$$ $C$ is an arbitrary normalization constant and $\kappa$ is the surface gravity of the $gnNHEK$ black hole. Under this set of diffeomorphisms the gauge field transforms as: $$\begin{aligned}
\delta_\xi \mathcal{A}_\mu=\left(\mathcal{O}\left(\left(\frac{1}{r}\right)^0\right),\mathcal{O}\left(\left(\frac{1}{r}\right)^1\right)\right)\end{aligned}$$ and thus $\delta_{\xi}$ may be elevated to a total symmetry of the action, i.e. $$\begin{aligned}
\delta_\xi\rightarrow\delta_{\xi+\Lambda},\end{aligned}$$ in accordance with [eq:mbc]{}. Switching to light cone coordinates $x^\pm=t\pm r^*$,[^7] we see that the set $\xi_n^\pm$ is well behaved on the $r\rightarrow\infty$ boundary and form a centerless Witt or $Diff(S^1)$ subalgebra: $$\begin{aligned}
i\left\{\mathbf{\xi}^\pm_m,\mathbf{\xi}^\pm_n\right\}=(m-n)\mathbf{\xi}^\pm_{m+n}.\end{aligned}$$
Energy Momentum and The Virasoro algebra {#sec:thermo}
----------------------------------------
We define the energy momentum tensor and $U(1)$ current of [eq:nhlcft]{} in their usual ways: $$\begin{aligned}
\label{eq:emt}
{\begin{split}}\left\langle T_{\mu\nu}\right\rangle=&\frac{2}{\sqrt{-g^{(2)}}}\frac{\delta S_{NHCFT}}{\delta g\indices{^{(2)}^\mu^\nu}}\\
=&\frac{r_+^2+a^2}{8\pi G}\left\{\partial_\mu\Phi\partial_\nu\Phi-2\nabla_\mu\partial_\nu\Phi+g\indices{^{(2)}_\mu_\nu}\left[2R^{(2)}-\frac12\nabla_\alpha\Phi\nabla^\alpha\Phi\right]\right\}\\
&+\frac{6 e^2 (r_+^2+a^2)}{\pi G}\left\{\partial_\mu B\partial_\nu B-\frac12{g\indices}{_\mu_\nu}\partial_\alpha B\partial^\alpha B\right\}~\mbox{and}\\
\left\langle J^{\mu}\right\rangle=&\frac{1}{\sqrt{-g^{(2)}}}\frac{\delta S_{NHCFT}}{\delta \mathcal{A}_\mu}=\frac{6 e^2 (r_+^2+a^2)}{\pi G}\frac{1}{\sqrt{-g^{(2)}}}\epsilon^{\mu\nu}\partial_\nu B
{\end{split}}\end{aligned}$$ Next, solving the equation of motions for the auxiliary fields: $$\begin{aligned}
\label{eq:eqmp}
{\begin{split}}\square_{g^{(2)}}\Phi=&R^{(2)}\\
\square_{g^{(2)}}B=&\epsilon^{\mu\nu}\partial_\mu \mathcal{A}_\nu
{\end{split}}\end{aligned}$$ using the metric [eq:2drwamet]{} and gauge field [eq:rw2dgf]{} and employeeing modified Unruh Vacuum boundary conditions (MUBC) [@unruh] $$\begin{aligned}
\label{eq:ubc}
\begin{cases}
\left\langle T_{++}\right\rangle=\left\langle J_{+}\right\rangle=0&r\rightarrow\infty,~{\ell}\rightarrow\infty\\
\left\langle T_{--}\right\rangle=\left\langle J_{-}\right\rangle=0&r\rightarrow r_+
\end{cases},\end{aligned}$$ we determine all relevant integration constants of [eq:emt]{} and [eq:eqmp]{}. For large $r$ and to $\mathcal{O}(\frac{1}{{\ell}})^2$, which we will denote as the single limit $r\to\infty$ in the remainder of this section, the resulting energy momentum tensor is dominated by one holomorphic component, $\left\langle T_{--}\right\rangle$. We are interested in the response of the energy momentum tensor and the $U(1)$ current to a total symmetry $\delta_{\xi^-_n+\Lambda}$, which after expanding in terms of the boundary fields [eq:as2dwa]{} and [eq:asgf]{} we obtain: $$\begin{aligned}
\begin{cases}
\delta_{\xi^-_n+\Lambda}\left\langle T_{--}\right\rangle=\xi^-_n\left\langle T_{--}\right\rangle'+2\left\langle T_{--}\right\rangle\left(\xi^-_n\right)'+\frac{r_+^2+a^2}{4\pi G}\left(\xi^-_n\right)'''+\mathcal{O}\left(\left(\frac{1}{r}\right)^3\right)\\
\delta_{\xi^-_n+\Lambda}\left\langle J_{-}\right\rangle=\mathcal{O}\left(\left(\frac{1}{r}\right)^3\right)
\end{cases}\end{aligned}$$ From the above we see that $\left\langle T_{--}\right\rangle$ transforms asymptotically as the energy momentum tensor of a one dimensional $CFT$ with center: $$\begin{aligned}
\label{eq:center}
\frac{c}{24\pi}=\frac{r_+^2+a^2}{4\pi G}\Rightarrow c=\frac{3A}{2\pi G},\end{aligned}$$ We should also note that the above central charge is in congruence with the 2-dimensional conformal/trace anomaly [@cft]: $$\begin{aligned}
\label{eq:tra}
\left\langle T\indices{_\mu^\mu}\right\rangle=-\frac{c}{24\pi}R^{(2)}\end{aligned}$$
Next, we define the quantum generators via the conserved charge: $$\begin{aligned}
\label{eq:ccppb}
{\mathcal{Q}}_n=\lim_{r\rightarrow\infty}\int dx^-\left\langle T_{--}\right\rangle\mathbf{\xi}^-_n,\end{aligned}$$ Whose algebraic structure is revealed by computing its response to a total symmetry, while compactifying the $x^-$ coordinate to a circle from $0\to2\pi/\kappa$: $$\begin{aligned}
\label{eq:ca}
\delta_{\xi^-_m+\Lambda}{\mathcal{Q}}_n=\left[{\mathcal{Q}}_m,{\mathcal{Q}}_n\right]=(m-n){\mathcal{Q}}_n+\frac{c}{12}m\left(m^2-1\right)\delta_{m+n,0},\end{aligned}$$ from the above we see that the quantum symmetry generators form a centrally extended Virasoro algebra with regulated/normalized zero-mode ${\mathcal{Q}}_0=\frac{A}{16\pi G}$.
$AdS_2/CFT_1$ and Entropy of Near Extremal Kerr {#sec:ent}
-----------------------------------------------
By employing the finite mass gauge, we have now shown that the near-extremal Kerr throat is holographically dual to a $CFT$ with center $$\begin{aligned}
\label{eq:cre}
c&=\frac{3A}{2\pi G}\end{aligned}$$ and lowest Virasoro eigen-mode $$\begin{aligned}
{\mathcal{Q}}_0&=\frac{A}{16\pi G}.\end{aligned}$$ We are now free to use the above results within the statistical Cardy Formula [eq:cf]{}: $$\begin{aligned}
S=2\pi\sqrt{\frac{c{\mathcal{Q}}_0}{6}}=\frac{A}{4G}=2\pi\left(GM^2+\sqrt{G^2M^4-(GJ)^2}\right),\end{aligned}$$ which is in agreement with the area law [eq:mal]{}, however derived without mixing results computed separately at near-extremality and extremality. Also the $c$ and ${\mathcal{Q}}_0$ extend smoothly to extremality in the limit as $a\to GM$ yielding: $$\begin{aligned}
\lim_{a\to GM}c=12J~\text{and}~\lim_{a\to GM}{\mathcal{Q}}_0=J/2\end{aligned}$$ which is the same value of the left central charge obtained in the Kerr/$CFT$ correspondence [@kerrcft] and together in the statistical Cardy formula the above values reproduce the extremal Kerr entropy. In addition, our derived zero-mode in [eq:cre]{} is in accordance with [@Button:2010kg] and relates to the irreducible mass of its black hole dual via:
\[ass:zerom\] The lowest Virasoro eigen-mode of a quantum $CFT$ is proportional to the irreducible mass of its black hole dual via the form $$\begin{aligned}
\label{eq:irrmq0}
\mathcal{Q}_0=GM_{irr}^2,\end{aligned}$$
This may be a general statement with a large avenue of application, though we are not aware of a rigorous proof at this time.
Near Extremal Black Hole Temperature {#sec:temp}
------------------------------------
To extract the temperature of the $gnNHEK$ horizon, we will focus on the gravitational part of [eq:nhlcft]{}, i.e. $$\begin{aligned}
\label{eq:nhlcftgrav}
S_{grav}=&\frac{(r_+^2+a^2)}{16\pi G\Xi}\int d^2x\sqrt{-g^{(2)}}\left\{-\Phi\square_{g^{(2)}}\Phi+2\Phi R^{(2)}\right\}.\end{aligned}$$ The energy momentum is given by: $$\begin{aligned}
\label{eq:emtgra}
{\begin{split}}\left\langle T_{\mu\nu}\right\rangle=&\frac{2}{\sqrt{-g^{(2)}}}\frac{\delta S_{NHCFT}}{\delta g\indices{^{(2)}^\mu^\nu}}\\
=&\frac{r_+^2+a^2}{8\pi G}\left\{\partial_\mu\Phi\partial_\nu\Phi-2\nabla_\mu\partial_\nu\Phi+g\indices{^{(2)}_\mu_\nu}\left[2R^{(2)}-\frac12\nabla_\alpha\Phi\nabla^\alpha\Phi\right]\right\}
{\end{split}}\end{aligned}$$ and following the steps [eq:emt]{} through [eq:ubc]{}, however focusing on the horizon limit $r\to r_+$, we are left with one holomorphic component given by: $$\begin{aligned}
\label{eq:hhf}
\left\langle T_{++}\right\rangle=-\frac{r_+^2+a^2}{32 \pi G }f'\left(r^+\right)^2.\end{aligned}$$ This is precisely the value of the Hawking Flux of the $gnNHEK$ metric weighted by the central charge [eq:center]{}: $$\begin{aligned}
\label{eq:htfhf}
\left\langle T_{++}\right\rangle=cHF=-c\frac{\pi}{12}\left(T_H\right)^2\Rightarrow,\end{aligned}$$ with Hawking temperature[@Jinwu; @caldarelli:1999xj]: $$\begin{aligned}
T_H=\frac{f'\left(r^+\right)}{4 \pi }.\end{aligned}$$ This is an interesting result, which hints that the $AdS_2/CFT_1$ correspondence constructed here contains information about both black hole entropy and temperature. Though the $\left\langle T_{++}\right\rangle$ component in the horizon limit is not precisely the Hawking Flux of the four dimensional parent black hole, but given prior knowledge of the central extension, it is possible to read of or extract the relevant information from the correspondence.
$gnNHEK$ and $AdS_2$ Quantum Gravity {#sec:2da}
====================================
We will now turn our attention to the extremal case $a=GM$ in [eq:knadsnhmkerr]{} for which the surface gravity vanishes identically. In other words, in contrast to the previous section we are interested in a formalism allowing $a=GM$ from the outset of the calculation, which is where the finite mass gauge will come in handy since in this limit the line element [eq:knadsnhmkerr]{} still exhibits an interesting Boyer-Lindquist structure. This specific case is cumbersome, as the Charge regulators of the previous section depended on $\kappa$ thus we will follow the seminal work of [@Castro:2009jf; @Hartman:2008dq; @Castro:2008ms; @Chen:2011gz; @Sen:2008vm; @Skenderis:2002wp]. The convenient decomposition of [eq:knadsnhmkerr]{} into two dimensional fields should allow for an off-shell analysis of the Einstein-Hilbert action and integrating out angular degrees of freedom should leave us with an alternate effective two dimensional theory. As already seen in the previous section, the resulting effective theory should hold computational significance relevant to the near horizon regime of the near-extremal and extremal Kerr black hole.
Bulk Action {#sec:2dBa}
-----------
Writing [eq:knadsnhmkerr]{} in terms of two dimensional fields and making the field redefinition $e^{-2\varphi(r)}\to e^{-2\varphi(r)}(r_+^2+a^2)$ we have $$\begin{aligned}
\label{eq:nhnek}
ds^2=\frac{1+\cos^2\theta}{2}\left[{g\indices}{^{(2)}_\mu_\nu}dx^\mu dx^\nu+e^{-2\varphi}(r_+^2+a^2)d\theta^2\right]+\frac{2\sin^2\theta}{1+\cos^2\theta}e^{-2\varphi}(r_+^2+a^2)\left[d\phi+\mathcal{A}\right]^2\end{aligned}$$ and substituting into the Einstein-Hilbert action $$\begin{aligned}
\label{}
S_{EH}=\frac{1}{16\pi G}\int d^4x\sqrt{-g}R\end{aligned}$$ and integrating out angular degrees of freedom we obtain $$\begin{aligned}
\label{eq:ehres}
{\begin{split}}S=&\frac{2\pi}{16\pi G}\int d^2x\left\{2 e^{-4 \varphi(r)}(r_+^2+a^2)^2 {\mathcal{A}}_t'(r)^2\right.\\
&\left.-2 e^{-2\varphi(r)}(r_+^2+a^2) \left(f''(r)-4f'(r)\varphi'(r)+f(r) \left(6\varphi'(r)^2-4 \varphi''(r)\right)\right)+2\right\},
{\end{split}}\end{aligned}$$ which may be recast into covariant form $$\begin{aligned}
\label{eq:nhagr}
{\begin{split}}S=\frac{(r_+^2+a^2)}{4G}\int d^2x\sqrt{-g^{(2)}}\left\{e^{-2\varphi}R^{(2)}+\frac{1}{(r_+^2+a^2)}+2\nabla_\mu e^{-\varphi}\nabla^\mu e^{-\varphi}-\frac{(r_+^2+a^2)}{2}e^{-4\varphi}{\mathcal}{F}^2\right\}\\
+\int d^2x\left\{\text{Total Derivative Terms}\cdots\right\}.
{\end{split}}\end{aligned}$$ Setting ${\ell}^2=(r_+^2+a^2)$ and dropping total derivatives we obtain the $AdS_2$ action: $$\begin{aligned}
\label{eq:nhagder}
{\begin{split}}S_{AdS_2}=\frac{{\ell}^2}{4G}\int d^2x\sqrt{-g^{(2)}}\left\{e^{-2\varphi}R^{(2)}+\frac{1}{{\ell}^2}+2\nabla_\mu e^{-\varphi}\nabla^\mu e^{-\varphi}-\frac{{\ell}^2}{2}e^{-4\varphi}{\mathcal}{F}^2\right\}.
{\end{split}}\end{aligned}$$ This theory is a classical effective gravity theory and exhibits regularization via a suitable choice of boundary counterterms. This process is equivalent to renormalizing the theory on the $CFT$ side to ensure finite energy momentum and current [@Skenderis:2002wp]. [eq:nhagder]{} exhibits three equations of motion obtained by variation with respect to ${g\indices}{^{(2)}_{\mu\nu}}$, ${\mathcal{A}}$ and $\varphi.$ Remembering that for any two dimensional Riemannian metric $R_{\mu\nu}-\frac12g_{\mu\nu}R=0$ we obtain: $$\begin{aligned}
\label{eq:daseqm1}
T_{\mu\nu}=&0 &Einstein,\\\label{eq:daseqm2}
\nabla_{\mu}{\mathcal}{F}^{\mu\nu}=&0 &Maxwell,\\\label{eq:daseqm3}
R^{(2)}-e^{-2\varphi}{\ell}^2{\mathcal}{F}^2=&0 &constant~Scalar.\end{aligned}$$ We will be interested in constructing the boundary counterterms to our above derived $AdS_2$ action by considering general solutions to [eq:daseqm1]{}-[eq:daseqm3]{}. A general set, as written in Gauss normal form, is given by $$\begin{aligned}
\label{eq:gngsln}
ds^2_{(2)}=e^{-2\varphi}d\rho^2+\gamma_{tt} dt^2~\text{and}~{\mathcal{A}}={\mathcal{A}}_\rho d\rho+{\mathcal{A}}_t dt,\end{aligned}$$ where $$\begin{aligned}
\label{eq:gngslneach}
{\begin{split}}\gamma_{tt}=&-\frac14e^{-2\varphi}\left(e^{\rho/{\ell}}-h(t)e^{-\rho/{\ell}}\right)^2,\\
{\mathcal{A}}_t=&\frac{1}{2{\ell}}e^{\rho/{\ell}}\left(1-\sqrt{h(t)}e^{-\rho/{\ell}}\right)^2,\\
{\mathcal{A}}_\rho=&0,
{\end{split}}\end{aligned}$$ and includes a free function $h(t)$. The $KLBH$ solution is recovered for the choice $$\begin{aligned}
\label{eq:hchoice}
h(t)=\frac{G^2M^2-a^2}{{\ell}^2}\end{aligned}$$ and coordinate redefinition: $$\begin{aligned}
\label{eq:coordef}
(r-GM)=\frac{{\ell}}{2}e^{\rho/{\ell}}\left(1+\frac{G^2M^2-a^2}{{\ell}^2}e^{-2\rho/{\ell}}\right).\end{aligned}$$ The extremal case corresponds to the case $h(t)=0$.
Boundary Counterterms {#sec:2dBact}
---------------------
Following [@Castro:2009jf; @Hartman:2008dq; @Castro:2008ms], we now determine the boundary counterterms within a well defined variational principle for [eq:nhagder]{}. The boundary contribution has the local form: $$\begin{aligned}
\label{eq:Sinct}
S_{ct}=\frac{{\ell}^2}{2G}\int dt\sqrt{-\gamma}\left\{e^{-2\varphi}K+\alpha e^{-\varphi}+\beta e^{-3\varphi}{\mathcal{A}}_a{\mathcal{A}}^a\right\},\end{aligned}$$ where the first term above is just the standard Gibbons-Hawking term for extrinsic curvature $K=\frac12\gamma^{tt}\sqrt{g_{\rho\rho}}\partial_\rho \gamma_{tt}=e^\varphi/{\ell}$, and $\alpha$ and $\beta$ are yet to be determined constants. Considering full variations including boundary terms of [eq:nhagder]{} we are left with solving the constraint equations: $$\begin{aligned}
\label{eq:coneq1}
{\begin{split}}\pi_{ab}\delta\gamma^{ab}=&0,\\
\pi^{a}\delta{\mathcal{A}}_{a}=&0,\\
\pi_{\varphi}\delta\varphi=&0,
{\end{split}}\end{aligned}$$ i.e. we require vanishing canonical momenta on the boundary. Expanding the momenta on the asymptotic $AdS_2$ boundary defined by the zeroth order fields: $$\begin{aligned}
\label{eq:ads2bcl}
{\begin{split}}\gamma\indices{^{(0)}_{tt}}=&-\frac14e^{-2\varphi^{(0)}}e^{2\rho/{\ell}},\\
{\mathcal{A}}\indices{^{(0)}_t}=&\frac{1}{2{\ell}}e^{\rho/{\ell}},\\
\varphi^{(0)}=&constant,
{\end{split}}\end{aligned}$$ and solving the constraint equations [eq:coneq1]{} we find: $$\begin{aligned}
\label{eq:absol}
\alpha=-\frac{1}{2{\ell}}~\text{and}~\beta=\frac{{\ell}}{2}.\end{aligned}$$ Substituting these values back into [eq:Sinct]{} and summarizing we obtain the total renormalized action: $$\begin{aligned}
\label{eq:fads2act}
{\begin{split}}S_{AdS_2}^{ct}=&\frac{{\ell}^2}{4G}\int d^2x\sqrt{-g^{(2)}}\left\{e^{-2\varphi}R^{(2)}+\frac{1}{{\ell}^2}+2\nabla_\mu e^{-\varphi}\nabla^\mu e^{-\varphi}-\frac{{\ell}^2}{2}e^{-4\varphi}{\mathcal}{F}^2\right\}+\\
&\frac{{\ell}^2}{2G}\int dt\sqrt{-\gamma}\left\{e^{-2\varphi}K-\frac{1}{2{\ell}} e^{-\varphi}+\frac{{\ell}}{2} e^{-3\varphi}{\mathcal{A}}_a{\mathcal{A}}^a\right\}.
{\end{split}}\end{aligned}$$ The above action is nearly identical to the one derived in [@Castro:2009jf], but differs in the coupling $\frac{{\ell}^2}{4G}$.
Boundary Currents, Asymptotic Symmetries and Central Extension {#sec:bemtccc}
--------------------------------------------------------------
The boundary energy momentum tensor and $U(1)$ current are defined as $$\begin{aligned}
\label{eq:emtsol}
{\begin{split}}T_{tt}=&\frac{2}{\sqrt{-\gamma}}\frac{\delta S_{AdS_2}^{ct}}{\delta \gamma\indices{^t^t}}=-\frac{{\ell}^2}{4G}\left(\frac{e^{-\varphi}}{{\ell}}\gamma_{tt}+{\ell}e^{-3\varphi}{\mathcal{A}}_t{\mathcal{A}}_t\right)\\
J_t=&\frac{1}{\sqrt{-\gamma}}\frac{\delta S_{AdS_2}^{ct}}{\delta \mathcal{A}_t}=\frac{{\ell}^2e^{-3\varphi}}{2G}\left(-e^{-\varphi}{\ell}^2n^\mu{\mathcal}{F}_{\mu t}+{\ell}{\mathcal{A}}_t\right),
{\end{split}}\end{aligned}$$ where $n^\mu$ is the radial ($\rho$) normal. We are interested in total symmetries on the $AdS_2$ boundary [eq:ads2bcl]{} preserving the conditions: $$\begin{aligned}
\label{eq:gmbc}
\delta g_{\rho\rho}=\delta g_{t\rho}=0,~\delta g_{tt}=0\cdot^{2\rho/{\ell}},~\delta{\mathcal{A}}=0,\end{aligned}$$ and gauge ${\mathcal{A}}_\rho=0$. A general set of diffeomorphisms preserving the above is given by: $$\begin{aligned}
\label{eq:gmbcdiff}
{\begin{split}}\epsilon=&\left[\xi(t)+2{\ell}^2\left(e^{2\rho/{\ell}}-h(t)\right)^{-1}\xi''(t)\right]\partial_t-{\ell}\xi'(t)\partial_\rho\\
\Lambda=&-2{\ell}e^{-\rho/{\ell}}\left(1+\frac{G^2M^2-a^2}{{\ell}^2}e^{-\rho/{\ell}}\right)^{-1}\xi''(t),
{\end{split}}\end{aligned}$$ where $\xi(t)$ is an undetermined function of time. Expanding the boundary energy momentum tensor of [eq:emtsol]{} in terms of boundary fields [eq:ads2bcl]{} and computing its response to a total symmetry we find[^8]: $$\begin{aligned}
\label{eq:bemtrp}
{\begin{split}}\delta_{\epsilon+\Lambda}T_{tt}=&2T_{tt}\xi'(t)+\xi(t)T'_{tt}+\frac{c}{12}{\ell}\xi'''(t)+\sqrt{h}\cdot{\mathcal}{O}\left(e^{\rho/{\ell}}\right),~\text{where}\\
c=&\frac{6{\ell}^2}{G},
{\end{split}}\end{aligned}$$ which is precisely the transformation law of an energy momentum tensor of a one dimensional $CFT$[^9]. we should note that the factor of ${\ell}$ is separated in the anomalous term to ensure proper units of the central charge given by: $$\begin{aligned}
c=\frac{6\left(r_+^2+a^2\right)}{G}=\frac{3A}{2\pi G},\end{aligned}$$ in accord with [eq:center]{}. Implementing Assertion \[ass:zerom\] we have $$\begin{aligned}
{\begin{split}}c=&\frac{3A}{2\pi G},\\
{\mathcal{Q}}_0=&\frac{A}{16\pi G},
{\end{split}}\end{aligned}$$ which together inside [eq:cf]{} gives $$\begin{aligned}
S=\frac{A}{4G}\end{aligned}$$ reproducing the standard area law.
In addition, we have avoided any regulators depending on factors of $\sim1/\kappa$ or $\sim1/h$ and thus, the above analysis may be repeated for the specific case $a=GM$ from the outset, for which [eq:bemtrp]{} becomes: $$\begin{aligned}
{\begin{split}}\delta_{\epsilon+\Lambda}T_{tt}=&2T_{tt}\xi'(t)+\xi(t)T'_{tt}+\frac{c}{12}{\ell}\xi'''(t),~\text{where}\\
c=&12J.
{\end{split}}\end{aligned}$$ The above result is precisely the left central charge of the Kerr/$CFT$ correspondence [@kerrcft] obtained from the $gnNHEK$ solution in the limit $a=GM.$
Conclusion and Comments {#sec:concom}
=======================
To conclude, we have analyzed quantum near-extremal Kerr black hole properties in the near horizon regime via the construction of an $AdS_2/CFT_1$ correspondence of the $gnNHEK$ metric, as outlined in Table \[tb:adscftc\], and extending our previous analysis [@ry; @Button:2010kg] to a new spacetime. The main results of our work includes the central charge $c=\frac{3A}{2\pi G}$, which was computed via a Lagrangian analysis of conserved currents of two different near horizon theories.
$CFT$ Black Hole
------------------------- ---------------------------
Conformal Group Asymptotic Symmetry Group
center $\frac{3A}{2\pi G}$
Hamiltonian eigen-value $GM^2_{irr}$
Regulator $\kappa_{gnNHEK}$
: Black-Hole/Near-Horizon-$CFT$ Duality[]{data-label="tb:adscftc"}
It is conceivable that other $AdS_2\times S^2$ gauges, exhibiting the field splitting of [eq:knadsnhm]{}, exist with relevance and physical connections to other classical near-extremal solutions. Many analogues to the $NHEK$ for charged rotating black holes with negative and positive cosmological consents have been shown to exist, see [@Compere:2012jk] for a comprehensive review, which suggests similar such analogues to the $gnNHEK$ solution and similar analysis of this note should be applicable. In particular, Assertion \[ass:zerom\] may be a useful tool in the asymptotic symmetry analysis of other extremal black holes, which have zero surface gravity and hence the need for a thermal Cardy formula [eq:tcf]{}. However extremal black holes in general have well defined horizons, which leads to finite non zero irreducible mass, thus allowing the implementation of a standard statistical Cardy formula [eq:cf]{} thus leading to the Bekenstein-Hawking entropy.
Acknowledgement {#acknowledgement .unnumbered}
===============
We thank Vincent Rodgers, Jacob Willig-Onwuachi, John Baker, Shanshan Rodriguez and the University of Iowa’s Diffeomorphisms and Geometry research group for enlightening discussions. L.R. is grateful to the University of Iowa and NASA Goddard Space Flight Center for their hospitality during the initial and final stages of this work.
This work is supported in part by the HHMI Undergraduate Science Education Award 52006298 and the Grinnell College Academic Affairs’ CSFS and MAP programs.
(1,0)[150]{}
[^1]: <[email protected]>
[^2]: <[email protected]>
[^3]: <[email protected]>
[^4]: <[email protected]>
[^5]: A similar identification can be found in [@Carlip:2011ax; @ChangYoung:2012kd].
[^6]: In [@Button:2010kg] it was shown that $\varphi_{lm}$ dies exponentially fast in time by analyzing the asymptotic behavior of its field equation. However we find the statement relating $\varphi_{lm}$ to a real gravitational field component, a stronger justification to neglect higher order terms in $l$ and $m$.
[^7]: Large $r$ behavior will be synonymous with large $x^+$ behavior.
[^8]: The factor $\frac{1}{12}$ in the anomaly is dependent upon the choice of conformal coordinates and normalization of ${\mathcal{Q}}$. For tortoise light cone coordinates and unit normalization, as in Section \[sec:qftnhnek\], the the factor is $\frac{1}{24\pi}$ [@cft; @Iso:2008sq].
[^9]: In contrast to [eq:clemt]{} and different boundary dependence.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Burgers vortices are explicit stationary solutions of the Navier-Stokes equations which are often used to describe the vortex tubes observed in numerical simulations of three-dimensional turbulence. In this model, the velocity field is a two-dimensional perturbation of a linear straining flow with axial symmetry. The only free parameter is the Reynolds number $\rm{Re} = \Gamma/\nu$, where $\Gamma$ is the total circulation of the vortex and $\nu$ is the kinematic viscosity. The purpose of this paper is to show that Burgers vortex is asymptotically stable with respect to general three-dimensional perturbations, for all values of the Reynolds number. This definitive result subsumes earlier studies by various authors, which were either restricted to small Reynolds numbers or to two-dimensional perturbations. Our proof relies on the crucial observation that the linearized operator at Burgers vortex has a simple and very specific dependence upon the axial variable. This allows to reduce the full linearized equations to a vectorial two-dimensional problem, which can be treated using an extension of the techniques developped in earlier works. Although Burgers vortices are found to be stable for all Reynolds numbers, the proof indicates that perturbations may undergo an important transient amplification if $\rm{Re}$ is large, a phenomenon that was indeed observed in numerical simulations.'
author:
- |
\
Thierry Gallay\
Institut Fourier\
Université de Grenoble I\
BP 74\
38402 Saint-Martin-d’Hères, France\
[[email protected]]{}
- |
\
Yasunori Maekawa\
Faculty of Science\
Kobe University\
1-1 Rokkodai, Nada-ku\
Kobe 657-8501, Japan\
[[email protected]]{}
date: 'February 9, 2010'
title: 'Three-dimensional stability of Burgers vortices'
---
Introduction {#sec.intro}
============
The axisymmetric Burgers vortex is an explicit solution of the three-dimensional Navier-Stokes equations which provides a simple and widely used model for the vortex tubes or filaments that are observed in turbulent flows [@Bu; @To]. Despite obvious limitations, due to oversimplified assumptions, this model describes in a correct way the fundamental mechanisms which are responsible for the persistence of coherent structures in three-dimensional turbulence, namely the balance between vorticity amplification due to stretching and vorticity dissipation due to viscosity. If one believes that vortex tubes play a significant role in the dynamics of turbulent flows, it is an important issue to determine their stability with respect to perturbations in the largest possible class. So far, this problem has been studied only for the axisymmetric Burgers vortex and for a closely related family of asymmetric vortices [@RS; @MKO].
As was shown by Leibovich and Holmes [@LH], one cannot hope to prove energetic stability of the Burgers vortex even if the circulation Reynolds number is very small. To tackle the stability problem, it is therefore necessary to have a closer look at the spectrum of the linearized operator. This is a relatively easy task if we restrict ourselves to [*two-dimensional*]{} perturbations. Assuming that the vortex tube is aligned with the vertical axis, this means that the perturbed velocity field lies in the horizontal plane and does not depend on the vertical variable. Under such conditions, the Burgers vortex is known to be stable for any value of the Reynolds number. This result was first established by Giga and Kambe [@GK] for $\rm{Re} \ll 1$ and then by Gallay and Wayne [@GW1] in the general case. Moreover, a lot is known about the spectrum of the linearized operator, which turns out to be purely discrete in a neighborhood of the origin in the complex plane. Using perturbative expansions, Robinson and Saffman [@RS] showed that all linear modes are exponentially damped for small Reynolds numbers. This property was then numerically verified by Prochazka and Pullin [@PP1] for $\rm{Re} \le 10^4$, and finally rigorously established in [@GW1].
The situation is much more complicated if we allow for arbitrary [*three-dimensional*]{} perturbations. In that case, it was shown by Rossi and Le Dizès [@RD] that the linearized operator does not have any eigenfunction with nontrivial dependence in the vertical variable. While this result precludes the existence of unstable eigenvalues, it also implies that stability cannot be deduced from such a simple analysis, and that continuous spectrum necessarily plays an important role. Unfortunately, the vertical dependence of the perturbed solutions is not easy to determine, as can be seen from the note [@Cr] where a few attempts are made in that direction. The only rigorous result so far is due to Gallay and Wayne [@GW2], who proved that the Burgers vortex is asymptotically stable with respect to three-dimensional perturbations in a fairly large class provided that the Reynolds number is sufficiently small. For larger Reynolds numbers, up to $\rm{Re} = 5000$, an important numerical work by Schmid and Rossi [@SR] indicates that all modes are exponentially damped by the linearized evolution, although significant short-time amplification can occur.
In this paper, we prove that the axisymmetric Burgers vortex is asymptotically stable with respect to [*three-dimensional*]{} perturbations for [*arbitrary values*]{} of the Reynolds number. As in [@GW2], we assume that the perturbations are nicely localized in the horizontal variables, but we do not impose any decay with respect to the vertical variable. Our approach is based on the fact that the linearized operator has a very simple dependence upon the vertical variable: the only term involving $x_3$ is the dilation operator $x_3 \partial_{x_3}$, which originates from the background straining field. This crucial property was already exploited in [@RD; @Cr; @SR], but we shall show that it allows to reduce the three-dimensional stability problem to a two-dimensional one, which can then be treated using an extension of the techniques developped in [@GW1]. Although the spectrum of the linearized operator remains stable for all Reynolds numbers, the estimates we have on the associated semigroup deteriorate as $\rm{Re}$ increases, in full agreement with the amplification phenomena observed in [@SR].
We now formulate our results in a more precise way. We start from the three-dimensional incompressible Navier-Stokes equations: $$\partial_t V + (V,\nabla )V \,=\, \nu\Delta V - \frac{1}{\rho} \nabla P~,
\qquad \nabla \cdot V \,=\, 0~, \label{eq.NS}$$ where $V = V(x,t) \in {\mathbb{R}}^3$ denotes the velocity field, $P = P(x,t)
\in {\mathbb{R}}$ is the pressure field, and $x = (x_1,x_2,x_3)^\top \in {\mathbb{R}}^3$ is the space variable. The parameters in are the kinematic viscosity $\nu > 0$ and the density $\rho > 0$. To obtain tubular vortices, we assume that the velocity $V$ can be decomposed as follows: $$V(x,t) \,=\, V^s(x) + U(x,t)~, \label{eq.velocity}$$ where $V^s$ is an axisymmetric straining flow given by the explicit formula $$V^s(x) \,=\, \frac{\gamma}{2} \begin{pmatrix} -x_1 \\ -x_2 \\ 2x_3
\end{pmatrix} \,\equiv\, \gamma Mx~, \qquad \hbox{where} \quad
M \,=\, \begin{pmatrix} -\frac12 & 0 & 0 \\ 0 & -\frac12 & 0 \\
0 & 0 & 1\end{pmatrix}~.
\label{def.strain}$$ Here $\gamma > 0$ is a parameter which measures the intensity of the strain. Note that $\nabla\cdot V^s = 0$, and that $V^s$ is a stationary solution of with the associated pressure $P^s
= -\frac12 \rho |V^s|^2$. Our goal is to study the evolution of the perturbed velocity field $U(x,t)$.
To simplify the notations, we shall assume henceforth that $\gamma = \nu = \rho = 1$. This can be achieved without loss of generality by replacing the variables $x$, $t$ and the functions $V$, $P$ with the dimensionless quantities $$\tilde x \,=\, \Bigl(\frac{\gamma}{\nu}\Bigr)^{1/2}x~, \qquad
\tilde t \,=\, \gamma t~, \qquad
\tilde V \,=\, \frac{V}{(\gamma \nu)^{1/2}}~, \qquad
\tilde P \,=\, \frac{P}{\rho \gamma \nu}~.$$ For further convenience, instead of considering the evolution of $V$ or $U$, we prefer working with the vorticity field $\Omega =
\nabla \times V = \nabla \times U$. Taking the curl of and using , , we obtain for $\Omega$ the evolution equation $$\partial_t \Omega + (U,\nabla )\Omega - (\Omega,\nabla )U \,=\,
L\Omega~, \qquad \nabla \cdot \Omega \,=\, 0~, \label{eq.vortex}$$ where $L$ is the differential operator defined by $$L\Omega \,=\, \Delta \Omega - (Mx,\nabla )\Omega + M\Omega~.
\label{def.L}$$
Under mild assumptions that will be specified below, the velocity field $U$ can be recovered from the vorticity $\Omega$ via the three-dimensional Biot-Savart law $$U(x) \,=\, -\frac{1}{4\pi}\int_{{\mathbb{R}}^3}\frac{(x-y)\times
\Omega(y)}{|x-y|^3} {\,{\rm d}}y ~=:~ (K_{3D}*\Omega)(x)~.
\label{def.3DBS}$$ In what follows we shall often encounter the particular situation where the velocity $U$ is two-dimensional and horizontal, namely $U(x) = (U_1(x_h),U_2(x_h),0)^\top$ where $x_h =
(x_1,x_2)^\top \in {\mathbb{R}}^2$. In that case the vorticity satisfies $\Omega(x) = (0,0,\Omega_3(x_h))^\top$, and the relation reduces to the two-dimensional Biot-Savart law $$U_h(x_h) \,=\, \frac{1}{2\pi}\int_{{\mathbb{R}}^2} \frac{(x_h-y_h)^\bot}
{|x_h-y_h|^2}\,\Omega_3(y_h) {\,{\rm d}}y_h ~=:~ (K_{2D}\star\Omega_3)(x_h)~,
\label{def.2DBS}$$ where $U_h =(U_1,U_2)^\top$ and $x_h^\bot = (-x_2,x_1)^\top$.
We can now introduce the [*Burgers vortices*]{}, which are explicit stationary solutions of of the form $\Omega = \alpha G$, where $\alpha \in {\mathbb{R}}$ is a parameter. The vortex profile is given by $$\quad~~ G(x) \,=\, \begin{pmatrix} 0 \\ 0 \\ g(x_h)\end{pmatrix}~,
\qquad \hbox{where} \quad
g(x_h) \,=\, \frac{1}{4\pi}\,e^{-|x_h|^2/4}~.
\label{def.g}$$ The associated velocity field $U = \alpha U^G$ can be obtained from the Biot-Savart law and has the following form $$U^G(x) \,=\, u^g(|x_h|^2)\begin{pmatrix} -x_2 \\ x_1 \\ 0\end{pmatrix}~,
\qquad \hbox{where} \quad u^g(r) \,=\, \frac{1}{2\pi r}
\Bigl(1-e^{-r/4}\Bigr)~. \label{def.u^g}$$ If $\Omega = \alpha G$, it is easy to verify that $\alpha =
\int_{{\mathbb{R}}^2}\Omega_3(x_h){\,{\rm d}}x_h$. This means that the parameter $\alpha \in {\mathbb{R}}$ represents the total circulation of the Burgers vortex $\alpha G$. In the physical literature, the quantity $|\alpha|$ is often referred to as the (circulation) Reynolds number.
The aim of this paper is to study the asymptotic stability of the Burgers vortices. We thus consider solutions of of the form $\Omega = \alpha G + \omega$, $U = \alpha U^G + u$, and obtain the following evolution equation for the perturbation: $$\partial_t\omega +(u,\nabla)\omega - (\omega,\nabla)u \,=\,
(L-\alpha \Lambda )\omega~, \qquad \nabla \cdot \omega \,=\,0~,
\label{eq.omega}$$ where $\Lambda$ is the integro-differential operator defined by $$\Lambda \omega \,=\, (U^G,\nabla )\omega - (\omega,\nabla )U^G +
(u,\nabla )G - (G,\nabla)u~. \label{def.Lambda}$$ Here and in the sequel, it is always understood that $u =
K_{3D}*\omega$.
An important issue is now to fix an appropriate function space for the admissible perturbations. Since the Burgers vortex itself is essentially a two-dimensional flow, it is natural to choose a functional setting which allows for perturbations in the same class, but we also want to consider more general ones. Following [@GW2], we thus assume that the perturbations are nicely localized in the horizontal variables, but merely bounded in the vertical direction. As we shall see below, this choice is more or less imposed by the particular form of the linear operator .
To specify the horizontal decay of the admissible perturbations, we first introduce two-dimensional spaces. Given $m \in [0,\infty]$, we denote by $\rho_m : [0,\infty) \to [1,\infty)$ the weight function defined by $$\label{def.weight}
\rho_m(r) \,=\, \begin{cases} 1 & \hbox{if} \quad m = 0\,, \\
(1+\frac{r}{4m})^m & \hbox{if} \quad 0 < m<\infty\,, \\
e^{r/4} & \hbox{if} \quad m = \infty\,.\end{cases}$$ We introduce the weighted $L^2$ space $$L^2(m) \,=\, \Bigl\{f\in L^2({\mathbb{R}}^2)~\Big|~\int_{{\mathbb{R}}^2} |f(x_h)|^2
\rho_m(|x_h|^2){\,{\rm d}}x_h < \infty \Bigr\}~, \label{def.L^2(m)}$$ which is a Hilbert space with a natural inner product. Using Hölder’s inequality, it is easy to verify that $L^2(m)
\hookrightarrow L^1({\mathbb{R}}^2)$ if $m > 1$. In that case, we also define the closed subspace $$L^2_0(m) \,=\, \Bigl\{f \in L^2(m)~\Big|~\int_{{\mathbb{R}}^2} f(x_h)
{\,{\rm d}}x_h = 0\Bigr\}~. \label{def.L^2_0(m)}$$
Next, we define the three-dimensional space $X(m)$ as the set of all $\phi : {\mathbb{R}}^3 \to {\mathbb{R}}$ for which the map $x_h \mapsto \phi(x_h,x_3)$ belongs to $L^2(m)$ for any $x_3 \in {\mathbb{R}}$, and is a bounded and continuous function of $x_3$. In other words, we set $$X(m) \,=\, BC({\mathbb{R}}\,;L^2(m))~, \qquad
X_0(m) \,=\, BC({\mathbb{R}}\,;L^2_0(m))~, \label{def.X(m)}$$ where “$BC({\mathbb{R}}\,;Y)$” denotes the space of all bounded and continuous functions from ${\mathbb{R}}$ into $Y$. Both $X(m)$ and $X_0(m)$ are Banach spaces equipped with the norm $$\label{def.Xnorm}
\|\phi\|_{X(m)} \,=\, \sup_{x_3\in{\mathbb{R}}}\|\phi(\cdot,x_3)\|_{L^2(m)}~.$$
Our goal is to study the stability of the Burgers vortex $\Omega =
\alpha G$ with respect to perturbations $\omega \in X(m)^3$. In fact, we can assume without loss of generality that $\omega$ belongs to the subspace $${\mathbb{X}}(m) \,=\, X(m) \times X(m) \times X_0(m) \,\subset\,
X(m)^3~, \label{def.mathX(m)}$$ which is invariant under the evolution defined by . This is a consequence of the following result, whose proof is postponed to Section \[subsec.lem.reduce\]:
\[lem.reduction\] Fix $m \in (1,\infty]$. If $\tilde\omega \in X(m)^3$ satisfies $\nabla \cdot \tilde\omega = 0$ in the sense of distributions, then there exists $\tilde \alpha \in {\mathbb{R}}$ such that $$\label{def.tildealpha}
\int_{{\mathbb{R}}^2}\tilde\omega_3(x_h,x_3){\,{\rm d}}x_h \,=\, \tilde\alpha~,
\quad \hbox{for all } x_3\in {\mathbb{R}}~.$$
In view of Lemma \[lem.reduction\], if $\Omega = \alpha G
+ \tilde\omega$ for some $\tilde\omega \in X(m)^3$, we can write $\Omega = (\alpha +\tilde\alpha)G + \omega$, where $\tilde
\alpha$ is given by and $\omega = \tilde
\omega - \tilde\alpha G$. Then $\omega \in {\mathbb{X}}(m)$ by construction, and we are led back to the stability analysis of the Burgers vortex $(\alpha +\tilde\alpha)G$ with respect to perturbations in ${\mathbb{X}}(m)$.
In what follows we always consider the solutions $\omega(x,t)$ of as ${\mathbb{X}}(m)$-valued functions of time, and we often denote by $\omega(\cdot,t)$ or simply $\omega(t)$ the map $x \mapsto
\omega(x,t)$. A minor drawback of our functional setting is that we cannot expect the solutions of to be continuous in time in the strong topology of ${\mathbb{X}}(m)$. This is because the operator $L$ defined in contains the dilation operator $-x_3\partial_{x_3}$, see Section \[subsec.semigroup.L\] below. To restore continuity, it is thus necessary to equip ${\mathbb{X}}(m)$ with a weaker topology. Following [@GW2], we denote by $X_{loc}(m)$ the space $X(m)$ equipped with the topology defined by the family of seminorms $$\|\phi\|_{X_n(m)} \,=\, \sup_{|x_3|\le n}\|\phi(\cdot,x_3)\|_{L^2(m)}~,
\qquad n\in {\mathbb{N}}~.$$ In analogy with , we set ${{\mathbb{X}}}_{loc}(m) = X_{loc}(m)
\times X_{loc}(m) \times X_{0,loc}(m)$, where $X_{0,loc}(m)$ is of course the space $X_0(m)$ equipped with the topology of $X_{loc}(m)$.
We are now able to formulate our main result:
\[thm.main1\] Fix $m \in (2,\infty]$ and $\alpha\in {\mathbb{R}}$. Then there exist $\delta = \delta(\alpha,m) > 0$ and $C =
C(\alpha,m) \ge 1$ such that, for any $\omega_0 \in {\mathbb{X}}(m)$ with $\nabla \cdot \omega_0 = 0$ and $\|\omega_0\|_{{\mathbb{X}}(m)}\le \delta$, Eq. has a unique solution $\omega \in
L^\infty({\mathbb{R}}_+\,;{\mathbb{X}}(m))\cap C([0,\infty)\,;{\mathbb{X}}_{loc}(m))$ with initial data $\omega_0$. Moreover, $$\|\omega(t)\|_{{\mathbb{X}}(m)} \,\le\, C\|\omega_0\|_{{\mathbb{X}}(m)}\,e^{-t/2}~,
\qquad \hbox{for all } t \ge 0~.
\label{est.thm.main1}$$
Theorem \[thm.main1\] shows that the Burgers vortex $\alpha G$ is [*asymptotically stable*]{} with respect to perturbations in ${\mathbb{X}}(m)$, for any value of the circulation $\alpha \in {\mathbb{R}}$. If one prefers to consider perturbations in the larger space $X(m)^3$, then our result means that the family $\{\alpha G\}_{\alpha
\in {\mathbb{R}}}$ of all Burgers vortices is asymptotically stable [*with shift*]{}, because the perturbations may then modify the circulation of the underlying vortex. The key point in the proof is to show that the linearized operator $L - \alpha\Lambda$ has a [*uniform spectral gap*]{} for all $\alpha \in {\mathbb{R}}$. This implies a uniform decay rate in time for the perturbations, as in . However, it should be emphasized that the constants $C$ and $\delta$ in Theorem \[thm.main1\] do depend on $\alpha$, in such a way that $C(\alpha,m) \to \infty$ and $\delta(\alpha,m) \to 0$ as $|\alpha| \to \infty$. This is in full agreement with the amplification phenomena numerically observed in [@SR].
The proof of Theorem \[thm.main1\] gives a more detailed information on the solutions of than what is summarized in . First of all, we can prove stability for any $m > 1$, but the exponential factor $e^{-t/2}$ in should then be replaced by $e^{-\eta t}$, where $\eta < (m-1)/2$ if $m
\le 2$. Next, thanks to parabolic smoothing, we can obtain decay estimates not only for $\omega(t)$ but also for its spatial derivatives. Finally, due to the particular structure of the linear operator $L - \alpha\Lambda$, it turns out that the horizontal part $\omega_h = (\omega_1,\omega_2)^\top$ of the vorticity vector has a faster decay than the vertical component $\omega_3$ as $t \to
\infty$. Thus, a more complete (but less readable) version of our result is as follows:
\[thm.main2\] Fix $m \in (1,\infty]$, $\alpha\in {\mathbb{R}}$, and take $\mu \in (1,\frac32)$, $\eta \in (0,\frac12]$ such that $2\mu < m+1$ and $2\eta < m-1$. Then there exist $\delta =
\delta(\alpha,m) > 0$ and $C = C(\alpha,m,\mu,\eta) > 1$ such that, for all initial data $\omega_0 \in {\mathbb{X}}(m)$ with $\nabla \cdot \omega_0 = 0$ and $\|\omega_0\|_{{\mathbb{X}}(m)}\le \delta$, Eq. has a unique solution $\omega \in L^\infty({\mathbb{R}}_+\,;{\mathbb{X}}(m))\cap C([0,\infty)\,;{\mathbb{X}}_{loc}(m))$. Moreover, for all $t > 0$, $$\begin{aligned}
\|\partial_x^\beta\omega_h(t)\|_{X(m)^2} \,&\le\,
\frac{C\|\omega_0\|_{{\mathbb{X}}(m)}}{a(t)^{|\beta|/2}} \,e^{-\mu t}~,
\label{est.thm.main2.1}\\
\|\partial_x^\beta\omega_3(t)\|_{X(m)} \,&\le\,
\frac{C\|\omega_0\|_{{\mathbb{X}}(m)}}{a(t)^{|\beta|/2}}\,e^{-\eta t}~,
\label{est.thm.main2.2}\end{aligned}$$ where $a(t) = 1-e^{-t}$ and $\beta \in {\mathbb{N}}^3$ is any multi-index of length $|\beta| = \beta_1 + \beta_2 + \beta_3 \le 1$.
The decay rates , are optimal when $\beta = 0$, but it turns out that vertical derivatives such as $\partial_{x_3}\omega_h(t)$ or $\partial_{x_3}
\omega_3(t)$ have a faster decay as $t \to \infty$, see Sections \[sec.linear\] and \[sec.nonlinear\] for more details. In any case, we believe that the optimal rates are those provided by the linear stability analysis, as in Proposition \[prop.linear\] below.
The rest of this paper is devoted to the proof of Theorems \[thm.main1\] and \[thm.main2\]. Before giving the details, we explain here the main ideas in an informal way. As was already mentioned, the main difficulty is to obtain good estimates on the solutions of the linearized equation $$\partial_t\omega \,=\, (L - \alpha\Lambda)\omega~, \qquad
\nabla \cdot \omega \,=\,0~.
\label{eq.linear}$$ Once this is done, the nonlinear terms in can be controlled using rather standard arguments, which are recalled in Section \[sec.nonlinear\]. To study , we use the fact that the operator $L - \alpha\Lambda$ depends on the vertical variable in a simple and very specific way. Indeed, it is easy to verify that $[\partial_{x_3},L] = -\partial_{x_3}$ and $[\partial_{x_3},\Lambda] = 0$. This key observation, which already plays a crucial role in the previous works [@RD; @Cr; @SR], implies the following identity: $$\partial_{x_3}^k \,e^{t(L-\alpha\Lambda)}\omega_0 \,=\,
e^{-kt}\,e^{t(L-\alpha\Lambda)}\partial_{x_3}^k \omega_0~,
\label{eq.semigroup}$$ for all $k \in {\mathbb{N}}$ and all $t \ge 0$. If we take $k \in {\mathbb{N}}$ sufficiently large, depending on $|\alpha|$, we can use to to show that $\partial_{x_3}^k \omega(t)$ decays exponentially as $t \to \infty$ if $\omega(t)$ is a solution of . Then, by an interpolation argument, we deduce that all expressions involving at least one vertical derivative play a negligible role in the long-time asymptotics, see Section \[sec.linear\] for more details. This “smoothing effect” in the vertical direction is due to the stretching properties of the linear flow .
As a consequence of these remarks, we can restrict our attention to those solutions of which are independent of the vertical variable $x_3$. We call this particular situation the [*vectorial 2D problem*]{}, and we study it in Section \[sec.vectorial\]. Note that the perturbations we consider here are two-dimensional in the sense that $\partial_{x_3} u = \partial_{x_3} \omega = 0$, but that all three components of $u$ or $\omega$ are possibly nonzero. This is in contrast with the purely two-dimensional case considered in [@GW1; @GW2], where in addition $u_3 = \omega_1 =
\omega_2 = 0$. Nevertheless, it is possible to show that the solutions of with $\partial_{x_3} \omega = 0$ converge exponentially to zero as $t \to \infty$, and that the decay rate is uniform in $\alpha$. Extending the techniques developped in [@GW1; @GW2], this can be done using spectral estimates and a detailed study of the eigenvalue equation $(L-\alpha\Lambda)\omega =
\lambda\omega$. It is then a rather straightforward task to complete the proof of Theorem \[thm.main1\] using the arguments presented above.
[**Remark.**]{} The vortex tubes observed in numerical simulations are usually not axisymmetric: in general, they rather exhibit an elliptical core region. A simple model for such asymmetric vortices is obtained by replacing the straining flow $V^s$ in with the nonsymmetric strain $V^s_\lambda(x) =
\gamma M_\lambda x$, where $\lambda \in (0,1)$ is an asymmetry parameter and $$M_\lambda \,=\, \begin{pmatrix} -\frac{1+\lambda}{2} & 0 & 0 \\ 0 &
-\frac{1-\lambda}{2} & 0 \\ 0 & 0 & 1\end{pmatrix}~.
\label{def.strain'}$$ Asymmetric Burgers vortices are then stationary solutions to , where the operator $L$ in the right-hand side is defined by with $M$ replaced by $M_\lambda$. Unlike in the symmetric case $\lambda = 0$, no explicit formula is available and proving the existence of stationary solutions is already a nontrivial task, except perhaps in the perturbative regime where either the asymmetry parameter $\lambda$ or the circulation number $\alpha$ is very small. In view of these difficulties, asymmetric Burgers vortices were first studied using formal asymptotic expansions and numerical calculations, see e.g. [@RS; @MKO; @PP2]. The mathematical theory is more recent, and includes several existence results which cover now the whole range of parameters $\lambda \in (0,1)$ and $\alpha \in {\mathbb{R}}$ [@GW2; @GW3; @M1; @M2]. In addition, the stability with respect to two-dimensional perturbations is known to hold at least for small values of the asymmetry parameter [@GW3; @M1]. However, the only result so far on three-dimensional stability is restricted to the particular case where the circulation number $\alpha$ is sufficiently small, depending on $\lambda$ [@GW2].
Using Theorem \[thm.main1\] and a simple perturbation argument, it is easy to show that asymmetric Burgers vortices are stable with respect to three-dimensional pertubations in the space ${\mathbb{X}}(m)$, provided that the asymmetry parameter $\lambda$ is small enough depending on the circulation number $\alpha$. This follows from the fact the the linearized operator at the symmetric Burgers vortex has a uniform spectral gap for all $\alpha \in {\mathbb{R}}$, and that the asymmetric Burgers vortex is $O(\lambda)$ close to the corresponding symmetric vortex in the topology of ${\mathbb{X}}(m)$, uniformly for all $\alpha \in {\mathbb{R}}$ [@GW3]. Although this stability result is new and not covered by [@GW2], it is certainly not optimal, and we prefer to postpone the study of the three-dimensional stability of asymmetric Burgers vortices to a future investigation.
Preliminaries {#sec.pre}
=============
In this preliminary section we collect a few basic estimates which will be used throughout the proof of Theorems \[thm.main1\] and \[thm.main2\]. They concern the semigroup generated by the linear operator , and the Biot-Savart law relating the velocity field to the vorticity. Most of the results were already established in [@GW2 Appendix A], and are reproduced here for the reader’s convenience.
As in [@GW2], we introduce the following generalization of the function spaces and . Given $m \in [0,\infty]$ and $p \in [1,\infty)$, we define the weighted $L^p$ space $$L^p(m) \,=\, \Bigl\{f\in L^p({\mathbb{R}}^2)~\Big| \|f\|_{L^p(m)}^p =
\int_{{\mathbb{R}}^2}|f(x_h)|^p \rho_m(|x_h|^2)^{p/2} {\,{\rm d}}x_h < \infty
\Bigr\}~,$$ and the corresponding three-dimensional space $$X^p(m) \,=\, BC({\mathbb{R}}\,;L^p(m))~, \qquad
\|\phi\|_{X^p(m)} \,=\, \sup_{x_3\in{\mathbb{R}}}\|\phi(\cdot,x_3)\|_{L^p(m)}~.$$ If $m > 2-\frac{2}{p}$, we also denote by $L^p_0(m)$ the subspace of all $f \in L^p(m)$ such that $\int_{{\mathbb{R}}}f{\,{\rm d}}x_h = 0$. In analogy with , we set ${\mathbb{X}}^p(m) \,=\, X^p(m) \times X^p(m)
\times X_0^p(m)$ , where $X_0^p(m) = BC({\mathbb{R}}\,;L^p_0(m))$.
The semigroup generated by $L$ {#subsec.semigroup.L}
------------------------------
If we decompose the vorticity $\omega$ into its horizontal part $\omega_h = (\omega_1,\omega_2)^\top$ and its vertical component $\omega_3$, it is clear from and that the linear operator $L$ has the following expression: $$L\omega \,=\, \begin{pmatrix} L_h\omega_h \\ L_3\omega_3\end{pmatrix}
\,=\, \begin{pmatrix} ({\mathcal{L}}_h + {\mathcal{L}}_3 -\frac{3}{2})\omega_h \\
({\mathcal{L}}_h + {\mathcal{L}}_3)\omega_3\end{pmatrix}~, \label{eq.Lexp}$$ where ${\mathcal{L}}_h$ is the two-dimensional Fokker-Planck operator $$\label{def.LLh}
{\mathcal{L}}_h \,=\, \Delta_h + \frac{x_h}{2}\cdot\nabla_h + 1 \,=\,
\sum_{j=1}^2\partial_{x_j}^2 + \sum_{j=1}^2\frac{x_j}{2}\partial_{x_j}
+ 1~,$$ and ${\mathcal{L}}_3 = \partial_{x_3}^2 - x_3\partial_{x_3}$ is a convection-diffusion operator in the vertical variable.
As is shown in [@GW0 appendix A], the operator ${\mathcal{L}}_h$ is the generator of a strongly continuous semigroup in $L^2(m)$ given by the explicit formula $$(e^{t{\mathcal{L}}_h}f)(x_h) \,=\, \frac{e^t}{4\pi a(t)} \int_{{\mathbb{R}}^2}
e^{-\frac{|x_h-y_h|^2}{4a(t)}} f(y_he^{t/2}){\,{\rm d}}y_h~, \qquad
t > 0~, \label{eq.semigroup.calL_h}$$ where $a(t)=1-e^{-t}$. Similarly, the operator ${\mathcal{L}}_3$ generates a semigroup of contractions in $BC({\mathbb{R}})$ given by $$(e^{t{\mathcal{L}}_3}f)(x_3) \,=\, \frac{1}{\sqrt{2\pi a(2t)}}\int_{{\mathbb{R}}}
e^{-\frac{|x_3 e^{-t}-y_3|^2}{2a(2t)}}f(y_3){\,{\rm d}}y_3~, \qquad
t > 0~,
\label{eq.semigroup.calL_3}$$ see [@GW2 Appendix A]. Note that the semigroup $e^{t{\mathcal{L}}_3}$ is not strongly continuous in the space $BC({\mathbb{R}})$ equipped with the supremum norm. This is mainly due to the dilation factor $e^{-t}$ in . However, if we equip $BC({\mathbb{R}})$ with the (weaker) topology of uniform convergence on compact sets, then the map $t \mapsto e^{t{\mathcal{L}}_3}f$ is continuous for any $f \in BC({\mathbb{R}})$. This observation is the reason for introducing the space $X_{loc}(m)$ in Section \[sec.intro\].
Since the operators ${\mathcal{L}}_h$ and ${\mathcal{L}}_3$ act on different variables, it is easy to obtain the semigroup generated by $L_3 = {\mathcal{L}}_h +
{\mathcal{L}}_3$ by combining the formulas and . We find $$(e^{tL_3}\phi)(x) \,=\, \frac{1}{\sqrt{2\pi a(2t)}}\int_{{\mathbb{R}}}
e^{-\frac{|x_3e^{-t}-y_3|^2}{2a(2t)}}\Bigl(e^{t{\mathcal{L}}_h}\phi(\cdot,y_3)
\Bigr)(x_h){\,{\rm d}}y_3~, \qquad t > 0~. \label{eq.semigroup.L_3}$$ In [@GW2 Proposition A.6], it is shown that this expression defines a uniformly bounded semigroup in $X(m)$ for any $m > 1$, and that the map $t \mapsto e^{tL_3}$ is strongly continous in the topology of $X_{loc}(m)$. Moreover, the subspace $X_0(m)$ is left invariant by $e^{tL_3}$ for any $t \ge 0$. Using these results and the relation , we conclude that the three-dimensional operator $L$ generates a uniformly bounded semigroup in the space ${\mathbb{X}}(m)$, given by $$e^{tL}\omega \,=\, \Bigl(e^{-3t/2}e^{tL_3}\omega_1\,,\,
e^{-3t/2}e^{tL_3}\omega_2\,,\,e^{tL_3}\omega_3\Bigr)^\top~,
\qquad t \ge 0~. \label{eq.semigroup.L}$$ As is easily verified, if $\nabla \cdot \omega = 0$, then $\nabla\cdot e^{tL}\omega = 0$ for all $t \ge 0$.
The asymptotic stability of the Burgers vortices relies heavily on the decay properties of the semigroup $e^{tL}$ as $t \to \infty$. In the proof of Theorems \[thm.main1\] and \[thm.main2\], we also use the smoothing properties of the operator $e^{tL}$ for $t > 0$, and in particular the fact that $e^{tL}$ extends to a bounded operator from ${\mathbb{X}}^p(m)$ into ${\mathbb{X}}^2(m)$ for all $p \in [1,2]$. All the needed estimated are collected in the following statement.
\[prop.semigroup.L\] Let $m\in (1,\infty]$, $p\in [1,2]$, and take $\eta \in (0,\frac12]$ such that $2\eta <
m - 1$. For any $\beta = (\beta_1,\beta_2,\beta_3) \in {\mathbb{N}}^3$, there exists $C > 0$ such that the following estimates hold: $$\begin{aligned}
\|\partial_x^\beta e^{tL_h} \omega_h \|_{X(m)^2} \,&\le\, \frac{C
e^{-(\frac{3}{2}+\beta_3) t}}{a(t)^{\frac{1}{p}-\frac{1}{2}+\frac{|\beta|}{2}}}
\|\omega_h\|_{X^p(m)^2}~, \label{est.prop.semigroup.L.1}\\
\|\partial_x^\beta e^{tL_3} \omega_3\|_{X(m)} \,&\le\, \frac{C
e^{-(\eta+\beta_3)t}}{a(t)^{\frac{1}{p}-\frac{1}{2}+\frac{|\beta|}{2}}}
\|\omega_3\|_{X^p(m)}~, \label{est.prop.semigroup.L.2}\end{aligned}$$ for any $\omega \in {\mathbb{X}}^p(m)$ and all $t > 0$. Here $a(t) = 1-e^{-t}$ and $|\beta| = \beta_1 + \beta_2 + \beta_3$.
[**Proof.**]{} We first assume that $m \in (1,\infty)$. If $p \in [1,2]$ and $\beta_h = (\beta_1,\beta_2) \in {\mathbb{N}}^2$, it is proved in [@GW0 Appendix A] that $$\|\partial_{x_h}^{\beta_h} e^{t{\mathcal{L}}_h} f\|_{L^2(m)} \,\le\,
\frac{C}{a(t)^{\frac{1}{p}-\frac{1}{2}+\frac{|\beta_h|}{2}}}
\|f\|_{L^p(m)}~, \qquad t > 0~,\label{est.semigroup_h.1}$$ for all $f \in L^p(m)$. If in addition $f \in L^p_0(m)$, we have the stronger estimate $$\|\partial_{x_h}^{\beta_h} e^{t{\mathcal{L}}_h} f\|_{L^2(m)} \,\le\,
\frac{C e^{-\eta t}}{a(t)^{\frac{1}{p}-\frac{1}{2}+\frac{|\beta_h|}{2}}}
\|f\|_{L^p(m)}~, \qquad t > 0~,\label{est.semigroup_h.2}$$ where $\eta > 0$ is as in Proposition \[prop.semigroup.L\]. On the other hand, using , we find by direct calculation $$\|\partial_{x_3}^{\beta_3}e^{t{\mathcal{L}}_3}f\|_{L^\infty({\mathbb{R}})} \,\le\,
\frac{C e^{-\beta_3 t}}{a(t)^\frac{\beta_3}{2}}
\|f\|_{L^\infty({\mathbb{R}})}~, \qquad t > 0~. \label{est.semigroup_3}$$ Here, as in , the stabilizing factor $e^{-\beta_3t}$ comes from the dilation operator $-x_3\partial_{x_3}$ which enters the definition of ${\mathcal{L}}_3$. Now, if we start from the representation and use the estimates –, we easily obtain , by a direct calculation, see [@GW2 Proposition A.6].
To complete the proof of Proposition \[prop.semigroup.L\], it remains to show that , still hold when $m = \infty$. If $t \in
(0,1)$, estimate is easily obtained by a direct calculation, based on the representation . Using this remark and the semigroup property of $e^{t{\mathcal{L}}_h}$, we conclude that it is sufficient to establish , in the particular case where $p = 2$ and $\beta_h = 0$. This in turns follows easily from the spectral properties of the generator ${\mathcal{L}}_h$. Indeed, it is well-known that ${\mathcal{L}}_h$ is a self-adjoint operator in $L^2(\infty)$ with purely discrete spectrum $\sigma({\mathcal{L}}_h) =
\{-\frac{k}{2}\,|\,k=0,1,2,\dots\}$. Moreover, the subspace $L^2_0(\infty)$ is precisely the orthogonal complement of the eigenspace corresponding to the zero eigenvalue, see for example [@GW1 Lemma 4.7]. It follows that $e^{t{\mathcal{L}}_h}$ is a semigroup of contractions in $L^2(\infty)$, and that $\|e^{t{\mathcal{L}}_h}f\|_{L^2(\infty)} \le e^{-t/2}\|f\|_{L^2(\infty)}$ for all $t \ge 0$ if $f \in L^2_0(\infty)$. This proves and , with $\eta = 1/2$. [$\Box$]{}
Estimates for the velocity fields {#subsec.est.velocity}
---------------------------------
If the velocity $u$ and the vorticity $\omega$ are related by the Biot-Savart law , we have $|u| \le J(|\omega|)$, where $J$ is the Riesz potential defined by $$J(\phi)(x) \,=\, \frac{1}{4\pi}\int_{{\mathbb{R}}^3}\frac{1}{|x-y|^2}
\,\phi(y){\,{\rm d}}y~, \qquad x \in {\mathbb{R}}^3~. \label{def.Riesz}$$ Since $\omega$ will typically belong to the Banach space ${\mathbb{X}}(m)$, we need estimates on the Riesz potential $J(\phi)$ for $\phi \in X(m)$. We start with a preliminary result:
\[lem.Riesz\] Let $p_1 \in [1,2)$, $p_2 \in [1,2]$, and assume that $\phi\in X^{p_1}(0)\cap X^{p_2}(0)$. If $q_1, q_2 \in [1,\infty]$ satisfy $$\label{def.q1q2}
\frac{2p_1}{2-p_1} \,<\, q_1 \,\le\, \infty~, \qquad
p_2 \,<\, q_2 \,<\, \frac{2p_2}{2-p_2}~,$$ then $J(\phi) = J_1(\phi)+J_2(\phi)$ with $J_i(\phi)\in X^{q_i}(0)$ for $i = 1,2$, and we have the following estimates $$\begin{aligned}
\|J_1(\phi)\|_{X^{q_1}(0)} \,&\le\, C(p_1,q_1)\|\phi\|_{X^{p_1}(0)}
~,\label{est.J1}\\
\|J_2(\phi)\|_{X^{q_2}(0)} \,&\le\, C(p_2,q_2)\|\phi\|_{X^{p_2}(0)}
~.\label{est.J2}\end{aligned}$$
[**Proof.**]{} We proceed as in [@GW2 Proposition A.9]. We first observe that $$\begin{aligned}
J(\phi)(x_h,x_3) \,&=\, \int_{|x_3-y_3|\ge 1}F(x_h;x_3,y_3){\,{\rm d}}y_3 +
\int_{|x_3-y_3|< 1}F(x_h;x_3,y_3){\,{\rm d}}y_3 \\
\,&=\, J_1(\phi)(x_h,x_3) + J_2(\phi)(x_h,x_3)~,\end{aligned}$$ where $$F(x_h;x_3,y_3) \,=\, \int_{{\mathbb{R}}^2}\frac{\phi(y_h,y_3)}{|x_h-y_h|^2
+(x_3-y_3)^2}{\,{\rm d}}y_h~, \qquad x_h \in {\mathbb{R}}^2~, \quad x_3,y_3 \in {\mathbb{R}}~.$$ For any $a \in {\mathbb{R}}$, let $f_a(y_h) = (a^2+|y_h|^2)^{-1}$. Then $f_a \in L^r({\mathbb{R}}^2)$ for any $r > 1$ and any $a \neq 0$, and there exists $C_r > 0$ such that $$\|f_a\|_{L^r({\mathbb{R}}^2)} \,\le\, \frac{C_r}{|a|^{2-\frac2r}}~.$$ Moreover, we have $F(\cdot\,;x_3,y_3) = \phi(\cdot,y_3)\star f_{x_3-y_3}$ by construction. Thus, if we take $1 \le p,q,r \le \infty$ such that $1+\frac1q = \frac1p+\frac1r$, we obtain using Young’s inequality $$\|F(\cdot\,;x_3,y_3)\|_{L^q({\mathbb{R}}^2)} \,\le\,
\|\phi(\cdot,y_3)\|_{L^p({\mathbb{R}}^2)} \|f_{x_3-y_3}\|_{L^r({\mathbb{R}}^2)}
\,\le\, \frac{C_r \|\phi(\cdot,y_3)\||_{L^p({\mathbb{R}}^2)}}{|x_3-
y_3|^{2-\frac2r}}~.$$
To estimate $J_1(\phi)$, we choose $p = p_1$, $q = q_1$. In view of , the corresponding exponent $r = r_1$ satisfies $2 < r_1 \le \infty$, so that $2-\frac2{r_1}
\in (1,2]$. By Minkowski’s inequality, we thus find $$\|J_1(\phi)(\cdot,x_3)\|_{L^{q_1}({\mathbb{R}}^2)} \,\le\, \int_{|x_3-y_3|\ge 1}
\|F(\cdot\,;x_3,y_3)\|_{L^{q_1}({\mathbb{R}}^2)}{\,{\rm d}}y_3 \,\le\,
C(r_1) \sup_{y_3 \in {\mathbb{R}}}\|\phi(\cdot,y_3)\|_{L^{p_1}({\mathbb{R}}^2)}~.$$ Taking the supremum over $x_3 \in {\mathbb{R}}$, we obtain . Similarly, to bound $J_2(\phi)$, we take $p = p_2$, $q = q_2$. Then $1 < r_2 < 2$, so that $2 - \frac2{r_2} \in (0,1)$. We thus obtain $$\|J_2(\phi)(\cdot,x_3)\|_{L^{q_2}({\mathbb{R}}^2)} \,\le\, \int_{|x_3-y_3|< 1}
\|F(\cdot\,;x_3,y_3)\|_{L^{q_2}({\mathbb{R}}^2)}{\,{\rm d}}y_3 \,\le\,
C(r_2) \sup_{y_3 \in {\mathbb{R}}}\|\phi(\cdot,y_3)\|_{L^{p_2}({\mathbb{R}}^2)}~,$$ and follows. Finally, the uniform continuity of $J_i(\phi)(\cdot,x_3)$ with respect to $x_3$ can be verified exactly as in the proof of [@GW2 Proposition A.9]. [$\Box$]{}
As an immediate consequence, we obtain the following useful statements.
\[prop.Riesz\] Let $\phi \in X(m)$ for some $m \in
(1,\infty]$. Then $J(\phi)\in X^q(0)$ for all $q\in (2,\infty)$, and there exists a positive constant $C = C(m,q)$ such that $$\|J(\phi)\|_{X^q(0)} \,\le\, C\|\phi\|_{X(m)}~.
\label{est.prop.Riesz}$$
[**Proof.**]{} If $m > 1$, we recall that $X(m) \hookrightarrow
X^p(0)$ for all $p \in [1,2]$. Thus we can apply Lemma \[lem.Riesz\] with $p_1 = 1$, $p_2 = 2$, and $q_1 = q_2 = q \in (2,\infty)$, and the result follows. [$\Box$]{}
\[cor.prop.Riesz\] Let $\phi_1$, $\phi_2 \in X(m)$ for some $m \in (1,\infty]$. Then $\phi_1J(\phi_2)\in X^p(m)$ for all $p\in (1,2)$, and there exists a positive constant $C = C(m,p)$ such that $$\|\phi_1J(\phi_2)\|_{X^p(m)} \,\le\, C\|\phi_1\|_{X(m)}
\|\phi_2\|_{X(m)}~.
\label{est.cor.prop.Riesz}$$
[**Proof.**]{} We proceed as in [@GW2 Corollary A.10]. Let $p \in (1,2)$, and take $q \in (2,\infty)$ such that $\frac1q = \frac1p -\frac12$. For any $x_3 \in {\mathbb{R}}$, we have by Hölder’s inequality $$\begin{aligned}
\|\phi_1(\cdot,x_3)J(\phi_2)&(\cdot,x_3)\|_{L^p(m)} \,=\,
\Bigl(\int_{{\mathbb{R}}^2}\rho_m(|x_h|^2)^{p/2}|\phi_1(x_h,x_3)|^p |J(\phi_2)
(x_h,x_3)|^p{\,{\rm d}}x_h\Bigr)^{1/p} \\
\,&\le\, \Bigl(\int_{{\mathbb{R}}^2}\rho_m(|x_h|^2)|\phi_1(x_h,x_3)|^2{\,{\rm d}}x_h
\Bigr)^{1/2}\Bigl(\int_{{\mathbb{R}}^2}|J(\phi_2)(x_h,x_3)|^q{\,{\rm d}}x_h\Bigr)^{1/q} \\
\,&=\, \|\phi_1(\cdot,x_3)\|_{L^2(m)} \|J(\phi_2)(\cdot,x_3)\|_{L^q(0)}~.\end{aligned}$$ Taking the supremum over $x_3 \in {\mathbb{R}}$ and using Proposition \[prop.Riesz\], we obtain . Finally, it is clear that the map $x_3 \mapsto \phi_1(\cdot,x_3)
J(\phi_2)(\cdot,x_3)$ is continuous from ${\mathbb{R}}$ into $L^p(m)$. [$\Box$]{}
We conclude this section with an estimate on the linear operator which will be needed in Section \[sec.linear\].
\[lem.after.Riesz\] Let $p \in [1,2]$ and $2-\frac2p < m
\le \infty$. For any $\beta \in {\mathbb{N}}^3$, there exists $C > 0$ such that $$\|\partial_x^\beta \Lambda \omega\|_{{\mathbb{X}}^p (m)} ~\le~
C \!\!\sum_{|\tilde\beta|\le |\beta|+1} \|\partial_x^{\tilde \beta}
\omega\|_{{\mathbb{X}}^p(m)}. \label{est.lem.after.Riesz}$$
[**Proof.**]{} It is sufficient to prove for $\beta = 0$. The general case easily follows if we use the Leibniz rule to differentiate $\Lambda \omega$ (we omit the details).
Assume thus that $\omega$ belongs to ${\mathbb{X}}^p(m)$, together with its first order derivatives. Since the function $U^G$ defined in is smooth and bounded (together with all its derivatives), it is clear that $$\|(U^G,\nabla)\omega\|_{{\mathbb{X}}^p(m)} + \|(\omega,\nabla)U^G\|_{{\mathbb{X}}^p(m)}
\,\le\, C\sum_{|\tilde \beta|\le 1}\|\partial_x^{\tilde \beta}\omega
\|_{{\mathbb{X}}^p(m)}~.$$ We now estimate the term $(u,\nabla)G = (K_{3D}*\omega,\nabla) G$, using the fact that $|K_{3D}*\omega| \le J(|\omega|)$. Since $|\omega| \in X^1(0) \cap X^p(0)$ by assumption, we can apply Lemma \[lem.Riesz\] with $p_1 = 1$, $q_1 = \infty$, $p_2 = p$, and $q_2\in (p,\frac{2p}{2-p})$. By Hölder’s inequality, we easily find $$\begin{aligned}
\|J_1(|\omega|) |\nabla G|\|_{X^p(m)} \,&\le\, C\|J_1(|\omega|)
\|_{X^\infty(0)} \,\le\, C\||\omega|\|_{X^1(0)} \,\le\,
C\|\omega\|_{{\mathbb{X}}^p(m)}~,\\
\|J_2(|\omega|) |\nabla G|\|_{X^p(m)} \,&\le\, C\|J_2(|\omega|)
\|_{X^{q_2}(0)} \,\le\, C\||\omega|\|_{X^p(0)} \,\le\,
C\|\omega\|_{{\mathbb{X}}^p(m)}~.\end{aligned}$$ We conclude that $\|(u,\nabla)G\|_{{\mathbb{X}}^p(m)} \le \|(K_{3D}*\omega,
\nabla)G\|_{{\mathbb{X}}^p(m)} \le C\|\omega\|_{{\mathbb{X}}^p(m)}$. In a similar way, commuting the derivative and the convolution operator, we obtain the estimate $\|(G,\nabla)u\|_{{\mathbb{X}}^p(m)} \le \|(G,\nabla)(K_{3D}
*\omega)\|_{{\mathbb{X}}^p(m)} \le C\|\nabla \omega\|_{{\mathbb{X}}^p(m)}$. This completes the proof. [$\Box$]{}
The vectorial 2D problem {#sec.vectorial}
========================
In this section we study the linearized equation $\partial_t \omega
= (L-\alpha\Lambda)\omega$ in the particular case where the vorticity $\omega$ does not depend on the vertical variable. As was explained in the introduction, this preliminary step is an essential ingredient in the linear stability proof which will be presented in Section \[sec.linear\].
If $\partial_{x_3}\omega = 0$, then ${\mathcal{L}}_3\omega = 0$, and the expression of the linear operator $L$ becomes significantly simpler. On the other hand, we know from that $$\Lambda \omega \,=\, \Lambda_1\omega - \Lambda_2\omega +
\Lambda_3\omega-\Lambda_4\omega~,
\label{eq.Lambda_dec}$$ where $$\begin{array}{l}
\Lambda_1 \omega \,=\, (U^G,\nabla)\omega \,=\, (U^G_h,\nabla_h)
\omega~, \\[2mm]
\Lambda_2 \omega \,=\, (\omega,\nabla)U^G \,=\,
(\omega_h,\nabla_h)U^G~,
\end{array} \qquad
\begin{array}{l}
\Lambda_3 \omega \,=\, (u,\nabla)G \,=\, (u_h,\nabla_h)G~, \\[2mm]
\Lambda_4 \omega \,=\, (G,\nabla)u \,=\, g\partial_{x_3}u~.
\end{array}
\label{def.Lambda_i}$$ Here $u = K_{3D}*\omega$ is the velocity field obtained from $\omega$ via the three-dimensional Biot-Savart law . Since $\partial_{x_3}\omega = 0$, we have $\partial_{x_3}u = 0$, hence $\Lambda_4 \omega = 0$ in our case. Moreover, it is easy to verify that $u = (u_h,u_3)$, where $u_h = K_{2D}\star\omega_3$. Thus, we see that $$(L-\alpha\Lambda)\omega \,=\, {\mathscr{L}}_\alpha \omega \,:=\,
\begin{pmatrix} ({\mathcal{L}}_h-\frac32)\omega_h - \alpha (\Lambda_1 -
\tilde\Lambda_2)\omega_h\\
{\mathcal{L}}_h \omega_3 - \alpha (\Lambda_1 + \tilde\Lambda_3)\omega_3
\end{pmatrix} \,\equiv\,
\begin{pmatrix} {\mathscr{L}}_{\alpha,h}\,\omega_h \\ {\mathscr{L}}_{\alpha,3}\,\omega_3
\end{pmatrix}~,
\label{def.calLambda}$$ where $\tilde \Lambda_2 \omega_h = (\omega_h,\nabla_h)U_h^G$ and $\tilde \Lambda_3 \omega_3 = (K_{2D}\star\omega_3,\nabla_h)g$.
For any $\alpha \in {\mathbb{R}}$ and any $m \in (1,\infty]$, the operator ${\mathscr{L}}_\alpha$ defined by is the generator of a strongly continuous semigroup in the space $L^2(m)^3$. This property can be established by a standard perturbation argument, see Lemma \[prop.vectorial.2D.local\] below. Our main goal here is to obtain accurate decay estimates for the semigroup $e^{t{\mathscr{L}}_\alpha}$ as $t \to \infty$. As is clear from , the evolutions for $\omega_h$ and $\omega_3$ are completely decoupled, so that we can consider the semigroups $e^{t{\mathscr{L}}_{\alpha,h}}$ and $e^{t{\mathscr{L}}_{\alpha,3}}$ separately. The main contribution of this section is:
\[prop.vectorial.2D\] Fix $m \in (1,\infty]$, $\alpha \in {\mathbb{R}}$, $\mu \in (0,\frac32)$, and take $\eta \in (0,\frac12]$ such that $1+2\eta < m$. Then there exists $C > 0$ such that $$\begin{aligned}
\|e^{t{\mathscr{L}}_{\alpha,h}}\omega_h\|_{L^2(m)^2} \,&\le\, C\,e^{-\mu t}
\|\omega_h\|_{L^2(m)^2}~, \qquad t \ge 0~,
\label{est.prop.vectorial.2D.1} \\
\|e^{t{\mathscr{L}}_{\alpha,3}}\omega_3\|_{L^2(m)} \,&\le\, C\,e^{-\eta t}
\|\omega_3\|_{L^2(m)}~, ~\qquad t \ge 0~,
\label{est.prop.vectorial.2D.2}\end{aligned}$$ for all $\omega \in L^2(m)^2 \times L^2_0(m)$.
Estimate was obtained in [@GW1 Proposition 4.12] for $m < \infty$, and the proof given there extends to the limiting case $m = \infty$ without additional difficulty. Remark that the decay rate $e^{-\eta t}$ is obtained using the fact that $\omega_3 \in L^2_0(m)$: If we only assume that $\omega_3 \in L^2(m)$ for some $m > 1$, then holds with $\eta = 0$. Note, however, that $\omega$ is not assumed to be divergence-free in this section.
From now on, we focus on the semigroup $e^{t{\mathscr{L}}_{\alpha,h}}$, which has not been studied yet. To prove , we use the same arguments as in [@GW1 Section 4.2]. We first establish a short time estimate:
\[prop.vectorial.2D.local\] Fix $m \in (1,\infty]$, $\alpha \in {\mathbb{R}}$, and $T > 0$. There exists $C = C(T,m,|\alpha|) > 0$ such that $$\sup_{0 \le t \le T}\Bigl(\|e^{t{\mathscr{L}}_{\alpha,h}}\omega_h\|_{L^2(m)^2}
+a(t)^\frac{1}{2}\|\nabla_h e^{t{\mathscr{L}}_{\alpha,h}}\omega_h\|_{L^2(m)^4}
\Bigr) \,\le\, C\|\omega_h\|_{L^2(m)^2}~,
\label{est.lem.local}$$ for all $\omega_h \in L^2(m)^2$. Here $a(t) = 1-e^{-t}$.
[**Proof.**]{} Given $\omega_h^0 \in L^2(m)^2$, the idea is to solve the integral equation $$\omega_h(t) \,=\, e^{t({\mathcal{L}}_h-\frac{3}{2})}\omega_h^0 -\alpha
\int_0^t e^{(t-s)({\mathcal{L}}_h-\frac{3}{2})}(\Lambda_1-\tilde\Lambda_2)
\omega_h(s){\,{\rm d}}s~, \qquad t \in [0,T]~,
\label{eq.integ.2Dvec}$$ by a fixed point argument in the space $X_T = \{\omega_h \in C([0,T],
L^2(m)^2\,|\, \|\omega_h\|_{X_T} < \infty\}$ defined by the norm $$\|\omega_h\|_{X_T} \,=\, \sup_{0\le t\le T}\|\omega_h(t)\|_{L^2(m)^2}
+ \sup_{0\le t\le T}a(t)^\frac{1}{2}\|\nabla_h \omega_h(t)
\|_{L^2(m)^4}~.$$ From we know that $\|e^{t({\mathcal{L}}_h-\frac{3}{2})}
\omega_h^0\|_{X_T} \le C_1\|\omega_h^0\|_{L^2(m)^2}$, for some $C_1 > 0$ independent of $T$. To estimate the integral term in , we first observe that the velocity field $U^G$ defined by satisfies $$\label{eq.UGbounds}
\sup_{x_h \in {\mathbb{R}}^2}(1+|x_h|)|U^G(x_h)| + \sup_{x_h \in {\mathbb{R}}^2}(1+|x_h|)^2
|\nabla_h U^G(x_h)| \,<\, \infty~.$$ In view of the definitions , we thus have $$\begin{aligned}
\|(1+|x_h|) \Lambda_1 \omega_h\|_{L^2(m)^2} \,&\le\,
C\|\nabla_h \omega_h\|_{L^2(m)^4}~,
\label{est.Lambda_1.subsec.2D}\\
\|(1+|x_h|)^2 \tilde\Lambda_2 \omega_h\|_{L^2(m)^2}\,&\le\,
C\|\omega_h\|_{L^2(m)^2}~.
\label{est.Lambda_2.subsec.2D}\end{aligned}$$ Using these estimates together with , we can bound $$\begin{aligned}
\Bigl\|\int_0^t e^{(t-s)({\mathcal{L}}_h-\frac{3}{2})}&(\Lambda_1-
\tilde \Lambda_2)\omega_h(s) {\,{\rm d}}s\Bigr\|_{L^2(m)^2} \\
\,&\le\, C \int_0^t e^{-\frac32(t-s)}\Bigl(\|\omega_h(s)\|_{L^2(m)^2} +
\|\nabla_h \omega_h(s)\|_{L^2(m)^4}\Bigr){\,{\rm d}}s \\
\,&\le\, C \|\omega_h\|_{X_T} \int_0^t e^{-\frac32(t-s)}
a(s)^{-\frac12}{\,{\rm d}}s \,\le\, C a(T)^{\frac12}\|\omega_h\|_{X_T}~. \end{aligned}$$ In a similar way, $$\begin{aligned}
\Bigl\|\nabla_h\int_0^t &e^{(t-s)({\mathcal{L}}_h-\frac{3}{2})}(\Lambda_1-\tilde
\Lambda_2) \omega_h(s){\,{\rm d}}s\Bigr\|_{L^2(m)^4} \label{est.nabla_int} \\
\,&\le\, C \int_0^t \frac{e^{-\frac32(t-s)}}{a(t-s)^{\frac12}}
\Bigl(\|\omega_h(s)\|_{L^2(m)^2} + \|\nabla_h \omega_h(s)\|_{L^2(m)^4}
\Bigr){\,{\rm d}}s \,\le\, C \|\omega_h\|_{X_T}~.
\nonumber \end{aligned}$$ Summarizing, we have shown that $\|\omega_h\|_{X_T} \le C_1
\|\omega_h^0\|_{L^2(m)^2} + C_2 |\alpha| a(T)^{1/2}\|\omega_h\|_{X_T}$, for some positive constants $C_1, C_2$. If we now take $T > 0$ small enough so that $C_2 |\alpha| a(T)^{1/2} \le 1/2$, we see that the right-hand side of is a strict contraction in $X_T$. We deduce that has a unique solution, which satisfies $\|\omega_h\|_{X_T} \le 2C_1\|\omega_h^0\|_{L^2(m)^2}$. Since $\omega_h(t) = e^{t{\mathscr{L}}_{\alpha,h}}\omega_h^0$ by construction, this proves for $T$ sufficiently small, and the general case follows due to the semigroup property. This concludes the proof. [$\Box$]{}
We next consider the essential spectrum of the semigroup $e^{t{\mathscr{L}}_{\alpha,h}}$, and begin with a few definitions. If $A$ is a bounded linear operator on a (complex) Banach space $X$, we define the essential spectrum $\sigma_{ess}(A\,;X)$ as the set of all $z \in {\mathbb{C}}$ such that $A-z$ is not a Fredholm operator with zero index, see [@Ka] or [@EN]. The essential spectral radius of $A$ in $X$ is given by $$r_{ess}(A\,;X) \,=\, \sup\Bigl\{|z|\,;\, z \in \sigma_{ess}(A\,;X)
\Bigr\} \,<\,\infty~.$$ If $|z| > r_{ess}(A\,;X)$, then either $z$ is in the resolvent set of $A$, or $z$ is an eigenvalue of $A$ with finite multiplicity, see [@EN Corollary IV.2.11]. In the latter case, we say that $z$ belongs to the discrete spectrum of $A$.
In what follows, we consider the linear operator ${\mathscr{L}}_{\alpha,h}$ as acting on the complexified space $L^2(m)^2$, i.e. the space of all $\omega_h : {\mathbb{R}}^2 \to {\mathbb{C}}^2$ such that $\|\omega_h\|_{L^2(m)^2}
< \infty$. Our first result shows that the essential spectral radius of the operator $e^{t{\mathscr{L}}_{\alpha,h}}$ in $L^2(m)^2$ does not depend on $\alpha$.
\[prop.essential.spectrum\] Let $m\in (1,\infty]$ and $\alpha \in {\mathbb{R}}$. Then for each $t > 0$ we have $$r_{ess}\Bigl(e^{t{\mathscr{L}}_{\alpha,h}}\,;L^2(m)^2\Bigr) \,=\,
r_{ess}\Bigl(e^{t{\mathscr{L}}_{0,h}}\,;L^2(m)^2\Bigr) \,=\,
e^{-(\frac{m}{2}+1)t}~.
\label{est.prop.essential.spectrum.1}$$
[**Proof.**]{} Since ${\mathscr{L}}_{0,h} = {\mathcal{L}}_h-\frac32$, the last equality in follows from [@GW0 Theorem A.1] if $m < \infty$. If $m = \infty$, then $e^{t{\mathcal{L}}_h}$ is a compact operator for any $t > 0$, hence $r_{ess}(e^{t{\mathscr{L}}_{0,h}}\,;L^2(\infty)^2) = 0$. To prove the first equality in , we fix $t > 0$. Our goal is to show that the linear operator $\Delta_\alpha(t) =
e^{t{\mathscr{L}}_\alpha,h} - e^{t({\mathcal{L}}_h-\frac32)}$ is compact in $L^2(m)^2$. By Weyl’s theorem, this will imply that both semigroups have the same essential spectrum, hence the same essential spectral radius. In view of we have, for all $\omega_h \in L^2(m)^2$, $$\Delta_\alpha(t)\omega_h \,=\, -\alpha \int_0^t e^{(t-s)({\mathcal{L}}_h
-\frac{3}{2})}(\Lambda_1-\tilde\Lambda_2)e^{s{\mathscr{L}}_{\alpha,h}}
\omega_h{\,{\rm d}}s~.
\label{def.Delta}$$ Let $w(x_h) = 1+|x_h|$. If $m < \infty$, it follows from and definition that $$\|w\,e^{t{\mathcal{L}}_h}\omega_h\|_{L^2(m)^2} \,\le\,
C \|e^{t{\mathcal{L}}_h}\omega_h\|_{L^2(m+1)^2} \,\le\,
C\|w\,\omega_h\|_{L^2(m)^2}~,
\label{eq.weightLL_h.1}$$ for all $\omega_h \in L^2(m)^2$ and all $t \ge 0$. If $m = \infty$, we know from [@GW3 Proposition 2.1] that $w (-{\mathcal{L}}_h+1)^{-1/2}$ is a bounded operator in $L^2(\infty)^2$, and since ${\mathcal{L}}_h$ is the generator of an analytic semigroup we easily obtain $$\|w\,e^{t{\mathcal{L}}_h}\omega_h\|_{L^2(m)^2} \,\le\,
C \|(-{\mathcal{L}}_h+1)^{1/2}e^{t{\mathcal{L}}_h}\omega_h\|_{L^2(m)^2} \,\le\,
\frac{C}{a(t)^{1/2}}\,\|\omega_h\|_{L^2(m)^2}~,
\label{eq.weightLL_h.2}$$ for all $t > 0$. Now, starting from and using either or together with , , and Lemma \[prop.vectorial.2D.local\], we find $$\begin{aligned}
\|w\,\Delta_\alpha(t)\omega_h\|_{L^2(m)^2} \,&\le\, C|\alpha|\int_0^t
\frac{e^{-\frac32(t-s)}}{a(t{-}s)^{1/2}}\Bigl(\|e^{s{\mathscr{L}}_{\alpha,h}}
\omega_h\|_{L^2(m)^2} + \|\nabla_h e^{s{\mathscr{L}}_{\alpha,h}}\omega_h
\|_{L^2(m)^4}\Bigr){\,{\rm d}}s \\
\,&\le\, C |\alpha| \|\omega_h\|_{L^2(m)^2} \int_0^t \frac{
e^{-\frac32(t-s)}}{a(t{-}s)^{1/2}a(s)^{1/2}}{\,{\rm d}}s \,\le\,
C |\alpha| \|\omega_h\|_{L^2(m)^2}~. \end{aligned}$$ Moreover, proceeding as in , we find $\|\nabla_h \Delta_\alpha(t) \omega_h\|_{L^2(m)^4} \le C |\alpha|
\|\omega_h \|_{L^2(m)^2}$. Thus we have shown that $w \Delta_\alpha(t)$ and $\nabla_h \Delta_\alpha(t)$ are bounded operators in $L^2(m)$. By Rellich’s criterion, we conclude that $\Delta_\alpha(t)$ is a compact operator in $L^2(m)^2$, for any $t > 0$. This completes the proof. [$\Box$]{}
In view of Proposition \[prop.essential.spectrum\], the spectrum of the semigroup $e^{t{\mathscr{L}}_{\alpha,h}}$ outside the disk of radius $e^{-(\frac{m}{2}+1)t}$ in the complex plane is purely discrete. By the spectral mapping theorem [@EN], to control that part of the spectrum it is sufficient to locate the eigenvalues of the generator ${\mathscr{L}}_{\alpha,h}$. Thus we look for nontrivial solutions of the eigenvalue problem $${\mathscr{L}}_{\alpha,h}\omega_h \,=\, \lambda \omega_h~,
\label{eq.eigenvalue}$$ where $\omega_h \in L^2(m)^2$ and $\lambda \in {\mathbb{C}}$ satisfies ${\mathop{\mathrm{Re}}}\lambda > -\frac{m}{2} - 1$. The following auxiliary result shows that the eigenfunctions $\omega_h$ always have a Gaussian decay at infinity.
\[prop.discrete.spectrum\] Let $m \in (1,\infty)$ and $\alpha\in{\mathbb{R}}$. If $\omega_h \in L^2(m)^2$ is a solution of with ${\mathop{\mathrm{Re}}}\lambda > -\frac{m}{2}
-1$, then $\omega_h\in L^2(\infty)^2$.
The proof of Proposition \[prop.discrete.spectrum\] is postponed to Section \[subsec.discrete\] below. Note that a similar result for the nonlocal operator ${\mathscr{L}}_{\alpha,3}$ has been obtained in [@GW1 Lemma 4.5], and plays a key role in the derivation of estimate . Thanks to Proposition \[prop.discrete.spectrum\], we only need to control the eigenvalues of ${\mathscr{L}}_{\alpha,h}$ in the Gaussian space $L^2(\infty)^2$. This is the last important step in the proof of Proposition \[prop.vectorial.2D\].
\[prop.discrete.spectrum.2\] If $\lambda$ is an eigenvalue of ${\mathscr{L}}_{\alpha,h}$ in $L^2(\infty)^2$, then ${\mathop{\mathrm{Re}}}\lambda
\le -\frac32$.
[**Proof.**]{} Assume that $\omega_h \in L^2(\infty)^2$ is a nontrivial solution of the eigenvalue problem , for some $\alpha \in {\mathbb{R}}$ and some $\lambda \in {\mathbb{C}}$. Using , we thus have $$\label{eq.eigenvalue2}
\lambda \omega_h \,=\, {\mathcal{L}}_h \omega_h -\frac32 \omega_h
-\alpha (U^G_h,\nabla_h)\omega_h + \alpha (\omega_h,\nabla_h)U^G_h~,$$ where the velocity field $U^G$ is defined in . Since ${\mathscr{L}}_{\alpha,h}$ is a relatively compact perturbation of ${\mathscr{L}}_{0,h} =
{\mathcal{L}}_h -\frac32$, both operators have the same domain, and it follows that $\omega_h$ belongs to the domain of ${\mathcal{L}}_h$. In particular, we have $\nabla_h \omega_h \in L^2(\infty)^4$ and $|x_h| \omega_h \in
L^2(\infty)^2$, see e.g. [@GW3 Section 2].
In the rest of the proof, we denote by $\langle \cdot,\cdot\rangle$ the inner product in the complexified space $L^2(\infty)^2$, namely $$\langle \omega_h^1,\omega_h^2\rangle \,=\, \int_{{\mathbb{R}}^2}
p(x_h)\omega_h^1(x_h)\cdot\overline{\omega_h^2(x_h)}{\,{\rm d}}x_h~,$$ where $p(x_h) = \rho_\infty(|x_h|^2) = e^{|x_h|^2/4}$. We also denote $\|\omega_h\|^2 = \langle \omega_h,\omega_h\rangle$. We recall that ${\mathcal{L}}_h$ is a selfadjoint operator in $L^2(\infty)^2$ which satisfies $-{\mathcal{L}}_h \ge 0$ on $L^2(\infty)^2$ and $-{\mathcal{L}}_h \ge 1/2$ on $L^2_0(\infty)^2$. For later use, we observe that the (unbounded) operator $\omega_h \mapsto (U^G_h,\nabla_h)\omega_h$ is skew-symmetric in $L^2(\infty)^2$, because the vector field $p(x_h)U^G(x_h)$ is divergence-free.
We now take the inner product of with $\omega_h$, and evaluate the real part of the result. Using the skew-symmetry of the operator $(U^G_h,\nabla_h)$, we easily obtain $$\begin{aligned}
\label{eq.spec1}
&{\mathop{\mathrm{Re}}}\lambda\,\|\omega_h\|^2 \,=\, \langle {\mathcal{L}}_h \omega_h,\omega_h\rangle
-\frac32\|\omega_h\|^2 + \alpha {\mathop{\mathrm{Re}}}\langle (\omega_h,\nabla_h)U^G_h,
\omega_h\rangle \\ \nonumber
\,&=\, \langle {\mathcal{L}}_h \omega_h,\omega_h\rangle -\frac32\|\omega_h\|^2
+ 2\alpha {\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^2}p(x_h)(x_h\cdot\omega_h)(x_h^\bot\cdot
\overline{\omega_h})(u^g)'(|x_h|^2){\,{\rm d}}x_h~,\end{aligned}$$ where $u^g(r)$ is defined in . On the other hand, it follows from that the scalar function $x_h\cdot\omega_h \in L^2(\infty)$ satisfies $$\lambda\,x_h\cdot\omega_h \,=\, {\mathcal{L}}_h(x_h\cdot\omega_h)
-2 x_h\cdot\omega_h -\alpha (U^G_h,\nabla_h)(x_h\cdot\omega_h)
-2\nabla_h \cdot \omega_h~.$$ Thus, proceeding as above and using the same notation $\langle \cdot,
\cdot\rangle$ for the inner product in $L^2(\infty)$, we find $$\label{eq.spec2}
{\mathop{\mathrm{Re}}}\lambda\,\|x_h\cdot\omega_h\|^2 \,=\,
\langle{\mathcal{L}}_h (x_h\cdot\omega_h),x_h\cdot\omega_h \rangle
-2 \|x_h\cdot\omega_h\|^2 -2{\mathop{\mathrm{Re}}}\langle \nabla_h\cdot\omega_h,
x_h\cdot\omega_h\rangle~.$$ Finally, the two-dimensional divergence $\nabla_h \cdot \omega_h
\in L^2_0(\infty)$ satisfies $$\label{eq.2Ddiv}
\lambda\,\nabla_h\cdot\omega_h \,=\, {\mathcal{L}}_h(\nabla_h\cdot\omega_h)
- \nabla_h\cdot\omega_h -\alpha (U^G_h,\nabla_h)(\nabla_h\cdot
\omega_h)~,$$ hence $$\label{eq.spec3}
{\mathop{\mathrm{Re}}}\lambda\,\|\nabla_h\cdot\omega_h\|^2 \,=\,
\langle{\mathcal{L}}_h (\nabla_h\cdot\omega_h),\nabla_h\cdot\omega_h \rangle
- \|\nabla_h\cdot\omega_h\|^2~.$$
Since $\nabla_h\cdot\omega_h \in L^2_0(\infty)$, it follows from that ${\mathop{\mathrm{Re}}}\lambda\,\|\nabla_h\cdot\omega_h\|^2 \le
-\frac32 \|\nabla_h\cdot\omega_h\|^2$. Thus we must have ${\mathop{\mathrm{Re}}}\lambda
\le -\frac32$, unless $\nabla_h\cdot\omega_h \equiv 0$. In the latter case, we deduce from that ${\mathop{\mathrm{Re}}}\lambda\,
\|x_h\cdot\omega_h\|^2 \le -2\|x_h\cdot\omega_h\|^2$, hence ${\mathop{\mathrm{Re}}}\lambda \le -2$ unless $x_h\cdot\omega_h \equiv 0$. But if this last condition is met, it follows from that ${\mathop{\mathrm{Re}}}\lambda\,\|\omega_h\|^2 \le -\frac32 \|\omega_h\|^2$, hence ${\mathop{\mathrm{Re}}}\lambda \le -\frac32$ because $\omega_h$ is not identically zero. Summarizing, we conclude that ${\mathop{\mathrm{Re}}}\lambda \le
-\frac32$ in all cases. [$\Box$]{}
[**Remark.**]{} Actually the conclusions of Proposition \[prop.discrete.spectrum.2\] can be slightly strengthened. First, in the invariant subspace where $\nabla_h \cdot
\omega_h = 0$, one can show that all eigenvalues of ${\mathscr{L}}_{\alpha,h}$ satisfy ${\mathop{\mathrm{Re}}}\lambda \le -2$. This follows from the proof above if we use in addition the fact that $\omega_h \in L^2_0(\infty)^2$, due to the divergence-free condition. The result is clearly sharp, because if $g(x_h)$ is defined by it is easy to verify that the function $\omega_h = x_h^\bot g(x_h)$ satisfies ${\mathscr{L}}_{\alpha,h}\omega_h =
-2\omega_h$ for any $\alpha \in {\mathbb{R}}$. On the other hand, if $\omega_h$ is a solution of such that $\nabla_h \cdot
\omega_h \neq 0$, we have ${\mathop{\mathrm{Re}}}\lambda < -\frac32$ if $\alpha \neq
0$. This follows from , because we know from [@GW0 Appendix A] that $$\langle{\mathcal{L}}_h (\nabla_h\cdot\omega_h),\nabla_h\cdot\omega_h
\rangle \,<\, -\frac12\|\nabla_h\cdot\omega_h\|^2~,$$ unless $\nabla_h \cdot\omega_h = (a_1x_1 + a_2x_2)g(x_h)$ for some $a_1, a_2 \in {\mathbb{C}}$. But this ansatz is not compatible with if $\alpha \neq 0$. In fact, using the techniques developped in [@M3] or [@GGN], it is possible to show that, given any $M > 0$, the eigenvalue equation restricted to the orthogonal complement of the space of all radially symmetric functions in $L^2(\infty)$ has no nontrivial solution such that ${\mathop{\mathrm{Re}}}\lambda \ge -M$, if $|\alpha|$ is sufficiently large depending on $M$.
It is now easy to conclude the proof of Proposition \[prop.vectorial.2D\]. As was already mentioned, we only need to prove that estimate holds for any $\mu < 3/2$. If $\rho_\alpha(m) > 0$ denotes the spectral radius of the operator $e^{{\mathscr{L}}_{\alpha,h}}$ in $L^2(m)^2$, this is equivalent to showing that $\log \rho_\alpha(m) \le -3/2$, see [@EN Proposition IV.2.2]. But that inequality follows immediately from Propositions \[prop.essential.spectrum\], \[prop.discrete.spectrum\], and \[prop.discrete.spectrum.2\], since $m > 1$. The proof of Proposition \[prop.vectorial.2D\] is now complete. [$\Box$]{}
Linear stability {#sec.linear}
================
Equipped with the results of the previous section, we now study the linearized equation in its full generality. Using Proposition \[prop.semigroup.L\] and a perturbation argument, it is not difficult to verify that the linear operator $L -
\alpha\Lambda$ generates a locally bounded semigroup in the space ${\mathbb{X}}(m)$ for any $\alpha \in {\mathbb{R}}$ and any $m \in (1,\infty]$, see Proposition \[prop.linear.local\] below. The goal of this section is to show that the semigroup $e^{t(L-\alpha\Lambda)}$ extends to a bounded operator from ${\mathbb{X}}^p(m)$ to ${\mathbb{X}}(m)$ for any $t > 0$ and any $p \in [1,2]$, and satisfies the following uniform estimates:
\[prop.linear\] Fix $m \in (1,\infty]$, $p\in [1,2]$, $\alpha\in {\mathbb{R}}$, and take $\mu \in (1,\frac32)$, $\eta \in (0,\frac12]$ such that $2\mu < m+1$ and $2\eta < m-1$. For any $\beta = (\beta_1,
\beta_2,\beta_3) \in {\mathbb{N}}^3$, there exists $C > 0$ such that $$\begin{aligned}
\|\partial_x^\beta(e^{t(L-\alpha\Lambda)}\omega_0)_h\|_{X(m)^2}
\,&\le\, \frac{C\,e^{-(\mu+\beta_3)t}}{a(t)^{\frac{1}{p}-\frac{1}{2}
+\frac{|\beta|}{2}}}\|\omega_0\|_{{\mathbb{X}}^p(m)}~,
\label{est.prop.linear.1}\\
\|\partial_x^\beta(e^{t(L-\alpha\Lambda)}\omega_0)_3\|_{X(m)}
\,&\le\, \frac{C\, e^{-(\eta+\beta_3)t}}{a(t)^{\frac{1}{p}-\frac{1}{2}
+\frac{|\beta|}{2}}}\|\omega_0\|_{{\mathbb{X}}^p(m)}~,
\label{est.prop.linear.2}\end{aligned}$$ for any $\omega_0\in {\mathbb{X}}^p(m)$ and all $t > 0$. Moreover, $\nabla\cdot
\omega_0 = 0$, then $\nabla\cdot e^{t(L-\alpha\Lambda)}\omega_0 = 0$ for all $t>0$.
The proof of this important result is divided into several steps.
Global existence and short time estimates {#subsec.est.local}
-----------------------------------------
We first prove that the linearized equation has a unique global solution in ${\mathbb{X}}(m)$.
\[prop.linear.local\] Fix $m\in (1,\infty]$, $p\in [1,2]$, and $\alpha\in {\mathbb{R}}$. Then, for any $\omega_0\in {\mathbb{X}}^p (m)$, Eq. has a unique solution $\omega\in L^\infty_{loc}
({\mathbb{R}}_+;{\mathbb{X}}(m))\cap C([0,\infty);{\mathbb{X}}_{loc}^p(m))$ with initial data $\omega_0$. Moreover, for any $\beta \in {\mathbb{N}}^3$, there exist positive constants $C_1$, $C_2$ (independent of $\alpha$) such that $$\|\partial_x^\beta\omega(t)\|_{{\mathbb{X}}(m)}\le \frac{C_1}{a(t)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}}\|\omega_0\|_{{\mathbb{X}}^p(m)}~, \qquad
\hbox{for} \quad 0 < t\le \frac{C_2}{|\alpha|^2+1}~,
\label{est.prop.linear.local}$$ where $a(t) = 1-e^{-t}$. Finally, if $\nabla\cdot \omega_0=0$, then $\nabla\cdot\omega(t)=0$ for all $t>0$.
[**Proof.**]{} We proceed as in the proof of Lemma \[prop.vectorial.2D.local\]. Let $e^{tL}$ be the semigroup generated by $L$, which is given by the explicit expression . The integral equation corresponding to is $$\omega (t) \,=\, e^{tL}\omega_0 - \alpha \int_0^t e^{(t-s)L}\Lambda
\omega(s) {\,{\rm d}}s \,=:\, (F\omega)(t)~, \quad t > 0~.
\label{eq.integral.prop.linear.local}$$ Given $k \in {\mathbb{N}}\setminus\{0\}$ and a sufficiently small $T \in (0,1]$, we shall solve in the Banach space $${\mathbb{U}}_{k,T} \,=\, \Bigl\{\omega \in L^\infty_{loc}((0,T);{\mathbb{X}}(m))\cap
C([0,T];{\mathbb{X}}^p_{loc}(m))~\Big|~ \|\omega\|_{k,T} < \infty\Bigr\}~,$$ equipped with the norm $$\|\omega\|_{k,T} \,=\, \sum_{|\beta|\le k} \Bigl(
\sup_{0<t<T}a(t)^{\frac{1}{p}-\frac{1}{2}+\frac{|\beta|}{2}}
\|\partial_x^\beta \omega(t)\|_{{\mathbb{X}}(m)} + \sup_{0<t<T}
a(t)^{\frac{|\beta|}{2}}\|\partial_x^\beta \omega(t)\|_{{\mathbb{X}}^p(m)}
\Bigr)~,$$ where $a(t) = 1 - e^{-t}$. If $\omega_0 \in {\mathbb{X}}^p(m)$, we know from Proposition \[prop.semigroup.L\] that the map $t \mapsto
e^{tL}\omega_0$ belongs to ${\mathbb{U}}_{k,T}$ for any $T > 0$, and that $\|e^{tL}\omega_0\|_{k,T} \le C_1\|\omega_0\|_{{\mathbb{X}}^p(m)}$ for some $C_1 > 0$ depending only on $k$, $m$, $p$.
Given $\omega \in {\mathbb{U}}_{k,T}$, we now estimate the integral term in . Using Proposition \[prop.semigroup.L\] and Lemma \[lem.after.Riesz\], we find $$\begin{aligned}
\|\partial_x^\beta e^{(t-s)L}\Lambda \omega(s)\|_{{\mathbb{X}}(m)} \,&\le\,
\frac{C\|\Lambda\omega(s)\|_{{\mathbb{X}}^p(m)}}{a(t-s)^{\frac{1}{p}-\frac{1}{2}
+\frac{|\beta|}{2}}} \,\le\, \frac{C\sum_{|\tilde \beta|\le 1}
\|\partial_x^{\tilde\beta}\omega(s)\|_{{\mathbb{X}}^p (m)}}{a(t-s)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}}\nonumber\\
\,&\le\, \frac{C\|\omega\|_{k,T}}{a(t-s)^{\frac{1}{p}-\frac{1}{2}
+\frac{|\beta|}{2}}\,a(s)^{\frac{1}{2}}}~, \qquad 0 < s < t~.
\label{est.proof.prop.linear.local.1}\end{aligned}$$ Similarly we have $\|\partial_x^\beta e^{(t-s)L}\Lambda \omega(s)
\|_{{\mathbb{X}}^p (m)} \le C a(t-s)^{-\frac{|\beta|}{2}}\,a(s)^{-\frac12}
\|\omega\|_{k,T}$ for $0 < s < t$. In the particular case where $\beta = 0$, it follows that $$\begin{aligned}
&\Bigl\|\int_0^t e^{(t-s)L}\Lambda \omega(s) {\,{\rm d}}s\Bigr\|_{{\mathbb{X}}(m)}
\,\le\, C a(t)^{1-\frac{1}{p}}\|\omega\|_{k,T}~,\label{eqint1}\\
&\Bigl\|\int_0^t e^{(t-s)L}\Lambda \omega(s){\,{\rm d}}s\Bigr\|_{{\mathbb{X}}^p(m)}
\,\le\, C a(t)^{\frac{1}{2}}\|\omega\|_{k,T}~, \qquad
0 < t \le T~.\label{eqint2} \end{aligned}$$
Assume now that $1 \le |\beta| \le k$. If $\beta' \le \beta$ and $|\beta'| = |\beta| - 1$, we have from Lemma \[lem.after.Riesz\] $$\|\partial_x^{\beta'} \Lambda \omega(s)\|_{{\mathbb{X}}(m)} \,\le\,
C \sum_{|\tilde \beta| = |\beta|}\|\partial_x^{\tilde \beta}
\omega(s)\|_{{\mathbb{X}}(m)} \,\le\, \frac{C \|\omega\|_{k,T}}{
a(s)^{\frac{1}{p}-\frac{1}{2} +\frac{|\beta|}{2}}}~,
\qquad 0 < s \le T~.$$ Thus, writing $\partial_x^\beta e^{(t-s)L} = \partial_x^{\beta-\beta'}
\partial_x^{\beta'} e^{(t-s)L} = \partial_x^{\beta-\beta'}
\,e^{(\frac{\beta_1'+\beta_2'}{2}-\beta_3')t}e^{(t-s)L}\partial_x^{\beta'}$, and using Proposition \[prop.semigroup.L\] again, we obtain $$\|\partial_x^\beta e^{(t-s)L}\Lambda \omega(s)\|_{{\mathbb{X}}(m)} \,\le\,
C \|\partial_x^{\beta-\beta'}e^{(t-s)L}\partial_x^{\beta'}\Lambda
\omega(s)\|_{{\mathbb{X}}(m)} \,\le\, \frac{C\|\omega\|_{k,T}}{a(t-s)^{\frac{1}{2}}
\,a(s)^{\frac{1}{p}-\frac{1}{2} +\frac{|\beta|}{2}}}~,
\label{est.proof.prop.linear.local.3}$$ for $0 < s < t$. Similarly, we have $\|\partial_x^\beta e^{(t-s)L}
\Lambda \omega(s)\|_{{\mathbb{X}}^p(m)} \le a(t-s)^{-\frac{1}{2}}
\,a(s)^{-\frac{|\beta|}{2}}\|\omega\|_{k,T}$. Combining and , we obtain the following estimate $$\begin{aligned}
\Bigl\|\partial_x^\beta &\int_0^t e^{(t-s)L}\Lambda \omega(s)
{\,{\rm d}}s\Bigr\|_{{\mathbb{X}}(m)}\nonumber \\
\,&\le\, C\Bigl(\int_0^\frac{t}{2} a(t-s)^{-\frac{1}{p}+\frac{1}{2}
-\frac{|\beta|}{2}}\,a(s)^{-\frac{1}{2}}{\,{\rm d}}s +
\int_\frac{t}{2}^t a(t-s)^{-\frac{1}{2}} \,a(s)^{-\frac{1}{p}
+\frac{1}{2}-\frac{|\beta|}{2}}{\,{\rm d}}s\Bigr)\|\omega\|_{k,T}
\label{eqint3}\\
\,&\le\, C a(t)^{1-\frac{1}{p}-\frac{|\beta|}{2}} \|\omega\|_{k,T}~,
\qquad 0 < t \le T~, \nonumber\end{aligned}$$ which generalizes . Similarly, the generalization of is $$\Bigl\|\partial_x^\beta \int_0^t e^{(t-s)L}\Lambda \omega(s)
{\,{\rm d}}s\Bigr\|_{{\mathbb{X}}^p(m)} \,\le\, C a(t)^{\frac{1}{2}-\frac{|\beta|}{2}}
\|\omega\|_{k,T}~, \qquad 0 < t \le T~. \label{eqint4}$$
Summarizing, we have shown that the linear map $F$ defined by satisfies the estimate $$\|F\omega\|_{k,T} \,\le\, C_1\|\omega_0\|_{{\mathbb{X}}^p(m)} +
\tilde C|\alpha| T^\frac{1}{2}\|\omega\|_{k,T}~, \qquad
\hbox{if} \quad 0 < T \le 1~,$$ where $\tilde C > 0$ depends only on $k$, $m$ and $p$. Arguing as in [@GW2 Corollary A.7 and Remark A.8], it is also straightforward to verify that $F\omega\in C([0,T];{\mathbb{X}}_{loc}^p(m))$ if $\omega \in
{\mathbb{U}}_{k,T}$. If we now assume that $T \le C_2(1+|\alpha|^2)^{-1}$, where $C_2 = 1/(4\tilde C^2)$, we see that $F$ is a strict contraction in ${\mathbb{U}}_{k,T}$. As a consequence, the integral equation has a unique fixed point $\omega \in {\mathbb{U}}_{k,T}$, which satisfies $\|\omega\|_{k,T} \le
2C_1\|\omega_0\|_{{\mathbb{X}}^p(m)}$. This proves that equation is locally well-posed in ${\mathbb{X}}^p(m)$, and since the local existence time $T$ is independent of the initial data, the solutions can be extended globally in time. Finally, since both operators $L$ and $\Lambda$ preserve the divergence-free condition, it is easy to check that, if $\nabla\cdot \omega_0=0$, then the solution $\omega$ of satisfies $\nabla \cdot
\omega(t)=0$ for all $t > 0$. This completes the proof. [$\Box$]{}
Decay estimates for the vertical derivatives {#subsec.est.x_3}
--------------------------------------------
Proposition \[prop.linear.local\] shows that the linearized equation is globally well-posed in the space ${\mathbb{X}}(m)$ for $m > 1$, but does not provide accurate estimates on the solution $\omega(t) = e^{t(L-\alpha\Lambda)}\omega_0$ for large times. In this section, we focus on the derivatives of $\omega(t)$ with respect to the vertical variable $x_3$. Using identity , we shall show that $\partial_{x_3}^k \omega(t)$ decays exponentially as $t \to \infty$, provided $k \in {\mathbb{N}}$ is large enough depending on $|\alpha|$. Albeit elementary, this observation plays a crucial role in the proof of Proposition \[prop.linear\], because it will allow us to simplify the study of the semigroup $e^{t(L-\alpha\Lambda)}$ by disregarding most of the terms involving a vertical derivative.
\[prop.derivative.x\_3\] Fix $m\in (1,\infty]$. There exist positive constants $C_3, C_4$ such that, for all $\alpha \in {\mathbb{R}}$, all $k \in {\mathbb{N}}$, and all $\omega_0 \in {\mathbb{X}}(m)$ with $\partial_{x_3}^k \omega_0 \in {\mathbb{X}}(m)$, the following estimate holds: $$\|\partial_{x_3}^k e^{t(L-\alpha\Lambda)}\omega_0\|_{{\mathbb{X}}(m)}\,\le\,
C_3 \,e^{(C_4(|\alpha|^2+1) - k)t} \|\partial_{x_3}^k \omega_0\|_{{\mathbb{X}}(m)}~,
\qquad t \ge 0~. \label{est.deriv.x_3}$$
[**Proof.**]{} In view of , it is sufficient to prove for $k = 0$. If $\omega_0 \in {\mathbb{X}}(m)$, we know from Proposition \[prop.linear.local\] that there exist constants $C_1 \ge 1$ and $C_2 > 0$, depending only on $m$, such that the solution $\omega(t) = e^{t(L-\alpha\Lambda)}
\omega_0$ of satisfies $\|\omega(t)\|_{{\mathbb{X}}(m)}
\le C_1 \|\omega_0\|_{{\mathbb{X}}(m)}$ for $t \in (0,t_0]$, where $t_0 =
C_2/(|\alpha|^2+1)$. Using the semigroup property, we can iterate this bound, and we easily obtain $$\|e^{t(L-\alpha\Lambda)}\omega_0\|_{{\mathbb{X}}(m)} \,\le\, C_3
\,e^{C_4(|\alpha|^2+1)t} \|\omega_0\|_{{\mathbb{X}}(m)}~, \qquad t \ge 0~,$$ where $C_3 = C_1$ and $C_4 = C_2^{-1}\log(C_1)$. This concludes the proof. [$\Box$]{}
Decomposition of the linearized operator {#subsec.decomposition}
----------------------------------------
Motivated by Proposition \[prop.derivative.x\_3\], we now decompose the linear operator $L - \alpha\Lambda$ as follows: $$L-\alpha\Lambda \,=\, {\mathscr{L}}_\alpha + {\mathcal{L}}_3 - \alpha H~,
\label{def.decompose}$$ where ${\mathscr{L}}_\alpha$ is defined in and ${\mathcal{L}}_3 =
\partial_{x_3}^2 - x_3\partial_{x_3}$. We recall that the operator ${\mathscr{L}}_\alpha$ does not involve any derivative with respect to the vertical variable $x_3$, and does not couple the horizontal and vertical components of $\omega = (\omega_h,\omega_3)^\top$. In view of –, the last term in has the following expression: $$H \,=\, \Lambda_3 - \tilde \Lambda_3 -\Lambda_4~,$$ where $\Lambda_3, \Lambda_4$ are defined in and $\tilde \Lambda_3$ after . More explicitly, we have $$H\omega \,=\, \begin{pmatrix} H_h\omega \\ H_3\omega\end{pmatrix}
\,=\, \begin{pmatrix} -g (K_{3D}*\partial_{x_3}\omega)_h \\
(K_{3D}*\omega-K_{2D}\star\omega_3,\nabla )g - g(K_{3D}*\partial_{x_3}
\omega )_3 \end{pmatrix}~, \label{def.Hexp}$$ where $K_{3D}$, $K_{2D}$ are the Biot-Savart kernels , , and $g$ is defined in . Here $\star$ denotes the convolution with respect to the horizontal variables, so that $$(K_{2D}\star\omega_3)(x_h,x_3) \,=\, \int_{{\mathbb{R}}^2}K_{2D}(x_h-y_h)
\,\omega_3(y_h,x_3){\,{\rm d}}y_h~.$$ Thus, unlike ${\mathscr{L}}_\alpha$, the operator $H$ involves vertical derivatives, and couples the horizontal and vertical components of $\omega$. As was already observed in Section \[sec.vectorial\], we have $H\omega = 0$ whenever $\partial_{x_3}\omega = 0$, see Proposition \[prop.H\] below.
Let $R_\alpha (t)$ denote the semigroup generated by the linear operator ${\mathscr{L}}_\alpha + {\mathcal{L}}_3$. In analogy with , we have the following representation: $$(R_\alpha(t)\omega)(x) \,=\, \frac{1}{\sqrt{2\pi a(2t)}}\int_{{\mathbb{R}}}
e^{-\frac{|x_3e^{-t}-y_3|^2}{2a(2t)}} \Bigl(e^{t{\mathscr{L}}_\alpha}\omega
(\cdot,y_3)\Bigr)(x_h){\,{\rm d}}y_3~, \quad t > 0~,
\label{eq.semigroup.R_alpha}$$ where $a(t) = 1-e^{-t}$ and $e^{t{\mathscr{L}}_\alpha}$ is the semigroup generated by ${\mathscr{L}}_\alpha$. Since $R_\alpha (t)$ does not couple the horizontal and vertical components of $\omega$, we can write $$R_{\alpha}(t)\omega \,=\, \begin{pmatrix}
R_{\alpha,h}(t)\omega_h \\ R_{\alpha,3}(t)\omega_3
\end{pmatrix}~,$$ where $R_{\alpha,h}(t)$ and $R_{\alpha,3}(t)$ are the semigroups generated by ${\mathscr{L}}_{\alpha,h} + {\mathcal{L}}_3$ and ${\mathscr{L}}_{\alpha,3} + {\mathcal{L}}_3$, respectively. Using the results of Section \[sec.vectorial\], we obtain the following estimates:
\[prop.R\_alpha\] Fix $m \in (1,\infty]$, $\alpha \in {\mathbb{R}}$, $\mu \in (1,\frac32)$, and take $\eta \in (0,\frac12]$ such that $2\eta < m-1$. Then there exists $C_5 > 0$ such that $$\begin{aligned}
\|R_{\alpha,h}(t)\omega_h\|_{X(m)^2} \,&\le\, C_5\,e^{-\mu t}
\|\omega_h\|_{X(m)^2}~, \label{est.R1}\\
\|R_{\alpha,3}(t)\omega_3\|_{X(m)} \,&\le\, C_5\,e^{-\eta t}
\|\omega_3\|_{X(m)}~, \label{est.R2}\end{aligned}$$ for all $\omega \in {\mathbb{X}}(m)$ and all $t \ge 0$.
[**Proof.**]{} Both estimates follow from the representation , Proposition \[prop.vectorial.2D\], and estimate . The calculations are straightforward, and can be omitted here. We just remark that, even if $\nabla\cdot\omega = 0$, the map $x_h \mapsto \omega_h(x_h,x_3)$ usually has a nonzero divergence for all values of $x_3 \in {\mathbb{R}}$. This is why Proposition \[prop.vectorial.2D\], hence also Proposition \[prop.R\_alpha\], was established without imposing any divergence-free condition. [$\Box$]{}
We conclude this section with a useful bound on the linear operator $H$.
\[prop.H\] Fix $m\in (1,\infty]$ and $\gamma \in
(0,1)$. There exists $C_6 > 0$ such that, for all $\omega \in {\mathbb{X}}(m)$ with $\partial_{x_3}\omega \in {\mathbb{X}}(m)$, one has $$\begin{aligned}
\|H_h\omega\|_{X(m)^2} \,&\le\, C_6\|\partial_{x_3}\omega\|_{{\mathbb{X}}(m)}~,
\label{est.H1}\\
\|H_3\omega\|_{X(m)} \,&\le\, C_6(\|\partial_{x_3}\omega\|_{{\mathbb{X}}(m)} +
\|\omega_h\|_{X(m)^2}^\gamma \|\partial_{x_3}\omega_h\|_{X(m)^2}^{1-\gamma})~.
\label{est.H2}\end{aligned}$$
[**Proof.**]{} We use the expression of the linear operator $H$. Since $\partial_{x_3}\omega \in {\mathbb{X}}(m)$, we know from Proposition \[prop.Riesz\] that $\partial_{x_3}u \equiv
K_{3D}*\partial_{x_3}\omega \in X^4(0)$. Thus, using Hölder’s inequality, we obtain $$\|g\,\partial_{x_3}u\|_{{\mathbb{X}}(m)} \,\le\, \|\partial_{x_3}u\|_{X^4(0)}
\Bigl(\int_{{\mathbb{R}}^2}\rho_m(|x_h|^2)^2 g(x_h)^4{\,{\rm d}}x_h\Bigr)^{1/4}
\,\le\, C \|\partial_{x_3}\omega\|_{{\mathbb{X}}(m)}~.$$ In particular, we have $\|H_h\omega\|_{X(m)^2} \le C\|\partial_{x_3}
\omega\|_{{\mathbb{X}}(m)}$.
We next consider the two-dimensional vector $I = (K_{3D}*\omega -
K_{2D}\star\omega_3)_h$ and estimate the term $(I,\nabla_h)g$. Using the definitions , , it is straightforward to verify that $I(x) = I_1(x) + I_2(x)$, where $$\begin{aligned}
I_1(x) \,&=\, \frac{1}{4\pi} \int_{{\mathbb{R}}^3} \frac{(x_h{-}y_h)^\bot}{
|x-y|^3}\,(\omega_3(y_h,y_3)-\omega_3(y_h,x_3)) {\,{\rm d}}y~, \\
I_2(x) \,&=\, \frac{1}{4\pi} \int_{{\mathbb{R}}^3} \frac{(x_3{-}y_3)}{|x-y|^3}
\,(\omega_h(y_h,y_3)-\omega_h(y_h,x_3))^\bot {\,{\rm d}}y~.\end{aligned}$$ Since $\nabla_h g (x_h)=- g(x_h) x_h/2$ and $|x_h\cdot
(x_h-y_h)^\bot|\leq |x_h||x_h-y_h|^{1-\sigma}|y_h|^\sigma$ for any $\sigma\in [0,1]$, we can bound $$\begin{aligned}
|(I_1,\nabla_h) g(x)| \,&\le\, C g(x_h) |x_h| \int_{|x_3-y_3|\ge 1}
\frac{|y_h|^\sigma}{|x-y|^{2+\sigma}}\,|\omega_3(y_h,y_3)-
\omega_3(y_h,x_3)|{\,{\rm d}}y \\
&\quad~ + C g(x_h) |x_h| \int_{|x_3-y_3|< 1}\frac{1}{|x-y|^{2}}
\,|\omega_3(y_h,y_3)-\omega_3(y_h,x_3)| {\,{\rm d}}y~.\end{aligned}$$ We now proceed like in the proof of Lemma \[lem.Riesz\]. Integrating first with respect to the horizontal variable $y_h \in {\mathbb{R}}^2$ and applying Hölder’s inequality, we obtain $$\begin{aligned}
|(I_1,\nabla_h) g(x)| \,&\le\, C g(x_h) |x_h| \int_{|x_3-y_3|\ge 1}
\frac{1}{|x_3-y_3|^{2+\sigma}} \||\cdot|^\sigma \{\omega_3(\cdot,y_3)-
\omega_3(\cdot,x_3)\}\|_{L^1({\mathbb{R}}^2)} {\,{\rm d}}y_3 \\
&\quad~ + C g(x_h) |x_h| \int_{|x_3-y_3|< 1}\frac{1}{|x_3-y_3|}
\|\omega_3(\cdot,y_3)-\omega_3(\cdot,x_3) \|_{L^2({\mathbb{R}}^2)}
{\,{\rm d}}y_3~.\end{aligned}$$ Assuming $0 < \sigma < m-1$, we have the estimate $\||\cdot|^\sigma f
\|_{L^1({\mathbb{R}}^2)} \le C\|f\|_{L^2(m)}$ for any $f \in L^2(m)$, hence $$\||\cdot|^{\sigma} \{\omega_3 (\cdot,y_3)-\omega_3 (\cdot,x_3)\}
\|_{L^1({\mathbb{R}}^2) } + \|\omega_3(\cdot,y_3)-\omega_3(\cdot,x_3)\|_{L^2({\mathbb{R}}^2)}
\,\le\, C |x_3-y_3| \|\partial_{x_3}\omega_3\|_{X(m)}~.$$ We conclude that $$\begin{aligned}
|(I_1,\nabla_h) g(x)| \,&\le\, C g(x_h)|x_h| \Bigl(\int_{|x_3-y_3|\ge 1}
\frac{1}{|x_3-y_3|^{1+\sigma}} {\,{\rm d}}y_3 + \int_{|x_3-y_3|< 1} {\,{\rm d}}y_3 \Bigr)
\|\partial_{x_3}\omega_3\|_{X(m)}\\
\,&\le\, C g(x_h)|x_h|\|\partial_{x_3}\omega_3\|_{X(m)}~,\end{aligned}$$ which gives the bound $\|(I_1,\nabla)g\|_{X(m)}\le C\|\partial_{x_3}
\omega_3\|_{X(m)}$.
Finally we consider the term $(I_2,\nabla_h)g$. Using again Hölder’s inequality, we obtain $$\begin{aligned}
|I_2(x)| \,&\le\, C\int_{|x_3-y_3|\ge 1}\frac{1}{|x-y|^2}|\omega_h
(y_h,y_3)-\omega_h(y_h,x_3)|{\,{\rm d}}y \\
&\quad~ + C \int_{|x_3-y_3|< 1}\frac{1}{|x-y|^2}|\omega_h (y_h,y_3)-
\omega_h (y_h,x_3)|{\,{\rm d}}y \\
\,&\le\, C \int_{|x_3-y_3|\ge 1}\frac{1}{|x_3-y_3|^2}\|\omega_h(\cdot,y_3)-
\omega_h(\cdot,x_3)\|_{L^1({\mathbb{R}}^2)} {\,{\rm d}}y_3 \\
&\quad~ + \int_{|x_3-y_3|< 1}\frac{1}{|x_3-y_3|}\|\omega_h(\cdot,y_3)-
\omega_h(\cdot,x_3) \|_{L^2({\mathbb{R}}^2)}{\,{\rm d}}y_3~.\end{aligned}$$ Since $L^2(m) \hookrightarrow L^p({\mathbb{R}}^2)$ for $p \in [1,2]$, we have $\|\omega_h (\cdot,y_3)-\omega_h (\cdot,x_3)\|_{L^p({\mathbb{R}}^2)^2} \le
2\|\omega_h\|_{X(m)^2}$ and $\|\omega_h (\cdot,y_3)-\omega_h
(\cdot,x_3)\|_{L^p({\mathbb{R}}^2)^2} \le |x_3-y_3|\|\partial_{x_3}\omega_h\|_{X(m)^2}$. In particular, for any $\gamma \in (0,1)$, $$\|\omega_h (\cdot,y_3)-\omega_h (\cdot,x_3)\|_{L^p({\mathbb{R}}^2)^2} \,\le\,
2^\gamma |x_3-y_3|^{1-\gamma} \|\omega_h \|_{X(m)^2}^\gamma
\|\partial_{x_3}\omega_h \|_{X(m)^2}^{1-\gamma}~.$$ Thus we obtain $$\|I_2\|_{L^\infty({\mathbb{R}}^3)^2} \,\le\, C \|\omega_h\|_{X(m)^2}^\gamma
\|\partial_{x_3}\omega_h\|_{X(m)^2}^{1-\gamma}~,$$ and conclude that $\|(I_2,\nabla_h)g\|_{X(m)} \le
C\|\omega_h\|_{X(m)^2}^\gamma \|\partial_{x_3}\omega_h\|_{X(m)^2}^{1-\gamma}$. This completes the proof of Proposition \[prop.H\]. [$\Box$]{}
Large time estimates {#subsec.est.large}
--------------------
In this section we complete the proof of Proposition \[prop.linear\]. Fix $m \in (1,\infty]$, $\alpha \in {\mathbb{R}}$, and assume that $\omega_0
\in {\mathbb{X}}^p(m)$ for some $p \in [1,2]$. Let $\omega(t) = e^{t(L-\alpha\Lambda)}
\omega_0$ be the solution of the linearized equation given by Proposition \[prop.linear.local\]. Take any $k \in {\mathbb{N}}$ such that $k > C_4(|\alpha|^2+1) + 1/2$, where $C_4$ is as in Proposition \[prop.derivative.x\_3\], and choose $t_0 > 0$ small enough so that estimate holds for all $t \in (0,t_0]$ and all $\beta \in {\mathbb{N}}^3$ with $|\beta| \le k$. Our goal is to control the solution $\omega(t)$ for $t \ge t_0$ and to establish the decay estimates , .
To this end, we first observe that $\omega(t)$ satisfies the integral equation $$\omega(t) \,=\,R_\alpha (t-t_0)\omega(t_0) - \alpha\int_{t_0}^t
R_\alpha (t-s) H\omega(s){\,{\rm d}}s~, \quad t \ge t_0~,
\label{eq.linear.integral}$$ where $R_\alpha(t)$ is the semigroup defined by . Fix $\bar\eta \in (0,1/2)$ such that $2\bar\eta < m-1$. By Proposition \[prop.R\_alpha\], we have $$\label{est.om1}
\|\omega(t)\|_{{\mathbb{X}}(m)} \,\le\, C_5\,e^{-\bar\eta(t-t_0)}\|\omega(t_0)\|_{{\mathbb{X}}(m)}
+ C_5|\alpha|\int_{t_0}^t e^{-\bar\eta(t-s)}\|H\omega(s)\|_{{\mathbb{X}}(m)}{\,{\rm d}}s~.$$ To estimate the term $\|H\omega(s)\|_{{\mathbb{X}}(m)}$, we first apply Proposition \[prop.H\] with $\gamma = 1/2$, and then the classical interpolation inequality $$\|\partial_{x_3}\omega\|_{{\mathbb{X}}(m)} \,\le\, C \|\omega\|_{{\mathbb{X}}(m)}^{1-1/k}
\,\|\partial_{x_3}^k\omega\|_{{\mathbb{X}}(m)}^{1/k}~.$$ Using in addition Young’s inequality, we conclude that, given any $\epsilon > 0$, there exists $C_\epsilon > 0$ such that $$\label{est.om2}
C_5 |\alpha|\,\|H\omega(s)\|_{{\mathbb{X}}(m)} \,\le\, \epsilon \|\omega(s)\|_{{\mathbb{X}}(m)}
+ C_\epsilon \|\partial_{x_3}^k\omega(s)\|_{{\mathbb{X}}(m)}~.$$ On the other hand, since $k > C_4(|\alpha|^2+1) + 1/2$, it follows from that $$\label{est.om3}
\|\partial_{x_3}^k\omega(s)\|_{{\mathbb{X}}(m)} \,\le\, C_3\,e^{-(s-t_0)/2}
\|\partial_{x_3}^k\omega(t_0)\|_{{\mathbb{X}}(m)}~, \quad s \ge t_0~.$$ Replacing and into , we easily obtain $$\|\omega(t)\|_{{\mathbb{X}}(m)} \,\le\, \Bigl(C_5\|\omega(t_0)\|_{{\mathbb{X}}(m)} +
C_\epsilon'\|\partial_{x_3}^k\omega(t_0)\|_{{\mathbb{X}}(m)}\Bigr)
\,e^{-\bar\eta(t-t_0)} + \epsilon \int_{t_0}^t e^{-\bar\eta(t-s)}
\|\omega(s)\|_{{\mathbb{X}}(m)}{\,{\rm d}}s~,$$ for some $C_\epsilon' > 0$. Applying now Gronwall’s lemma, and using to bound $\|\omega(t_0)\|_{{\mathbb{X}}(m)}$ and $\|\partial_{x_3}^k\omega(t_0)\|_{{\mathbb{X}}(m)}$ in terms of $\omega_0$, we see that $\|\omega(t)\|_{{\mathbb{X}}(m)} \le C\,e^{-\eta t}\|\omega_0\|_{{\mathbb{X}}^p(m)}$ for $t \ge t_0$, where $\eta = \bar \eta - \epsilon$. Finally, using again to control the solution for $t < t_0$, we conclude that there exists $C_7 > 0$ such that $$\label{est.om4}
\|\omega(t)\|_{{\mathbb{X}}(m)} \,\equiv\, \|e^{t(L-\alpha\Lambda)}\omega_0\|_{{\mathbb{X}}(m)}
\,\le\, \frac{C_7\,e^{-\eta t}}{a(t)^{\frac1p-\frac12}}
\,\|\omega_0\|_{{\mathbb{X}}^p(m)}~,$$ for all $t > 0$. Since $\epsilon > 0$ was arbitrary, estimate holds for any $\eta \in (0,1/2)$ such that $2\eta < m-1$.
To conclude the proof, it remains to find the optimal decay rates for $\|\omega_h(t)\|_{{\mathbb{X}}(m)}$, $\|\omega_3(t)\|_{{\mathbb{X}}(m)}$ (including the value $\eta = 1/2$ if $m > 2$), and to establish , for $\beta \neq
0$ too. First, combining , and using again for short times, we easily obtain $$\label{est.om5}
\|\partial_{x_3}\omega(t)\|_{{\mathbb{X}}(m)} \,\equiv\, \|\partial_{x_3}
e^{t(L-\alpha\Lambda)}\omega_0\|_{{\mathbb{X}}(m)} \,\le\, \frac{C\,e^{-(\eta+1)t}}
{a(t)^{\frac1p}}\,\|\omega_0\|_{{\mathbb{X}}^p(m)}~,$$ for all $t > 0$. Moreover, if $m > 2$, we know from Proposition \[prop.R\_alpha\] that holds with $\bar
\eta = 1/2$. Thus, applying Proposition \[prop.H\] to estimate $\|H\omega(s)\|_{{\mathbb{X}}(m)}$ and using , , we find that $\|\omega(t)\|_{{\mathbb{X}}(m)}$ decays like $e^{-t/2}$ as $t \to
\infty$, hence holds with $\eta = 1/2$ if $m >
2$.
Next, to obtain a faster decay estimate for the horizontal component $\omega_h$, we use and . Instead of , we find $$\label{est.om6}
\|\omega_h(t)\|_{X(m)^2} \,\le\, C\,e^{-\mu(t-t_0)}
\|(\omega(t_0))_h\|_{X(m)^2} + C|\alpha|\int_{t_0}^t e^{-\mu(t-s)}
\|\partial_{x_3}\omega(s)\|_{{\mathbb{X}}(m)}{\,{\rm d}}s~,$$ for any $\mu \in (1,\frac32)$. Since $\|\partial_{x_3}\omega(t)\|_{{\mathbb{X}}(m)}
\le Ce^{-(\eta+1)t} \|\partial_{x_3}\omega_0\|_{{\mathbb{X}}(m)}$ by , , we conclude that $\|\omega_h(t)\|_{X(m)^2}$ decays like $e^{-\mu t}$ as $t \to \infty$, provided $\mu < 1+\eta$. In other words, if $\mu \in (1,\frac32)$ satisfies $2\mu < m+1$, we have $$\label{est.om7}
\|\omega_h(t)\|_{X(m)^2} \,\equiv\, \|(e^{t(L-\alpha\Lambda)}\omega_0)_h
\|_{X(m)^2} \,\le\, C\,e^{-\mu t}(\|(\omega_0)_h\|_{X(m)^2}
+ \|\partial_{x_3}\omega_0\|_{{\mathbb{X}}(m)})~,$$ for all $t > 0$. Using the arguments leading to and proceeding as in Proposition \[prop.linear.local\], we can also derive the following short time estimate, which complements : $$\|\partial_x^\beta\omega_h(t)\|_{X(m)^2} \,\le\, \frac{C_1}{a(t)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}}\Bigl(\|(\omega_0)_h\|_{X^p(m)^2}
+ \|\partial_{x_3}\omega_0\|_{{\mathbb{X}}^p(m)}\Bigr)~, \quad
0 < t\le \frac{C_2}{|\alpha|^2{+}1}~.
\label{est.om8}$$
Finally, to obtain decay estimates for the derivative $\partial_x^
\beta \omega(t)$, where $\beta \in {\mathbb{N}}^3$, we can restrict ourselves to $t \ge 2t_1$, where $t_1 > 0$ is small enough so that the short time estimates , hold for $0 < t \le 2t_1$. In view of , we have the identity $$\partial_x^\beta e^{t(L-\alpha\Lambda)}\omega_0 \,=\,
e^{-\beta_3(t-t_1)}~\partial_{x_h}^{\beta_h} e^{t_1(L-\alpha\Lambda)}~
e^{(t-2t_1)(L-\alpha\Lambda)}~\partial_{x_3}^{\beta_3} e^{t_1(L-\alpha\Lambda)}
\omega_0~.$$ Using the short time estimates , with $p = 2$ to bound the first operator $\partial_{x_h}^{\beta_h} e^{t_1(L-\alpha\Lambda)}$, then the long-time estimates , or to treat the middle term $e^{(t-2t_1)(L-\alpha\Lambda)}$, and finally again to bound the last term $\partial_{x_3}^{\beta_3} e^{t_1(L-\alpha\Lambda)}\omega_0$, we easily obtain and , together with the following estimate $$\|\partial_x^\beta(e^{t(L-\alpha\Lambda)}\omega_0)_h\|_{X(m)^2}
\,\le\, \frac{C\,e^{-(\mu+\beta_3)t}}{a(t)^{\frac{1}{p}-\frac{1}{2}
+\frac{|\beta|}{2}}}\Bigl(\|(\omega_0)_h\|_{X^p(m)^2}
+ \|\partial_{x_3}\omega_0\|_{{\mathbb{X}}^p(m)}\Bigr)~, \quad t > 0,
\label{est.prop.linear.3}$$ which will also be used in the next section. This concludes the proof of Proposition \[prop.linear\]. [$\Box$]{}
Nonlinear stability {#sec.nonlinear}
===================
In this section we consider the nonlinear stability of the Burgers vortex and prove Theorems \[thm.main1\] and \[thm.main2\]. Our starting point is the perturbation equation , which is equivalent to the integral equation $$\omega(t) \,=\, e^{t(L-\alpha\Lambda)}\omega_0 + \sum_{j=1}^2
\int_0^t e^{(t-s)(L-\alpha\Lambda)} N_j(\omega (s),\omega (s)){\,{\rm d}}s~,
\qquad t \ge 0~,
\label{eq.omega.integral}$$ where $N_1(v,w) = (K_{3D}*v,\nabla)w$, $N_2(v,w) = (v,\nabla)K_{3D}*w$, and $K_{3D}$ is the Biot-Savart kernel . We first establish the following result, which already implies Theorem \[thm.main1\].
\[prop.nonlinear.stability\] Fix $m\in (1,\infty]$, $\alpha\in {\mathbb{R}}$, and take $\eta \in (0,\frac12]$ such that $2\eta < m-1$. Then there exist $\delta = \delta(\alpha,m,\eta) > 0$ and $C = C(\alpha,m,
\eta) > 0$ such that, for any $\omega_0\in {\mathbb{X}}(m)$ with $\nabla\cdot
\omega_0=0$ and $\|\omega_0\|_{{\mathbb{X}}(m)} \le \delta$, Eq. has a unique solution $\omega\in L^\infty({\mathbb{R}}_+;{\mathbb{X}}(m))\cap C([0,\infty);{\mathbb{X}}_{loc}(m))$, which satisfies $$\label{eq.decay2}
\|\partial_x^\beta \omega(t)\|_{{\mathbb{X}}(m)} \,\le\,
\frac{C\|\omega_0\|_{{\mathbb{X}}(m)}}{a(t)^\frac{|\beta|}{2}}
\,e^{-\eta t}~, \qquad t>0~,$$ for any multi-index $\beta \in {\mathbb{N}}^3$ of length $|\beta|\le 1$.
[**Proof.**]{} Let ${\mathbb{U}}$ be the Banach space of all $\omega
\in L^\infty({\mathbb{R}}_+;{\mathbb{X}}(m))\cap C([0,\infty);{\mathbb{X}}_{loc}(m))$ such that $\nabla\cdot \omega(t)=0$ for all $t > 0$ and $\|\omega\|_{\mathbb{U}}< \infty$, where $$\|\omega\|_{\mathbb{U}}\,=\, \sum_{|\beta|\le 1}\sup_{t>0} a(t)^\frac{|\beta|}{2}
e^{\eta t}\|\partial_x^\beta \omega(t)\|_{{\mathbb{X}}(m)}~.$$ Given $\omega_0 \in {\mathbb{X}}(m)$ such that $\nabla\cdot\omega_0=0$, we denote by $\Phi : {\mathbb{U}}\to {\mathbb{U}}$ the nonlinear map defined by $$\Phi(\omega)(t) \,=\, e^{t(L-\alpha\Lambda)}\omega_0 +
\sum_{j=1}^2\Phi_j(\omega,\omega)(t)~, \quad t > 0~,
\label{def.Phi}$$ where $\Phi_1$, $\Phi_2$ are the following bilinear operators: $$\Phi_j(\omega,\tilde\omega)(t) \,=\, \int_0^t e^{(t-s)(L-\alpha\Lambda)}
N_j(\omega(s),\tilde\omega(s)){\,{\rm d}}s~, \quad j = 1,2~.
\label{def.Phi_j}$$ If $\|\omega_0\|_{{\mathbb{X}}(m)}$ is sufficiently small, we shall show that the map $\Phi$ is a strict contraction in the ball $B_K = \{\omega \in {\mathbb{U}}\,|\,
\|\omega\|_{\mathbb{U}}\le K\}$ for some suitable $K > 0$. It will follow that $\Phi$ has a unique fixed point $\omega$ in $B_K$, which by construction is the desired solution of .
Since $\omega_0 \in {\mathbb{X}}(m)$ and $\nabla\cdot\omega_0=0$, Proposition \[prop.linear\] shows that the map $t \mapsto
e^{t(L-\alpha\Lambda)}\omega_0$ belongs to ${\mathbb{U}}$, and satisfies the estimate $$\|e^{t(L-\alpha\Lambda)}\omega_0\|_{\mathbb{U}}\,\le\, C_1\|\omega_0\|_{{\mathbb{X}}(m)}~,$$ for some $C_1 > 0$ (depending on $m$, $\alpha$, $\eta$). On the other hand, if $v,w \in {\mathbb{X}}(m)$, Corollary \[cor.prop.Riesz\] implies that $N_1(v,w)$ and $N_2(v,w)$ belong to $X^p(m)^3$ for any $p \in (1,2)$, and satisfy the bound $$\|N_1(v,w)\|_{X^p(m)^3} + \|N_2(v,w)\|_{X^p(m)^3} \,\le\,
C\|v\|_{{\mathbb{X}}(m)}\|\nabla w\|_{{\mathbb{X}}(m)}~,$$ for some $C > 0$ (depending on $m$ and $p$). If in addition $\nabla\cdot
v = 0$, then denoting $u = K_{3D}*v$ we find $$\int_{{\mathbb{R}}^2} (N_1(v,v) + N_2(v,v))_3{\,{\rm d}}x_h \,=\, \int_{{\mathbb{R}}^2} \nabla_h
\cdot (v_h u_3 - u_h v_3){\,{\rm d}}x_h \,=\, 0~,$$ for all $x_3 \in {\mathbb{R}}$, hence $N_1(v,v) + N_2(v,v) \in {\mathbb{X}}^p(m)$. As a consequence, if $\omega, \tilde \omega \in {\mathbb{U}}$, we have $N_j(\omega(t),
\tilde\omega(t)) \in X^p(m)^3$ for $j = 1,2$ and all $t > 0$, and using Proposition \[prop.linear\] again we obtain the following estimate for the bilinear operators $\Phi_j$: $$\begin{aligned}
\Bigl\|\sum_{j=1}^2 \partial_x^\beta &\Phi_{j}(\omega,\tilde\omega)
(t)\Bigr\|_{{\mathbb{X}}(m)} \,\le\, \sum_{j=1}^2 \int_0^t \|\partial_x^\beta
e^{(t-s)(L-\alpha\Lambda)}N_j(\omega(s),\tilde\omega(s))\|_{{\mathbb{X}}(m)}{\,{\rm d}}s\\
\,&\le\, C \sum_{j=1}^2 \int_0^t \frac{e^{-\eta(t-s)}}{a(t{-}s)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}}\|N_j(\omega(s),\tilde\omega(s))
\|_{X^p(m)^3}{\,{\rm d}}s\\
\,&\le\, C \int_0^t\frac{e^{-\eta(t-s)}}{a(t{-}s)^{\frac{1}{p}-\frac{1}{2}
+\frac{|\beta|}{2}}} \|\omega(s)\|_{{\mathbb{X}}(m)}\|\nabla\tilde
\omega(s)\|_{{\mathbb{X}}(m)} {\,{\rm d}}s\\
\,&\le\, C \int_0^t \frac{e^{-\eta (t-s)} e^{-2\eta s}}{a(t{-}s)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}a(s)^\frac{1}{2}}{\,{\rm d}}s \,\|\omega\|_{\mathbb{U}}\|\tilde \omega\|_{\mathbb{U}}\,\le\, \frac{Ce^{-\eta t}}{a(t)^{\frac{1}{p}
+\frac{|\beta|}{2}-1}}\|\omega\|_{\mathbb{U}}\|\tilde\omega\|_{\mathbb{U}}~.\end{aligned}$$ Since we also know that $N_1(\omega(t),\omega(t)) + N_2(\omega(t),
\omega(t))$ belongs to ${\mathbb{X}}^p(m)$ for all $t > 0$ and is divergence-free, we have shown that $\Phi$ maps ${\mathbb{U}}$ into ${\mathbb{U}}$, and that there exists $C_2 > 0$ (depending on $|\alpha|$, $m$, and $\eta$) such that $$\|\Phi(\omega)\|_{\mathbb{U}}\,\le\, C_1\|\omega_0\|_{{\mathbb{X}}(m)} + C_2\|\omega\|_{\mathbb{U}}^2~,
\quad
\|\Phi(\omega) - \Phi(\tilde\omega)\|_{\mathbb{U}}\,\le\, C_2(\|\omega\|_{\mathbb{U}}+
\|\tilde\omega\|_{\mathbb{U}})\|\omega-\tilde\omega\|_{\mathbb{U}}~,$$ for all $\omega, \tilde \omega \in {\mathbb{U}}$. We now take $K > 0$ such that $2C_2K < 1$, and assume that $\|\omega_0\|_{{\mathbb{X}}(m)} \le K/(2C_1)$. Then the estimates above show that $\Phi$ is a strict contraction in the ball $B_K$, hence has a unique fixed point $\omega \in B_K$ which, of course, satisfies . Moreover $\|\omega\|_{\mathbb{U}}\le 2C_1\|\omega_0\|_{{\mathbb{X}}(m)}$, hence holds with $C = 2C_1$. This concludes the proof. [$\Box$]{}
[**Remark.**]{} The size $\delta$ of the local basin of attraction of the Burgers vortex $\alpha G$ in ${\mathbb{X}}(m)$ depends a priori on $\alpha$, $m$, and $\eta$. However, as announced in Theorem \[thm.main2\], the dependence on the decay rate $\eta$ can easily be removed by the following (standard) argument. Given $m > 1$, we first choose $\eta =
\bar\eta(m) = \min(\frac12,\frac{m-1}{4})$ and apply Proposition \[prop.nonlinear.stability\] with that value of $\eta$. We thus obtain a constant $\bar\delta > 0$ depending only on $\alpha$ and $m$ such that, for any $\omega_0\in {\mathbb{X}}(m)$ with $\nabla\cdot \omega_0=0$ and $\|\omega_0\|_{{\mathbb{X}}(m)} \le \bar\delta$, Eq. has a unique solution $\omega \in
L^\infty({\mathbb{R}}_+;{\mathbb{X}}(m)) \cap C([0,\infty);{\mathbb{X}}_{loc}(m))$, which converges exponentially to zero as $t \to \infty$. In particular, given any $\eta \in (0,\frac12]$ such that $2\eta < m-1$, there exists $T = T(\eta) > 0$ such that $\|\omega(t)\|_{{\mathbb{X}}(m)} \le \delta$ for all $t \ge T$, where $\delta = \delta(\alpha,m,\eta)$ is the constant given by Proposition \[prop.nonlinear.stability\]. By uniqueness of the solution, we conclude that $\omega$ satisfies for any admissible value of $\eta$.
In view of Proposition \[prop.nonlinear.stability\] and the remark that follows, the proof of Theorem \[thm.main2\] will be complete once we have established the improved decay estimate for the horizontal component $\omega_h$. A convenient way to do so is to repeat the proof of Proposition \[prop.nonlinear.stability\] using a different function space, which incorporates a faster decay rate as $t \to
\infty$. Given $\mu \in (1,1+\eta)$, where $\eta \in (0,\frac12]$ is as in Proposition \[prop.nonlinear.stability\], we introduce the space ${\mathbb{V}}\subset {\mathbb{U}}$ defined by the norm $$\|\omega\|_{\mathbb{V}}\,=\, \sum_{k=0,1}\sum_{|\beta|\le 1}\Bigl(
\sup_{t>0} a(t)^\frac{k}{2} e^{(\mu +k\eta)t}\|\partial_{x_3}^k
\partial_x^\beta \omega_h (t)\|_{X(m)^2} + \sup_{t>0}a(t)^\frac{k}{2}
e^{(\eta+k)t}\|\partial_{x_3}^k\partial_x^\beta \omega_3(t)\|_{X(m)}
\Bigr)~.$$ As in the remark above, we can assume here (without loss of generality) that $\|\partial_x^\beta\omega_0\|_{{\mathbb{X}}(m)}$ is finite and arbitrarily small, for all $\beta \in {\mathbb{N}}^3$ with $|\beta| \le 1$. Using Proposition \[prop.linear\], we thus obtain $$\|e^{t(L-\alpha\Lambda)}\omega_0\|_{\mathbb{V}}\,\le\, C_3\sum_{|\beta|\le 1}
\|\partial_x^\beta\omega_0\|_{{\mathbb{X}}(m)}~,$$ for some $C_3 > 0$. On the other hand, if $v,w \in {\mathbb{X}}(m)$, the following estimates hold for any $p \in (1,2)$: $$\begin{aligned}
\|N_{1,h}(v,w)\|_{X^p(m)^2} \,&\le\, C\|v\|_{{\mathbb{X}}(m)}
\|\nabla w_h\|_{X(m)^2}~, \\
\|N_2(v,w)\|_{X^p(m)^3} \,&\le\, C(\|v_h\|_{X(m)^2}\|\nabla_h w\|_{{\mathbb{X}}(m)} +
C\|v_3\|_{X(m)}\|\partial_{x_3} w\|_{{\mathbb{X}}(m)})~,\\
\|\partial_{x_3}N_j(v,w)\|_{X^p(m)^3} \,&\le\, C(\|\partial_{x_3}v\|_{{\mathbb{X}}(m)}
\|\nabla w\|_{{\mathbb{X}}(m)} + \|v\|_{{\mathbb{X}}(m)}\|\partial_{x_3}\nabla w\|_{{\mathbb{X}}(m)})~.\end{aligned}$$
We now estimate the bilinear operators $\Phi_j(\omega,\tilde\omega)$ for $\omega, \tilde\omega \in {\mathbb{V}}$. First, using , we find for $t \ge 1$: $$\begin{aligned}
\nonumber
\|\partial_x^\beta &\Phi_{1,h}(\omega,\tilde\omega)(t)\|_{X(m)^2}
\,\le\, \int_0^t \|\partial_x^\beta \{e^{(t-s)(L-\alpha\Lambda)}
N_1(\omega(s),\tilde\omega(s))\}_h\|_{X(m)^2}{\,{\rm d}}s\\ \nonumber
\,&\le\, C \int_0^t \frac{e^{-\mu(t-s)}}{a(t{-}s)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}} (\|N_{1,h}(\omega(s),\tilde
\omega(s))\|_{X^p(m)^2} + \|\partial_{x_3}N_1(\omega(s),\tilde
\omega(s))\|_{X^p(m)^3}){\,{\rm d}}s\\ \nonumber
\,&\le\, C \int_0^t \frac{e^{-\mu(t-s)}}{a(t{-}s)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}}(\|\omega(s)\|_{{\mathbb{X}}(m)}\|\nabla
\tilde\omega_h(s)\|_{(X(m))^2}\\ \nonumber
& \hspace{3.8cm} +\|\partial_{x_3}\omega(s)\|_{{\mathbb{X}}(m)}\|\nabla \tilde
\omega(s)\|_{{\mathbb{X}}(m)} +\|\omega(s)\|_{{\mathbb{X}}(m)}\|\partial_{x_3}\nabla\tilde
\omega(s)\|_{{\mathbb{X}}(m)}){\,{\rm d}}s\\ \label{est.Phi1}
\,&\le\, C \int_0^t \frac{e^{-\mu(t-s)} e^{-(\mu+\eta)s}}{a(t{-}s)^{
\frac{1}{p}-\frac{1}{2}+\frac{|\beta|}{2}} a(s)^\frac{1}{2}}{\,{\rm d}}s
~\|\omega\|_{\mathbb{V}}\|\tilde\omega\|_{\mathbb{V}}\,\le\, C e^{-\mu t}
\|\omega\|_{\mathbb{V}}\|\tilde\omega\|_{\mathbb{V}}~.\end{aligned}$$ In the last inequality, we have used the definition of the norm in ${\mathbb{V}}$ and the fact that $\mu + \eta < 1 + 2\eta$. The bound also holds for $t < 1$, and can easily be established using instead of .
Next, to bound $\partial_{x_3}\Phi_{1,h}(\omega,\tilde\omega)$, we recall that $\partial_{x_3}e^{t(L-\alpha\Lambda)} = e^{-t}e^{t(L-\alpha\Lambda)}
\partial_{x_3}$. Applying , we find $$\begin{aligned}
\|\partial_{x_3}\partial_x^\beta &\Phi_{1,h}(\omega,\tilde\omega)(t)\|_{X(m)^2}
\,\le\, \int_0^t e^{-(t-s)}\|\partial_x^\beta \{e^{(t-s)(L-\alpha\Lambda)}
\partial_{x_3}N_1(\omega(s),\tilde\omega(s))\}_h\|_{X(m)^2}{\,{\rm d}}s\\
\,&\le\, C \int_0^t \frac{e^{-(\mu+1)(t-s)}}{a(t{-}s)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}} \|\partial_{x_3} N_1(\omega(s),
\tilde\omega(s))\|_{X^p(m)^3}{\,{\rm d}}s\\
\,&\le\, C \int_0^t \frac{e^{-(\mu+1)(t-s)}e^{-(\mu+\eta)s}}{a(t{-}s)^{
\frac{1}{p}-\frac{1}{2}+\frac{|\beta|}{2}}a(s)^\frac{1}{2}}{\,{\rm d}}s
~\|\omega\|_{\mathbb{V}}\|\tilde\omega\|_{\mathbb{V}}\,\le\, \frac{Ce^{-(\mu+\eta)t}}{a(t)^{
\frac{1}{p}+\frac{|\beta|}{2}-1}}\,\|\omega\|_{\mathbb{V}}\|\tilde\omega\|_{\mathbb{V}}~.\end{aligned}$$ Similarly, for $k = 0,1$, we can estimate $\partial_{x_3}^k\Phi_{2,h}
(\omega,\tilde\omega)$ as follows: $$\begin{aligned}
\|\partial_{x_3}^k\partial_x^\beta &\Phi_{2,h}(\omega,\tilde\omega)
(t)\|_{X(m)^2} \,\le\, \int_0^t e^{-k(t-s)}\|\partial_x^\beta
\{e^{(t-s)(L-\alpha\Lambda)} \partial_{x_3}^k N_2(\omega(s),\tilde
\omega(s))\}_h\|_{X(m)^2}{\,{\rm d}}s\\
\,&\le\, C \int_0^t \frac{e^{-(\mu+k)(t-s)}}{a(t{-}s)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}} \|\partial_{x_3}^k N_{2}(\omega(s),
\tilde\omega(s))\|_{X^p(m)^3}{\,{\rm d}}s\\
\,&\le\, C \int_0^t \frac{e^{-(\mu+k)(t-s)}e^{-(\mu+\eta)s}}{a(t{-}s)^{
\frac{1}{p}-\frac{1}{2}+\frac{|\beta|}{2}}a(s)^\frac{k}{2}}{\,{\rm d}}s
~\|\omega\|_{\mathbb{V}}\|\tilde\omega\|_{\mathbb{V}}\,\le\, \frac{Ce^{-(\mu+k\eta)t}}{
a(t)^{\frac{1}{p}+\frac{|\beta|}{2}+\frac{k}{2}-\frac{3}{2}}}
\|\omega\|_{\mathbb{V}}\|\tilde\omega\|_{\mathbb{V}}~.\end{aligned}$$ Finally, using , we obtain for the vertical components of $\Phi_j(\omega,\tilde\omega)$: $$\begin{aligned}
\|\partial_{x_3}^k\partial_x^\beta &\Phi_{j,3}(\omega,
\tilde\omega)(t)\|_{X(m)} \,\le\, \int_0^t e^{-k(t-s)}\|\partial_x^\beta
\{e^{(t-s)(L-\alpha\Lambda)} \partial_{x_3}^k N_j(\omega(s),\tilde
\omega(s))\}_3\|_{X(m)}{\,{\rm d}}s\\
\,&\le\, C \int_0^t \frac{e^{-(\eta+k)(t-s)}}{a(t{-}s)^{\frac{1}{p}
-\frac{1}{2}+\frac{|\beta|}{2}}}\|\partial_{x_3}^k N_j(\omega(s),
\tilde\omega(s))\|_{X^p(m)^3}{\,{\rm d}}s\\
\,&\le\, C \int_0^t \frac{e^{-(\eta+k)(t-s)}e^{-(k+2\eta)s}}{a(t{-}s)^{
\frac{1}{p}-\frac{1}{2}+\frac{|\beta|}{2}}a(s)^\frac{k}{2}}{\,{\rm d}}s
~\|\omega\|_{\mathbb{V}}\|\tilde\omega\|_{\mathbb{V}}\,\le\, \frac{Ce^{-(\eta+k)t}}{
a(t)^{\frac{1}{p}+\frac{|\beta|}{2} +\frac{k}{2}-\frac{3}{2}}}
\|\omega\|_{\mathbb{V}}\|\tilde\omega\|_{\mathbb{V}}~.\end{aligned}$$ Summarizing, we have shown that $\Phi$ defined by maps ${\mathbb{V}}$ into ${\mathbb{V}}$ and satisfies the following bounds: $$\begin{aligned}
\|\Phi(\omega)\|_{\mathbb{V}}\,&\le\, C_3\sum_{|\beta|\le 1}\|\partial_x^\beta
\omega_0\|_{{\mathbb{X}}(m)} + C_4 \|\omega\|_{\mathbb{V}}^2~, \\
\|\Phi(\omega)- \Phi(\tilde\omega)\|_{\mathbb{V}}\,&\le\, C_4(\|\omega\|_{\mathbb{V}}+\|\tilde\omega\|_{\mathbb{V}})\|\omega-\tilde\omega\|_{\mathbb{V}}~, \end{aligned}$$ for all $\omega, \tilde \omega \in {\mathbb{V}}$. If $K = 2C_3\sum_{|\beta|\le 1}
\|\partial_x^\beta\omega_0\|_{{\mathbb{X}}(m)}$ is sufficiently small, it follows that $\Phi$ is a strict contraction in the ball $\tilde B_K = \{\omega
\in {\mathbb{V}}\,|\, \|\omega\|_{\mathbb{V}}\le K\}$, hence has a unique fixed point there. Denoting by $\omega(t)$ the solution of given by Proposition \[prop.nonlinear.stability\], this implies that $t \mapsto \omega(t+T)$ belongs to $\tilde B_K$ if $T > 0$ is sufficiently large. In particular, $\omega(t)$ satisfies for some suitable $C > 0$. The proof of Theorem \[thm.main2\] is now complete. [$\Box$]{}
Appendix
========
Proof of Lemma \[lem.reduction\] {#subsec.lem.reduce}
--------------------------------
\[subsec.reduc\]
Let $\chi \in C_0^\infty({\mathbb{R}}^2)$ be a cut-off function such that $\chi(x_h) = 1$ if $|x_h|\le 1$ and $\chi(x_h) = 0$ if $|x_h| \ge 2$. Given $R > 0$, we denote $\chi_R(x_h) = \chi(x_h/R)$, so that $|\nabla_h\chi_R(x_h)| \le C/R$. For any $x_3 \in {\mathbb{R}}$, we define $$f(x_3) \,=\, \int_{{\mathbb{R}}^2}\tilde\omega_3(x_h,x_3){\,{\rm d}}x_h~, \qquad
f_R(x_3) \,=\, \int_{{\mathbb{R}}^2}\tilde\omega_3(x_h,x_3)\chi_R(x_h){\,{\rm d}}x_h~.$$ Since $\tilde\omega_3\in X(m)$ for some $m > 1$, it is easy to verify that $\|f-f_R\|_{L^\infty({\mathbb{R}})}\to 0$ as $R\to \infty$. On the other hand, for any test function $\psi \in C_0^\infty({\mathbb{R}})$, we have $$\label{eq.hReq}
\Bigl|\int_{{\mathbb{R}}}f(x_3)\frac{{{\rm d}}\psi}{{{\rm d}}x_3}(x_3) {\,{\rm d}}x_3\Bigr| \,\le\,
\Bigl|\int_{{\mathbb{R}}}f_R(x_3)\frac{{{\rm d}}\psi}{{{\rm d}}x_3}(x_3) {\,{\rm d}}x_3\Bigr|
+ \|f-f_R\|_{L^\infty ({\mathbb{R}})} \Bigl\|\frac{{{\rm d}}\psi}{{{\rm d}}x_3}\Bigr\|_{L^1({\mathbb{R}})}~.$$ The last term in the right-hand side converges to zero as $R \to
\infty$. To treat the other term, we observe that $$\int_{{\mathbb{R}}}f_R(x_3)\frac{{{\rm d}}\psi}{{{\rm d}}x_3}(x_3) {\,{\rm d}}x_3 \,=\,
\int_{{\mathbb{R}}^3}\tilde\omega_3(x_h,x_3)\chi_R(x_h)\frac{{{\rm d}}\psi}{{{\rm d}}x_3}
(x_3) {\,{\rm d}}x_3 \,=\, \langle \tilde \omega_3\,,\,\frac{\partial\phi_R}{
\partial x_3}\rangle~,$$ where $\phi_R(x_h,x_3) = \chi_R(x_h)\psi(x_3)$ and $\langle \cdot,\cdot
\rangle$ denotes the duality pairing of $\mathcal{D}'({\mathbb{R}}^3)$ and $C_0^\infty({\mathbb{R}}^3)$. Now, since $\nabla\cdot\tilde \omega = 0$ in the sense of distributions, we have $$\langle \tilde \omega_3\,,\,\frac{\partial\phi_R}{\partial x_3}
\rangle \,=\, -\langle \frac{\partial\tilde\omega_3}{\partial x_3}
\,,\, \phi_R\rangle \,=\, \langle \nabla_h \cdot\tilde \omega_h
\,,\, \phi_R\rangle \,=\, -\langle \tilde \omega_h
\,,\, \nabla_h \phi_R\rangle~,$$ so that $$\int_{{\mathbb{R}}}f_R(x_3)\frac{{{\rm d}}\psi}{{{\rm d}}x_3}(x_3){\,{\rm d}}x_3 \,=\,
-\int_{{\mathbb{R}}^3} \tilde\omega_h(x_h,x_3)\cdot \nabla_h\chi_R(x_h)
\psi(x_3) {\,{\rm d}}x_h {\,{\rm d}}x_3~.$$ Using the inclusion $L^2(m) \hookrightarrow L^1({\mathbb{R}}^2)$ and the definition of the space $X(m)$, we thus find $$\Bigl|\int_{{\mathbb{R}}}f_R(x_3)\frac{{{\rm d}}\psi}{{{\rm d}}x_3}(x_3){\,{\rm d}}x_3\Bigr|
\,\le\, \frac{C}{R}\,\|\tilde\omega_h\|_{X(m)^2}\|\psi\|_{L^1({\mathbb{R}})}
~\xrightarrow[R \to \infty]{}~ 0~.$$ Returning to , we conclude that the left-hand side vanishes for all $\psi \in C_0^\infty({\mathbb{R}})$, hence $\frac{{{\rm d}}f}{{{\rm d}}x_3}=0$ in the sense of distributions. Since $f \in BC({\mathbb{R}})$, it follows that $f$ is identically constant, which is the desired result. [$\Box$]{}
[**Remark.**]{} If $\omega(x,t)$ is any solution of that is integrable with respect to the horizontal variables, we can define $$\phi(x_3,t) \,=\, \int_{{\mathbb{R}}^2}\omega_3(x_h,x_3,t){\,{\rm d}}x_h~,
\qquad x_3 \in {\mathbb{R}}~, \quad t \ge 0~.$$ As was observed in [@GW2], this quantity satisfies a remarkably simple equation $$\label{eq.phi}
\partial_t \phi(x_3,t) + x_3\partial_{x_3}\phi(x_3,t) \,=\,
\partial_{x_3}^2\phi(x_3,t)~,$$ which can be solved explicitly. However, if $\omega(\cdot,t) \in
X(m)^3$ for some $m > 1$ with $\nabla\cdot\omega(\cdot,t) = 0$, Lemma \[lem.reduction\] shows that $\phi(x_3,t)$ does not depend on $x_3$, and then implies that $\phi(x_3,t)$ is also independent of $t$. Thus, as was already mentioned, we can restrict ourselves to the particular case where $\phi \equiv 0$ without loss of generality. Being unaware of this simple observation, the authors of [@GW2] have stated their stability result in a seemingly more general form, allowing (apparently) for nontrivial functions $\phi(x_3,t)$, but thanks to Lemma \[lem.reduction\] (which also holds in the slightly different functional setting of [@GW2]) the simpler presentation adopted here in Theorem \[thm.main1\] is exactly as general.
Proof of Proposition \[prop.discrete.spectrum\]. {#subsec.discrete}
------------------------------------------------
This final section is devoted to the proof of Proposition \[prop.discrete.spectrum\], which shows that eigenfunctions of ${\mathscr{L}}_{\alpha,h}$ corresponding to eigenvalues outside the essential spectrum have a Gaussian decay at infinity. For the nonlocal operator ${\mathscr{L}}_{\alpha,3}$, the same result was established in [@GW1 Lemma 4.5] using ODE techniques, but we prefer using here a more flexible method based on weighted $L^2$ estimates. In fact, we shall consider a more general elliptic problem of the form $$-{\mathcal{L}}f + F(x,f,\nabla f) + \lambda f \,=\, h~, \qquad x\in {\mathbb{R}}^n~,
\label{eq.prop.appendix.prop.reduction}$$ where the unknown is the vector-valued function $f = (f_1,\dots,f_N)^\top$. Here and below we denote by ${\mathcal{L}}= \Delta+\frac{x}{2}\cdot\nabla
+\frac{n}{2}$ the analog of operator in dimension $n$. The data of the problem are the functions $F : {\mathbb{R}}^n\times \mathbb{C}^N
\times \mathbb{C}^{nN} \to \mathbb{C}^N$ and $h : {\mathbb{R}}^n \to \mathbb{C}^N$, and the complex number $\lambda$.
For $m \in [0,\infty]$, we denote by $L^2(m)$, $H^1(m)$ the following complex Hilbert spaces on ${\mathbb{R}}^n$: $$\begin{aligned}
L^2(m) \,&=\,\Bigl\{f\in L^2({\mathbb{R}}^n,\mathbb{C})~\Big|~\int_{{\mathbb{R}}^n}|f(x)|^2
\rho_m(|x|^2){\,{\rm d}}x < \infty\Bigr\}~,\\
H^1(m) \,&=\, \Bigl\{f\in L^2(m)~\Big|~\partial_{x_j}f\in L^2(m)
\quad \hbox{for }j = 1,\dots,n\Bigr\}~,\end{aligned}$$ where $\rho_m$ is the weight function defined by . Our main result is:
\[prop.appendix.prop.reduction\] Let $m \in [0,\infty)$, $\lambda\in \mathbb{C}$, $h\in L^2(\infty)^N$, and assume that $F$ is a continuous function satisfying $$|F(x,p,Q)| \,\le\, A(x)|p| + B(x)|Q|~, \qquad \hbox{for all}~(x,p,Q)
\in {\mathbb{R}}^n\times \mathbb{C}^N\times \mathbb{C}^{nN}~,
\label{con.prop.appendix.prop.reduction2}$$ where $A$ and $B$ are bounded, nonnegative functions such that $$\lim_{R\to \infty}\sup_{|x|\ge R} A(x) \,=\, \lim_{R\to
\infty}\sup_{|x|\ge R} B(x) \,=\, 0~.
\label{con.prop.appendix.prop.reduction}$$ If ${\mathop{\mathrm{Re}}}\lambda > \frac{n}{4}-\frac{m}{2}$, then any solution $f\in
H^1(m)^N$ of satisfies $f\in H^1(\infty)^N$.
[**Proof.**]{} The proof is a simple modification of [@KM Proposition 12], which in turn is inspired by a recent work of Fukuizumi and Ozawa [@FO] where decay estimates are obtained for solutions of the Haraux-Weissler equation. For $k\ge 1$, $\epsilon>0$, and $\theta
\in [0,m]$, we define the weight functions $$\xi_{k,\epsilon}(x) \,=\, e^{\frac{(1-\epsilon)k|x|^2}{4k+|x|^2}}~, \qquad
\zeta_{\theta}(x) \,=\, (1+|x|^2)^\theta~, \qquad x \in {\mathbb{R}}^n~.$$ Multiplying both sides of by $\zeta_\theta\xi_{k,\epsilon} \bar f$ and integrating by parts the real part of the resulting expression, we obtain the identity $$\begin{aligned}
\int_{{\mathbb{R}}^n} \zeta_\theta \xi_{k,\epsilon} &|\nabla f|^2 {\,{\rm d}}x + {\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n} \bar f \cdot (\nabla ( \zeta_\theta \xi_{k,\epsilon} ),
\nabla )f {\,{\rm d}}x + \int_{{\mathbb{R}}^n}|f|^2 \frac{x}{4} \cdot \nabla
(\zeta_\theta \xi_{k,\epsilon} ) {\,{\rm d}}x \nonumber \\
\,&=\,-{\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n} \zeta_\theta \xi_{k,\epsilon} \bar f \cdot
F(x,f(x),\nabla f(x)) {\,{\rm d}}x+ \Bigl(\frac{n}{4} -{\mathop{\mathrm{Re}}}\lambda
\Bigr)\int_{{\mathbb{R}}^n}\zeta_\theta \xi_{k,\epsilon} |f|^2 {\,{\rm d}}x
\label{eq.prop.appendix.prop.reduction2}\\
&\quad~ + {\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n}\zeta_\theta \xi_{k,\epsilon}\bar f\cdot h
{\,{\rm d}}x~.\nonumber\end{aligned}$$ Clearly, $$\label{eq.nablaxi}
\nabla\xi_{k,\epsilon}(x) \,=\, \frac{8(1-\epsilon)k^2x}{(4k+|x|^2)^2}
\,\xi_{k,\epsilon}(x)~, \qquad
\nabla\zeta_\theta(x) \,=\, \frac{2\theta x}{1+|x|^2}\,\zeta_\theta(x)~.$$ Thus, the second term in the left-hand side of can be written in the following way: $$\begin{aligned}
{\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n} &\bar f \cdot (\xi_{k,\epsilon} \nabla\zeta_\theta,
\nabla)f {\,{\rm d}}x + {\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n} \bar f \cdot (\zeta_\theta\nabla
\xi_{k,\epsilon},\nabla)f {\,{\rm d}}x\\
\,&=\, -\int_{{\mathbb{R}}^n} |f|^2\,\nabla\cdot\Bigl(\frac{\theta x\zeta_\theta
\xi_{k,\epsilon}}{1+|x|^2}\Bigr){\,{\rm d}}x + {\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n} \bar f \cdot
(\zeta_\theta\nabla \xi_{k,\epsilon},\nabla)f {\,{\rm d}}x\\
\,&=\, -\int_{{\mathbb{R}}^n} |f|^2\,\xi_{k,\epsilon}\,x \cdot\nabla \frac{\theta
\zeta_\theta }{1+|x|^2} {\,{\rm d}}x - \int_{{\mathbb{R}}^n}|f|^2\frac{\theta
\zeta_\theta }{1+|x|^2}\,x \cdot\nabla \xi_{k,\epsilon} {\,{\rm d}}x \\
&\quad~ -n\theta\int_{{\mathbb{R}}^n} \frac{\zeta_\theta \xi_{k,\epsilon}}{1+|x|^2}
|f|^2 {\,{\rm d}}x + {\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n} \frac{8(1-\epsilon) k^2\zeta_\theta
\xi_{k,\epsilon}}{(4k+|x|^2)^2}\,\bar f \cdot (x, \nabla) f {\,{\rm d}}x~.\end{aligned}$$ To bound this quantity from below, we observe that $$\int_{{\mathbb{R}}^n} |f|^2\,\xi_{k,\epsilon} x \cdot \nabla \frac{\theta
\zeta_\theta }{1+|x|^2} {\,{\rm d}}x\le 2\theta^2\int_{{\mathbb{R}}^n} \frac{\zeta_\theta
\xi_{k,\epsilon}}{1+|x|^2}|f|^2 {\,{\rm d}}x~.$$ Moreover, for each $\eta_1>0$, $$\begin{aligned}
- {\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n} &\frac{8(1-\epsilon) k^2\zeta_\theta \xi_{k,\epsilon}}
{(4k+|x|^2)^2}\,\bar f \cdot (x, \nabla) f {\,{\rm d}}x \,\le\,
\int_{{\mathbb{R}}^n} \frac{2(1-\epsilon)k \zeta_\theta \xi_{k,\epsilon}}{4k+|x|^2}
|xf| |\nabla f| {\,{\rm d}}x \\
\,&\le\, (1-\eta_1) \int_{{\mathbb{R}}^n} \zeta_\theta \xi_{k,\epsilon} |\nabla f|^2
{\,{\rm d}}x + \frac{(1-\epsilon)^2}{1-\eta_1} \int_{{\mathbb{R}}^n} \frac{ k^2 \zeta_\theta
\xi_{k,\epsilon}|xf|^2 }{(4k+|x|^2)^2} {\,{\rm d}}x~.\end{aligned}$$ Thus, using the expression of $\nabla \xi_{k,\epsilon}$, we find $$\begin{aligned}
{\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n} \bar f \cdot (\nabla (\zeta_\theta \xi_{k,\epsilon}),
\nabla)f &{\,{\rm d}}x
\,\ge\, -C \int_{{\mathbb{R}}^n} \frac{\zeta_\theta \xi_{k,\epsilon}}{1+|x|^2}
|f|^2 {\,{\rm d}}x - \int_{{\mathbb{R}}^n} \frac{8(1-\epsilon)\theta k^2 \zeta_\theta
\xi_{k,\epsilon} |xf|^2}{(4k+|x|^2)^2(1+|x|^2)}{\,{\rm d}}x
\label{eq.prop.appendix.prop.reduction3}\\
\,&-\, (1-\eta_1) \int_{{\mathbb{R}}^n} \zeta_\theta \xi_{k,\epsilon} |\nabla f |^2
dx - \frac{(1-\epsilon)^2}{1-\eta_1} \int_{{\mathbb{R}}^n} \frac{ k^2 \zeta_\theta
\xi_{k,\epsilon}|xf|^2 }{(4k+|x|^2)^2} {\,{\rm d}}x~, \nonumber\end{aligned}$$ where $C = n\theta+\theta^2$ does not depend on $k$ and $\epsilon$. We next consider the third term in the left-hand side of , which satisfies $$\int_{{\mathbb{R}}^n}|f|^2 \frac{x}{4}\cdot\nabla (\zeta_\theta\xi_{k,\epsilon})
{\,{\rm d}}x \,=\, \frac{\theta}{2} \int_{{\mathbb{R}}^n} \frac{\zeta_\theta
\xi_{k,\epsilon}}{1+|x|^2}\,|xf|^2{\,{\rm d}}x + 2(1-\epsilon)\int_{{\mathbb{R}}^n}
\frac{k^2 \zeta_\theta \xi_{k,\epsilon} |xf|^2}{(4k+|x|^2)^2} {\,{\rm d}}x~.
\label{eq.prop.appendix.prop.reduction4}$$ To estimate the right-hand side of , we use and obtain, for each $\eta_2>0$, $$\begin{aligned}
- {\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n}\zeta_\theta \xi_{k,\epsilon} \bar f \cdot F(x,f(x),
&\nabla f(x)){\,{\rm d}}x \,\le\, \int_{{\mathbb{R}}^n}\zeta_{\theta}\xi_{k,\epsilon}
A|f|^2 {\,{\rm d}}x + \int_{{\mathbb{R}}^n}\zeta_{\theta}\xi_{k,\epsilon}B|f||\nabla f|
{\,{\rm d}}x \nonumber\\
\,&\le\, \int_{{\mathbb{R}}^n} \zeta_{\theta} \xi_{k,\epsilon}\Bigl(A+\frac{B^2}
{4\eta_2}\Bigr)|f|^2 {\,{\rm d}}x + \eta_2 \int_{{\mathbb{R}}^n} \zeta_{\theta}\xi_{k,\epsilon}
|\nabla f|^2 {\,{\rm d}}x~. \label{eq.prop.appendix.prop.reduction5} \end{aligned}$$ Finally, for each $\eta_3>0$, we have $${\mathop{\mathrm{Re}}}\int_{{\mathbb{R}}^n}\zeta_\theta \xi_{k,\epsilon} \bar f\cdot h {\,{\rm d}}x \,\le\,
\eta_3 \int_{{\mathbb{R}}^n}\zeta_\theta \xi_{k,\epsilon}|f|^2 {\,{\rm d}}x +
\frac{1}{4\eta_3} \int_{{\mathbb{R}}^n}\zeta_\theta \xi_{k,\epsilon}|h|^2 {\,{\rm d}}x~.
\label{eq.prop.appendix.prop.reduction6}$$ Substituting – into , we arrive at our basic inequality: $$\begin{aligned}
(\eta_1-&\eta_2)\int_{{\mathbb{R}}^n} \zeta_{\theta} \xi_{k,\epsilon} |\nabla f|^2
{\,{\rm d}}x + \int_{{\mathbb{R}}^n} \frac{(1-\epsilon) k^2
\zeta_{\theta} \xi_{k,\epsilon} |x f|^2}{(4k+|x|^2)^2} \Bigl(
\frac{1-2\eta_1+\epsilon}{1-\eta_1} - \frac{8 \theta}{1+|x|^2}\Bigr)
{\,{\rm d}}x \nonumber \\
\,&\le\, \int_{{\mathbb{R}}^n} \zeta_{\theta} \xi_{k,\epsilon}
\Bigl\{\Bigl(\frac{C}{1+|x|^2} + \frac{n}{4} - {\mathop{\mathrm{Re}}}\lambda + A + \frac{B^2}
{4\eta_2} +\eta_3 -\frac{\theta}{2}\Bigr) |f|^2 + \frac{1}{4\eta_3}
|h|^2 \Bigr\} {\,{\rm d}}x~. \label{ineq.appendix.prop.reduction}\end{aligned}$$
To exploit , we first take $\eta_1=\eta_2=\frac{1}{2}$ and $\theta=m$. Using and the assumption that ${\mathop{\mathrm{Re}}}\lambda > \frac{n}{4}-\frac{m}{2}$, we see that there exists $R > 0$ independent of $k\ge 1$ such that, if $\eta_3>0$ is sufficiently small, the following inequality holds: $$\epsilon(1-\epsilon)\int_{{\mathbb{R}}^n} \frac{ k^2 \zeta_{\theta}
\xi_{k,\epsilon} |xf|^2}{(4k+|x|^2)^2} {\,{\rm d}}x \le C \int_{|x|\le R}
\zeta_{\theta}\xi_{k,\epsilon} |f|^2 {\,{\rm d}}x +\frac{1}{4\eta_3}\int_{{\mathbb{R}}^n}
\zeta_{\theta}\xi_{k,\epsilon} |h|^2 {\,{\rm d}}x~,$$ where the constant $C > 0$ is independent of $k\ge 1$. Thus, taking the limit $k \to \infty$ and using Fatou’s lemma, we obtain $$\frac{\epsilon(1{-}\epsilon)}{16}\int_{{\mathbb{R}}^n} (1+|x|^2)^m e^{\frac{1-
\epsilon}{4}|x|^2} |xf|^2 {\,{\rm d}}x \,\le\, C(R) \int_{|x|\le R} |f|^2 {\,{\rm d}}x
+ \frac{1}{4\eta_3}\int_{{\mathbb{R}}^n} (1+|x|^2)^m e^{\frac{1-\epsilon}{4}|x|^2}
|h|^2 {\,{\rm d}}x~,$$ which shows that $e^{\frac{1-\epsilon}{8}|x|^2}f \in L^2({\mathbb{R}}^2)$ for any $\epsilon>0$. Next we choose $\eta_1=\frac{1}{4}$, $\eta_2=\frac{1}{8}$, $\eta_3=1$, and $\theta=0$ in . Taking again the limit $k \to \infty$ and using Lebesgue’s dominated convergence theorem, we find $$\frac{1}{8} \int_{{\mathbb{R}}^n} e^{\frac{1-\epsilon}{4}|x|^2}|\nabla f|^2 {\,{\rm d}}x +
\frac{1{-}\epsilon}{24} \int_{{\mathbb{R}}^n} e^{\frac{1-\epsilon}{4}|x|^2} |xf|^2
{\,{\rm d}}x \le C \int_{{\mathbb{R}}^n} e^{\frac{1-\epsilon}{4}|x|^2} |f|^2 {\,{\rm d}}x
+\frac{1}{4}\int_{{\mathbb{R}}^n} e^{\frac{1-\epsilon}{4}|x|^2} |h|^2 {\,{\rm d}}x~,$$ where the constant $C > 0$ does not depend on $\epsilon > 0$. This inequality shows that $$\frac{1}{8}\int_{{\mathbb{R}}^n} \!e^{\frac{1-\epsilon}{4}|x|^2}|\nabla f|^2 {\,{\rm d}}x +
\frac{1{-}\epsilon}{48} \int_{{\mathbb{R}}^n} \!e^{\frac{1-\epsilon}{4}|x|^2} |xf|^2
{\,{\rm d}}x \,\le\, C \int_{|x|\le R'} \!e^{\frac{1-\epsilon}{4}|x|^2}|f|^2
{\,{\rm d}}x + \frac{1}{4}\int_{{\mathbb{R}}^n} \!e^{\frac{1-\epsilon}{4}|x|^2}|h|^2 {\,{\rm d}}x\,,$$ for some $R'>0$ independent of $\epsilon > 0$. Taking now the limit $\epsilon\to 0$, we conclude that $f \in H^1(\infty)$, which is the desired result. [$\Box$]{}
[**Proof of Proposition \[prop.discrete.spectrum\].**]{} We consider the eigenvalue equation , which can be written in the form $$\begin{aligned}
-{\mathcal{L}}_h\omega_h + \alpha \Lambda_1\omega_h - \alpha\tilde\Lambda_2
\omega_h +\Bigl(\lambda +\frac{3}{2}\Bigr)\omega_h \,=\, 0~,
\label{eq.prop.discrete.spectrum.1}\end{aligned}$$ where ${\mathcal{L}}_h$ is given by and the operators $\Lambda_1$, $\tilde \Lambda_2$ are defined at the beginning of Section \[sec.vectorial\]. We recall that $|\Lambda_1 \omega_h| \le
|U^G_h| |\nabla_h \omega_h|$ and $|\tilde \Lambda_2 \omega_h| \le
|\nabla_h U^G_h| |\omega_h|$, where the velocity profile $U^G_h$ satisfies . Assume that ${\mathop{\mathrm{Re}}}\lambda > -\frac{m}{2}-1$ and let $\omega_h \in H^1(m)^2$ be a solution to . Applying Proposition \[prop.appendix.prop.reduction\] with $n=N=2$, $F(x,f,\nabla f) = \alpha \Lambda_1 f - \alpha\tilde\Lambda_2 f$, and $h=0$, we obtain $\omega_h \in H^1(\infty)^2$. This completes the proof of Proposition \[prop.discrete.spectrum\]. [$\Box$]{}
[99]{}
J. M. Burgers, A mathematical model illustrating the theory of turbulence, Adv. Appl. Mech. [**1**]{} (1948), 171–199.
A. Carpio, Asymptotic behavior for the vorticity equations in dimensions two and three, Commun. in PDE [**19**]{} (1994), 827–872.
D. G. Crowdy, A note on the linear stability of Burgers vortex, Stud. Appl. Math. [**100**]{} (1998), 107–126.
G.-H. Cottet, quations de Navier-Stokes dans le plan avec tourbillon initial mesure, C. R. Acad. Sci. Paris S[é]{}r. I Math. [**303**]{} (1986), 105–108.
K.-J. Engel and R. Nagel, One-Parameter semigroups for linear evolution equations, Graduate Texts in Mathematics, Springer, 2000.
R. Fukuizumi and T. Ozawa, On a decay property of solutions to the Haraux-Weissler equation, J. Diff. Equations [**221**]{} (2006), 134–142.
J. D. Gibbon, A. S. Fokas, and C. R. Doering, Dynamically stretched vortices as solutions of the 3D Navier-Stokes equations, Physica D [**132**]{} (1999), 497–510.
J. Jim[é]{}nez, H. K. Moffatt, C. Vasco, The structure of the vortices in freely decaying two-dimensional turbulence, J. Fluid Mech. [**313**]{} (1996), 209–222.
I. Gallagher, Th. Gallay, and F. Nier, Spectral asymptotics for large skew-symmetric perturbations of the harmonic oscillator, Int. Math. Res. Notices [**2009**]{} (2009), 2147–2199.
Th. Gallay and C. E. Wayne, Invariant manifold and the long-time asymptotics of the Navier-Stokes and vorticity equations on ${\bf R}^2$, Arch. Rational Mech. Anal. [**163**]{} (2002), 209–258.
Th. Gallay and C. E. Wayne, Global Stability of vortex solutions of the two dimensional Navier-Stokes equation, Comm. Math. Phys. [**255**]{} (2005), 97–129.
Th. Gallay and C. E. Wayne, Three-dimensional stability of Burgers vortices : the low Reynolds number case, Phys. D [**213**]{} (2006), 164–180.
Th. Gallay and C. E. Wayne, Existence and stability of asymmetric Burgers vortices, J. Math. Fluid Mech. [**9**]{} (2007), 243–261.
Y. Giga and M.-H. Giga, , Kyoritsu 1999 (in Japanese), English version to be published by Birkhäuser.
Y. Giga and T. Kambe, Large time behavior of the vorticity of two dimensional viscous flow and its application to vortex formation, Comm. Math. Phys. [**117**]{}, (1988) 549–568.
Y. Kagei and Y. Maekawa, On asymptotic behaviors of solutions to parabolic systems modelling chemotaxis, to appear.
T. Kato, Perturbation Theory for Linear Operators, Springer, 1966.
S. Kida and K. Ohkitani, Spatiotemporal intermittency and instability of a forced turbulence, Phys. Fluids A. [**4**]{} (1992), 1018–1027.
S. Leibovich and Ph. Holmes, Global stability of the Burgers vortex, Phys. Fluids [**24**]{} (1981), 548–549.
T. S. Lundgren, Strained spiral vortex model for turbulent fine structure, Phys. Fluids [**25**]{} (1982), 2193–2203.
H. K. Moffatt, S. Kida and K. Ohkitani, Stretched vortices-the sinews of turbulence; large-Reynolds-number asymptotics, J. Fluid Mech. [**259**]{} (1994), 241–264.
Y. Maekawa, On the existence of Burgers vortices for high Reynolds numbers, J. Math. Anal. Appl. [**349**]{} (2009), 181–200.
Y. Maekawa, Existence of asymmetric Burgers vortices and their asymptotic behavior at large circulations, Math. Model Methods Appl. Sci. [**19**]{}, (2009) 669–705.
Y. Maekawa, Spectral properties of the linearization at the Burgers vortex in the high rotation limit, to appear in J. Math. Fluid Mech.
A. Prochazka and D. I. Pullin, On the two-dimensional stability of the axisymmetric Burgers vortex, Phys. Fluids. [**7**]{} (1995), 1788–1790.
A. Prochazka and D. I. Pullin, Structure and stability of non-symmetric Burgers vortices, J. Fluid Mech. [**363**]{} (1998), 199–228.
A. C. Robinson and P. G. Saffman, Stability and Structure of stretched vortices, Stud. Appl. Math. [**70**]{} (1984), 163–181.
M. Rossi and S. Le Diz[è]{}s, Three-dimensional temporal spectrum of stretched vortices, Phys. Rev. Lett. [**78**]{} (1997), 2567–2569.
P. J. Schmid and M. Rossi, Three-dimensional stability of a Burgers vortex, J. Fluid Mech. [**500**]{} (2004), 103–112.
A. A. Townsend, On the fine-scale structure of turbulence, Proc. R. Soc. A [**208**]{} (1951), 534–542.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
,A. Al-Haydari$^a$, V. M. Braun$^b$, S. Collins$^b$, M. Göckeler$^b$, G. N. Lacagnina$^c$, M. Panero$^{b, d}$, A. Schäfer$^b$, and G. Schierholz$^{b, e}$\
$^a$ Department of Physics, Faculty of Science, Taiz University, Taiz, Yemen Republic\
$^b$ Institute for Theoretical Physics, University of Regensburg, 93040 Regensburg, Germany\
$^c$ INFN, Sezione di Milano, 20133 Milano, Italy\
$^d$ Institute for Theoretical Physics, ETH Zürich, 8093 Zürich, Switzerland\
$^e$ Deutsches Elektronen-Synchrotron DESY, 22603 Hamburg, Germany
title: 'Matrix elements of heavy-light mesons from a fine lattice'
---
The details of our calculation of the matrix elements and the results for the form factors at $q^2 = 0$ are given in Ref. [@alhaydari].
![The solid lines denote our results for the form factors ($f_+$: upper line, $f_0$: lower line) of the decay $B \rightarrow \pi l \nu $. The dashed lines denote our error bounds. The squares denote $f_+$, circles denote $f_0$ from other recent lattice calculations (red: quenched, [@Abada:2000ty], green: $N_f = 2+1$, [@Bailey:2008wp], magenta: $N_f = 2+1$, [@Dalgic:2006dt])](poster_compare_B_to_pi_extrapolated_curves.eps){width=".60\textwidth"}
.
\[fig:decays\]
We observe a relatively good agreement of our form factors with other lattice calculations.
Our results for the charmonium spectrum are in very good agreement with a previous quenched calculation at the same lattice spacing [@Choe:2003wx] and in agreement with a recent calculation with two heavy flavors [@Ehmann:2007hj]. Since we work on fine lattices where discretization effects are under good control for charmonia it is also of interest to calculate the charmonium decay constants. Our result for the decay constant of the $J/\psi$ is lower than the result of 399(4) MeV from quenched anisotropic lattices of [@Dudek:2006ej] and of 413(40) MeV from lattices with $N_f = 2$ of [@Dimopoulos:2008ee]. The experimental value is 411(7) MeV.
[**Acknowledgement:**]{} We thank J. Simone for useful discussions.
[99]{} QCDSF collaboration, A. Al-Haydari [*et al.*]{}, in preparation. A. Abada [*et al.*]{}, Nucl. Phys. B [**619**]{}, 565 (2001) \[arXiv:hep-lat/0011065\]. J. Bailey [*et al.*]{}, arXiv:0811.3640 \[hep-lat\]. E. Dalgic [*et al.*]{}, Phys. Rev. D [**73**]{} (2006) 074502 \[Erratum-ibid. D [**75**]{} (2007) 119906\] \[arXiv:hep-lat/0601021\]. S. Choe [*et al.*]{} \[QCD-TARO Collaboration\], JHEP [**0308**]{}, 022 (2003) \[arXiv:hep-lat/0307004\]. C. Ehmann and G. Bali, PoS [**LAT2007**]{}, 094 (2007) \[arXiv:0710.0256 \[hep-lat\]\].
J. J. Dudek, R. G. Edwards and D. G. Richards, Phys. Rev. D [**73**]{}, 074507 (2006) \[arXiv:hep-ph/0601137\]. P. Dimopoulos, C. McNeile, C. Michael, S. Simula and C. Urbach \[ETM Collaboration\], arXiv:0810.1220 \[hep-lat\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Many-body localisation in interacting quantum systems can be cast as a disordered hopping problem on the underlying Fock-space graph. A crucial feature of the effective Fock-space disorder is that the Fock-space site energies are strongly correlated – maximally so for sites separated by a finite distance on the graph. Motivated by this, and to understand the effect of such correlations more fundamentally, we study Anderson localisation on Cayley trees and random regular graphs, with maximally correlated disorder. Since such correlations suppress short distance fluctuations in the disorder potential, one might naively suppose they disfavour localisation. We find however that there exists an Anderson transition, and indeed that localisation is more robust in the sense that the critical disorder scales with graph connectivity $K$ as $\sqrt{K}$, in marked contrast to $K\ln K$ in the uncorrelated case. This scaling is argued to be intimately connected to the stability of many-body localisation. Our analysis centres on an exact recursive formulation for the local propagators as well as a self-consistent mean-field theory; with results corroborated using exact diagonalisation.'
author:
- Sthitadhi Roy
- 'David E. Logan'
bibliography:
- 'refs.bib'
title: '[Localisation on certain graphs with strongly correlated disorder]{}'
---
Disorder-induced localisation of non-interacting quantum particles – the phenomenon of Anderson localisation (AL) – has been one of the most profound discoveries in physics [@anderson1958absence]. Its robustness to interactions in quantum many-body systems has lately been a major research theme, under the banner of many-body localisation (MBL) [@gornyi2005interacting; @basko2006metal; @oganesyan2007localisation; @pal2010many] (see Refs. [@nandkishore2015many; @alet2018many; @abanin2019colloquium] for reviews and further references). MBL systems fall outside the paradigm of conventional statistical mechanics allowing for novel quantum phases, and are thus of fundamental interest.
Efforts to understand the MBL phase and the accompanying MBL transition have ranged from extensive numerical studies [@kjall2014many; @luitz2015many; @alet2018many] and phenomenological treatments [@vosk2015theory; @potter2015universal; @goremykina2019analytically; @dumitrescu2018kosterlitz; @morningstar2019renormalization; @morningstar2020manybody] to studying the problem directly on the Fock space [@logan1990quantum; @altshuler1997quasiparticle; @MonthusGarel2010PRB; @pietracaprina2016forward; @logan2019many; @roy2018exact; @roy2018percolation; @roy2019self; @pietracaprina2019hilbert; @ghosh2019manybody; @roy2020fock]. One virtue of the latter is that the problem can be cast as a disordered hopping problem on the Fock-space graph, thus offering the prospect of exploiting techniques and understandings developed for AL. However, MBL on Fock space is fundamentally different from conventional AL on high-dimensional graphs, due to the presence of *maximal correlations* in the effective Fock-space disorder: the statistical correlation between two Fock-space site energies, scaled by their variance, approaches its maximum value of unity in the thermodynamic limit, for any pair separated by a finite Hamming distance on the Fock-space graph. This was found to be a necessary condition for MBL to exist [@roy2020fock].
Motivated by this, here we ask a fundamental question: what is the fate of AL on random graphs with maximally correlated disorder? In parallel to the case of Fock-space disorder, the correlation between the disordered site energies of any two sites separated by a finite distance on the graph takes it maximum value in the thermodynamic limit. In suppressing fluctuations in the site-energies, one might naively suppose these correlations would strongly favour delocalisation; indeed it is not *a priori* obvious that a localised phase must exist in such a case. Nevertheless, not only do we find inexorably a localised phase and an Anderson transition, but also that the scaling of the critical disorder with graph connectivity is qualitatively different to that for the standard model with uncorrelated disorder. These models thus introduce a novel class of AL problems with intimate connections to the problem of MBL on Fock space.
Concretely, we consider a disordered tight-binding model on a rooted Cayley tree (as well as on random regular graphs (RRG) which are locally tree-like). For uncorrelated disorder, such models have served as archetypes for studying a range of phenomena such as localisation transitions, multifractality, and glassy dynamics on complex high-dimensional graphs [@abou-chacra1973self; @chalker1990anderson; @luca2014anderson; @altshuler2016multifractal; @tikhonov2016anderson; @garciamata2017scaling; @sonner2017multifractality; @biroli2018delocalization; @kravtsov2018nonergodic; @tikhonov2019critical; @savitz2019anderson; @garciamata2020two; @tarzia2020manybody; @biroli2017delocalized; @biroli2020anomalous; @detomasi2020subdiffusion]. The model Hamiltonian is $$H = \Gamma\sum_{\braket{i,j}}[\ket{i}\bra{j}+\mathrm{h.c.}]+W\sum_{i}\epsilon_i\ket{i}\bra{i}
\label{eq:ham}$$ in the position basis $\{\ket{i}\}$, where $\braket{i,j}$ denotes a sum over nearest neighbour pairs. We denote the branching number of the tree by $K$ and the total number of generations in a finite-sized tree by $L$; the total number of sites in the tree is $N\sim K^L$. The set of correlated random site-energies, $\{\epsilon_i\}$, is fully specified by a $N$-dimensional joint distribution. To mimic the case of many-body systems on Fock space [@welsh2018simple; @logan2019many; @roy2020fock], we take these distributions to be multivariate Gaussians, $\mathcal{N}(\bm{0},\mathbf{C})$, characterised completely by the covariance matrix $\mathbf{C}$ [@roy2020fock]. Taking a cue from disordered interacting local Hamiltonians, we consider the matrix elements $C_{ij}$ to depend only on the distance $\ell_{ij}$ between a pair of sites. To impose the maximally correlated limit, we consider $$C_{ij}^{{\phantom\dagger}} = \braket{\epsilon_i^{{\phantom\dagger}}\epsilon_j^{{\phantom\dagger}}}=f(\ell_{ij}/L);~~~\lim_{x\to0}f(x)=1.
\label{eq:maxcorr}$$ The functional form of $f$ does not qualitatively affect our results, but for concreteness in numerical calculations we take $C_{ij} = \exp[-\ell_{ij}/\lambda L]$ with $\lambda=1$ [^1]. The choice of the argument of $f$ is motivated by the form of correlations in the Fock-space disorder of disordered many-body systems; for $p$-local Hamiltonians the analogous $f$ was shown to be a $p^\mathrm{th}$-order polynomial of $\ell_{ij}/\ln N_\mathcal{H}$, $N_\mathcal{H}$ being the Fock-space dimension [@roy2020fock].
{width="\linewidth"}
Our analysis centres on the local Feenberg self-energy $S_i(\omega)\equiv X_i(\omega)-i\Delta_i(\omega)$, defined via the local propagator as $G_i(\omega)=[\omega^+ - \epsilon_i - S_i(\omega)]^{-1}$ with $\omega^+=\omega+i\eta$ ($\eta =0^{+}$). We focus on the imaginary part of the self-energy, $\Delta_i(\omega)$, as it serves as a probabilistic order parameter for a localisation transition. Physically, $\Delta_i(\omega)$ gives the rate of loss of probability from site $i$ into states of energy $\omega$. In a delocalised phase $\Delta_i(\omega)$ is finite, whereas in a localised phase it vanishes $\propto\eta$ (with $y_{i}(\omega)=\Delta_{i}(\omega)/\eta$ finite), both with unit probability. These characteristics of $\Delta_i(\omega)$ have long been used successfully to understand Anderson transitions [@anderson1958absence; @economous1972existence; @abou-chacra1973self; @ThoulessReview1974; @Licciardello+Economou1975; @logan1985anderson; @*DELPGWPRB1987]; and, more recently, MBL transitions on Fock space [@logan2019many; @roy2019self; @roy2020fock].
We focus on the self-energy of the root site ($i=0$) of the rooted Cayley tree. $S_{0}(\omega)$ is given exactly by $$S_0^{{\phantom\dagger}}(\omega) = \Gamma^2\sum_{i_1\in{\mathsf{N}}[0]}[\omega^+-W\epsilon_{i_1}^{{\phantom\dagger}}-S_{i_1}^{(0)}]^{-1},
\label{eq:S0def}$$ with the sum over all sites in the first generation, and $S_{i_1}^{(0)}$ the self-energy of site $i_1$ with the root site removed. One could in principle now approximate the self-energy on the right-hand side of Eq. by a typical $S_{\mathrm{typ}}$, and obtain the distribution of $S_0$ self-consistently [@logan2019many; @roy2019self; @roy2020fock]. Here however we go far beyond such a treatment, addressing Eq. to arbitrarily high orders via an exact recursive method. We first sketch the formulation, focussing on the localised phase, in particular its stability and self-consistency; whence the quantity of interest is $y_0(\omega)$.
From Eq. , $y_0^{{\phantom\dagger}}(\omega)$ can be expressed as $$y_0^{{\phantom\dagger}} = \sum_{i_1\in{\mathsf{N}}[0]}\frac{\Gamma^2}{{\Omega}_{i_1}^2}\left[1+y_{i_1}^{(0)}\right];~{\Omega}_{i_1}^{{\phantom\dagger}} = \omega-W\epsilon_{i_1}^{{\phantom\dagger}}-X_{i_1}^{(0)}.
\label{eq:y0}$$ This is a recursion relation, which can be iterated as $$y_0^{{\phantom\dagger}} = \sum_{i_1\in{\mathsf{N}}[0]}\frac{\Gamma^2}{{\Omega}_{i_1}^2}\left[1+\sum_{i_2\in{\mathsf{N}}[i_1]}\frac{\Gamma^2}{{\Omega}_{i_2}^2}\left[1+\sum_{i_3\in{\mathsf{N}}[i_2]}\frac{\Gamma^2}{{\Omega}_{i_3}^2}[1+\cdots\right.\right..
\label{eq:y0-recursion}$$ In Eq. , for any site $i_n$ on generation $n$ of the tree, ${\Omega}_{i_n}=\omega-W\epsilon_{i_n}-X_{i_n}^{(i_{n-1})}(\omega)$, with $X_{i_n}^{(i_{n-1})}$ the real part of the self-energy of site $i_n$ with its (unique) neighbour $i_{n-1}$ on the previous generation removed. As for the imaginary part of the self-energy, a recursion relation for the real part can also be derived from Eq. . This leads to a recursion relation for ${\Omega}_{i_n}$, $${\Omega}_{i_n}^{{\phantom\dagger}} = \omega-W\epsilon_{i_n}^{{\phantom\dagger}}-\sum_{i_{n+1}\in{\mathsf{N}}[i_{n}]}\frac{\Gamma^2~~~}{{\Omega}_{i_{n+1}}^2},
\label{eq:ve-recursion}$$ with the boundary condition ${\Omega}_{i_L}=\omega-W\epsilon_{i_L}$ for a tree with $L$ generations. Eqs. , comprise the complete set of recursion relations required to compute $y_0(\omega)$ to all orders. We now make key conceptual points about the stability of the localised phase or lack thereof, and describe our results.
Note that by evaluating $y_0(\omega)$ using Eq. for many disorder realisations, one can generate its entire distribution $P_{y_0}$, and also compute its typical value via $\ln {y_{0,\mathrm{typ}}^{{\phantom\dagger}}}= \int dy_0~P_{y_0}(y_0)\ln y_0$. A stable localised phase is indicated by ${y_{0,\mathrm{typ}}^{{\phantom\dagger}}}$ taking a finite value independent of system size; whereas the delocalised phase is identified via a systematic growth of ${y_{0,\mathrm{typ}}^{{\phantom\dagger}}}$ with system size, such that it diverges in the thermodynamic limit. The disorder strength separating these two behaviours, if present, is the critical disorder. Numerical results for the localisation phase digaram so obtained for a $K=2$ Cayley tree with maximally correlated disorder are shown in Fig. \[fig:recursive-tree\]. Considering the band centre $\omega=0$ as an example (panel (a)), $\ln {y_{0,\mathrm{typ}}^{{\phantom\dagger}}}$ is independent of $L$ for $W>W_c$ whereas it diverges with $L$ for $W<W_c$; thus showing that a localisation transition is indeed present in the model. The phase diagram similarly obtained in the entire $\omega$-$W$ plane is given in Fig. \[fig:recursive-tree\](b), which shows the presence of mobility edges in the spectrum. Finally, Fig. \[fig:recursive-tree\](c), the distribution of $y_0$ is shown for a representative disorder in the localised phase, and shows excellent agreement with a Lévy distribution characteristic of a localised phase, $P_{y_0}(y_0) = \sqrt{\kappa/\pi}~y_0^{-3/2}e^{-\kappa/y_0}$ with scale parameter $\kappa$.
The stability of the localised phase can also be understood as the convergence of the recursion relation in Eq. . The series for $y_0$ can be organised as $$y_0^{{\phantom\dagger}} = \sum_{l=1}^\infty \phi_l^{{\phantom\dagger}}; ~\phi_l^{{\phantom\dagger}}=
\sum_{i_1\in{\mathsf{N}}[0]}\frac{\Gamma^2}{{\Omega}_{i_1}^2}
\sum_{i_2\in{\mathsf{N}}[i_1]}
\frac{\Gamma^2}{{\Omega}_{i_2}^2}
\cdots\sum_{i_l\in{\mathsf{N}}[i_{l-1}]}\frac{\Gamma^2}{{\Omega}_{i_l}^2},
\label{eq:y0-series}$$ with $\phi_l$ the total contribution to $y_0$ from all sites on the $l^\mathrm{th}$ generation. Diagrammatically, it is the total contribution to $y_0$ from all $K^l$ paths of length $2l$, each of which goes from the root site to a unique site in the $l^\mathrm{th}$ generation and retraces itself back to the root site [^2]. For the series in Eq. to converge in the thermodynamic limit, $\phi_l$ must decrease sufficiently fast with increasing $l$. This suggests that the distributions $P_{\phi_l}$ of $\phi_l$, should evolve with $l$ in a qualitatively different manner in the delocalised and localised phases. Calculating $P_{\phi_l}$ shows that this is indeed so, as shown in Fig. \[fig:recursive-conv\](a)-(b). For strong disorder (localised phase), the vast bulk of the distribution shifts rapidly to smaller values with increasing $l$, while in the delocalised phase the support of the $P_{\phi_l}$ moves to larger values with increasing $l$. This is itself indicative of the convergence of the series in the localised phase and otherwise in the delocalised. To further quantify the convergence, one can define $y_0^{\scriptstyle{[l]}}\equiv\sum_{n=1}^l\phi_n$ and study its typical value, $y_{0,\mathrm{typ}}^{\scriptstyle{[l]}}$, as a function of $l$ and $W$. Representative results at $\omega=0$ are shown in Fig. \[fig:recursive-conv\](c). For weak disorder, $y_{0,\mathrm{typ}}^{\scriptstyle{[l]}}$ grows rapidly with $l$, whereas for strong disorder it saturates to its converged value in the localised phase; again clearly showing the presence of a localisation transition.
Two further remarks should be made. First, the recursive formulation also treats the *real* parts of all self-energies exactly. One can however make the simplifying approximation of neglecting them – Anderson’s ‘upper limit approximation’ [@anderson1958absence; @abou-chacra1973self]. For the tree with correlated disorder this approximation again predicts the presence of a transition, albeit naturally at a higher $W_c$ [@supp]. Second, the terms appearing in the series in Eq. but with $X_{i_{n}}^{\scriptstyle{(i_{n-1})}}=0$ (i.e. ${\Omega}_{i_{n}}\equiv\omega-W\epsilon_{i_{n}}$) are precisely those appearing in the Forward Approximation [@pietracaprina2016forward]. By including the contribution of non-local propagators to the local propagator in an exact, *fully* renormalised fashion, the recursive formulation is a significant technical advance.
. []{data-label="fig:recursive-conv"}](recursive-conv.pdf){width="\linewidth"}
Correlations in the $\epsilon_i$’s preclude an exact analytic solution for the distribution of $y_0$ from Eq. . One can nevertheless perform a self-consistent mean-field calculation analytically at leading order in the renormalised perturbation series [@logan2019many; @roy2020fock; @roy2019self] (here illustrated for $\omega=0$). Here $y_0$ depends only on the site energies of its neighbours, $\{i_1\}$. Since $\ell_{0i_1}=1$, the maximally correlated limit implies the conditional distribution $P(\epsilon_{i_1}|\epsilon_0)=\delta(\epsilon_{i_1}-\epsilon_0)$ in the thermodynamic limit. The distribution of $y_0$ can thus be simply calculated as $P_{y_{0}}(y_{0})=\int d\epsilon_{0}~P(\epsilon_0^{{\phantom\dagger}})~\delta\big(y_0^{{\phantom\dagger}}-K\Gamma^2(1+{y_{0,\mathrm{typ}}^{{\phantom\dagger}}})/[W^2\epsilon_0^2]\big)$. Since the univariate distribution $P(\epsilon_0)$ is a standard Normal, this yields $P_{y_0}(y_0,{y_{0,\mathrm{typ}}^{{\phantom\dagger}}}) = \sqrt{\kappa/\pi}~e^{-\kappa/y_0}y_0^{-3/2}$ where $\kappa = K(1+{y_{0,\mathrm{typ}}^{{\phantom\dagger}}})\Gamma^2/2W^2$. Remarkably and reassuringly, the distribution indeed has the Lévy form, just as obtained numerically by summing the entire series (Fig. \[fig:recursive-tree\](c)).
Self-consistency can now be imposed by requiring $\ln{y_{0,\mathrm{typ}}^{{\phantom\dagger}}}=\int dy_0P_{y_0}(y_0,{y_{0,\mathrm{typ}}^{{\phantom\dagger}}})\ln y_0$; the solution of which is ${y_{0,\mathrm{typ}}^{{\phantom\dagger}}}=2e^\gamma K\Gamma^2(W^2-2e^\gamma K\Gamma^2)^{-1}$, with $\gamma$ the Euler-Mascheroni constant. Since $y_0$ is necessarily non-negative, self-consistency of the localised phase requires $W\ge W_c$, with [^3] $$W_c^{{\phantom\dagger}} = \sqrt{2}e^{\gamma/2}\Gamma\sqrt{K}.
\label{eq:Wc-sc}$$ This $W_{c}\propto \sqrt{K}$ scaling is qualitatively different from that arising for uncorrelated disorder, where $W_c\propto K\ln K$ [@abou-chacra1973self]; and stems intrinsically from the maximal correlations in the disorder.
We turn now to results arising for RRGs, via exact diagonalisation (ED) of tight-binding Hamiltonians Eq. with maximally correlated disorder Eq. . Our motivation here is twofold. First, while results above were for a rooted Cayley tree, we expect them to hold qualitatively for other random graphs. Second, it is important to corroborate the results with other independent measures of localisation. Cayley trees are not moreover readily amendable to ED, since a finite fraction of sites live on the boundary; this issue is sidestepped by considering RRGs, which are locally tree-like but contain long loops.
In the following we consider RRGs with a coordination number $Z=K+1 =3$; denoting the total number of sites in the RRG by $N$. In accordance with the form of the covariance matrix for the Cayley tree, we take $C_{ij}=\exp[-\ell_{ij}\ln K/\ln N]$. The quantities studied will be the level spacing ratios, and $\Delta_i$ computed directly. We focus on the middle of the spectrum ($\omega =0$) and consider 50-100 eigenstates therein.
For an ordered set of eigenvalues $\{E_n\}$, the level spacing ratio is $r_n=\min[s_n,s_{n+1}]/\max[s_n,s_{n+1}]$ with $s_n=E_n-E_{n-1}$. In an ergodic phase the distribution of $r_n$ follows the Wigner-Dyson surmise with mean $\overline{r}\simeq0.53$, while in a localised phase the distribution is Poisson with $\overline{r}\simeq 0.386$. Results for $\overline{r}$ *vs* $W$ are shown in Fig. \[fig:rrg-ed\](a), and show clearly a localisation transition. A scaling collapse of the data for various $N$ onto a common function of $(W-W_c)N^{1/\nu}$ yields a critical disorder strength of $W_c\simeq 6.8$ and $\nu\simeq 4.6$. Note that the $W_c$ estimated is remarkably close to that obtained above numerically for the $K=2$ Cayley tree.
![ ED results for a $K=2$ RRG with maximally correlated disorder. (a) Mean level spacing ratio *vs* $W$ shows a crossing for different $N$. Data collapse onto a common function of $(W-W_c)N^{1/\nu}$ yields $W_c\simeq 6.8$ and $\nu\simeq4.6$ (inset). (b) Typical value $\Delta_{{\mathrm{typ}}}(\omega =0)$ computed exactly from Eq. . In the delocalised \[localised\] phase it is independent of \[decays with\] $N$. Dashed line shows extrapolation to $N\to\infty$. Grey shaded regions in (a), (b) denote the estimated critical region. (c)-(d) Distributions of $\Delta$ and $y=\Delta/\eta$ in the delocalised and localised phases respectively. Grey shaded regions show best fits to Log-Normal and Lévy distributions respectively.[]{data-label="fig:rrg-ed"}](rrg-ed.pdf){width="\linewidth"}
From the set of exact eigenvalues $\{E_n\}$ and eigenstates $\{\ket{\psi_n}\}$, $\Delta_i(\omega)$ can be computed as $$\Delta_i(\omega) = \mathrm{Im}[G_i^{-1}(\omega)]-\eta,~~G_i=\sum_n\frac{\vert\braket{\psi_n|i}\vert^2}{\omega+i\eta-E_n}.
\label{eq:Delta-RRG}$$ As $\Delta$ is finite with unit probability in the delocalised phase, $\Delta_\mathrm{typ}$ should converge to a finite value with increasing $N$; while in a localised phase $\Delta \propto\eta$ vanishes with unit probability, so $\Delta_\mathrm{typ}$ should decrease with $N$. This behaviour is indeed found, see Fig. \[fig:rrg-ed\](b). To estimate numerically the critical $W_{c}$, we posit $\Delta_\mathrm{typ}=\Delta_{\mathrm{typ},N\to\infty}+ a/N^\beta$ and extrapolate the data to the thermodynamic limit. As shown in Fig. \[fig:rrg-ed\](b), the vanishing of $\Delta_{\mathrm{typ},N\to\infty}$ gives a $W_c$ consistent with that obtained from level statistics. In the localised phase, the distribution of $y=\Delta/\eta$ is again in very good agreement with a Lévy distribution (see Fig. \[fig:rrg-ed\](d)). In the delocalised phase by contrast, $\Delta$ is qualitatively different, and appears to be log-normally distributed (Fig. \[fig:rrg-ed\](c)).
As above, whether for a Cayley tree or RRG, we find a one-parameter Lévy distribution for $y=\Delta/\eta$ in the localised phase. Importantly, it is thus universal: distributions for different $W>W_{c}$ can be collapsed onto a universal form by scaling the self-energy as $y/y_\mathrm{typ}$ [@supp]. Further, the distribution can be directly connected to that of wavefunction amplitudes, the moments of which (via generalised IPRs) probe the divergence of the localisation length, $\xi$, as $W\to W_{c}$ [@supp]. Within our mean-field theory, we find $\xi\sim(W-W_c)^{-1}$ with an exponent of $1$.
We turn now to the $K\to\infty$ limit. For any one-body problem to remain well-defined in this limit, the hopping must be rescaled as $\Gamma = \Gamma_\ast/\sqrt{K}$. The mean-field theory then yields a finite critical $W_c=\sqrt{2}e^{\gamma/2}\Gamma_\ast$; in stark contrast to the case of uncorrelated disorder where, despite rescaling $\Gamma$, $W_{c}/\Gamma_\ast\propto\sqrt{K}\ln K$ thus precludes localisation as $K\to\infty$. For MBL on Fock space, in a system containing $L$ real-space sites, the effective connectivity on the Fock-space graph scales as $K\sim L$, and the effective Fock-space disorder as $W_\mathrm{FS}\sim\sqrt{L}W_{\mathrm{t}}$ (with $W_{\mathrm{t}}\sim\mathcal{O}(1)$) [@roy2020fock; @logan2019many]. Rescaling all energies by $\sqrt{L}$, as required to attain a well-defined thermodynamic limit $L\to \infty$, again leads [@roy2020fock] to a finite critical $W_{\mathrm{t},c}$, in direct parallel to the $K\to\infty$ limit of the present problem. The existence of an MBL phase thus provides an indirect but complementary argument for the $\sqrt{K}$ scaling of $W_c$.
In summary, we have studied AL on Cayley trees and RRGs with maximally correlated on-site disorder, mimicking the effective Fock-space disorder of MBL systems. While such correlations might be thought to disfavour localisation by suppressing site-energy fluctuations, we find both that an Anderson transition is present, and that scaling of the critical disorder with graph connectivity is qualitatively different from that of uncorrelated disorder, with correlations favouring localisation. Our results address a new class of AL problems, and shed light on the crucial role played by correlations in Fock-space disorder in stabilising MBL. Many questions arise as to what further aspects of MBL can be captured by AL problems with maximally correlated disorder. One such is the multifractal character of wavefunctions, and its possible connection to the anomalous statistics of MBL wavefunctions on Fock space; and our preliminary results indeed suggest the presence of multifractal eigenstates on RRGs. Looking further afield, understanding the effect of maximal correlations on glassy dynamics on such graphs is also immanently important.
We thank J. T. Chalker, A. Duthie, and A. Lazarides for useful discussions and comments on the manuscript. This work was in part supported by EPSRC Grant No. EP/S020527/1.
[^1]: The algorithm for constructing the correlated energies is described in the supplementary material [@supp]
[^2]: On a tree, there exists a unique shortest path between any pair of sites. For a site on generation $l$, the length of the corresponding path between it and the root site is $l$.
[^3]: While this analysis focuses on the localised phase, self-consistency for the delocalised phase commensurately breaks down at the same $W_{c}$ as in Eq. [@logan2019many; @roy2019self].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We formulate the data analysis problem for the detection of the Newtonian waveform from an inspiraling compact-binary by a network of arbitrarily oriented and arbitrarily distributed laser interferometric gravitational wave detectors. We obtain for the first time the relation between the optimal statistic and the magnitude of the network correlation vector, which is constructed from the matched network-filter. This generalizes the calculation reported in an earlier work (gr-qc/9906064), where the detectors are taken to be coincident.'
address:
- '(1) Department of Physics and Astronomy, P. O. Box 913, Cardiff University, CF24 3YB, United Kingdom'
- |
(2) Inter-University Centre for Astronomy and Astrophysics, Post Bag 4, Ganeshkhind,\
Pune 411007, India
- '(3) Max Planck Institut fur Gravitationphysik, Albert Einstein Institut, Am Muhlenberg 1, Golm, D-14476, Germany'
author:
- 'Sukanta Bose[^1](1), Archana Pai[^2](2), and Sanjeev V. Dhurandhar[^3](2,3)'
title: 'Detection of gravitational waves from inspiraling compact binaries using a network of interferometric detectors[^4]'
---
= 10000
ł
Introduction {#sec:intro}
============
Inspiraling compact binaries form prime candidates for detection by earth-based interferometric gravitational-wave (GW) detectors owing to the well understood waveform (chirp) emitted by them. Searching for chirps using a network of such detectors is gaining importance due to (a) its superior sensitivity [*vis a vis*]{} that of a constituent detector \[1\] and (b) improving feasibility for a real-time computational search. Here, we formulate the problem of how to optimally detect the Newtonian chirp using a network of arbitrarily orientated and arbitrarily located detectors. This extends a similar study in Ref. \[2\] of a network with coincident detectors.
We use the maximum likelihood method for optimizing the detection problem.\[3\] A single likelihood ratio (LR) is deduced for the entire network. A super-threshold value for the maximized likelihood ratio (MLR) implies a detection. The MLR is obtained by maximizing the LR over the eight parameters that determine the Newtonian chirp: the distance $r$ to the binary, the inital phase $\delta$ of the waveform, the polarization angle $\psi$, the inclination angle $\epsilon$ of the binary orbit, the time of arrival $t_a$ at a fiducial detector (fide), the source-direction angles $\{\phi, \theta\}$, and the chirp time $\xi$. In principle, this can always be done numerically using a grid in the eight dimensional parameter space. In practice, however, such a strategy is computationally unfeasible and wasteful. We show that maximization of the LR over four parameters, $\{r,\delta,\psi,\epsilon\}$, can be performed analytically using the symmetries in detector responses. This allows us to scan this parameter subspace [*continuously*]{}. Further, the Fast Fourier Transform (FFT) can be used to maximize LR over $t_a$, as in the case of a single detector. Such an analytic maximization and the FFT allow us to save substantially on computational costs. Numerical maximization is required over the remaining parameters, $\{\phi, \theta, \xi\}$, which we discuss in a future work. Here, we follow the convention laid out in Ref. \[2\].
The signal
==========
There are four distinctly different reference frames of interest, associated with the source, wave, fide, and a representative detector in the network. Physical quantities in these frames are related by orthogonal transformations, ${\cal O}_{\rm k}$, which are defined in terms of three sets of Euler angles that specify the orientation of one frame with respect to another.\[4\] Let ${\sf x}$ be an arbitrary three-dimensional real vector. Then, ${\sf x}_{\rm wave} = {\cal O}_1(\psi, \epsilon, 0) {\sf x}_{\rm source}$, ${\sf x}_{\rm fide} = {\cal O}_2(\p, \t, 0) {\sf x}_{\rm wave}$, and ${\sf x}_{\rm detector} = {\cal O}^{-1}_3(\a, \b, \g) {\sf x}_{\rm fide}$, Here, the source axes have been chosen in accordance with Ref. \[5\].
The wave tensor $w_{ij}$ associated with any source can be expanded in terms of the STF-2 tensors ${\cal Y}^{ij}_{2n}$ in an arbitrary frame as \[2\]: \[wstf2a\] w\^[ij]{}(t) = , where $i$, $j$ denote spatial indices, and $h_+$ and $h_\times$ are the two GW polarizations in the transverse-traceless gauge, as measured in some given frame. The expansion coefficients $T_{\pm 2}{}^n$ are the Gel’fand functions,\[2\] which depend on the Euler angles through which one must rotate that frame into the frame in which $w^{ij}$ is being analyzed. The above form suggests the definitions, $e_L^{ij} = \sqrt{8\pi /15}\>T_{2}{}^n {\cal Y}^{ij}_{2n}$ and $e_R^{ij} = \sqrt{8\pi /15}\>T_{-2}{}^n {\cal Y}^{ij}_{2n}$, for the left- and right-circular polarization tensors, respectively. They obey, $e_L^{ij\>*} = e_R^{ij}$, $e_{L,\> R}^{ij} \>e_{L,\> R\> ij}^* =1$, and $e_{L,\> R}^{ij} \>e_{R,\>L\> ij}^* = 0$, in any frame. Thus, \[wS\] w\^[ij]{} (t)= [Re]{} 2 , where $\kappa ={\sqrt \xi} / r$ (up to a normalization factor) and $R\equiv (h_+(t) + ih_\times (t))/(2\kappa)$. For a chirp, we define $R$ in the source frame.\[5\] Then $h_{+,\times}$ are GW amplitudes for a face-on binary (i.e., for $\e=0$), and $R$ depends only on $\{\d,t_a,\xi\}$.
The response amplitude (i.e., the signal) in the $I$-th detector is the scalar product $s^I = w^{ij} d_{ij}^I$, which depends on projections of $e_{L,\> R}^{
ij}$ onto the $I$-th detector tensor, $d_{ij}^I$. One such projection defines the extended beam-pattern function: \[Fdef\] F\^I = e\_L\^[ij]{}d\_[ij]{}\^I T\_[2]{}\^p (, , 0) [D]{}\_p\^I ,p=2 , which corresponds to the left-circular polarization. Above, \[LI\] [D]{}\^[I]{}\_[p]{} T\_[p]{}\^n (,,0)d\^[I]{}\_[ij]{} [Y]{}\^[ij]{}\_[2n]{} =ig\^I T\_[p]{}\^n (, , 0) (T\^[I\*]{}\_[2n]{} - T\^[I\*]{}\_[-2n]{} ) , where $T^{I}_{\pm 2n} =T_{\pm 2n}(\a_{I},\b_{I},\g_{I})$ and $d_{ij}^I=g^I
(n_{1i}^I n_{1j}^I -n_{2i}^In_{2j}^I)$, with ${\bf n}^I_{1,2}$ being unit vectors along the two arms of the $I$-th interferometer, respectively. Also, $g^I$ is the detector’s noise power spectral density.\[6\] Then, \[sigW\] s\^[I]{}(t)= 2 ( F\^[I\*]{} R\^[I]{} ) 2 ( F\^[I\*]{} S\^[I]{} e\^[i]{}) where $R^{I}$ is defined via Eq. (\[wS\]) and $S^{I}$ is independent of $\d$.
The optimal network statistic {#sec:detect}
=============================
Under the Neyman-Pearson decision criterion,\[3\] the optimal network statistic is the network LR, $\l$. If the noise in each detector is additive and independent of the noise in any other detector in the network, then $\l$ reduces to a product of the individual detector LR’s.\[2\] Further, for Gaussian noise,\[7\] the logarithmic likelihood ratio (LLR), $\ln \lambda$, simplifies to the following sum of LLR’s of $N$ individual detectors \[2\]: = \_[I=1]{}\^N s\_[I]{}, x\_[I]{} \_[I]{} - [1 2]{}\_[I=1]{}\^N s\_[I]{}, s\_[I]{} \_[I]{} = [b]{} \_[I=1]{}\^N z\_[I]{}, x\_[I]{}\_[I]{} - [12]{} [b]{}\^2 , \[LRb\] where ${\sf b}\equiv 2\kappa (\sum_{I=1}^N \|F_{I}\|^2)^{1/2}$ and $z_{I} = s_{I}/{\sf b}$. Above, $r$ appears only in ${\sf b}$.
Maximizing $\ln\lambda$ with respect to ${\sf b}$ and $\d$ gives, $\ln \lambda |_{\hat{\sf b},\hat{\d}} =
\left|\sum_{I=1}^N Q_{I} C_{I}^*\right|^2 /2$, where $Q_{I} \equiv 2\kappa F_{I}/{\sf b}$ and $C_{I}^* \equiv \langle S_{I}, x_{I}
\rangle_{I}$.\[2\] This shows that the network vector ${\sf S}$, with $S_{I}$’s as its components, is the matched network-filter. Also, $\ln \lambda |_{\hat{\sf b},
\hat{\d}}$ is a function of six parameters, namely, $\{\psi,\e,t_a,\p,\t,\xi\}$. To extend these results to the case where the detectors are arbitrarily located, note that the dependence of $\ln \lambda |_{\hat{\sf b},\hat{\d}}$ on $\{\psi, \epsilon\}$ can be isolated. This is because the network vector ${\sf Q}$, with $Q^I$’s as its components, is: = [ ]{}\^[-1]{} ( T\_[2]{}\^[-2]{} (,, 0) [D]{}\_[-2]{} + T\_[2]{}\^[2]{} (,, 0) [D]{}\_[2]{} ) Q\^[-2]{} \_[-2]{} + Q\^2 \_[2]{} , where ${\sf D}_p$ define network vectors with ${D}^I_p$ as their components; ${\hat {\sf D}}_p$ are their normalized counterparts. Thus, $Q_2 = {\hat {\sf D}}_{2}\cdot {\sf Q} = Q^{+2} + Q^{-2} {\hat {\sf D}}_{2}
\cdot {\hat {\sf D}}_{-2}$. Hence, $\{{\hat {\sf D}}_{2}, \>{\hat {\sf D}}_{-2}\}$ define a two-dimensional complex plane, ${\cal P}$ (the helicity space, a subspace of ${\cal C}^N$), on which a metric $g_{pq}$ can be defined. Then, $Q_p = g_{pq} Q^q$ with $p,q
=\pm 2$. The $N$-dimensional correlation vector ${\sf C}$, in general, lies outside ${\cal P}$. However, ${\sf Q}$ lies totally in ${\cal P}$. Thus, the statistic reduces to |\_[,]{} = | [Q]{} \^\* |\^2 /2= | [Q]{} \^\*\_[P]{} |\^2 /2 , where ${\sf C}_{\cal P}$ is the projection of ${\sf C}$ onto the helicity space.
Maximization of $\ln\lambda |_{\hat{\sf b},\hat{\d}}$ over $\{\psi,\e\}$ is achieved by aligning ${\sf Q}$ along ${\sf C}_{\cal P}$. This requires that $Q^{\pm 2} =
{C^{\pm 2}_{\cal P} / {\parallel {{\sf C}_{\cal P}} \parallel}_{\cal P}}$, which implies: ${Q^{+2} / {Q^{-2}}} =
{{C^{+2}_{\cal P}}/ C^{-2}_{\cal P}}$. Since the RHS above can take any value in the complex plane, we need to prove that ${Q^{+2} / {Q^{-2}}}$ obeys the same property. To this end, note that, = [[T\_[2]{}\^[-2]{} (,, 0)]{} ]{} = ([[1 - ]{} ]{})\^2 ([4i ]{}) , which can indeed attain any value on the Argand plane. The values of $\psi$ and $\e$ that maximize the statistic are, ${\hat \psi} = \arg (x)/4$ and ${\hat \e} = \cos^{-1} [(1-\sqrt{\|x\|}\>)/(1+\sqrt{\|x\|}\>)]$ where $x \equiv C^{+2}_{\cal P} / C^{{-2}}_{\cal P}$. Thus, the statistic maximized over these four parameters is, \[st\] |\_[,,,]{} = [\_[P]{} ]{}\^[2]{} /2. Let ${\hat V}^\pm$ denote a pair of orthonormal, complex basis vectors on ${\cal P}$. Then, \[st2\] [\_[P]{} ]{}\^[2]{} = [\^[+]{} ]{}\^[2]{} +[\^[-]{} ]{}\^[2]{} = (c\_0\^+)\^2 + (c\_[/2]{}\^+)\^2 + (c\_0\^-)\^2 + (c\_[/2]{}\^-)\^2 , where $C^\pm = {\sf C}_{\cal P} \cdot {\hat V}^\pm =c_0^\pm + ic_{\pi/2}^\pm$. It can be verified that the network statistic is, therefore, a sum of the squares of four Gaussian random variables with constant variance. This simplifies the computation of thresholds and detection probabilities. A network filter is constructed as follows: For a given $\xi$ compute the Newtonian chirp for $t_a = 0$. Then, for a given direction $\{\phi,\theta\}$, use the appropriate time-delays with respect to fide to time-displace the chirp at each detector. This collection of time-displaced chirps constitute the network filter. Also, ${\hat{t}}_a$ is obtained by shifting the network filter ‘rigidly’ on the time axis, which can be done efficiently using FFT. The bank of filters on $\{\phi,\theta,\xi\}$ can be obtained by correlating two neighboring normalized filters in the usual way. This work is now in progress.
Conclusion
==========
We have given here a formulation to optimally detect the Newtonian chirp with a network of detectors in which the noise is additive, Gaussian, and uncorrelated between detectors. We have shown how this can be done efficiently by analytically maximizing the LR over four parameters and using FFT to maximize over the time-of-arrival. In a future work we hope to address key issues such as required computational power for such a search and also estimate errors in parameter values.
SB acknowledges support from PPARC grant PPA/G/S/1997/00276. AP acknowledges support from a CSIR grant and from AEI, Potsdam, for a 3-month visit.
B. F. Schutz, in [it The detection of gravitational waves]{}, ed. D. G. Blair (Cambridge University Press, Cambridge, 1992), pp. 406–452.
S. Bose, S. V. Dhurandhar, and A. Pai, Pramana [**53**]{} (1999) 1125. (gr-qc/9906064)
C. W. Helstrom, [*Statistical Theory of Signal Detection*]{} (Pergamon, London, 1968).
In defining the Euler angles, we follow the convention of H. Goldstein, [*Classical Mechanics*]{} (Addison-Wesley Publishing Co., Reading, 1980), 2nd ed..
K. S. Thorne, in [*300 Years of Gravitation*]{}, eds. S. Hawking and W. Israel (Cambridge University Press, Cambridge, 1987).
In Ref. \[2\], $W^{(I)}$ is the symbol used for denoting a quantity proportional to $F^I$.
Rare instances of non-Gaussian noise can be vetoed by techniques similar to those applied for a single detector; see, e.g., B. Allen [*et al.*]{}, Phys. Rev. Lett. [**83**]{} (1999) 1498.
[^1]: Electronic address: [*[email protected]*]{}
[^2]: Electronic address: [*[email protected]*]{}
[^3]: Electronic address: [*[email protected]*]{}
[^4]: Based on talk given at GWDAW-99, Rome, in December 1999.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this note, it is shown that closed string graviton scattering amplitudes obey the same subleading soft limit as field theory graviton scattering amplitudes. The result is derived using a combination of recent results and methods including the subleading soft expansion for type I open string gluon disk scattering amplitudes, the single value projection for multiple $\zeta$ values, and the KLT relations in field and string theory.'
author:
- 'Burkhard U. W. Schwab'
bibliography:
- 'soft.bib'
title: A Note on Soft Factors for Closed String Scattering
---
= 1
[emailempty@@email[@]{}]{}
[affiliation[Brown University\
Department of Physics\
182 Hope St, Providence, RI, 02912]{}]{}
[keywords[Gauge theory, Cluster algebras]{}]{} [pdfsubject[Supercluster algebras]{}]{} [preprint[Brown-HET-1660]{}]{}
preprint
**title**
authors
*affiliation1*
email
Introduction and Discussion {#sec:introduction}
===========================
@Cachazo:2014fwa’s suggestion of a new soft theorem for graviton scattering sparked a surge of interest in the low energy behaviour of these amplitudes [@Cachazo:2014fwa]. Originally treated to leading order by Weinberg [@Weinberg:1965nx; @Weinberg:1964ew], it was shown by @Gross:1968in that there are universal subleading contributions [@Gross:1968in]. @Low:1958sn showed that there is a very similar looking soft expansion for Yang Mills theory [@Low:1958sn].
The new soft theorem is usually stated as follows. The $N$ graviton tree level amplitude $M_{N}$ factorises in the presence of a soft particle with $k_{N}=q\to0$ (and polarisation tensor $\epsilon_{\mu\nu}$) into a soft part and the $(N-1)$ point amplitude $$M_{N} = \left(S^{(0)}_{\rm g} + S^{(1)}_{\rm g} + S^{(2)}_{\rm g}\right)M_{N-1} + \mathcal{O}(q).$$ The colour ordered $N$ gluon Yang Mills tree level amplitude $A_{N}$ behaves in a very similar way $$A_{N} = \left(S^{(0)}_{\rm YM} + S^{(1)}_{\rm YM}\right)A_{N-1} + \mathcal{O}(q).\label{eq:softyang}$$ For graviton scattering, when taking the soft particle to be the $N$^th^ particle, the soft factors are $$S^{(0)}_{\rm g} = \sum_{i=1}^{N-1}\frac{\epsilon_{\mu\nu}p^{\mu}_{i}p^{\nu}_{i}}{q.p_{i}},\qquad S^{(1)}_{\rm g} = \sum_{i=1}^{N-1}\frac{\epsilon_{\mu\nu}p^{\mu}_{i}(q_{\rho}J^{\rho \nu}_{i})}{q.p_{i}},\qquad S^{(2)}_{\rm g} = \sum_{i=1}^{N-1}\frac{\epsilon_{\mu\nu}(q_{\lambda}J^{\lambda \mu}_{i})(q_{\rho}J^{\rho \nu}_{i})}{q.p_{i}}\label{eq:softgrav}$$ while the analogous factors for Yang Mills theory are $$S^{(0)}_{\rm YM} = \frac{\epsilon_{\mu}p^{\mu}_{1}}{q.p_{1}} - \frac{\epsilon_{\mu}p^{\mu}_{n}}{q.p_{n}},\qquad S^{(1)}_{\rm YM} =\frac{\epsilon_{\mu}q_{\nu}J^{\mu\nu}_{1}}{q.p_{1}} - \frac{\epsilon_{\mu}q_{\nu}J^{\mu\nu}_{n}}{q.p_{n}}.$$
The operator $J_{i}^{\mu\nu}$ denotes the total angular momentum operator of the $i$^th^ particle. $S^{(1)}_{\rm g}$, $S^{(2)}_{\rm g}$, and $S^{(1)}_{\rm YM}$ are therefore derivative operators as opposed to the leading order soft gluon and graviton factors, which are only multiplication operators. The leading and subleading soft graviton factors as well as the soft gluon factors are universal and the soft graviton factor is protected from quantum corrections.
At approximately the same time when these ideas were first developed (in fact, three years before), the group of symmetries at null infinity of asymptotically flat space was investigated by Bondi, van der Burg, Metzner, and Sachs [@Bondi:1962px; @Sachs:1962wk]. This symmetry group is known as the BMS group $${\rm BMS} = T \ltimes SL(2,\mathbb{C})$$ and consists of the semidirect product of the Abelian group $T$ of “supertranslations” (which are in no way related with supersymmetry) and the group $SL(2,\mathbb{C})$ of non-singular transformations of the two sphere at infinity. The latter can be extended to also contain singular transformations – called superrotations – to form a Virasoro algebra [@Barnich:2013axa; @Barnich:2011mi; @Barnich:2009se; @Barnich:2011ct]. This line of thought has only recently been put forward by @Barnich:2011ct. While Low already connected the soft gluon factors with gauge symmetry, it took fifty years for the development of a logical connection between BMS and the soft graviton factor. Recently, such a relation has been presented in a series of papers [@He:2014laa; @Kapec:2014opa; @Strominger2013; @Strominger:2013jfa]. Note that the work by @He:2014laa not only connects the supertranslations with Weinberg’s soft factor, but the subsequent work by @Kapec:2014opa also makes use of the superrotations of the extended BMS group and connect these with the subleading term. Thus the gravitational soft factors are protected by symmetries of the (classical) gravitational $\mathcal{S}$ matrix. The work by @Cachazo:2014fwa mentioned before [@Cachazo:2014fwa] also introduced a novel sub-subleading term – $S^{(2)}_{\rm g}$ in – which seems universal but is not thought to be protected by any symmetry.
This connection has seen many checks and there have been notable extensions for subleading soft theorems. The latter have been discussed by @Casali:2014xpa for Yang Mills theory [@Casali:2014xpa] in the more modern language of the spinor helicity formalism following [@Cachazo:2014fwa]. @Bern:2014oka as well as @He:2014bga have shown that the subleading soft factor in Yang Mills theory receives corrections and so does the gravitational subleading soft factor at the loop level [@He:2014bga; @Bern:2014oka; @Bianchi:2014gla], an alternative prescription has been offered by @Cachazo:2014dia in [@Cachazo:2014dia] which however clashes with the conventional way the soft theorem is understood and used (see also [@Bonocore:2014wua]). The universality of the subleading and subsubleading soft factors for arbitrary dimensions in gravity and Yang Mills theory have been checked in [@Schwab:2014xua; @Afkhami-Jeddi:2014fia; @Zlotnikov:2014sva; @Kalousios:2014uva]. @Du:2014eca have shown that there is a nontrivial connection between the Yang Mills subleading soft factor and the gravity subleading soft factor using the field theory KLT relations [@Du:2014eca]. These relations impose a novel nontrivial constraint on the field theory KLT kernel. @Broedel:2014fsa as well as @Bern:2014vva [@Broedel:2014bza; @Broedel:2014fsa; @Bern:2014vva] have shown that the form of the soft factors is highly constrained (see also [@Larkoski:2014hta]). A derivation of the symmetry principle in ambitwistor string theory has been given by @Geyer:2014lca. A theory at null infinity describing the observed characteristics of graviton scattering in the soft limit was given by @Adamo2014 in [@Adamo2014; @Geyer:2014lca]. There have been multiple results on supersymmetric theories as well as QCD [@Liu:2014vva; @Rao:2014zaa; @Luo:2014wea]. On the other hand, a novel extension of BMS has been put forward in [@Campiglia:2014yka].
In essence this note should be treated as an appendix to [@Schwab:2014fia]. In this paper it was shown that single trace type I superstring gluon scattering amplitudes satisfy the same subleading soft theorem as field theory Yang Mills scattering amplitudes. As a tree level result, it is independent of the dimension, chosen compactification or the amount of supersymmetry. In the last section of the paper, a way to use the string theory Kawai-Lewellen-Tye (KLT) relations [@Kawai:1985xq] $$\mathcal{M} = (-1)^{N-3}\kappa^{N-2}\mathcal{A}^{T}.\mathcal{S}.\mathcal{A}$$ was mentioned to produce the closed string version of the soft theorem from the result for open strings. In the last equation, $\mathcal{M}$ are either type I or type two closed string amplitudes, $\mathcal{A}$ is a vector of colour ordered type I open string disk amplitudes, and $\mathcal{S}$ is the string theory phase matrix. The notation which will be made more explicit in the text. Proving the soft graviton theorem for string theory from this perspective involves an additional step. The phase matrix $\mathcal{S}$, which is necessary to combine two open string amplitudes into a closed string amplitude in the KLT prescription must be shown to behave correctly under the soft limit, too.
In the meantime, the above mentioned result by @Du:2014eca became available. Using this result it is possible to take a field theory path to the string theory result. To do so, another string theory result connecting type I open string scattering amplitudes with heterotic string scattering amplitudes and closed string scattering amplitudes [@Stieberger:2014hba] due to @Stieberger:2014hba has to be used. However, ref. [@Du:2014eca] only showed the four dimensional case, so here a version which is independent of the number of dimensions is presented as a side product. It will turn out that combining these three results will make for a rather elegant proof of the subleading soft graviton theorem for closed string scattering amplitudes.
It will only be shown that the subleading soft factor $S^{(1)}_{\rm g}$ can be recovered in this way. A field theory proof that the subsubleading contribution in gravity can be obtained from KLT and the Yang Mills soft theorem is still pending. A moment of reflection will make clear that the subsubleading term in gravity will depend on subsubleading expressions from Yang Mills amplitudes which are not thought to be universal. A possible way to attack this problem would be to use the expansion for MHV amplitudes given in [@He:2014bga] to show that the subsubleading term in gravity can be recovered from KLT at least in the MHV case.
This note’s organisation is as follows. In sec. \[sec:preliminaries\], the preliminaries will be presented. A review of [@Schwab:2014fia] is given in ssec. \[sec:soft-behaviour-open\]. Ssec. \[sec:single-value-proj\] provides an introduction to the single value projection which connects type I gluon scattering amplitudes with heterotic string gluon scattering amplitudes [@Brown:2013gia; @Stieberger:2014hba]. A byproduct will be the subleading soft theorem for the heterotic string. The derivation of the theorem for gravity from Yang Mills theory via the KLT relations in field theory is presented in \[sec:grav-soft-fact\]. In this subsection, an extension of the derivation to any dimension is given. The derivation of the subleading soft theorem for closed string graviton scattering, which is very short, is presented in sec. \[sec:closed-soft\].
Preliminaries {#sec:preliminaries}
=============
The following three subsections provide background for the closed string graviton result and extend some results of previous papers.
Soft behaviour of open string disk amplitudes {#sec:soft-behaviour-open}
---------------------------------------------
In [@Schwab:2014fia] it was shown that the open string disk amplitudes behave just like the field theory amplitudes in the limit of a single momentum becoming soft. Clearly, this result is reasonable since string theory should not deviate from the field theoretic behaviour in this low energy regime.
An open string disk amplitude ${\cal A}_{N} = {\cal A}(1,\ldots,N)$ with $N$ particles is given by essentially two pieces. Employing the notation of [@Stieberger:2014hba] the first piece is a vector $A$ of colour ordered Yang Mills scattering amplitudes. The entries of this vector are denoted by $A_{\sigma} = A(1,\sigma_{2,N-2},N-1,N)$ where $\sigma\in S_{N-3}$ is a permutation of particles $2$ to $N-2$ $$\sigma: (2,\ldots, N-2) \to (2_{\sigma},\ldots, (N-2)_{\sigma}).$$ This vector is dotted into the period matrix $F_{\pi\sigma}$ where $\pi,\sigma \in S_{N-3}$ such that $${\cal A}_{\pi} =F_{\pi\sigma}A_{\sigma} \label{eq:period}$$ where $\pi$ denotes the colour ordering of the string disk amplitude. The form of the period matrix $F$ and can be found in numerous publications, see e.g., [@Mafra:2011nv; @Mafra:2011nw].
In the presence of a soft string momentum $k_{N-2}=q\to0$, here taken to be particle $N-2$, the string disk amplitude factorises $${\cal A}(1,\pi(2,\ldots,N-3,q),N-2,N-1) \to (S^{(0)}_{\rm YM} + S^{(1)}_{\rm YM}){\cal A}(1,\pi'(2,\ldots,N-2),N-1,N).$$ The numbering of the $N$ particles on the left hand side amplitude was changed for ease of comparison. The $S^{(i)}_{\rm YM}$ are the leading $(i=0)$ and subleading $(i=1)$ gluon soft factors from field theory as in . Note that the Yang Mills soft factor for colour ordered amplitudes always depends on the particles adjacent to the soft particle and only on these.
Single value projection {#sec:single-value-proj}
-----------------------
The next, vital ingredient to introduce is the single value projection $\operatorname{sv}$ [@Brown:2013gia; @Stieberger:2014hba]. The map $\operatorname{sv}$ is a homomorphism between the Hopf algebra of multiple $\zeta$ values (MZVs) defined by $$\zeta_{n_{1},\ldots,n_{r}} = \sum_{0<k_{1}<\ldots<k_{r}}\prod_{\ell =1}^{r}k_{\ell}^{-n_{\ell}}$$ and Brown’s single valued multiple $\zeta$ values $\zeta^{\operatorname{sv}}_{n_{1},\ldots,n_{r}}$ defined as the evaluation of single valued (multiple) polylogarithms at $1$. As an example: Single valued (simple) polylogarithms $$D_{m}(z)=\Re\left(i^{m+1}\left[\sum_{k=1}^{m}\frac{(-\log|z|)^{m-k}}{(m-k)!}\operatorname{Li}_{k}(z) - \frac{(-\log|z|)^{m}}{2m!}\right]\right)\label{eq:zagier_ramakrish}$$ are a generalisation of the Bloch-Wigner-function $$D(z) = \Im(\operatorname{Li}_{2}(z)) + \arg(1-z)\log |z|$$ and were studied in slightly different forms by, e.g., Ramakrishnan [@MR862642], Zagier [@MR1032949], Wojtkowiak [@MR1051830] and most recently Brown [@Brown:2013gia]. In the Ramakrishnan-Zagier form it is $D_{2}(z)=D(z)$. The single-value projection $$\operatorname{sv}: \zeta_{n_{1},\ldots,n_{r}}\mapsto \zeta^{\operatorname{sv}}_{n_{1},\ldots,n_{r}}$$ satisfies various relationships which derive from the functional equations satisfied by the single-valued polylogarithms. For example it is easy to see that $$\operatorname{sv}(\zeta_{2}) = \operatorname{sv}(\operatorname{Li}_{2}(1)) = D(1) = \zeta^{\operatorname{sv}}_{2} = 0.$$ Further examples and relationships can be found in the papers by Brown as well as in [@Stieberger:2013wea]. Note that scalars with respect to the Hopf algebra of multiple $\zeta$-values pass through the single value projection.
The relationship between (colour ordered) type I $N$ gluon disk scattering amplitudes $\mathcal{A}$ and the corresponding heterotic string scattering amplitudes $\mathcal{A}^{\rm HET}$ is governed by $\operatorname{sv}$. In [@Stieberger:2014hba] it was shown that these amplitudes are directly connected via the single value projection (and a rescaling $\alpha'\to \alpha'/4$), i.e., $${\cal A}^{\rm HET}_{\pi} = \operatorname{sv}({\cal A}_{\pi}) = \operatorname{sv}(F_{\pi\sigma}) A_{\sigma}.\label{eq:heterotic}$$ This is the single value projection of the matrix equation . The action of $\operatorname{sv}$ on $F_{\pi\sigma}$ is to take the $\zeta$ value expansion of $F_{\pi\sigma}$ and replace every occurrence of a MZV by the corresponding single valued MZV. Since the soft factors are scalar with respect to $\operatorname{sv}$, the relation $${\cal A}^{\rm HET}_{N+1,\pi} \to \Big(S^{(0)}_{\rm YM}+S^{(1)}_{\rm YM}\Big)_{\pi(q+1),\pi(q-1)}{\cal A}^{\rm HET}_{N,\pi'},\label{eq:hetsoft}$$ where $\pi\in S_{N-2}$, $\pi'\in S_{N-3}$, follows almost trivially. As before, $q$ is the soft momentum and the subscript $\pi(q+1),\pi(q-1)$ indicates that the soft factors will depend on the colour ordering in the same way as they depend on the colour ordering in Yang Mills or type I disk scattering amplitudes — via particles adjacent to the soft particle. Since the field theory limit $\alpha'\to 0$ of $\mathcal{A}^{\rm HET}$ are Yang Mills gluon scattering amplitudes, it is not surprising that the heterotic string scattering amplitudes behave in this way. Despite that, the result will be an important stepping stone for the closed string scattering amplitudes which will be investigated in the next section.
The result in has another implication [@Stieberger:2014hba]. The KLT relations split closed string scattering amplitudes into a right moving and a left moving part. Using as a basis for the left moving part, let $$\widetilde{{\cal A}}_{N,\sigma} = {\cal A}(1,\sigma(2,\ldots,N-2),N,N-1)$$ be a basis for the right moving part. The KLT relations connect (type I and type II) $N$ graviton string scattering amplitudes with a sum of products of open string amplitudes $${\cal M} = (-1)^{N-3}\kappa^{N-2}\sum_{\rho,\sigma\in S_{N-3}}{\cal A}_{\rho}{\cal S}_{\rho,\sigma}\widetilde{{\cal A}}_{\sigma}$$ where [@BjerrumBohr:2010hn] $${\cal S}_{\rho,\sigma} = \prod_{j=2}^{N-2}\sin\left(s_{1,j_{\rho}} + \sum_{k=2}^{j-1}\theta(j_{\rho},k_{\rho}) s_{j_{\rho},k_{\rho}}\right)$$ with $\theta(j_{\rho},k_{\rho}) = 1$ if the ordering of $j_{\rho}$ and $k_{\rho}$ is the same in both $\rho$ and $\sigma$, otherwise it is zero.[^1] The scattering amplitudes $\mathcal{A}$, $\widetilde{\mathcal{A}}$, and $\mathcal{M}$ on both sides of the equation have each the same amount of legs $N$. The variables $s_{ij}:= \alpha' p_{i}.p_{j}$ are the usual dimensionless kinematic invariants. Using the $\operatorname{sv}$ projection and some manipulations it is possible to show that closed string graviton scattering amplitudes can be given in terms of only one “stringy” factor of ${\cal A}^{\rm HET}$ and contributions of the field theory KLT relations, i.e., $${\cal M}_{N} = (-1)^{N-3}\kappa^{N-2} A^{T}.S_{\rm FT}.{\cal A}^{\rm HET}.\label{eq:KLThet}$$ In the last equation, $A^{T}$ is the transpose of the vector of colour ordered Yang Mills scattering amplitudes appearing in , while $S_{\rm FT}$ is the KLT-kernel $$S_{\rho,\sigma} = \prod_{j=2}^{N-2}\left(s_{1,j_{\rho}} + \sum_{k=2}^{j-1}\theta(j_{\rho},k_{\rho}) s_{j_{\rho},k_{\rho}}\right)\label{eq:kltkernel}$$ subject to a basis change of the scattering amplitudes and $s_{ij} = p_{i}.p_{j}$ here. This basis change can be implemented by a matrix $D_{\sigma,\rho}$ which connects the field theory amplitudes $A(1,\sigma(2,\ldots,N-2),N-1,N)$ with the amplitudes $\widetilde{A}(1,\rho(2,\ldots,N-2),N,N-1)$, i.e., $A_{\sigma} = {D_{\sigma,\rho}}\widetilde{A}_{\rho}$ and $$S_{\rm FT,\rho,\sigma} = \sum_{\gamma\in S_{n-3}}S_{\rho,\gamma}D_{\gamma,\sigma}.$$
Graviton soft factors from KLT {#sec:grav-soft-fact}
------------------------------
The KLT relations have proven to be a very valuable tool for investigations into the classical gravity $\mathcal{S}$ matrix. With the growing understanding of the gauge theory $\mathcal{S}$ matrix the KLT relations give guidance on how to generalise results from the gauge side to gravity. In a recent paper [@Du:2014eca], it was shown that the field theory KLT relations can also be used to prove a connection between the soft factors in gauge theory and gravity.
One important caveat is in order. The results in [@Du:2014eca] were derived using methods available only in $d=4$. The KLT relations hold in any dimension, and the subleading soft theorem equally well extends to any dimension [@Schwab:2014xua; @Afkhami-Jeddi:2014fia; @Kalousios:2014uva; @Zlotnikov:2014sva]. Trivial considerations then show that the work by @Du:2014eca can be extended to arbitrary dimensions. In the process of a review of the result, this generalisation will be derived.
The proof in [@Du:2014eca] uses an alternative form of the KLT relations as its starting point. Using the notation introduced in ssec. \[sec:single-value-proj\], but now with field theory amplitudes (and stripped gravitational coupling) $$M_{N} = (-1)^{N-3}\sum_{t=2}^{N-2}\sum_{\sigma,\beta\in S_{N-4}} A_{N}(1,t,\sigma_{2,N-2},N-1,N) S^{p_{1}}_{t,t}S_{\sigma,\beta}^{p_{N-1}}\widetilde{A}_{N}(t,1,N-1,\beta,N)\label{eq:altKLT}.$$ This expression is valid in any dimension. Note the appearance of yet another basis $\widetilde{A}$ for the field theory amplitudes on the right hand side. Later on, it will be necessary to use a basis transformation to bring it into the form of .
The KLT-kernel $S_{\sigma,\beta}^{p_{N-1}}$ is built around particle $N-1$ which takes the place of particle $1$ as given in . More compactly, this can be written as $$M_{N} = (-1)^{N-3}\sum_{t=2}^{N-2}\sum_{\sigma,\beta\in S_{N-3}} A_{N}(t,\sigma) S^{p_{1}}_{t,t}S_{\sigma,\beta}^{p_{N-1}}\widetilde{A}_{N}(t,\beta).$$ A quick calculation will show that the leading contribution to the soft limit follows very naturally from this form.[^2] First, send $p_{1}\to 0$. Then $S_{t,t} = p_{1}.p_{t} =: s_{1t}$ and $$\begin{aligned}
A_{N}(1,t,\sigma,N-1,N) &\to (S^{(0)}_{\rm YM} + S^{(1)}_{\rm YM})_{t,N}A_{N-1}(t,\sigma,N-1,N)\\
\widetilde{A}_{N}(t,1,N-1,\beta,N) &\to (S^{(0)}_{\rm YM} + S^{(1)}_{\rm YM})_{N-1,t}A_{N-1}(t,N-1,\sigma,N)\end{aligned}$$ where the subscripts denote the particles adjacent to particle $1$ in the given colour ordering. After the expansion, the leading order term can be assembled for each $t$ as $S^{(0)}_{t,N}s_{1t}S^{(0)}_{N-1,t}$. This expression multiplies $$(-1)^{N-4}\sum_{\sigma,\beta\in S_{N-3}} A_{N-1}(t,\sigma,N-1,N)S_{\sigma,\beta}^{N-1}\widetilde{A}_{N-1}(t,N-1,\beta,N)\label{eq:kltrep}$$ in the sum over $t$. But is just another KLT representation of the gravity amplitude $M_{N-1}(2,\ldots,N)$ and therefore in particular independent of $t$. Then, using conservation of momentum it is easy to show that $$\sum_{t=2}^{N-2} S^{(0)}_{t,N}s_{1t}S^{(0)}_{N-1,t} = -\sum_{t=2}^{N}\frac{\epsilon.p_{t}\epsilon.p_{t}}{s_{1t}} \xlongrightarrow{\epsilon_{\mu}\epsilon_{\nu} \to \epsilon_{\mu\nu}} - S^{(0)}_{\rm g}$$ which is Weinberg’s soft graviton factor. The minus sign cancels the missing minus sign in from the definition of $M_{N}$. As usual for KLT in dimensions other than four, the gluon polarisation vectors $\epsilon_{\mu}$ will have to be exchanged for a graviton polarisation tensor $\epsilon_{\mu\nu}$ to get the correct answer.
The subleading term is more subtle, and as was shown in [@Du:2014eca], requires a non-trivial identity for the KLT kernel to hold. The same identity $$\sum_{t=2}^{N-2}\sum_{\sigma,\beta\in S_{N-4}}D[2,\tilde\sigma,N-1,N|t,\sigma,N-1,N]S_{\sigma,\beta}^{N-1}\Delta_{t} C[t,N-1,\beta,N|2,N-1,\tilde\beta,N] =0 \label{eq:identity}$$ where $$\Delta_{t} = \epsilon_{\mu}p_{1\nu}\left(\frac{\epsilon.p_{t}J^{\mu\nu}_{N-1}}{s_{1,N-1}} + \frac{\epsilon.p_{N}J^{\mu\nu}_{t}}{s_{1,N}} + \frac{s_{1t}\epsilon.p_{N}J^{\mu\nu}_{N-1}}{s_{1,N-1}s_{1,N}} - \frac{\epsilon.p_{t}J^{\mu\nu}_{N}}{s_{1,N}} - \frac{\epsilon.p_{N-1}J^{\mu\nu}_{t}}{s_{1,N-1}} - \frac{s_{1t}\epsilon.p_{N-1}J^{\mu\nu}_{N}}{s_{1,N-1}s_{1,N}}\right)$$ can be shown to arise in the calculation valid for any dimension. The base change matrices $D$ and $C$ are used to transform the basis of colour-ordered Yang-Mills amplitudes via $$A_{N-1}(t,\tilde\sigma,N-1,N) = \sum_{\sigma\in S_{N-4}}A_{N-1}(s,\tilde\sigma,N-1,N)D[s,\tilde\sigma,N-1,N|t,\sigma,N-1,N]$$ where $s$ is arbitrary ($s=2$ in ) and similarly for $\widetilde{A}_{N-1}$ and $C$. The operators $\Delta_{t}$ are more involved in this representation than in the spinor-helicity form given in [@Du:2014eca] and depend on the angular momentum operators of particles $N-1$, $N$ and $t$. In this note, no attempts will be made at proving this result. It appears however that the calculation in momentum variables is in fact easier to perform than in spinor helicity variables. Under the above assumption, it can be shown that the subleading order in the soft expansion of the terms in the KLT relation actually lead to the subleading soft graviton factor as given in , i.e., $$\begin{aligned}
(-1)^{N-3}\sum_{t=2}^{N-1}\sum_{\sigma,\beta\in S_{N-4}}&\left[\Big(S^{(0)}_{N-1,t}s_{1t}S^{(1)}_{t,n}A_{N-1}\Big)S_{\sigma,\beta}^{N-1}\widetilde{A}_{N-1} + A_{N-1}S_{\sigma,\beta}^{N-1}\Big(S^{(0)}_{N-1,t}s_{1t}S^{(1)}_{t,n}\widetilde{A}_{N-1}\Big)\right]{\notag\\}&= \sum_{t=2}^{N} \frac{\epsilon.p_{t}\epsilon_{\mu}p_{1\nu}J^{\mu\nu}_{t}}{s_{1t}}M_{N-1} \xlongrightarrow{\epsilon_{m}\epsilon_{\nu}\to\epsilon_{\mu\nu}}S^{(1)}_{\rm g}M_{N-1}\end{aligned}$$
With this, all that is left to do is to assemble the results of this section into the promised proof.
Closed string scattering soft factor {#sec:closed-soft}
====================================
This rather short section contains the actual proof of the subleading soft theorem for closed string graviton scattering amplitudes. To perform the final step and show that closed string graviton scattering amplitudes behave like field theory graviton scattering amplitudes in the subleading soft limit, it is only necessary to synthesise the results in ssecs. \[sec:soft-behaviour-open\], \[sec:single-value-proj\], and \[sec:grav-soft-fact\].
First, revisit equation . The right hand side can be written more verbosely in the following form $$\mathcal{M}_{N} = (-1)^{N-3} \kappa^{N-2} \sum_{\sigma,\beta\in S_{N-3}} A(1,\sigma_{2,N-2},N-1,N)S_{\rm FT}[\sigma|\beta] \operatorname{sv}\mathcal{A}(1,\beta_{2,N-2},N-1,N).$$ The KLT field theory kernel $S_{\rm FT}$ is related to the KLT kernel $S$ used in sec. \[sec:grav-soft-fact\] via a basis transformation which solely depends on combinations of kinematic data $s_{ij}$. Therefore, this expression can be transformed into the alternative representation of the KLT relations in using basis transformations which do not interfere with the single value projection, i.e., $$\mathcal{M}_{N} = (-1)^{N-3} \kappa^{N-2} \sum_{t=2}^{N-2}\sum_{\sigma,\beta\in S_{N-3}}A_{N}(t,\sigma)S_{t,t}S_{\sigma,\beta}\operatorname{sv}\widetilde{\mathcal{A}}_{N}(t,\beta).\label{eq:altKLThet}$$ This is possible since the heterotic string amplitudes $\mathcal{A}^{\rm HET}$ satisfy the same Kleiss-Kuijf [@Kleiss:1988ne] and Bern-Carrasco-Johansson [@Bern:2008qj] identities as the field theory amplitudes. Furthermore, the base change matrices connecting right moving $\widetilde{\mathcal{A}}$ and left moving $\mathcal{A}$ are dimensionless [@Stieberger:2014hba]. In particular, they do not depend on $\alpha'$. It follows that the leading and subleading soft limit follow in the same way as in sec. \[sec:grav-soft-fact\]. Specifically, combining ,, and the soft gluon limit for particle $1$ $$\mathcal{M}_{N} =
(-1)^{N-3} \kappa^{N-2} \sum_{t=2}^{N-2}\sum_{\sigma,\beta\in S_{N-3}}(S^{(0)} + S^{(1)})_{N,t}A_{N-1}(t,\sigma)s_{1t}S_{\sigma,\beta}(S^{(0)} + S^{(1)})_{t,N-1}\operatorname{sv}\widetilde{\mathcal{A}}_{N-1}(t,\beta)$$ But the above expression is exactly the same expression as the field theory KLT, which means that all calculations in \[sec:grav-soft-fact\] hold in this case. Thus $$\mathcal{M}_{N}\overset{p_{1}\to 0}{\longrightarrow}\kappa(S^{(0)}_{\rm g} + S^{(1)}_{\rm g})\mathcal{M}_{N-1} + \mathcal{O}(p_{1})$$ which ends the proof that closed string theory tree level amplitudes feature the same soft limit as their field theory descendants.
In conclusion, all string and field theories connected through KLT relations, the single value projection and the field theory limit $\alpha'\to 0$ must be behaving in the same way under the subleading soft limit at tree level. This encompasses gauge field theory, type I open strings, and heterotic string theory gauge scattering amplitudes. On the gravitational side, it was shown here that the behaviour of (type I, type II, heterotic and bosonic) closed string graviton scattering can be determined from that of open string gluon amplitudes. The field theory limit of these amplitudes coincides with the known gravitational result as it should be. It remains to be seen whether there is an equivalent symmetry statement behind these results as in the field theory case.
#### Acknowledgements
The author would like to thank Stephan Stieberger for suggesting the use of the $\operatorname{sv}$ projection. The author is also grateful for discussions with Tim Adamo, Steven Avery, Johannes Brödel, Matteo Rosso and Oliver Schlotterer. This work is supported by the US Department of Energy under contract DE-FG02-11ER41742.
[^1]: The constant $\kappa$ is the gravitational coupling constant. Note that ${\cal A}$ does not contain any factors of $g_{\rm YM}$.
[^2]: The reader will have to excuse the profusion of quantities denoted by the letter $S$. Most of the time, the meaning of the letter should be unambiguous as soft factors always bear a bracketed superscript denoting their level in the soft expansion while KLT kernels always come with subscript indices.
| {
"pile_set_name": "ArXiv"
} |
This paper had been withdrawed
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show a Lichnerowicz-Obata type estimate for the first eigenvalue of the Laplacian of $n$-dimensional closed Riemannian manifolds with an almost parallel $p$-form ($2\leq p \leq n/2$) in $L^2$-sense, and give an almost decomposition result of the manifold under some pinching conditions when $2\leq p<n/2$.'
address: 'Graduate School of Mathematics, Nagoya University, Chikusa-Ku Nagoya, 464-8602, Japan'
author:
- Masayuki Aino
title: 'Lichnerowicz-Obata Estimate, Almost Parallel $p$-form and Almost Product Manifolds'
---
Introduction
============
In this paper we give an estimate for the first eigenvalue of the Laplacian of closed Riemannian manifolds with positive Ricci curvature and an almost parallel form, and give a pinching result about the almost equality case.
One of the most famous theorem about the estimate of the first eigenvalue of the Laplacian is the Lichnerowicz-Obata theorem. Lichnerowicz showed the optimal comparison result for the first eigenvalue when the Riemannian manifold has positive Ricci curvature, and Obata showed that the equality of the Lichnerowicz estimate implies that the Riemannian manifold is isometric to the standard sphere. In the following, $\lambda_k(g)$ denotes the $k$-th eigenvalue of the Laplacian $\Delta:=-{\mathop{\mathrm{tr}}\nolimits}_g {\mathop{\mathrm{Hess}}\nolimits}$ acting on functions.
Take an integer $n\geq 2$. Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold. If ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-1) g$, then $\lambda_1(g)\geq n$. The equality holds if and only if $(M,g)$ is isometric to the standard sphere of radius $1$.
Petersen [@Pe1], Aubry [@Au] and Honda [@Ho] showed the stability result of the Lichnerowicz-Obata theorem. In the following, $d_{GH}$ denotes the Gromov-Hausdorff distance and $S^n$ denotes the $n$-dimensional standard sphere of radius $1$. (see Definition \[DGH\] for the definition of the Gromov-Hausdorff distance).
\[PA\] For given an integer $n\geq 2$ and a positive real number $\epsilon>0$, there exists $\delta(n,\epsilon)>0$ such that if $(M,g)$ is an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-1) g$ and $\lambda_n(g)\leq n+\delta$, then $d_{GH}(M,S^n)\leq \epsilon$.
Note that Petersen considered the pinching condition on $\lambda_{n+1}(g)$, and Aubry and Honda improved it independently.
We mention some improvements of the Lichnerowicz estimate when the Riemannian manifold has a special structure. If $(M,g)$ is a real $n$-dimensional Kähler manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-1)g$, then the Lichnerowicz estimate is improved as follows: $$\label{kae}
\lambda_1(g)\geq 2(n-1).$$ See [@Be Theorem 11.49] for the proof. If $(M,g)$ is a real $n$-dimensional quaternionic Kähler manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-1)g$, then we have $$\label{qk}
\lambda_1(g)\geq \frac{2n+8}{n+8}(n-1).$$ See [@AM] for the proof. For these cases, the Riemannian manifold $(M,g)$ has a non-trivial parallel $2$ and $4$-form, respectively. When $(M,g)$ is an $n$-dimensional product Riemannian manifold $(N_1\times N_2,g_1+g_2)$ with ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-1)g$, then we have $$\lambda_1(g)\geq \min_{i\in\{1,2\}}\left\{\frac{\dim N_i}{\dim N_i-1}\right\}(n-1),$$ and $M$ has a non-trivial parallel form if either $N_1$ or $N_2$ is orientable.
Grosjean [@gr] gave a unified proof of the improvements of the Lichnerowicz estimate when the Riemannian manifold has a non-trivial parallel form.
\[grosjean\] Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold. Assume that ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-p-1)g$ and that there exists a nontrivial parallel $p$-form on $M$ $(2\leq p\leq n/2)$. Then, we have $$\label{grs}
\lambda_1(g)\geq n-p.$$
Moreover, if $p<n/2$ and if in addition $M$ is simply connected, then the equality in $(\ref{grs})$ implies that $(M,g)$ is isometric to a product $S^{n-p}\times (X,g')$, where $(X,g')$ is some $p$-dimensional closed Riemannian manifold.
We give several remarks on this theorem.
- Grosjean also showed this type theorem when $M$ has a convex smooth boundary.
- Though Grosjean originally assumed the manifold is orientable, the assumption can be easily removed by taking the orientable double covering.
- If $M$ is simply connected, $p=n/2$ and $n\geq 6$, then it is not difficult to show that the equality in (\[grs\]) also implies that $M$ is isometric to a product $S^{n/2}\times X$ (see Corollary \[p3d\]).
- If $(M,g)$ is either a Kähler manifold or a quaternionic Kähler manifold, then the estimate (\[kae\]) or (\[qk\]) is better.
- If there exists a non-trivial parallel $p$-form $\omega$ ($1\leq p\leq n-1$) on an $n$-dimensional Riemannian manifold $(M,g)$, then $\omega(x)\in \bigwedge^p T^\ast_x M$ ($x\in M$) is invariant under the Holonomy action, and so the Holonomy group coincides with neither $\mathrm{SO}(n)$ nor $\mathrm{O}(n)$.
The main aim of this paper is to show the almost version of Grosjean’s result. We also give the almost version of the estimate (\[kae\]) in Appendix B.
We first note that, for a closed Riemannian manifold $(M,g)$, there exists a non zero $p$-form $\omega$ with $\|\nabla \omega\|_2^2\leq \delta\|\omega\|_2^2$ for some $\delta>0$ if and only if $\lambda_1(\Delta_{C,p})\leq \delta$ holds, where $\lambda_1(\Delta_{C,p})$ is defined by $$\lambda_1(\Delta_{C,p}):=\inf\left\{\frac{\|\nabla \omega\|_2^2}{\|\omega\|_2^2}: \omega\in\Gamma(\bigwedge^p T^\ast M) \text{ with }\omega\neq 0\right\}.$$
Let us state our eigenvalue estimate.
For given integers $n\geq 4$ and $2\leq p \leq n/2$, there exists a constant $C(n,p)>0$ such that if $(M,g)$ is an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-p-1)g$, then we have $$\lambda_1(g)\geq n-p-C(n,p)\lambda_1(\Delta_{C,p})^{1/2}.$$
We immediately have the following corollary:
For given integers $n\geq 4$ and $2\leq p \leq n/2$, there exists a constant $C(n,p)>0$ such that if $(M,g)$ is an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-p-1)g$ and $$\frac{n(n-p-1)}{n-1}\leq \lambda_1(g)\leq n-p,$$ then we have $$\lambda_1(\Delta_{C,p})\geq \left(\frac{n-p-\lambda_1(g)}{C(n,p)}\right)^2.$$
Note that we always have the lower bound on the eigenvalue of the Laplacian $\lambda_1(g)\geq n(n-p-1)/(n-1)$ if ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-p-1)g$ by the Lichnerowicz estimate. An upper bound on $C(n,p)$ is computable. However, we do not know the optimal value of it.
We next state the eigenvalue pinching result.
For given integers $n\geq 5$ and $2\leq p < n/2$ and a positive real number $\epsilon>0$, there exists $\delta=\delta(n,p,\epsilon)>0$ such that if $(M,g)$ is an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-p-1)g$, $$\lambda_{n-p+1}(g)\leq n-p+\delta$$ and $$\lambda_1(\Delta_{C,p})\leq \delta,$$ then $M$ is orientable and $$d_{GH}(M,S^{n-p}\times X)\leq \epsilon,$$ where $X$ is some compact metric space.
In fact, we prove that there exist constants $C(n,p)>0$ and $\alpha(n)>0$ such that $$d_{GH}(M,S^{n-p}\times X)\leq C(n,p)\delta^{\alpha(n)}$$ under the assumption of Main Theorem 2. One can easily find the explicit value of $\alpha(n)$ (see Notation \[order\] and Theorem \[MT2\]). However, it might be far from the optimal value. By the Gromov’s pre-compactness theorem, we can take $X$ to be a geodesic space. However, we lose the information about the convergence rate in that case.
Based on Theorem \[PA\], one might expect that we can replace the assumption “$\lambda_{n-p+1}(g)\leq n-p+\delta$” in Main Theorem 2 to the weaker assumption “$\lambda_{n-p}(g)\leq n-p+\delta$”. However, an example shows that we cannot do it even if $\delta=0$ (see Proposition \[p3e\]). Instead of that, we have the following theorems:
For given integers $n\geq 4$ and $2\leq p \leq n/2$, there exists a constant $C(n,p)>0$ such that if $(M,g)$ is an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-p-1)g$, then we have $$\lambda_1(g)\geq n-p-C(n,p)\lambda_1(\Delta_{C,n-p})^{1/2}.$$
For given integers $n\geq 5$ and $2\leq p < n/2$ and a positive real number $\epsilon>0$, there exists $\delta=\delta(n,p,\epsilon)>0$ such that if $(M,g)$ is an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-p-1)g$, $$\lambda_{n-p}(g)\leq n-p+\delta$$ and $$\lambda_1(\Delta_{C,n-p})\leq \delta,$$ then we have $$d_{GH}(M,S^{n-p}\times X)\leq \epsilon,$$ where $X$ is some compact metric space.
Note that the assumption “$\lambda_1(\Delta_{C,n-p})\leq \delta$” is equivalent to the assumption “$\lambda_1(\Delta_{C,p})\leq \delta$” if the manifold is orientable. In particular, we have the following corollary:
For given integers $n\geq 5$ and $2\leq p < n/2$ and a positive real number $\epsilon>0$, there exists $\delta=\delta(n,p,\epsilon)>0$ such that if $(M,g)$ is an $n$-dimensional orientable closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-p-1)g$, $$\lambda_{n-p}(g)\leq n-p+\delta$$ and $$\lambda_1(\Delta_{C,p})\leq \delta,$$ then we have $$d_{GH}(M,S^{n-p}\times X)\leq \epsilon,$$ where $X$ is some compact metric space.
We would like to point out that our work was motivated by Honda’s spectral convergence theorem [@Ho2], which asserts the continuity of the eigenvalues of the connection Laplacian $\Delta_{C,p}$ acting on $p$-forms with respect to the non-collapsing Gromov-Hausdorff convergence assuming the two-sided bound on the Ricci curvature. By virtue of his theorem, we can generalize our main theorems to Ricci limit spaces under such assumptions. See Appendix A for detail. Note that we show our main theorems without the non-collapsing assumption, i.e., without assuming the lower bound on the volume of the Riemannian manifold.
Our work was also motivated by the Cheeger-Colding almost splitting theorem (see [@Ch Theorem 9.25]), whose conclusion is the Gromov-Hausdorff approximation to a product $\mathbb{R}\times X$. As the almost splitting theorem, we need to show the almost Pythagorean theorem under the assumption of Main Theorem 2. One step of the proof (Lemma \[p54i\]) is similar to the final step of the almost splitting theorem [@Ch Lemma 9.16].
The structure of this paper is as follows.
In section 2, we recall some basic definitions and facts, and give calculations of differential forms.
In section 3, we assume that the Riemannian manifold has a non-trivial parallel $p$-form. We give an easy proof of the formula used by Grosjean to prove Theorem \[grosjean\].
In section 4, we estimate the error terms of the Grosjean’s formula when the Riemannian manifold has a non-trivial almost parallel $p$-form. As a consequence, we prove Main Theorem 1 and Main Theorem 3.
In section 5, we prove Main Theorem 2 and Main Theorem 4. In subsection 5.1, we list some useful techniques for pinching problems. In subsection 5.2, we show some pinching conditions on the eigenfunctions along geodesics under the assumption $\lambda_k(g)\leq n-p+\delta$ and $\lambda_1(\Delta_{C,p})\leq \delta$. In subsection 5.3, we show that similar results hold under the assumption $\lambda_k(g)\leq n-p+\delta$ and $\lambda_1(\Delta_{C,n-p})\leq \delta$. In subsection 5.4, we show that the eigenfunctions are almost cosine functions in some sense under our pinching condition. In subsection 5.5, we construct an approximation map and show Main Theorem 2 except for the orientability. In subsection 5.6, we give some lemmas to prove the remaining part of main theorems. In subsection 5.7, we show the orientability of the manifold under the assumption of Main Theorem 2, and complete the proof of it. In subsection 5.8, we show that the assumption of Main Theorem 4 implies that $\lambda_{n-p+1}(g)$ is close to $n-p$, and complete the proof of Main Theorem 4.
In Appendix A, we discuss Ricci limit spaces. We show a gap theorem of the first eigenvalue of the Laplacian acting on $n$-forms for $n$-dimensional unorientable closed Riemannian manifolds. As a consequence, we show the stability of unorientability under the non-collapsing Gromov-Hausdorff convergence assuming the two-sided bound on the Ricci curvature and the upper bound on the diameter. This enable us to generalize our main theorems to Ricci limit spaces under such assumptions.
In Appendix B, we give the almost version of the estimate (\[kae\]) assuming that there exists a $2$-form $\omega$ which satisfies that $\|\nabla \omega\|_2$ and $\|J_\omega^2+{\mathop{\mathrm{Id}}\nolimits}\|_2$ are small, where $J_\omega\in\Gamma(T^\ast M\otimes T M)$ is defined so that $\omega=g(J_\omega\cdot,\cdot)$.
[**Acknowledgments**]{}. I am grateful to my supervisor, Professor Shinichiroh Matsuo for his advice. I also thank Professor Shouhei Honda for helpful discussions about the orientability of Ricci limit spaces. I thank Shunsuke Kano for the discussions about the examples. The works in section 3 were done during my stay at the University of Côte d’Azur. I would like to thank Professor Erwann Aubry for his warm hospitality. This work was supported by JSPS Overseas Challenge Program for Young Researchers and by JSPS Research Fellowships for Young Scientists (JSPS KAKENHI Grant Number JP18J11842).
Preliminaries
=============
Basic Definitions
-----------------
We first recall some basic definitions and fix our convention.
\[Dhau\] Let $(X,d)$ be a metric space. For each point $x_0\in X$, subsets $A,B\subset X$ and $r>0$, define $$\begin{aligned}
d(x_0,A):=&\inf\{d(x_0,a):a\in A\},\\
B_{r}(x_0):=&\{x\in X: d(x,x_0)<r\},\\
B_{r}(A):=&\{x\in X:d(x,A)<r\},\\
d_{H,d}(A,B):=&\inf\{\epsilon>0:A\subset B_{\epsilon}(B) \text{ and } B\subset B_{\epsilon}(A)\}\end{aligned}$$ We call $d_{H,d}$ the Hausdorff distance.
The Hausdorff distance defines a metric on the collection of compact subsets of $X$.
\[DGH\] Let $(X,d_X),(Y,d_Y)$ be metric spaces. Define $$\begin{aligned}
d_{GH}(X,Y):=\inf\Big\{d_{H,d}(X,Y): &\text{ $d$ is a metric on $X\coprod Y$ such that}\\
&\qquad\qquad\quad\text{$d|_X=d_X$ and $d|_Y=d_Y$}\Big\}.\end{aligned}$$
The Gromov-Hausdorff distance defines a metric on the set of isometry classes of compact metric spaces (see [@Pe3 Proposition 11.1.3]).
\[hap\] Let $(X,d_X),(Y,d_Y)$ be metric spaces. We say that a map $f\colon X\to Y$ is an $\epsilon$-Hausdorff approximation map for $\epsilon>0$ if the following two conditions hold.
- For all $a,b\in X$, we have $|d_X(a,b)-d_Y(f(a),f(b))|< \epsilon$,
- $f(X)$ is $\epsilon$-dense in $Y$, i.e., for all $y\in Y$, there exists $x\in X$ with $d_Y(f(x),y)< \epsilon$.
If there exists an $\epsilon$-Hausdorff approximation map $f\colon X\to Y$, then we can show that $d_{GH}(X,Y)\leq 3\epsilon/2$ by considering the following metric $d$ on $X\coprod Y$:
[align\*]{} &d\_X(a,b)&& (a,bX),\
& +\_[xX]{}(d\_X(a,x)+d\_Y(f(x),b))&&(aX,bY),\
&d\_Y(a,b)&&(a,bY).
If $d_{GH}(X,Y)< \epsilon$, then there exists a $2\epsilon$-Hausdorff approximation map from $X$ to $Y$.
Let $C(u_1,\ldots,u_l)>0$ denotes a positive function depending only on the numbers $u_1,\ldots,u_l$. For a set $X$, ${\mathop{\mathrm{Card}}\nolimits}X$ denotes a cardinal number of $X$.
Let $(M,g)$ be a closed Riemannian manifold. For any $p\geq 1$, we use the normalized $L^p$-norm: $$\|f\|_p^p:=\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M |f|^p\,d\mu_g,$$ and $\|f\|_{\infty}:=\mathop{\mathrm{sup~ess}}\limits_{x\in M}|f(x)|$ for a measurable function $f$ on $M$. We also use this notation for tensors. We have $\|f\|_p\leq \|f\|_q$ for any $p\leq q \leq \infty$.
Let $\nabla$ denotes the Levi-Civita connection. Throughout in this paper, $0=\lambda_0(g)< \lambda_1(g) \leq \lambda_2(g) \leq\cdots \to \infty$ denotes the eigenvalues of the Laplacian $\Delta=-\sum_{i,j}g^{ij}\nabla_i \nabla_j$ acting on functions. We sometimes identify $TM$ and $T^\ast M$ using the metric $g$. Given points $x,y\in M$, let $\gamma_{x,y}$ denotes one of minimal geodesics with unit speed such that $\gamma_{x,y}(0)=x$ and $\gamma_{x,y}(d(x,y))=y$. For given $x\in M$ and $u\in T_x M$ with $|u|=1$, let $\gamma_{u}\colon \mathbb{R}\to M$ denotes the geodesic with unit speed such that $\gamma_u(0)=x$ and $\dot{\gamma}_u(0)=u$.
For any $x\in M$ and $u\in T_x M$ with $|u|=1$, put $$t(u):=\sup\{t\in\mathbb{R}_{>0}: d(x,\gamma_u(t))=t\},$$ and define the interior set $I_x\subset M$ at $x$ (see also [@Sa p.104]) by $$I_x:=\{\gamma_u (t): u\in T_x M \text{ with $|u|=1$ and } 0\leq t< t(u)\}.$$ Then, $I_x$ is open and ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus I_x)=0$ [@Sa III Lemma 4.4]. For any $y\in I_x\setminus \{x\}$, the minimal geodesic $\gamma_{x,y}$ is uniquely determined. The function $d(x,\cdot)\colon M\to \mathbb{R}$ is differentiable in $I_x\setminus\{x\}$ and $\nabla d(x,\cdot)(y)=\dot{\gamma}_{x,y}(d(x,y))$ holds for any $y\in I_x\setminus \{x\}$ [@Sa III Proposition 4.8].
Let $V$ be an $n$-dimensional real vector space with an inner product $\langle,\rangle$. We define inner products on $\bigwedge^k V$ and $V\otimes \bigwedge^k V$ as follows: $$\begin{split}
&\langle v_1\wedge\ldots\wedge v_k,w_1\wedge \ldots\wedge w_k\rangle=\det \{\langle v_i,w_j\rangle\}_{i,j},\\
&\langle v_0\otimes v_1\wedge\ldots\wedge v_k,w_0\otimes w_1\wedge \ldots\wedge w_k\rangle=\langle v_0,w_0\rangle \det \{\langle v_i,w_j\rangle \}_{i,j},
\end{split}$$ for $v_0,\ldots,v_k,w_0,\ldots,w_k\in V$. For $\alpha\in V$ and $\omega\in \bigwedge^k V$, there exists unique $\iota(\alpha)\omega\in \bigwedge^{k-1} V$ such that $\langle\iota(\alpha)\omega,\eta\rangle=\langle\omega,\alpha \wedge \eta\rangle$ holds for any $\eta\in \bigwedge^{k-1} V $. If $k=0$, we define $\iota(\alpha)\omega=0$ and $\bigwedge^{-1}V=\{0\}$. Then, $\iota$ defines a bi-linear map: $$\iota\colon V\times \bigwedge^k V\to \bigwedge^{k-1} V.$$ By identifying $V$ and $V^\ast$ using $\langle,\rangle$, we also use the notation $\iota$ for the bi-linear map: $$\iota\colon V^\ast \times \bigwedge^k V\to \bigwedge^{k-1} V.$$
For any Riemannian manifold $(M,g)$, we define operators $\nabla^\ast \colon \Gamma(T^\ast M\otimes \bigwedge^k T^\ast M)\to \Gamma(\bigwedge^k T^\ast M)$ and $d^\ast \colon \Gamma(\bigwedge^k T^\ast M)\to \Gamma(\bigwedge^{k-1}T^\ast M)$ by $$\begin{aligned}
\nabla^\ast(\alpha\otimes \beta):&=-{\mathop{\mathrm{tr}}\nolimits}_{T^\ast M} \nabla(\alpha\otimes \beta)
=-\sum_{i=1}^n \left(\nabla_{e_i}\alpha\right)(e_i)\cdot \beta-\sum_{i=1}^n\alpha(e_i)\cdot\nabla_{e_i}\beta.\\
d^\ast \omega:&=-\sum_{i=1}^n\iota(e_i)\nabla_{e_i}\omega\end{aligned}$$ for all $\alpha\otimes\beta\in \Gamma(T^\ast M\otimes\bigwedge^k T^\ast M)$ and $\omega\in\Gamma(\bigwedge^k T^\ast M)$, where $n=\dim M$ and $\{e_1,\ldots,e_n\}$ is an orthonormal basis of $TM$. If $M$ is closed, then we have $$\begin{aligned}
\int_M \langle T,\nabla \alpha\rangle\,d\mu_g&=\int_M \langle \nabla^\ast T, \alpha\rangle\,d\mu_g,\\
\int_M \langle \omega,d\eta \rangle\,d\mu_g&=\int_M \langle d^\ast \omega, \eta \rangle\,d\mu_g\end{aligned}$$ for all $T\in\Gamma(T^\ast M\otimes\bigwedge^k T^\ast M)$, $\alpha\in\Gamma(\bigwedge^k T^\ast M)$, $\omega\in\Gamma(\bigwedge^k T^\ast M)$ and $\eta\in\Gamma(\bigwedge^{k-1} T^\ast M)$ by the divergence theorem. The Hodge Laplacian $\Delta\colon \Gamma(\bigwedge^k T^\ast M)\to\Gamma(\bigwedge^k T^\ast M)$ is defined by $$\Delta:=d d^\ast +d^\ast d.$$
For an $n$-dimensional Riemannian manifold $(M,g)$, we can take orthonormal basis of $TM$ only locally in general. However, for example, the tensor $$\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i} \nabla f)\omega\in \Gamma(T^\ast M\otimes \bigwedge^{k-1} T^\ast M)\quad (f\in C^\infty(M),\,\omega\in \Gamma(\bigwedge^k T^\ast M))$$ is defined independently of the choice of the orthonormal basis $\{e_1,\ldots,e_n\}$ of $TM$, where $\{e^1,\ldots,e^n\}$ denotes its dual. Thus, we sometimes use such notation without taking a particular orthonormal basis.
Finally, we list some important notation. Let $(M,g)$ be a closed Riemannian manifold.
- $d$ denotes the Riemannian distance function.
- ${\mathop{\mathrm{Ric}}\nolimits}$ denotes the Ricci curvature tensor.
- ${\mathop{\mathrm{diam}}\nolimits}$ denotes the diameter.
- ${\mathop{\mathrm{Vol}}\nolimits}$ or $\mu_g$ denotes the Riemannian volume measure.
- $\|\cdot\|_p$ denotes the normalized $L^p$-norm for each $p\geq 1$, which is defined by $$\|f\|_p^p:=\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M |f|^p\,d\mu_g$$ for any measurable function $f$ on $M$.
- $\|f\|_{\infty}$ denotes the essential sup of $|f|$ for any measurable function $f$ on $M$.
- $\nabla$ denotes the Levi-Civita connection.
- $\nabla^2$ denotes the Hessian for functions.
- $\Delta\colon \Gamma(\bigwedge^k T^\ast M)\to\Gamma(\bigwedge^k T^\ast M)$ denotes the Hodge Laplacian defined by $\Delta:=d d^\ast +d^\ast d$. We frequently use the Laplacian acting on functions. Note that $\Delta=-{\mathop{\mathrm{tr}}\nolimits}_g \nabla^2$ holds for functions under our sign convention.
- $0=\lambda_0(g)< \lambda_1(g) \leq \lambda_2(g) \leq\cdots \to \infty$ denotes the eigenvalues of the Laplacian acting on functions.
- $\gamma_{x,y}\colon [0,d(x,y)]\to M$ denotes one of minimal geodesics with unit speed such that $\gamma_{x,y}(0)=x$ and $\gamma_{x,y}(d(x,y))=y$ for any $x,y\in M$.
- $\gamma_{u}\colon \mathbb{R}\to M$ denotes the geodesic with unit speed such that $\gamma_u(0)=x$ and $\dot{\gamma}(0)=u$ for any $x\in M$ and $u\in T_x M$ with $|u|=1$.
- $I_x$ denotes the interior set at $x\in M$. We have ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus I_x)=0$. We have that $\gamma_{x,y}$ is uniquely determined and $\nabla d(x,\cdot)=\dot{\gamma}_{x,y}(d(x,y))$ holds for any $y\in I_x\setminus\{x\}$.
- $\Delta_{C,k}=\nabla^\ast \nabla\colon \Gamma(\bigwedge^k T^\ast M)\to \Gamma(\bigwedge^k T^\ast M)$ denotes the connection Laplacian acting on $k$-forms.
- $0\leq \lambda_1(\Delta_{C,k}) \leq \lambda_2(\Delta_{C,k}) \leq\cdots \to \infty$ denotes the eigenvalues of the connection Laplacian $\Delta_{C,k}$ acting on $k$-forms.
- $S^n(r)$ denotes the $n$-dimensional standard sphere of radius $r$.
- $S^n:=S^n(1)$.
Note that the lowest eigenvalue of the Laplacian $\Delta$ acting on function is always equal to $0$, and so we start counting the eigenvalues of it from $i=0$. This is not the case with the connection Laplacian $\Delta_{C,k}$ acting on $k$-forms, and so we start counting the eigenvalues of it from $i=1$. For any $i\in\mathbb{Z}_{>0}$, we have $$\lambda_i(\Delta_{C,0})=\lambda_{i-1}(g).$$
Calculus of Differential Forms
------------------------------
In this subsection, we recall some facts about differential forms, and do some calculations.
We first recall the decomposition: $$T^\ast M\otimes \bigwedge^k T^\ast M=T^{k,1}M\oplus\bigwedge^{k+1} T^\ast M\oplus \bigwedge^{k-1} T^\ast M.$$ See also [@Se Section 2].
Let $V$ be an $n$-dimensional real vector space with an inner product $\langle,\rangle$. We put $$\begin{split}
&P_1\colon V\otimes \bigwedge^k V\to \bigwedge^{k+1} V,\quad P_1(\alpha\otimes \omega):=\left(\frac{1}{k+1}\right)^\frac{1}{2}\alpha\wedge\omega,\\
&P_2\colon V\otimes \bigwedge^k V\to \bigwedge^{k-1} V,\quad P_2(\alpha\otimes \omega):=\left(\frac{1}{n-k+1}\right)^\frac{1}{2}\iota(\alpha)\omega,\\
&Q_1\colon \bigwedge^{k+1} V\to V\otimes \bigwedge^k V,\quad Q_1(\zeta):=\left(\frac{1}{k+1}\right)^\frac{1}{2}\sum_{i=1}^n e^i\otimes\iota(e^i)\zeta,\\
&Q_2\colon \bigwedge^{k-1} V\to V\otimes \bigwedge^k V,\quad Q_2(\eta):=\left(\frac{1}{n-k+1}\right)^\frac{1}{2}\sum_{i=1}^n e^i\otimes e^i\wedge\eta,
\end{split}$$ where $\{e^1,\ldots,e^n\}$ is orthonormal basis of $V$. Then, we have
- ${\mathop{\mathrm{Im}}\nolimits}Q_1\bot {\mathop{\mathrm{Im}}\nolimits}Q_2$,
- $P_i\circ Q_i={\mathop{\mathrm{Id}}\nolimits}$ for each $i=1,2$,
- $Q_1$ and $Q_2$ preserve the norms,
- $Q_i\circ P_i\colon V\otimes \bigwedge^k V\to V\otimes \bigwedge^k V$ is symmetric and $(Q_i\circ P_i)^2=Q_i\circ P_i$ for each $i=1,2$.
Therefore, $Q_i\circ P_i$ is the orthogonal projection $V\otimes \bigwedge^k V\to {\mathop{\mathrm{Im}}\nolimits}Q_i$. Since $\bigwedge^{k+1} V\cong {\mathop{\mathrm{Im}}\nolimits}Q_1$ and $\bigwedge^{k-1} V \cong{\mathop{\mathrm{Im}}\nolimits}Q_2$, we can regard $\bigwedge^{k+1} V$ and $\bigwedge^{k-1} V$ as subspaces of $V\otimes \bigwedge^k V$.
Take an $n$-dimensional Riemannian manifold $(M,g)$ and consider the case when $V=T^\ast_x M$ ($x\in M$). We can take a sub-bundle $T^{k,1}M$ of $T^\ast M\otimes \bigwedge^k T^\ast M$ such that $$T^\ast M\otimes \bigwedge^k T^\ast M=T^{k,1}M\oplus\bigwedge^{k+1} T^\ast M\oplus \bigwedge^{k-1} T^\ast M$$ is an orthogonal decomposition. Then, for $\omega\in\Gamma(\bigwedge^k T^\ast M)$, we can decompose $\nabla \omega\in \Gamma(T^\ast M\otimes\bigwedge^k T^\ast M)$, the $\bigwedge^{k+1} T^\ast M$-component is equal to $\left(1/(k+1)\right)^\frac{1}{2}d\omega$ and the $\bigwedge^{k-1} T^\ast M$-component is equal to $-\left(1/(n-k+1)\right)^{1/2} d^\ast \omega$. Let $T(\omega)$ denotes the remaining part ($T\colon \Gamma(\bigwedge^k T^\ast M)\to \Gamma(T^{k,1}M)$). Then, we have $$\nabla \omega=T(\omega)+ \left(\frac{1}{k+1}\right)^\frac{1}{2} Q_1(d\omega)-\left(\frac{1}{n-k+1}\right)^\frac{1}{2}Q_2(d^\ast w).$$ Therefore, we get $$\label{2b}
|\nabla\omega|^2=|T(\omega)|^2+\frac{1}{k+1} |d\omega|^2+\frac{1}{n-k+1}|d^\ast \omega|^2.$$ If $d^\ast \omega=0$ and $T(\omega)=0$, then $\omega$ is called a Killing k-form (see also [@Se Definition 2.1]).
We next recall the Bochner-Weitzenböck formula.
\[p2a\] Let $(M,g)$ be an $n$-dimensional Riemannian manifold. We define a homomorphism $\mathcal{R}_k\colon \bigwedge^k T^\ast M\to \bigwedge^k T^\ast M$ as $$\mathcal{R}_k \omega=-\sum_{i,j}e^i\wedge \iota(e_j)\left(R(e_i,e_j)\omega\right)$$ for any $\omega\in\bigwedge^k T^\ast M$, where $\{e_1,\ldots,e_n\}$ is an orthonormal basis of $TM$, $\{e^1,\ldots,e^n \}$ is its dual and $R(e_i,e_j)\omega$ is defined by $$R(e_i,e_j)\omega=\nabla_{e_i}\nabla_{e_j}\omega-\nabla_{e_j}\nabla_{e_i}\omega-\nabla_{[e_i,e_j]}\omega\in \Gamma(\bigwedge^k T^\ast M).$$
Note that if $k=1$, then we have $\mathcal{R}_1 \omega={\mathop{\mathrm{Ric}}\nolimits}(\omega,\cdot)$ for any $\omega\in\Gamma(T^\ast M)$.
The Bochner-Weitzenböck formula is stated as follows:
\[p2b\] For any $\omega\in\Gamma (\bigwedge^k T^\ast M)$, we have $$\Delta\omega=\nabla^\ast \nabla \omega+\mathcal{R}_k \omega.$$
In particular, we have the following theorem when $k=1$:
For any $\omega\in\Gamma(T^\ast M)$, we have $$\Delta \omega =\nabla^\ast \nabla \omega + {\mathop{\mathrm{Ric}}\nolimits}(\omega,\cdot).$$
Let us do some calculations of differential forms.
\[p2c\] Let $(M,g)$ be a Riemannian manifold of dimension $n$. Take a vector field $X\in\Gamma(TM)$, a $p$-form $\omega\in\Gamma(\bigwedge^p T^\ast M)$ $(p\geq 1)$ and a local orthonormal bases $\{e_1,\ldots,e_n\}$ of $TM$.
- We have $$\mathcal{R}_{p-1}(\iota(X)\omega)=\iota(X) \mathcal{R}_p \omega+\iota({\mathop{\mathrm{Ric}}\nolimits}(X))\omega+2\sum_{i=1}^n\iota(e_i)(R(X,e_i)\omega).$$
- We have $$\Delta (\iota(X)\omega)=\iota(\Delta X)\omega+\iota(X)\Delta \omega
+2\sum_{i=1}^n\iota(e_i) (R(X,e_i)\omega)-2\sum_{i=1}^n\iota(\nabla_{e_i}X) (\nabla_{e_i}\omega).$$
- We have $$\sum_{i=1}^n\iota(e_i) (R(X,e_i)\omega)
=-\nabla_X d^\ast \omega +d^\ast \nabla_X \omega+\sum_{i,j=1}^n \langle \nabla_{e_j} X, e_i\rangle\iota(e_j)\nabla_{e_i}\omega.$$
Let $\{e^1,\ldots,e^n\}$ be the dual basis of $\{e_1,\ldots,e_n\}$.
We first show (i). If $p=1$, both sides are equal to $0$. Let us assume $p\geq 2$. We have $$\label{2c}
\begin{split}
&\iota({\mathop{\mathrm{Ric}}\nolimits}(X))\omega\\
=&\frac{1}{(p-1)!}\sum_{i,i_1,\ldots,i_{p-1}}\omega(R(X,e_i)e_i,e_{i_1},\cdots,e_{i_{p-1}})e^{i_1}\wedge\cdots\wedge e^{i_{p-1}}\\
=&\frac{-1}{(p-1)!}\sum_{i,i_1,\ldots,i_{p-1}} (R(X,e_i)\omega)(e_i,e_{i_1},\ldots,e_{i_n})e^{i_1}\wedge\cdots\wedge e^{i_{p-1}}\\
&-\frac{1}{(p-1)!}\sum_{i,i_1,\ldots,i_{p-1}} \sum_{l=1}^{p-1} \omega(e_i,e_{i_1},\cdots,R(X,e_i)e_{i_l},\ldots,e_{i_{p-1}})e^{i_1}\wedge\cdots\wedge e^{i_{p-1}}\\
=&-\sum_{i=1}^n\iota(e_i)(R(X,e_i)\omega)\\
&-\frac{1}{(p-1)!}\sum_{i,i_1,\ldots,i_{p-1}} \sum_{l=1}^{p-1} \omega(e_i,e_{i_1},\cdots,R(X,e_i)e_{i_l},\ldots,e_{i_{p-1}})e^{i_1}\wedge\cdots\wedge e^{i_{p-1}}
\end{split}$$ We calculate the second term. $$\begin{split}
-&\frac{1}{(p-1)!}\sum_{i,i_1,\ldots,i_{p-1}} \sum_{l=1}^{p-1} \omega(e_i,e_{i_1},\cdots,R(X,e_i)e_{i_l},\ldots,e_{i_{p-1}})e^{i_1}\wedge\cdots\wedge e^{i_{p-1}}\\
=&\frac{1}{(p-1)!}\sum_{l=1}^{p-1} \sum_{i,j,i_1,\ldots,i_{p-1}}\langle R(e_j,e_{l_l})X,e_i\rangle\omega(e_i,e_j,e_{i_1},\cdots,\widehat{e_{i_l}},\ldots,e_{i_{p-1}})\\
&\qquad\qquad\qquad \qquad\qquad\qquad\qquad\qquad\qquad\qquad
e^{i_l}\wedge e^{i_1}\wedge\cdots\wedge\widehat{e^{i_l}}\wedge\cdots\wedge e^{i_{p-1}}\\
=&\sum_{j,k} e^k\wedge\iota(e_j)\iota(R(e_j,e_{k})X)\omega\\
=&\sum_{j,k} e^k\wedge\iota(e_j)R(e_j,e_{k})(\iota(X)\omega)-\sum_{j,k} e^k\wedge\iota(e_j)\iota(X)R(e_j,e_{k})\omega\\
=&\mathcal{R}_{p-1}(\iota(X)\omega)-\iota(X)\mathcal{R}_{p}\omega
-\sum_{j=1}^n \iota(e_j)(R(X,e_j)\omega)
\end{split}$$ Combining this and (\[2c\]), we get (i).
Let us show (ii). We have $$\nabla^\ast \nabla \iota(X)\omega
=\iota(\nabla^\ast \nabla X)\omega
-2\sum_{i} \iota(\nabla_{e_i}X)\nabla_{e_i}\omega+\iota(X)\nabla^\ast \nabla\omega.$$ Thus, by (i), we get $$\begin{split}
\Delta ( \iota(X)\omega)
=&\nabla^\ast \nabla \iota(X)\omega+\mathcal{R}_{p-1}\iota(X)\omega\\
=&\iota(\nabla^\ast \nabla X)\omega-2\sum_{i} \iota(\nabla_{e_i}X)\nabla_{e_i}\omega+\iota(X)\nabla^\ast \nabla\omega+\mathcal{R}_{p-1}\iota(X)\omega\\
=&\iota(\Delta X)\omega+\iota(X)\Delta \omega
+2\sum_{i=1}^n\iota(e_i) (R(X,e_i)\omega)-2\sum_{i=1}^n\iota(\nabla_{e_i}X) (\nabla_{e_i}\omega).
\end{split}$$ This gives (ii).
Finally, we show (iii). We have $$\begin{split}
\sum_{i=1}^n \iota(e_i)(R(X,e_i)\omega)
=&\sum_{i=1}^n \iota(e_i)\left(\nabla_{X}\nabla_{e_i}\omega-\nabla_{e_i}\nabla_X\omega-\nabla_{\nabla_{X} e_i}\omega+\nabla_{\nabla_{e_i}X}\omega\right)\\
=&-\nabla_{X}d^\ast \omega+d^\ast \nabla_X\omega+\sum_{i,j=1}^n \langle \nabla_{e_j} X, e_i\rangle\iota(e_j)\nabla_{e_i}\omega.
\end{split}$$ This gives (iii).
When $\omega$ is parallel, we have the following corollary.
\[p2d\] Let $(M,g)$ be a Riemannian manifold of dimension $n$. Take a vector field $X\in\Gamma(TM)$ and a parallel $p$-form $\omega\in\Gamma(\bigwedge^p T^\ast M)$ $(p\geq 1)$.
- We have $$\mathcal{R}_{p-1}(\iota(X)\omega)=\iota({\mathop{\mathrm{Ric}}\nolimits}(X))\omega.$$
- We have $$\Delta (\iota(X)\omega)=\iota(\Delta X)\omega.$$
Finally, we give some easy equations for later use. Let $(M,g)$ be a Riemannian manifold of dimension $n$. Take a local orthonormal basis $\{e_1,\ldots,e_n\}$ of $TM$. Let $\{e^1,\ldots,e^n\}$ be its dual. For any $\omega,\eta\in\Gamma(\bigwedge^k T^\ast M)$, we have $$\sum_{i=1}^n \langle e^i\wedge \omega, e^i\wedge \eta \rangle=(n-k)\langle\omega,\eta\rangle,
\quad \sum_{i=1}^n \langle \iota(e_i)\omega, \iota(e_i) \eta \rangle=k\langle\omega,\eta\rangle.$$ For any $\alpha_1,\ldots,\alpha_k\in \Gamma(T^\ast M)$, we have $$Q_1(\alpha_1\wedge\cdots\wedge\alpha_k)=\left(\frac{1}{k}\right)^{1/2}\sum_{i=1}^k(-1)^{i-1}\alpha_i\otimes\alpha_1\wedge\cdots\wedge\widehat{\alpha_i}\wedge\cdots\wedge\alpha_k.$$ Since $Q_1$ preserves the norms, we have $$\label{q1k}
\begin{split}
&k\left|\alpha_1\wedge\cdots\wedge\alpha_k\right|^2\\
=&\left|\sum_{i=1}^k(-1)^{i-1}\alpha_i\otimes\alpha_1\wedge\cdots\wedge\widehat{\alpha_i}\wedge\cdots\wedge\alpha_k\right|^2
\end{split}$$ for any $\alpha_1,\ldots,\alpha_k\in \Gamma(T^\ast M)$.
Suppose that $M$ is oriented. For any $k$, the Hodge star operator $\ast\colon \bigwedge^k T^\ast M\to \bigwedge^{n-k} T^\ast M$ is defined so that $$\langle\ast\omega,\eta \rangle V_g=\omega\wedge\eta$$ for all $\omega\in\Gamma(\bigwedge^k T^\ast M)$ and $\eta\in\Gamma(\bigwedge^{n-k} T^\ast M)$, where $V_g$ denotes the volume form on $(M,g)$. For any $\alpha\in\Gamma(T^\ast M)$, $\omega\in\Gamma(\bigwedge^k T^\ast M)$ and $\eta\in\Gamma(\bigwedge^{k-1} T^\ast M)$, we have $$\begin{aligned}
\langle\ast(\omega \wedge \alpha),\eta\rangle V_g&=\omega \wedge \alpha \wedge \eta,\\
\langle\iota(\alpha)\ast \omega,\eta\rangle V_g=\langle\ast \omega,\alpha\wedge \eta\rangle V_g&=\omega\wedge \alpha \wedge \eta.\end{aligned}$$ Thus, we get $$\label{hstar2}
\ast(\omega \wedge \alpha)=\iota(\alpha)\ast \omega.$$ Therefore, for any $\alpha,\beta\in\Gamma(T^\ast M)$ and $\omega,\eta\in\Gamma(\bigwedge^k T^\ast M)$, we have $$\begin{split}
&\langle\iota (\alpha)\omega,\iota(\beta)\eta\rangle
=\langle\omega,\alpha \wedge \iota(\beta)\eta\rangle\\
=&-\langle\beta \wedge \omega,\alpha \wedge \eta\rangle+\langle\alpha,\beta\rangle\langle\omega,\eta\rangle
=-\langle\iota(\beta)\ast \omega,\iota(\alpha)\ast \eta\rangle+\langle\alpha,\beta\rangle\langle\omega,\eta\rangle,
\end{split}$$ and so $$\label{hstar}
\langle\iota (\alpha)\omega,\iota(\beta)\eta\rangle+\langle\iota(\beta)\ast \omega,\iota(\alpha)\ast \eta\rangle=\langle\alpha,\beta\rangle\langle\omega,\eta\rangle.$$
Parallel $p$-form
=================
In this section, we consider Riemannian manifolds with a non-trivial parallel differential form. The reader who is interested only in the proof of the main theorems can skip this section.
Bochner-Reilly-Grosjean Formula
-------------------------------
The aim of this subsection is to give an easy proof of what Grosjean called a new Bochner-Reilly formula [@gr Proposition 3.1] when the Riemannian manifold has a non-trivial parallel $p$-form $\omega$. In section 4, we estimate the error terms when the manifold has no boundary and $\omega$ is not parallel.
\[p3a\] Let $(M,g)$ be a compact $n$-dimensional Riemannian manifold possibly with a smooth boundary $(\partial M,g')$, and let $\nu$ be the outward unit normal vector field. For any $f\in C^\infty(M)$ and any parallel $p$-form $\omega$ $(1\leq p \leq n-1)$ on $M$, we have $$\begin{split}
&\int_M |T (\iota(\nabla f)\omega)|^2\,d\mu_g\\
=&\frac{p-1}{p}\int_M\langle\iota(\nabla f)\omega, \iota(\nabla\Delta f)\omega\rangle \,d\mu_g-\int_M \langle\iota({\mathop{\mathrm{Ric}}\nolimits}(\nabla f))\omega,\iota(\nabla f)\omega\rangle\,d\mu_g\\
&-\frac{1}{p}\int_{\partial M} \langle \iota(\nu)d(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle\,d\mu_{g'}
+\int_{\partial M} \langle\nabla_{\nu}(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle\,d\mu_{g'}.
\end{split}$$
Since $d^\ast \iota(\nabla f) \omega=-d^\ast d^\ast(f\omega)=0$, we have $$\label{3a}
\begin{split}
&\int_M \langle\iota({\mathop{\mathrm{Ric}}\nolimits}(\nabla f))\omega,\iota(\nabla f)\omega\rangle\,d\mu_g\\
=&\int_M \langle\mathcal{R}_{p-1}(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle\,d\mu_g\\
=&\int_M \langle\Delta(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle\,d\mu_g
-\int_M \langle\nabla^\ast\nabla(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle\,d\mu_g\\
=&\int_M \langle d(\iota(\nabla f)\omega),d(\iota(\nabla f)\omega)\rangle\,d\mu_g
-\int_M \langle\nabla(\iota(\nabla f)\omega),\nabla(\iota(\nabla f)\omega)\rangle\,d\mu_g\\
-&\int_{\partial M} \langle \iota(\nu)d(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle\,d\mu_{g'}
+\int_{\partial M} \langle\nabla_{\nu}(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle\,d\mu_{g'}
\end{split}$$ by Corollary \[p2d\] (i), Bochner-Weitzenböck formula and the divergence theorem. By (\[2b\]) and Corollary \[p2d\] (ii), we have $$\label{3b}
\begin{split}
&\int_M \langle d(\iota(\nabla f)\omega),d(\iota(\nabla f)\omega)\rangle\,d\mu_g
-\int_M \langle\nabla(\iota(\nabla f)\omega),\nabla(\iota(\nabla f)\omega)\rangle\,d\mu_g\\
=&\frac{p-1}{p}\int_M \langle d(\iota(\nabla f)\omega),d(\iota(\nabla f)\omega)\rangle\,d\mu_g-\int_M |T(\iota(\nabla f)\omega)|^2\,d\mu_g\\
=&\frac{p-1}{p}\int_M \langle \iota(\nabla\Delta f)\omega),\iota(\nabla f)\omega\rangle\,d\mu_g
\\
&\quad +\frac{p-1}{p}\int_{\partial M} \langle \iota(\nu)d(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle\,d\mu_{g'}-\int_M |T(\iota(\nabla f)\omega)|^2\,d\mu_g
\end{split}$$ By (\[3a\]) and (\[3b\]), we get the proposition.
Estimate and Equality Case
--------------------------
In this subsection, we give more general result than Theorem \[grosjean\] without assuming positive Ricci curvature.
For any closed Riemannian manifold $(M,g)$, we define $$\Omega_1(g)=\sup\left\{ \frac{\int_M {\mathop{\mathrm{Ric}}\nolimits}(\nabla f, \nabla f)\,d\mu_g}{\int_M (\Delta f)^2\,d\mu_g}: \text{ $f$ is a non-constant function on $M$}\right\}.$$
By the Bochner formula, we always have $$\Omega_1(g)\leq\frac{n-1}{n},$$ where $n=\dim M$, and $$\|\Delta f\|_2^2\leq \frac{1}{1-\Omega_1(g)}\|\nabla^2 f\|_2^2$$ for all $f\in C^\infty(M)$. Since $$\|\nabla^2 f+\frac{\Delta f}{n}g\|_2^2=\|\nabla^2 f\|_2^2-\frac{1}{n}\|\Delta f\|_2^2,$$ we have $$\|\nabla^2 f\|_2^2\leq \frac{n}{n-\frac{1}{1-\Omega_1(g)}}\|\nabla^2 f+\frac{\Delta f}{n}g\|_2^2$$ for all $f\in C^\infty(M)$ if $\Omega_1(g)< (n-1)/n$. If $\Omega_1(g)=(n-1)/n$, then $\Omega_1(g)$ is attained by a non-constant function $f\in C^\infty(M)$ such that $\nabla^2 f+(\Delta f/n)g=0$, and so $(M,g)$ is isometric to $S^n$ with a rotationally symmetric metric by the Tashiro theorem [@T1] (see also [@Ai Property A]). If ${\mathop{\mathrm{Ric}}\nolimits}_g\geq k g$ ($k>0$), we easily get $\Omega_1(g)>0$ and $$\label{3c}
\lambda_1(g)\geq\frac{k}{\Omega_1(g)}.$$ If ${\mathop{\mathrm{Ric}}\nolimits}_g= k g$ ($k>0$), then we have $$\label{3ca}
\lambda_1(g)=\frac{k}{\Omega_1(g)}.$$ The following proposition is the main result of this subsection.
\[p3c\] Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold. Assume that there exists a non-trivial parallel $p$-form $\omega$ on $M$ $(1\leq p\leq n/2)$. Then, we have $$\label{3d}
\Omega_1(g)\leq \frac{n-p-1}{n-p}.$$
Moreover, if either $p\neq \frac{n}{2}$ or $n\geq 6$ and if in addition $M$ is simply connected, then the equality in $(\ref{3d})$ implies that $(M,g)$ is isometric to a product $(S^{n-p},g_r)\times (X,g')$, where $g_r$ is some rotationally symmetric metric on $S^{n-p}$ and $(X,g')$ is a $p$-dimensional closed Riemannian manifold.
We first show (\[3d\]). By taking the two-sheeted orientable Riemannian covering of $(M,g)$ if necessary, we can assume that $(M,g)$ is oriented. Then, we have $$\begin{split}
\int_M \langle\iota({\mathop{\mathrm{Ric}}\nolimits}(\nabla f))\omega,\iota(\nabla f)\omega\rangle\,d\mu_g
\leq &\frac{p-1}{p}\int_M\langle\iota(\nabla\Delta f)\omega, \iota(\nabla f)\omega\rangle \,d\mu_g,\\
\int_M \langle\iota({\mathop{\mathrm{Ric}}\nolimits}(\nabla f))\ast\omega,\iota(\nabla f)\ast\omega\rangle\,d\mu_g
\leq &\frac{n-p-1}{n-p}\int_M\langle\iota(\nabla\Delta f)\ast\omega, \iota(\nabla f)\ast\omega\rangle \,d\mu_g.
\end{split}$$ Thus, we get $$\begin{split}
\int_M {\mathop{\mathrm{Ric}}\nolimits}(\nabla f,\nabla f)\,d\mu_g
\leq &\frac{n-p-1}{n-p}\int_M(\Delta f)^2 \,d\mu_g
\end{split}$$ by (\[hstar\]). This implies the estimate (\[3d\]).
We next consider the equality case. Suppose that $M$ is simply connected. Let $$TM=\bigoplus_{i=1}^k E_i$$ be the irreducible decomposition of the holonomy representation, and let $$(M,g)=(M_1,g_1)\times\cdots\times (M_k,g_k)$$ be the corresponding de Rham decomposition. There exist non-negative integers $p_1,\ldots,p_k\in\mathbb{Z}_{\geq 0}$ such that $p_1+\cdots+p_k=p$ and the $\bigwedge^{p_1} E^1\otimes\cdots\otimes\bigwedge^{p_k} E^k$-component of $\omega$ is non-zero and parallel, where $E^i$ is the sub-bundle of $T^\ast M$ that corresponds to $E_i$. Thus, we can assume $\omega\in\Gamma(\bigwedge^{p_1} E^1\otimes\cdots\otimes\bigwedge^{p_k} E^k)$.
Take $i$ with $p_i\neq 0$. Let us show that there exists a non-trivial parallel $p_i$-form on $M_i$. Take some $x\in M$ and decompose $\omega_x$ as $$\omega_x=\sum_{j=1}^l \eta_j\wedge\gamma_j,$$ where $\eta_j\in \bigwedge^{p_i} E^i_x$ and $$\gamma_j\in\bigwedge^{p_1} E^1_x\otimes\cdots\otimes\bigwedge^{p_{i-1}} E^{i-1}_x\otimes\bigwedge^{p_{i+1}} E^{i+1}_x\otimes\cdots\otimes\bigwedge^{p_k} E^k_x$$ with $\langle\gamma_j,\gamma_k\rangle=\delta_{jk}$ for all $j,k\in\{1,\ldots,l\}$. Then, $\eta_j$ is invariant under the holonomy representation of $M_i$ for each $j$. Thus, $\eta_j$ defines a parallel $p_i$-form on $M_i$. Therefore, there exists a non-trivial parallel $p_i$-form $\omega_i$ on $M_i$. Then, the eigenspace of the symmetric form on $T M_i$ $$\langle\iota(\cdot)\omega_i, \iota(\cdot)\omega_i\rangle$$ is invariant under the the holonomy representation. Since $T M_i$ is irreducible, there exists a positive number $\mu_i>0$ such that $$\label{wnondeg}
\langle\iota(\cdot)\omega_i, \iota(\cdot)\omega_i\rangle=\mu_i \langle\cdot,\cdot\rangle.$$ Thus, we get $$\begin{split}
\int_{M_i}{\mathop{\mathrm{Ric}}\nolimits}(\nabla f,\nabla f)\,d\mu_{g_i}\leq \frac{p_i-1}{p_i}\int_{M_i} (\Delta f)^2 \,d\mu_{g_i}
\end{split}$$ for all $f\in C^\infty (M_i)$ by Proposition \[p3a\], and so $$\label{3e}
\Omega_1(g_i)\leq\frac{p_i-1}{p_i}.$$ By considering $\ast \omega$, we also have $$\label{3f}
\Omega_1(g_i)\leq\frac{\dim M_i-p_i-1}{\dim M_i-p_i}$$ if $p_i\neq \dim M_i$.
By (\[3e\]), (\[3f\]) and [@Ai Proposition 2.4], we get $$\label{3g}
\Omega_1(g)=\max\{\Omega_1(g_1),\ldots,\Omega_1(g_k)\}\leq \max_i \left\{\frac{\overline{p}_i-1}{\overline{p}_i}\right\},$$ where we put
[align\*]{} &{p\_i,M\_i-p\_i}&& (p\_i0,M\_i),\
&M\_i &&(p\_i=0,M\_i).
Suppose that $$\label{3h}
\Omega_1(g)= \frac{n-p-1}{n-p}$$ and either $p\neq \frac{n}{2}$ or $n\geq 6$. Without loss of generality, we can assume that $\dim M_1=\max_i\{\dim M_i\}$. If $\dim M_1< n-p$, then we get $$\Omega_1(g)\leq\frac{n-p-2}{n-p-1} < \frac{n-p-1}{n-p}$$ by (\[3g\]). This contradicts to (\[3h\]), and so we have $\dim M_1\geq n-p$.
We consider the following three cases:
- $n-p<\dim M_1<n$,
- $\dim M_1=n-p$,
- $\dim M_1=n$.
We first suppose that $n-p<\dim M_1<n$. Then, $p_2+\cdots+p_k\leq n-\dim M_1<p$, and so $p_1\neq 0$. Moreover, we have $p_1\leq p<\dim M_1$. Thus, we have $$\overline{p}_1=\min\{p_1,\dim M_1- p_1\}\leq \frac{\dim M_1}{2}< \frac{n}{2}\leq n-p.$$ Since $\dim M_i<n-p$ for all $i=\{2,\ldots,k\}$, we get $$\Omega_1(g)< \frac{n-p-1}{n-p}$$ by (\[3g\]). This contradicts to (\[3h\]).
We next suppose that $\dim M_1=n$. Then, we have $M=M_1$. Since we have $\Omega_1(g)\leq (p-1)/p$ and $p\leq n-p$, we get $$\label{3i}
p=n-p=n/2\geq 3$$ by (\[3h\]). Since there exists a non-trivial parallel $p$-form, the holonomy group of $(M,g)$ is not equal to $\mathrm{SO}(n)$. If ${\mathop{\mathrm{Ric}}\nolimits}\leq 0$, then $\Omega_1(g)=0$, and so we have one of the following by the Berger classification theorem:
- $(M,g)$ is a Kähler manifold,
- $(M,g)$ is a quaternionic Kähler manifold,
- $(M,g)$ is a symmetric space.
If $(M,g)$ is Kähler manifold, then there exists a Kähler form. Thus, we get $\Omega_1(g)\leq \frac{1}{2}<\frac{n-p-1}{n-p}$ by (\[3g\]) and (\[3i\]). This contradicts to (\[3h\]). If $(M,g)$ is quaternionic Kähler manifold of dimension $n=4d$ ($d\geq 2$), then $(M,g)$ is a positive Einstein manifold $${\mathop{\mathrm{Ric}}\nolimits}=c g \quad (c>0)$$ by $\Omega_1(g)>0$ and [@Be Theorem 14.39]. Thus, we get $$\Omega_1(g)=\frac{c}{\lambda_1(g)}\leq \frac{1}{2}\frac{d+2}{d+1}\leq\frac{2}{3}<\frac{3}{4}$$ by (\[qk\]) and (\[3ca\]). This contradicts to (\[3h\]) and $n-p=p=2d\geq 4$. Finally, we suppose that $(M,g)$ is a symmetric space. Since $(M,g)$ has a non-trivial parallel $p$-form, we have $M\neq S^n$. Thus, by [@bms Theorem 1.1], there exists no non-parallel Killing $(p-1)$-form on $(M,g)$, and so $T(\iota(\nabla f)\omega)\neq 0$ for any non-constant function $f\in C^\infty(M)$. Since $\Omega_1(g)$ is attained by some smooth function [@Ai Lemma 2.1], we get $\Omega_1(g)<(p-1)/p$ by (\[wnondeg\]) and Proposition \[p3a\]. This contradicts to (\[3h\]).
Therefore, we get $\dim M_1=n-p$. Put $(X,g')=(M_2,g_2)\times\cdots\times (M_k,g_k)$. Then, we have $$\Omega_1(g)=\max\{\Omega_1(g_1),\Omega_1(g')\}$$ by [@Ai Proposition 2.4], and so either $
\Omega_1(g_1)=(n-p-1)/(n-p)
$ or $
\Omega_1(g')=(n-p-1)/(n-p).
$ If $
\Omega_1(g_1)=(n-p-1)/(n-p)
$ (resp. $
\Omega_1(g')=(n-p-1)/(n-p)
$), then $(M_1,g_1)$ (resp. $(X,g')$) is isometric to $(S^{n-p},g_r)$, where $g_r$ is rotationally symmetric metric on $S^{n-p}$.
As a corollary, we get the following:
\[p3d\] Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold. Assume that ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-p-1)g$ and there exists a non-trivial parallel $p$-form on $M$ $(2\leq p\leq n/2)$. Then, we have $$\label{3j}
\lambda_1(g)\geq n-p.$$
Moreover, if either $p\neq n/2$ or $n\geq 6$ and if in addition $M$ is simply connected, the equality in (\[3j\]) implies that $(M,g)$ is isometric to a product $S^{n-p}\times X$, where $X$ is a $p$-dimensional closed Riemannian manifold.
By (\[3c\]) and (\[3d\]), we get $$\label{3k}
\lambda_1(g)\geq\frac{1}{\Omega_1(g)}(n-p-1)\geq n-p.$$ This implies (\[3j\]).
Suppose that $M$ is simply connected, $\lambda_1(g)=n-p$ and either $p\neq \frac{n}{2}$ or $n\geq6$. Then, we have $\Omega_1(g)=(n-p-1)/(n-p)$ by (\[3k\]), and so $(M,g)$ is isometric to a product $(S^{n-p},g_r)\times (X,g')$, where $g_r$ is some rotationally symmetric metric on $S^{n-p}$ and $(X,g')$ is a $p$-dimensional closed Riemannian manifold by Proposition \[p3c\]. Since we have ${\mathop{\mathrm{Ric}}\nolimits}_{g_r}\geq (n-p-1)g_r$, ${\mathop{\mathrm{Ric}}\nolimits}_{g'}\geq (n-p-1)g'$ and $n-p=\lambda_1(g)=\min\{\lambda_1(g_r),\lambda_1(g')\}$, we get that either $(S^{n-p},g_r)$ or $(X,g')$ is isometric to $S^{n-p}(1)$ by the Lichnerowicz-Obata theorem (Theorem 1).
If we assume more strong condition on eigenvalues, then the assumption that the manifold is simply connected can be removed.
\[p3d2\] Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold. Assume that ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-p-1)g$ and there exists a non-trivial parallel $p$-form on $M$ $(2\leq p< n/2)$. If $$\lambda_{n-p+1}(g)= n-p,$$ then $(M,g)$ is isometric to a product $S^{n-p}\times (X,g')$, where $(X,g')$ is a $p$-dimensional closed Riemannian manifold.
Let $f_k$ be the $k$-th eigenfunction of the Laplacian on $S^{n-p}$. Note that the functions $f_1,\ldots,f_{n-p+1}$ are height functions.
By Corollary \[p3d\], the universal cover $(\widetilde{M},\tilde{g})$ of $(M,g)$ is isometric to a product $S^{n-p}\times (X,g')$, where $(X,g')$ is a $p$-dimensional closed Riemannian manifold. We regard the function $f_i$ as a function on $\widetilde{M}$. Since $\lambda_{n-p+1}(g)= n-p$, each $f_i\in C^\infty(\widetilde{M})$ ($i=1,\ldots,n-p+1$) is a pull back of some function on $M$. Thus, the covering transformation preserves $f_1,\ldots,f_{n-p+1}$. Therefore, the covering transformation does not act on $S^{n-p}$, and so we get the corollary.
The almost version of this corollary is Main Theorem 2.
Examples
--------
In this subsection, we show that the assumption of Corollary \[p3d2\] is optimal in some sense by giving examples.
Take a positive odd integer $p$ with $p\geq 3$ and a positive integer $n$ with $n> 2p$. Put $a:=\sqrt{(p-1)/(n-p-1)}$. We define an equivalence relation $\sim$ on $S^{n-p}\times S^p(a)$ as follows: $$\begin{split}
&((x_0,\ldots,x_{n-p}),(y_0,\ldots,y_p))\sim ((x'_0,\ldots,x'_{n-p}),(y'_0,\ldots,y'_p))\\
\Leftrightarrow &\text{ there exists $k\in \mathbb{Z}$ such that}\\
&((x'_0,\ldots,x'_{n-p}),(y'_0,\ldots,y'_p))=(((-1)^k x_0, x_1,\ldots,x_{n-p}),(-1)^k(y_0,\ldots,y_p))
\end{split}$$ for any $((x_0,\ldots,x_{n-p}),(y_0,\ldots,y_p)), ((x'_0,\ldots,x'_{n-p}),(y'_0,\ldots,y'_p))\in S^{n-p}\times S^p(a)$. Then, we have the following:
\[p3e\] We have the following properties:
- $(M,g)=(S^{n-p}\times S^p(a))/\sim$ is an $n$-dimensional closed Riemannian manifold with a non-trivial parallel $p$-form.
- ${\mathop{\mathrm{Ric}}\nolimits}= (n-p-1)g$.
- $\lambda_{n-p}(g)=n-p$.
- $(M,g)$ is not isometric to any product Riemannian manifolds.
Let $\omega$ be the volume form on $S^p(a)$. Since the action on $S^{n-p}\times S^p(a)$ preserves $\omega$, there exists a non-trivial parallel $p$-form on $(M,g)$. We also denote it by $\omega$. Since the action on $S^{n-p}\times S^p(a)$ preserves the function $$x_i \colon S^{n-p}\times S^p(a)\to \mathbb{R},\,((x_0,\ldots,x_{n-p}),(y_0,\ldots,y_p))\mapsto x_i$$ for each $i=1,\ldots,n-p$, we have $\lambda_{n-p}(g)=n-p$.
Suppose that $(M,g)$ is isometric to a product $(M^{n-k}_1,g_1)\times (M^{k}_2,g_2)$ ($k\leq n-k$) for some $(n-k)$ and $k$-dimensional closed Riemannian manifolds $(M_1,g_1)$ and $(M_2,g_2)$. Since we have the irreducible decomposition $T_{(x,y)} M\cong T_x S^{n-p}\oplus T_y S^p(a)$ of the restricted holonomy action, we get $k=p$. Since $\lambda_1(g)=n-p$, we have that $(M_1,g_1)$ is isometric to $S^{n-p}$. Thus, we get $\lambda_{n-p+1}(g)=n-p$. However the action on $S^{n-p}\times S^p(a)$ does not preserve the function $$x_0\colon S^{n-p}\times S^p(a)\to \mathbb{R},\,((x_0,\ldots,x_{n-p}),(y_0,\ldots,y_p))\mapsto x_0,$$ and so $\lambda_{n-p+1}(g)\neq n-p$. This is a contradiction.
We next define an equivalence relation $\sim'$ on $S^{n-p}\times S^p(a)$ as follows: $$\begin{split}
&((x_0,\ldots,x_{n-p}),(y_0,\ldots,y_p))\sim' ((x'_0,\ldots,x'_{n-p}),(y'_0,\ldots,y'_p))\\
\Leftrightarrow &\text{ there exists $k\in \mathbb{Z}$ such that}\\
&((x'_0,\ldots,x'_{n-p}),(y'_0,\ldots,y'_p))=(((-1)^k x_0, (-1)^k x_1,x_2,\ldots,x_{n-p}),(-1)^k(y_0,\ldots,y_p))
\end{split}$$ for any $((x_0,\ldots,x_{n-p}),(y_0,\ldots,y_p)), ((x'_0,\ldots,x'_{n-p}),(y'_0,\ldots,y'_p))\in S^{n-p}\times S^p(a)$. Similarly to Proposition \[p3e\], we have the following proposition:
\[p3f\] We have the following properties:
- $(M',g')=(S^{n-p}\times S^p(a))/\sim'$ is an $n$-dimensional closed orientable Riemannian manifold with a non-trivial parallel $p$-form.
- ${\mathop{\mathrm{Ric}}\nolimits}= (n-p-1)g'$.
- $\lambda_{n-p-1}(g')=n-p$.
- $(M',g')$ is not isometric to any product Riemannian manifolds.
Almost Parallel $p$-form
========================
In this section, we show Main Theorem 1 and Main Theorem 3. Recall that $\lambda_1(\Delta_{C,p})$ denotes the first eigenvalue of the connection Laplacian acting on $p$-forms, and $$\Delta_{C,p}:=\nabla^\ast\nabla \colon \Gamma(\bigwedge^p T^\ast M)\to\Gamma(\bigwedge^p T^\ast M).$$ It is enough to show Main Theorem 1 when $\lambda_1(\Delta_{C,p})\leq 1$. Note that we always have $$\lambda_1(\Delta_{C,1})\geq 1$$ if ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-1)g$.
Error Estimates
---------------
In this subsection, we give error estimates about Proposition \[p3a\]. Lemma \[p4e\] (vii) corresponds to Proposition \[p3a\]. We list the assumptions of this subsection. We mention that most techniques in this paper can be used under the assumption ${\mathop{\mathrm{Ric}}\nolimits}_g\geq -Kg$ and ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$.
In this subsection, we assume the following:
- $(M,g)$ is an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}_g\geq -Kg$ and ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$ for some positive real numbers $K>0$ and $D>0$.
- $1\leq k \leq n-1$.
- A $k$-form $\omega\in \Gamma(\bigwedge^k T^\ast M)$ satisfies $\|\omega\|_2=1$, $\|\omega\|_\infty\leq L_1$ and $\|\nabla \omega\|_2^2\leq \lambda$ for some $L_1>0$ and $0\leq \lambda\leq 1$.
- A function $f\in C^\infty(M)$ satisfies $\|f\|_{\infty}\leq L_2\|f\|_2$, $\|\nabla f\|_{\infty}\leq L_2\|f\|_2$ and $\|\Delta f\|_2\leq L_2\|f\|_2$ for some $L_2>0$.
Note that we have $$\label{4a0}
\|\nabla^2 f\|_2^2=\|\Delta f\|_2^2-\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M {\mathop{\mathrm{Ric}}\nolimits}(\nabla f,\nabla f)\,d\mu_g\leq (1+K)L^2_2\|f\|_2^2$$ by the Bochner formula.
We first show the following:
\[p4c\] There exists a positive constant $C(n,K,D)>0$ such that $\||\omega|-1\|_2\leq C \lambda^{1/2}$ holds.
Put $$\overline{\omega}:=\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M |\omega| \,d\mu_g.$$ Since we have $|\omega|\in W^{1,2}(M)$, we get $$\||\omega|-\overline{\omega}\|_2^2\leq \frac{1}{\lambda_1(g)}\|\nabla|\omega|\|_2^2\leq \frac{1}{\lambda_1(g)}\|\nabla\omega\|_2^2\leq\frac{\lambda}{\lambda_1(g)}$$ by the Kato inequality. Thus, by the Li-Yau estimate [@SY p.116], we have $$\||\omega|-\overline{\omega}\|_2\leq C\lambda^{1/2},$$ and so $$|1-\overline{\omega}|=\left|\|\omega\|_2-\|\overline{\omega}\|_2\right|\leq \||\omega|-\overline{\omega}\|_2\leq C\lambda^{1/2}.$$ Therefore, we get $
\||\omega|-1\|_2\leq C\lambda^{1/2}.
$
Let us give error estimates about Proposition \[p3a\].
\[p4d\] There exists a positive constant $C=C(n,k,K,D,L_1,L_2)>0$ such that the following properties hold:
- We have $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M |d^{\ast}(\iota(\nabla f)\omega)|^2\,d\mu_g
\leq C\|f\|_2^2\lambda.$$
- We have $$\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \Big(\langle \iota({\mathop{\mathrm{Ric}}\nolimits}(\nabla f))\omega,\iota(\nabla f)\omega\rangle -\langle \mathcal{R}_{k-1}(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle \Big)\,d\mu_g\right|
\leq C\|f\|_2^2\lambda^{1/2}.$$
- We have $$\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \Big(\langle\Delta(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle
-\langle
\iota(\nabla \Delta f)\omega,\iota(\nabla f)\omega\rangle\Big) \,d\mu_g\right|
\leq C\|f\|_2^2\lambda^{1/2}.$$
- We have $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \left|\nabla (\iota(\nabla f)\omega)-\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\omega\right|^2 \,d\mu_g\leq C\|f\|_2^2\lambda.$$
- We have $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \left|d (\iota(\nabla f)\omega)-\sum_{i=1}^n e^i\wedge \iota(\nabla_{e_i}\nabla f)\omega\right|^2 \,d\mu_g\leq C\|f\|_2^2\lambda.$$
- We have $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M |\nabla (\iota(\nabla f)\omega) |^2\,d\mu_g\leq C\|f\|_2^2.$$
- We have $$\begin{aligned}
&\Bigg|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \iota({\mathop{\mathrm{Ric}}\nolimits}(\nabla f))\omega,\iota(\nabla f)\omega\rangle
\,d\mu_g\\
&\quad-
\frac{k-1}{k} \frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \iota(\nabla\Delta f)\omega,\iota(\nabla f)\omega\rangle \,d\mu_g+\|T(\iota(\nabla f)\omega)\|_2^2\Bigg|
\leq C\|f\|_2^2\lambda^{1/2}.\end{aligned}$$
- If $M$ is oriented and $1\leq k\leq n/2$, then we have $$\begin{aligned}
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M {\mathop{\mathrm{Ric}}\nolimits}(\nabla f,\nabla f)|\omega|^2\,d\mu_g\\
\leq &
\frac{n-k-1}{n-k} \frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \nabla\Delta f,\nabla f\rangle|\omega|^2 \,d\mu_g
-\|T(\iota(\nabla f)\omega)\|_2^2
-\|T(\iota(\nabla f)\ast\omega)\|_2^2\\
&\qquad\qquad -\left(\frac{n-k-1}{n-k} -\frac{k-1}{k} \right)\|d(\iota(\nabla f)\omega)\|^2_2
+C\|f\|_2^2\lambda^{1/2}.\end{aligned}$$
Although an orthonormal basis $\{e_1,\ldots,e_n\}$ of $TM$ is defined only locally, $\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\omega$ and $\sum_{i=1}^n e^i\wedge \iota(\nabla_{e_i}\nabla f)\omega$ are well-defined as tensors.
We first prove (i). Since $$d^\ast (f\omega)=-\iota(\nabla f)\omega +f d^\ast \omega$$ and $d^\ast\circ d^\ast=0$, we have $$d^\ast (\iota(\nabla f)\omega)=-\iota(\nabla f)d^\ast \omega.$$ Thus, we get $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M |d^{\ast}(\iota(\nabla f)\omega|^2\,d\mu_g
\leq C\|\nabla f\|_{\infty}^2 \|\nabla \omega\|_2^2
\leq C\|f\|_2^2 \lambda.$$
To prove (ii) and (iii), we estimate following terms: $$\begin{aligned}
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle\iota(\nabla f)\Delta \omega,\iota(\nabla f) \omega\rangle\,d\mu_g,\\
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle\iota(\nabla f)\nabla^\ast \nabla \omega,\iota(\nabla f) \omega\rangle\,d\mu_g,\\
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle \iota(\nabla f) \mathcal{R}_k \omega, \iota(\nabla f)\omega\rangle\,d\mu_g,\\
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle \sum_{i=1}^n\iota(\nabla_{e_i}\nabla f) (\nabla_{e_i}\omega),\iota(\nabla f)\omega\rangle\,d\mu_g,\\
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle\sum_{i=1}^n\iota(e_i)(R(\nabla f,e_i)\omega),\iota(\nabla f) \omega\rangle\,d\mu_g.\end{aligned}$$
We have $$\begin{aligned}
&\int_M \langle\iota(\nabla f)\Delta \omega,\iota(\nabla f) \omega\rangle\,d\mu_g\\
=&\int_M \langle d \omega,d (d f \wedge\iota(\nabla f) \omega)\rangle\,d\mu_g+\int_M \langle d^\ast\omega,d^\ast(d f \wedge\iota(\nabla f) \omega)\rangle\,d\mu_g\end{aligned}$$ and $$\begin{aligned}
&|\langle d \omega,d (d f \wedge\iota(\nabla f) \omega)\rangle|\\
=&|\langle d \omega, \sum_{i=1}^n d f\wedge e^i \wedge\left(\iota(\nabla_{e_i}\nabla f) \omega+\iota(\nabla f) \nabla_{e_i}\omega \right)\rangle|\\
\leq& C|\nabla \omega||\nabla f|(|\nabla^2 f||\omega|+|\nabla f||\nabla \omega|),\\
&|\langle d^\ast \omega,d^\ast (d f \wedge\iota(\nabla f) \omega)\rangle|\\
=&|\langle d^\ast \omega, \sum_{i=1}^n \iota(e_i)\left( \nabla_{e_i} d f\wedge \iota(\nabla f) \omega+ d f\wedge \iota(\nabla_{e_i} \nabla f) \omega+
d f\wedge\iota(\nabla f) \nabla_{e_i}\omega \right)\rangle|\\
\leq& C|\nabla \omega||\nabla f|(|\nabla^2 f||\omega|+|\nabla f||\nabla \omega|).\end{aligned}$$ Thus, we get $$\label{4a}
\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle\iota(\nabla f)\Delta \omega,\iota(\nabla f) \omega\rangle\,d\mu_g\right|
\leq C\|f\|_2^2\lambda^{1/2}.$$
We have $$\begin{aligned}
&\int_M \langle\iota(\nabla f)\nabla^\ast\nabla \omega,\iota(\nabla f) \omega\rangle\,d\mu_g\\
=&\int_M \langle \nabla \omega,\nabla (d f \wedge\iota(\nabla f) \omega)\rangle\,d\mu_g\end{aligned}$$ and $$\begin{aligned}
&|\langle \nabla \omega,\nabla (d f \wedge\iota(\nabla f) \omega)\rangle|\\
=& C|\nabla \omega||\nabla f|(|\nabla^2 f||\omega|+|\nabla f||\nabla \omega|).\end{aligned}$$ Thus, we get $$\label{4aa}
\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle\iota(\nabla f)\nabla^\ast\nabla \omega,\iota(\nabla f) \omega\rangle\,d\mu_g\right|
\leq C\|f\|_2^2\lambda^{1/2}.$$
By Theorem \[p2b\], (\[4a\]) and (\[4aa\]), we have $$\label{4b}
\begin{split}
&\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle \iota(\nabla f) \mathcal{R}_k \omega, \iota(\nabla f)\omega\rangle\,d\mu_g\right|\\
\leq & \frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\left(\left|\int_M \langle \iota(\nabla f)\Delta \omega, \iota(\nabla f) \omega\rangle\,d\mu_g\right|+ \left|\int_M \langle \iota(\nabla f)\nabla^\ast\nabla \omega, \iota(\nabla f)\omega\rangle\,d\mu_g\right|\right)\\
\leq & C\|f\|_2^2\lambda^{1/2}.
\end{split}$$
Since $$|\langle \sum_{i=1}^n\iota(\nabla_{e_i}\nabla f) (\nabla_{e_i}\omega),\iota(\nabla f)\omega\rangle|
\leq C|\omega||\nabla f| |\nabla \omega||\nabla^2 f|,$$ we have $$\label{4c}
\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle \sum_{i=1}^n\iota(\nabla_{e_i}\nabla f) (\nabla_{e_i}\omega),\iota(\nabla f)\omega\rangle\,d\mu_g\right|
\leq C \|f\|_2^2\lambda^{1/2}.$$
To estimate $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle\sum_{i=1}^n\iota(e_i)(R(\nabla f,e_i)\omega),\iota(\nabla f) \omega\rangle\,d\mu_g,$$ we estimate the following terms: $$\begin{aligned}
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle \nabla_{\nabla f} d^\ast \omega,\iota(\nabla f) \omega\rangle\,d\mu_g,\\
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle d^\ast \nabla_{\nabla f} \omega, \iota(\nabla f)\omega\rangle\,d\mu_g,\\
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle \sum_{i,j=1}^n \langle \nabla_{e_j} \nabla f, e_i\rangle\iota(e_j)\nabla_{e_i}\omega,\iota(\nabla f)\omega\rangle\,d\mu_g.\end{aligned}$$ We have $$\begin{split}
\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle \nabla_{\nabla f} d^\ast \omega,\iota(\nabla f) \omega\rangle\,d\mu_g\right|&
=\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle d^\ast \omega,\nabla^\ast (d f\otimes\iota(\nabla f) \omega)\rangle\,d\mu_g\right|\\
&\leq C\|f\|_2^2\lambda^{1/2},\\
\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle d^\ast \nabla_{\nabla f}\omega,\iota(\nabla f) \omega\rangle\,d\mu_g\right|&
=\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle \nabla \omega, d f \otimes d (\iota(\nabla f) \omega)\rangle\,d\mu_g\right|\\
&\leq C\|f\|_2^2\lambda^{1/2}
\end{split}$$ and $$\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle \sum_{i,j=1}^n \langle \nabla_{e_j} \nabla f, e_i\rangle\iota(e_j)\nabla_{e_i}\omega,\iota(\nabla f)\omega\rangle\,d\mu_g\right|\\
\leq C\|f\|_2^2\lambda^{1/2}.$$ Thus, by Lemma \[p2c\] (iii), we get $$\label{4d}
\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)} \int_M \langle\sum_{i=1}^n\iota(e_i)(R(\nabla f,e_i)\omega),\iota(\nabla f) \omega\rangle\,d\mu_g\right|
\leq C\|f\|_2^2\lambda^{1/2}.$$
By (\[4a\]), (\[4b\]), (\[4c\]), (\[4d\]) and Lemma \[p2c\], we get (ii) and (iii).
Since $$\nabla (\iota(\nabla f)\omega)-\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\omega
=\sum_{i=1}^n e^i\otimes\iota(\nabla f)\nabla_{e_i}\omega,$$ we get (iv).
Since $$d (\iota(\nabla f)\omega)-\sum_{i=1}^n e^i\wedge \iota(\nabla_{e_i}\nabla f)\omega
=\sum_{i=1}^n e^i\wedge\iota(\nabla f)\nabla_{e_i}\omega,$$ we get (v).
Since $$\nabla (\iota(\nabla f)\omega)
=\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\omega+\sum_{i=1}^n e^i\otimes\iota(\nabla f)\nabla_{e_i}\omega,$$ we get (vi).
By Theorem \[p2b\] and (\[2b\]), we have $$\begin{split}
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \mathcal{R}_{k-1}(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle \,d\mu_g\\
=&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle (\Delta-\nabla^\ast\nabla)(\iota(\nabla f)\omega),\iota(\nabla f)\omega\rangle \,d\mu_g\\
=&\frac{k-1}{k} \frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M |d(\iota(\nabla f)\omega)|^2\,d\mu_g\\
&+\frac{n-k+1}{n-k+2} \frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M |d^\ast(\iota(\nabla f)\omega)|^2\,d\mu_g-\|T(\iota(\nabla f)\omega)\|_2^2\\
=& \frac{k-1}{k} \frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \Delta (\iota(\nabla f)\omega), \iota(\nabla f)\omega\rangle\,d\mu_g\\
&+\left(\frac{n-k+1}{n-k+2}-\frac{k-1}{k} \right)\|d^\ast(\iota(\nabla f)\omega)\|_2^2-\|T(\iota(\nabla f)\omega)\|_2^2.
\end{split}$$ Thus, by (i), (ii) and (iii), we get (vii) Finally, we prove (viii). Suppose that $M$ is oriented and $1\leq k\leq n/2$. Since $\nabla (\ast \omega)=\ast\nabla \omega$, we have $$\begin{aligned}
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \iota({\mathop{\mathrm{Ric}}\nolimits}(\nabla f))\ast\omega,\iota(\nabla f)\ast\omega\rangle
\,d\mu_g\\
\leq &
\frac{n-k-1}{n-k} \frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \iota(\nabla\Delta f)\ast\omega,\iota(\nabla f)\ast\omega\rangle \,d\mu_g-\|T(\iota(\nabla f)\ast\omega)\|_2^2+C\|f\|_2^2\lambda^{1/2}\end{aligned}$$ by (vii). Thus, by (\[hstar\]), (i), (iii) and (vii), we get $$\begin{aligned}
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M {\mathop{\mathrm{Ric}}\nolimits}(\nabla f,\nabla f)|\omega|^2\,d\mu_g\\
\leq &
\frac{n-k-1}{n-k} \frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \nabla\Delta f,\nabla f\rangle|\omega|^2 \,d\mu_g
-\|T(\iota(\nabla f)\omega)\|_2^2
-\|T(\iota(\nabla f)\ast\omega)\|_2^2\\
&\qquad -\left(\frac{n-k-1}{n-k} -\frac{k-1}{k} \right)
\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \iota(\nabla\Delta f)\omega,\iota(\nabla f)\omega\rangle \,d\mu_g
+C\|f\|_2^2\lambda^{1/2}\\
\leq &
\frac{n-k-1}{n-k} \frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \nabla\Delta f,\nabla f\rangle|\omega|^2 \,d\mu_g
-\|T(\iota(\nabla f)\omega)\|_2^2
-\|T(\iota(\nabla f)\ast\omega)\|_2^2\\
&\qquad\qquad -\left(\frac{n-k-1}{n-k} -\frac{k-1}{k} \right)\|d(\iota(\nabla f)\omega)\|^2_2
+C\|f\|_2^2\lambda^{1/2}.\end{aligned}$$ This gives (viii).
Eigenvalue Estimate
-------------------
In this subsection, we complete the proofs of Main Theorem 1 and Main Theorem 3.
We need the following $L^\infty$ estimates.
\[Linfes\] Take an integer $n\geq 2$ and positive real numbers $K>0$, $D>0$, $\Lambda>0$. Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq-Kg$ and ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$. Then, we have the following:
- For any function $f\in C^\infty(M)$ and any $\lambda\geq 0$ with $\Delta f=\lambda f$ and $\lambda\leq \Lambda$, then we have $\|\nabla f\|_\infty\leq C(n,K,D,\Lambda)\|f\|_2$ and $\|f\|_\infty\leq C(n,K,D,\Lambda)\|f\|_2$.
- For any $p$-form $\omega\in \Gamma\left(\bigwedge^p T^\ast M\right)$ and any $\lambda\geq 0$ with $\Delta_{C,p} \omega=\lambda \omega$ and $\lambda\leq \Lambda$, then we have $\|\omega\|_\infty\leq C(n,K,D,\Lambda)\|\omega\|_2$.
By the gradient estimate for eigenfunctions [@Pe1 Theorem 7.3], we get (i).
Let us show (ii). Since we have $$\Delta |\omega|^2=2\langle \Delta_{C,p} \omega, \omega \rangle-2|\nabla \omega|^2\leq 2 \Lambda |\omega|^2,$$ we get $\|\omega\|_\infty\leq C$ by [@Pe3 Proposition 9.2.7] (see also Proposition 7.1.13 and Proposition 7.1.17 in [@Pe3]). Note that our sign convention of the Laplacian is different from [@Pe3].
We use the following proposition not only for the proofs of Main Theorem 1 and Main Theorem 3 but also for other main theorems.
\[p4e\] For given integers $n\geq 4$ and $2\leq p \leq n/2$, there exists a constant $C(n,p)>0$ such that the following property holds. Let $(M,g)$ be an $n$-dimensional closed oriented Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-p-1)g$. Suppose that an integer $i\in\mathbb{Z}_{>0}$ satisfies $\lambda_i(g)\leq n-p+1$, and there exists an eigenform $\omega$ of the connection Laplacian $\Delta_{C,p}$ acting on $p$-forms with $\|\omega\|_2=1$ corresponding to the eigenvalue $\lambda$ with $0\leq \lambda\leq 1$. Then, we have $$\begin{aligned}
&\frac{n-p-1}{n-p}\lambda_i(g)\left(\lambda_i(g)-(n-p)\right)\|f_i\|^2\\
\geq&\|T(\iota(\nabla f_i)\omega)\|_2^2
+\|T(\iota(\nabla f_i)\ast\omega)\|_2^2\\
&+\left(\frac{n-p-1}{n-p} -\frac{p-1}{p} \right)\|d(\iota(\nabla f_i)\omega)\|^2_2
-C\lambda^{1/2}\|f_i\|_2^2,\end{aligned}$$ where $f_i$ denotes the $i$-th eigenfunction of the Laplacian acting on functions.
By Lemma \[p4d\] (viii), we have $$\begin{aligned}
&\frac{n-p-1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle\nabla f_i,\nabla f_i\rangle|\omega|^2\,d\mu_g\\
\leq&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M {\mathop{\mathrm{Ric}}\nolimits}(\nabla f_i,\nabla f_i)|\omega|^2\,d\mu_g\\
\leq&\frac{n-p-1}{n-p}\frac{\lambda_i(g)}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle\nabla f_i,\nabla f_i\rangle|\omega|^2\,d\mu_g
-\|T(\iota(\nabla f_i)\omega)\|_2^2
-\|T(\iota(\nabla f_i)\ast\omega)\|_2^2\\
&\qquad\qquad -\left(\frac{n-p-1}{n-p} -\frac{p-1}{p} \right)\|d(\iota(\nabla f_i)\omega)\|^2_2
+C\lambda^{1/2}\|f_i\|_2^2.\end{aligned}$$ Thus, we get the proposition by Lemma \[p4c\].
If $M$ is orientable, we get the theorem immediately by Proposition \[p4e\]. If $M$ is not orientable, we get the theorem by considering the two-sheeted orientable Riemannian covering $\pi\colon (\widetilde{M},\tilde{g})\to (M,g)$ because we have $
\lambda_1(g)\geq\lambda_1(\tilde{g})
$ and $
\lambda_1(\Delta_{C,p},g)\geq \lambda_1(\Delta_{C,p},\tilde{g}).
$
Similarly, we get Main Theorem 3 because $\lambda_1(\Delta_{C,p},g)=\lambda_1(\Delta_{C,n-p},g)$ holds if the manifold is orientable.
Pinching
========
In this section, we show the remaining main theorems. Main Theorem 2 is proved in subsection 5.5 except for the orientability, and the orientability is proved in subsection 5.7. Main Theorem 4 is proved in subsection 5.8. We list assumptions of this section.
\[asu1\] Throughout in this section except for subsection 5.1, we assume the following:
- $n\geq 5$ and $2\leq p < n/2$.
- $(M,g)$ is an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}_g\geq (n-p-1)g$.
- $C=C(n,p)>0$ denotes a positive constant depending only on $n$ and $p$.
- $\delta>0$ satisfies $\delta\leq \delta_0$ for sufficiently small $\delta_0=\delta_0(n,p)>0$.
Note that, for given real numbers $a,b$ with $0<b<a$ and a positive constant $C>0$, we can assume that $$C \delta^a\leq\delta^b.$$ For most subsections, we list additional assumptions at the beginning of them.
Useful Techniques
-----------------
In this subsection, we list some useful techniques for the pinching problems.
The following lemma is a variation of the Cheng-Yau estimate. See [@Ai2 Lemma 2.10] for the proof (see also [@Ch Theorem 7.1]).
\[chya\] Take an integer $n\geq 2$ and positive real numbers $K>0$, $D>0$, $\Lambda>0$ and $0<\epsilon_1 \leq1$. Then, there exists a positive constant $ C(n,K,D,\Lambda)>0$ such that the following property holds. Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq-Kg$ and ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$. Take a function $$f\in \bigoplus_{\lambda_j(g)\leq \Lambda} \mathbb{R} f_j,$$ where $f_j$ denotes the $j$-th eigenfunction of the Laplacian acting on functions. Let $p\in M$ be a maximum point of $f$. Then, we have $$|\nabla f|^2(x)\leq \frac{C}{\epsilon_1}\left(f(p)-f(x)+\epsilon_1\|f\|_2\right)^2$$ for all $x\in M$.
The following theorem is an easy consequence of the Bishop-Gromov inequality.
\[bigr\] Given a positive integer $n\geq2$ and positive real numbers $K>0$ and $D>0$, there exists a positive constant $ C(n,K,D)>0$ such that the following property holds. Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq -Kg$ and ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$. Then, for any $p\in M$ and $0<r\leq D+1$, we have $r^n {\mathop{\mathrm{Vol}}\nolimits}(M)\leq C{\mathop{\mathrm{Vol}}\nolimits}(B_r(p))$.
The following theorem is due to Cheeger-Colding [@CC2] (see also [@Pe3 Theorem 7.1.10]). By this theorem, we get integral pinching conditions along the geodesics under the integral pinching condition for a function on $M$.
\[seg\] Given an integer $n\geq 2$ and positive real numbers $K>0$ and $D>0$, there exists a positive constant $ C(n,K,D)>0$ such that the following property holds. Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq -Kg$ and ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$. For any non-negative measurable function $h\colon M\to \mathbb{R}_{\geq 0}$, we have $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)^2}\int_{M\times M} \frac{1}{d(y_1,y_2)}\int_0^{d(y_1,y_2)} h\circ \gamma_{y_1,y_2}(s) \,dsdy_1dy_2\leq \frac{C}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M h\,d\mu_g.$$
The book [@Pe3] deals with the segment $c_{y_1,y_2}\colon[0,1]\to M$ for each $y_1,y_2\in M$, defined to be $c_{y_1,y_2}(0)=y_1$, $c_{x,y}(1)=y_2$ and $\nabla_{\partial /\partial t} \dot{c}=0$. We have $c_{x,y}(t)=\gamma_{x,y}(t d(x,y))$ for all $t\in[0,1]$ and $$d(y_1,y_2)\int_0^1 h\circ c_{y_1,y_2}(t) \,d t=\int_0^{d(y_1,y_2)} h\circ \gamma_{y_1,y_2}(s) \,d s.$$
After getting integral pinching conditions along the geodesics, we use the following lemma to get $L^\infty$ error estimate along them. The proof is standard (c.f. [@CC2 Lemma 2.41]).
\[trif\] Take positive real numbers $l,\epsilon>0$ and a non-negative real number $r\geq 0$. Suppose that a smooth function $u\colon [0,l]\to \mathbb{R}$ satisfies $$\int_0^l |u''(t)+r^2 u(t)| \,dt\leq\epsilon.$$ Then, we have $$\begin{split}
\left|u(t)-u(0) \cos r t- \frac{u'(0)}{r} \sin r t\right|&\leq \epsilon\frac{\sinh rt}{r},\\
\left|u'(t)+ r u(0)\sin r t- u'(0)\cos r t\right|&\leq \epsilon+\int_0^t\left|u(s)-u(0)\cos r s-\frac{u'(0)}{r}\sin r s\right|\,ds,
\end{split}$$ for all $t\in [0,l]$, where we defined $$\begin{aligned}
\frac{1}{r}\sin r t:=t,\quad
\frac{1}{r}\sinh r t:=t\end{aligned}$$ if $r=0$.
The following lemma is standard.
\[cosi\] For all $t\in \mathbb{R}$, we have $$1-\frac{1}{2}t^2\leq \cos t\leq 1-\frac{1}{2}t^2+\frac{1}{24}t^4.$$ For any $t\in [-\pi,\pi]$, we have $\cos t\leq 1-\frac{1}{9}t^2$, and so $|t|\leq3(1-\cos t)^{1/2}$. For any $t_1,t_2 \in [0,\pi]$, we have $|t_1-t_2|\leq3|\cos t_1-\cos t_2|^{1/2}$.
Finally, we recall some facts about the geodesic flow. Let $(M,g)$ be a closed Riemannian manifold and let $U M$ denotes the sphere bundle defined by $$U M:=\{u\in TM:|u|=1\}.$$ There exists a natural Riemannian metric $G$ on $UM$, which is the restriction of the Sasaki metric on $TM$ (see [@Sa p.55]). The Riemannian volume measure $\mu_G$ satisfies $$\int_{UM} F\,d\mu_G=\int_M \int_{U_p M} F(u)\, d\mu_0(u) \,d\mu_g(p)$$ for any $F\in C^\infty(U M)$, where $\mu_0$ denotes the standard measure on $U_p M\cong S^{n-1}$. The geodesic flow $\phi_t\colon U M\to U M$ ($t\in\mathbb{R}$) is defined by $$\phi_t(u):=\left.\frac{\partial}{\partial s}\right|_{s=t}\gamma_u (s)\in U_{\gamma_u(t)} M$$ for any $u\in U M$. Though $\phi_t$ does not preserve the metric $G$ in general, it preserves the measure $\mu_G$. This is an easy consequence of [@Sa Lemma 4.4], which asserts that the geodesic flow on $T M$ preserve the natural symplectic structure on $T M$. We use the following lemma.
\[geofl\] Let $(M,g)$ be a closed Riemannian manifold. For any $f\in C^\infty (M)$ and $l>0$, we have $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M f \,d\mu_g=\frac{1}{l{\mathop{\mathrm{Vol}}\nolimits}(UM)}\int_{UM}\int_0^l f\circ\gamma_u(t)\,d t\,d\mu_G(u).$$
Let $\pi\colon UM\to M$ be the projection. Since the geodesic flow $\phi_t$ preserves the measure $\mu_G$, we have $$\begin{aligned}
&\frac{1}{l{\mathop{\mathrm{Vol}}\nolimits}(UM)}\int_{UM}\int_0^l f\circ\gamma_u(t)\,d t\,d\mu_G(u)\\
=&\frac{1}{l{\mathop{\mathrm{Vol}}\nolimits}(M){\mathop{\mathrm{Vol}}\nolimits}(S^{n-1})}\int_{UM}\int_0^l f\circ\pi\circ\phi_t(u)\,d t\,d\mu_G(u)\\
=&\frac{1}{l{\mathop{\mathrm{Vol}}\nolimits}(M){\mathop{\mathrm{Vol}}\nolimits}(S^{n-1})}\int_0^l\int_{UM} f\circ\pi(u)\,d\mu_G(u)\,d t\\
=&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M f \,d\mu_g.\end{aligned}$$ This gives the lemma.
This kind of lemma was used by Colding [@Co1] to prove that the almost equality of the Bishop comparison theorem implies the Gromov-Hausdorff closeness to the standard sphere.
Estimates for the Segments
--------------------------
The goal of this subsection is to give error estimates along the geodesics.
In this subsection, we assume the following in addition to Assumption \[asu1\].
- $1\leq k\leq n-p+1$
- $f_i\in C^\infty(M)$ ($i\in\{1,\ldots,k\}$) is an eigenfunction of the Laplacian acting on functions with $\|f_i\|_2^2=1/(n-p+1)$ corresponding to the eigenvalue $\lambda_i$ with $0<\lambda_i\leq n-p+\delta$ such that $$\int_M f_i f_j\,d\mu_g=0$$ holds for any $i\neq j$.
- $\omega\in\Gamma(\bigwedge^p T^\ast M)$ is an eigenform of the connection Laplacian $\Delta_{C,p}$ with $\|\omega\|_2=1$ corresponding to the eigenvalue $\lambda$ with $0\leq \lambda \leq \delta$.
Note that we have $\|\omega\|_\infty\leq C$, $\|f_i\|_\infty \leq C $ and $\|\nabla f_i\|_\infty \leq C$ for all $i\in\{1,\ldots,k\}$ (see Lemma \[Linfes\]). By Main Theorem 1, we have $$\lambda_i\geq n-p-C(n,p)\delta^{1/2}$$ for all $i$. If the manifold is orientable, one can regard $f_i$ as the $i$-th eigenfunction and $\lambda_i$ as $\lambda_i(g)$. However, for the unorientable case, our assumption is convenient when we consider $f_i\circ \pi$, where $\pi: (\widetilde{M},\tilde{g})\to (M,g)$ denotes the two-sheeted orientable Riemannian covering.
We first list some basic consequences of our pinching condition.
\[p5c\] For any $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots,f_{k}\}$, we have
- $\|\iota(\nabla f)\omega\|_2^2\leq C\delta^{1/2}\|f\|_2^2$,
- $\|\nabla(\iota(\nabla f)\omega)\|_2^2\leq C\delta^{1/2}\|f\|_2^2$,
- $\|(|\nabla^2 f|^2-\frac{1}{n-p}|\Delta f|^2)|\omega|^2\|_1\leq C\delta^{1/4}\|f\|_2^2$.
It is enough to consider the case when $M$ is orientable.
We first assume that $f=f_i$ for some $i=1,\ldots,k$. Then, we have $$\label{5a0}
\begin{split}
&\|d(\iota(\nabla f)\omega)\|^2_2\leq C\delta^{1/2}\|f\|_2^2,\\
&\|d^\ast (\iota(\nabla f)\omega)\|^2_2 \leq C\delta^{1/2}\|f\|_2^2,\quad
\|T(\iota(\nabla f)\omega)\|_2^2\leq C\delta^{1/2}\|f\|_2^2,\\
&\|d^\ast (\iota(\nabla f)\ast \omega)\|^2_2 \leq C\delta^{1/2}\|f\|_2^2,\quad
\|T(\iota(\nabla f)\ast\omega)\|_2^2 \leq C\delta^{1/2}\|f\|_2^2
\end{split}$$ by Lemma \[p4d\] (i) and Proposition \[p4e\]. Thus, by (\[2b\]), we get $$\label{5a}
\|\nabla (\iota(\nabla f)\omega)\|^2_2\leq C\delta^{1/2}\|f\|_2^2$$ and $$\label{5b}
\|\nabla (\iota(\nabla f)\ast\omega)\|^2_2\leq \frac{1}{n-p} \|d (\iota(\nabla f)\ast \omega)\|^2_2+C\delta^{1/2}\|f\|_2^2.$$ Moreover, by Lemma \[p4d\] (iii), we have $$\label{5c}
\begin{split}
\|\iota(\nabla f)\omega\|_2^2
=&\frac{1}{\lambda_i}\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle \iota(\nabla \Delta f)\omega, \iota(\nabla f)\omega\rangle\,d\mu_g\\
\leq& C\|d(\iota(\nabla f)\omega)\|^2_2+C\|d^\ast (\iota(\nabla f)\omega)\|^2_2+C\delta^{1/2}\|f\|_2^2\\
\leq& C\delta^{1/2}\|f\|_2^2.
\end{split}$$
For any $f=a_1 f_1+\cdots + a_k f_k\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots,f_{k}\}$, we have (\[5a0\]), (\[5a\]), (\[5b\]), (\[5c\]). For example, we have $$\begin{split}
\|\nabla (\iota(\nabla f)\omega)\|_2\leq&\sum_{i=1}^k |a_k|\|\nabla (\iota(\nabla f_i)\omega)\|_2\leq C\delta^{1/4}\sum_{i=1}^k |a_k|\|f_i\|_2\leq C\delta^{1/4}\|f\|_2,\\
\|\nabla (\iota(\nabla f)\ast\omega)\|_2^2=&\frac{1}{n-p}\|d(\iota(\nabla f)\ast\omega)\|_2^2+\frac{1}{p+2}\|d^\ast(\iota(\nabla f)\ast\omega)\|_2^2+\|T(\iota(\nabla f)\ast\omega)\|_2^2\\
\leq& \frac{1}{n-p}\|d(\iota(\nabla f)\ast\omega)\|_2^2+C\delta^{1/2}\|f\|_2^2.
\end{split}$$ Thus, we get (i) and (ii) by (\[5a\]) and (\[5c\]).
Finally, we prove (iii). Take arbitrary $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots,f_{k}\}$. We have $$\label{5ca}
\begin{split}
&\left|\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\ast\omega\right|^2\\
=&\sum_{i=1}^n \langle \iota(\nabla_{e_i} \nabla f)\ast\omega,\iota(\nabla_{e_i} \nabla f)\ast\omega\rangle
=|\nabla^2 f|^2|\omega|^2-\left|\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\omega\right|^2.
\end{split}$$ Thus, we have $$\begin{split}
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \left||\nabla(\iota(\nabla f)\ast \omega)|^2-|\nabla^2 f|^2|\omega|^2\right|\,d\mu_g\\
\leq &\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \left||\nabla(\iota(\nabla f)\ast \omega)|^2-\left|\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\ast\omega\right|^2\right| \,d\mu_g\\
&\qquad+\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \left|\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\omega\right|^2\,d\mu_g,
\end{split}$$ and so we get $$\label{5d}
\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \left||\nabla(\iota(\nabla f)\ast \omega)|^2-|\nabla^2 f|^2|\omega|^2\right|\,d\mu_g
\leq C\delta^{1/2}\|f\|_2^2$$ by (ii) and Lemma \[p4d\] (iv) and (vi). We have $$\label{5e}
\begin{split}
&\left|\sum_{i=1}^n e^i\wedge \iota(\nabla_{e_i}\nabla f)\ast\omega\right|^2\\
=&\sum_{i=1}^n |\iota(\nabla_{e_i} \nabla f)\ast\omega|^2-\sum_{i,j=1}^n \langle \iota(e_i)\iota(\nabla_{e_j}\nabla f)\ast \omega, \iota(e_j)\iota(\nabla_{e_i}\nabla f)\ast \omega \rangle\\
=&|\nabla^2 f|^2|\omega|^2-\left|\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\omega\right|^2\\
&\qquad -\sum_{i,j,k,l=1}^n \nabla^2 f(e_i,e_k)\nabla^2 f(e_j,e_l)\langle e^i\wedge e^l \wedge \omega,
e^j\wedge e^k \wedge \omega \rangle
\end{split}$$ by (\[5ca\]) and (\[hstar2\]). Since $$\begin{aligned}
\langle e^i\wedge e^l \wedge \omega,
e^j\wedge e^k \wedge \omega \rangle
=&(\delta_{i j}\delta_{k l}-\delta_{i k}\delta_{j l})|\omega|^2
-\delta_{i j}\langle \iota(e_k)\omega,\iota(e_l)\omega\rangle\\
&+\delta_{i k}\langle \iota(e_j)\omega,\iota(e_l)\omega\rangle
+\langle e^l\wedge \omega,e^j\wedge e^k\wedge\iota(e_i)\omega\rangle,\end{aligned}$$ we have $$\label{5f}
\begin{split}
&\sum_{i,j,k,l=1}^n \nabla^2 f(e_i,e_k)\nabla^2 f(e_j,e_l)\langle e^i\wedge e^l \wedge \omega,
e^j\wedge e^k \wedge \omega \rangle\\
=&|\nabla^2 f|^2|\omega|^2-(\Delta f)^2|\omega|^2
-\sum_{i=1}^n | \iota(\nabla_{e_i}\nabla f)\omega|^2
-\sum_{i=1}^n\Delta f \langle \iota(\nabla_{e_i}\nabla f)\omega,\iota(e_i)\omega\rangle\\
&\qquad+\sum_{j,k,l=1}^n \nabla^2 f(e_j,e_l)\langle e^l\wedge \omega,e^j\wedge e^k\wedge\iota(\nabla_{e_k} \nabla f)\omega\rangle.
\end{split}$$ By (\[5e\]) and (\[5f\]), we get $$\begin{split}
\left|\sum_{i=1}^n e^i\wedge \iota(\nabla_{e_i}\nabla f)\ast\omega\right|^2
=&(\Delta f)^2|\omega|^2+\sum_{i=1}^n\Delta f \langle \iota(\nabla_{e_i}\nabla f)\omega,\iota(e_i)\omega\rangle\\
-&\sum_{j,k,l=1}^n \nabla^2 f(e_j,e_l)\langle e^l\wedge \omega,e^j\wedge e^k\wedge\iota(\nabla_{e_k} \nabla f)\omega\rangle,
\end{split}$$ and so $$\label{5g}
\left|\left|\sum_{i=1}^n e^i\wedge \iota(\nabla_{e_i}\nabla f)\ast\omega\right|^2-(\Delta f)^2|\omega|^2\right|
\leq C|\nabla^2 f| |\omega|\left|\sum_{i=1}^n e^i\otimes \iota(\nabla_{e_i}\nabla f)\omega\right|$$ By (\[5g\]), (ii) and Lemma \[p4d\], we get $$\label{5h}
\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \left|
|d (\iota(\nabla f)\ast\omega)|^2-(\Delta f)^2|\omega|^2\right|\,d\mu_g
\leq C\delta^{1/4}\|f\|_2^2.$$ Since we have $$|\nabla (\iota(\nabla f)\ast\omega)|^2\geq \frac{1}{n-p} |d (\iota(\nabla f)\ast \omega)|^2$$ at each point by (\[2b\]), we get (iii) by (\[5b\]), (\[5d\]) and (\[5h\]).
We use the following notation.
\[np5d\] Take $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots,f_{k}\}$ with $\|f\|_2^2=1/(n-p+1)$ and put $$\begin{aligned}
h_0&:=|\nabla^2 f|^2, \quad h_1:=||\omega|^2-1|, \quad h_2:=|\nabla \omega|^2,\\
h_3&:=|\iota(\nabla f)\omega |^2, \quad h_4:=|\nabla (\iota(\nabla f)\omega)|^2,\quad h_5:=\left|\sum_{i=1}^n e^i\otimes\iota(\nabla_{e_i}\nabla f)\omega\right|^2\\
h_6&:=\left||\nabla^2 f|^2-\frac{1}{n-p}(\Delta f)^2\right||\omega|^2.\end{aligned}$$ For each $y_1\in M$, we define $$\begin{aligned}
D_f(y_1):=&\Big\{y_2\in I_{y_1}\setminus\{y_1\}:\frac{1}{d(y_1,y_2)}\int_{0}^{d(y_1,y_2)} h_0\circ \gamma_{y_1,y_2}(s)\,d s\leq \delta^{-1/50} \text{ and}\\
&\qquad \quad\frac{1}{d(y_1,y_2)}\int_{0}^{d(y_1,y_2)} h_i\circ \gamma_{y_1,y_2}(s)\,d s\leq \delta^{1/5} \text{ for all $i=1,\ldots,6$}
\Big\},\\
Q_f:=&\{y_1\in M: {\mathop{\mathrm{Vol}}\nolimits}(M\setminus D_f(y_1))\leq\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(M)\},\\
E_f(y_1):=&\Big\{u\in U_{y_1} M: \frac{1}{\pi }\int_{0}^{\pi} h_0\circ \gamma_u(s)\,d s\leq \delta^{-1/50} \text{ and }\frac{1}{\pi}\int_{0}^{\pi} h_i\circ \gamma_u (s)\,d s\leq \delta^{1/5} \\
&\qquad \quad\qquad \quad\qquad \quad\qquad \quad\qquad \quad\qquad \quad\qquad \quad\text{ for all $i=1,\ldots,6$}
\Big\},\\
R_f:=&\{y_1\in M: {\mathop{\mathrm{Vol}}\nolimits}(U_{y_1} M\setminus E_f(y_1))\leq\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(U_{y_1}M)\}.\end{aligned}$$
Now, we use the segment inequality and Lemma \[geofl\]. We show that we have the integral pinching condition along most geodesics.
\[p5d\] Take $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots,f_{k}\}$ with $\|f\|_2^2=1/(n-p+1)$. Then, we have the following properties:
- ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus Q_f)\leq C\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(M).$
- ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus R_f)\leq C\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(M).$
We have $\|h_i\|_1\leq C\delta^{1/4}$ for all $i=1,\ldots,6$ by the assumption, Lemma \[p4c\], Lemma \[p4d\] (iv) and Lemma \[p5c\], and we have $\|h_0\|_1\leq C$ by (\[4a0\]).
For any $y_1\in M\setminus Q_f$, we have ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus D_f(y_1))>\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(M)$, and so we have either $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M\frac{1}{d(y_1,y_2)}\int_0^{d(y_1,y_2)}h_0\circ \gamma_{y_1,y_2}(s)\,d s \,d y_2\geq \frac{1}{7}\delta^{-1/100}$$ or $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M\frac{1}{d(y_1,y_2)}\int_0^{d(y_1,y_2)}h_i\circ \gamma_{y_1,y_2}(s)\,d s \,d y_2\geq\frac{1}{7}\delta^{21/100}$$ for some $i=1,\ldots,6$. Thus, we get either $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \int_M\frac{1}{d(y_1,y_2)}\int_0^{d(y_1,y_2)}h_0\circ \gamma_{y_1,y_2}(s)\,d s \,d y_1\,d y_2\geq \frac{1}{49}\delta^{-1/100}{\mathop{\mathrm{Vol}}\nolimits}(M\setminus Q_f)$$ or $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \int_M\frac{1}{d(y_1,y_2)}\int_0^{d(y_1,y_2)}h_i\circ \gamma_{y_1,y_2}(s)\,d s \,d y_1\,d y_2\geq \frac{1}{49}\delta^{21/100}{\mathop{\mathrm{Vol}}\nolimits}(M\setminus Q_f)$$ for some $i=1,\ldots,6$. Therefore, we get (i) by the segment inequality (Theorem \[seg\]).
Similarly, we get (ii) by Lemma \[geofl\].
Under the pinching condition along the geodesic, we get the following:
\[p5e\] Take $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots,f_{k}\}$ with $\|f\|_2^2=1/(n-p+1)$. Suppose that a geodesic $\gamma\colon [0,l]\to M$ satisfies one of the following:
- There exist $x\in M$ and $y\in D_f(x)$ such that $l=d(x,y)$ and $\gamma=\gamma_{x,y}$,
- There exist $x\in M$ and $u\in E_f(x)$ such that $l=\pi$ and $\gamma=\gamma_u$.
Then, we have $$||\omega|^2(s)-1|\leq C\delta^{1/10},\quad |\iota(\nabla f)\omega|(s)\leq C\delta^{1/10}$$ for all $s\in [0,l]$, and at least one of the following:
- $\frac{1}{l}\int_0^l|\nabla^2 f|\circ \gamma(s)\,d s\leq C\delta^{1/250}$,
- There exists a parallel orthonormal basis $\{E^1(s),\ldots,E^n(s)\}$ of $T_{\gamma(s)}^\ast M$ along $\gamma$ such that $$|\omega-E^{n-p+1}\wedge\cdots\wedge E^n|(s)\leq C\delta^{1/25}$$ for all $s\in[0,l]$, and $$\frac{1}{l}\int_0^l|\nabla^2 f+f\sum_{i=1}^{n-p}E^i\otimes E^i|(s)\, d s\leq C\delta^{1/200},$$ where we write $|\cdot|(s)$ instead of $|\cdot|\circ\gamma(s)$.
In particular, for both cases, there exists a parallel orthonormal basis $\{E^1(s),\ldots,E^n(s)\}$ of $T_{\gamma(s)}^\ast M$ along $\gamma$ such that $$\frac{1}{l}\int_0^l|\nabla^2 f+f\sum_{i=1}^{n-p}E^i\otimes E^i|(s)\, d s\leq C\delta^{1/250}.$$
Moreover, if we put $$\dot{\gamma}^E:=\sum_{i=1}^{n-p} \langle\dot{\gamma},E_i\rangle E_i,$$ where $\{E_1,\ldots,E_n\}$ denotes the dual basis of $\{E^1,\ldots,E^n\}$, then $|\dot{\gamma}^E|$ is constant along $\gamma$, and $$\begin{split}
\left|f\circ \gamma(s)-f(\gamma(s_0))\cos (|\dot{\gamma}^E|(s-s_0))-\frac{1}{|\dot{\gamma}^E|}\langle\nabla f,\dot{\gamma}(s_0)\rangle\sin (|\dot{\gamma}^{E}|(s-s_0))\right|&\leq C\delta^{1/250},\\
\left| \langle \nabla f, \dot{\gamma}(s)\rangle+f(\gamma(s_0))|\dot{\gamma}^{E}|\sin (|\dot{\gamma}^{E}|(s-s_0))-\langle\nabla f,\dot{\gamma}(s_0)\rangle\cos (|\dot{\gamma}^{E}|(s-s_0))\right|&\leq C\delta^{1/250}
\end{split}$$ for all $s,s_0\in[0,l]$.
We first show the first assertion. Since $\frac{d}{d s}|\omega|^2(s)=2\langle\nabla_{\dot{\gamma}}\omega,\omega\rangle$, we have $$\begin{aligned}
\left||\omega|^2(s)-|\omega|^2(0)\right|
=&\left|\int_0^s \frac{d}{d s}|\omega|^2(t)\,d t\right|\\
\leq& 2 \left(\int_0^s |\nabla \omega|^2 (t)\,d t\right)^{1/2} \left(\int_0^s |\omega|^2 (t)\, d t\right)^{1/2}
\leq C\delta^{1/10}\end{aligned}$$ for all $s\in[0,l]$. Since we have $\int_0^l||\omega|^2-1|\, d t \leq \delta^{1/5}$, we get $||\omega|^2(s)-1|\leq C\delta^{1/10}$. In particular, $|\omega|(s)\geq 1/2$, and so $$\label{5i}
\frac{1}{l}\int_0^l\left||\nabla^2 f|^2-\frac{1}{n-p}(\Delta f)^2\right|(s)\,d s\leq 2\delta^{1/5}.$$ Similarly, we have $|\iota(\nabla f)\omega|(s)\leq C\delta^{1/10}$ for all $s\in [0,l]$.
We show the remaining assertions. Put $$\begin{aligned}
A_1:=&\left\{s\in [0,l]:\left|\sum_{i=1}^n e^i\otimes\iota(\nabla_{e_i}\nabla f)\omega\right|^2(s)>\delta^{1/10}\right\},\\
A_2:=&\left\{s\in [0,l]:\left||\nabla^2 f|^2-\frac{1}{n-p}(\Delta f)^2\right|(s)>\delta^{1/10}\right\},\\
A_3:=&\left\{s\in [0,l]:|\nabla^2 f|(s)<\delta^{1/250}\right\}.\end{aligned}$$ Then, we have $H^1(A_1)\leq \delta^{1/10}l$ and $H^1(A_2)\leq 2\delta^{1/10} l$, where $H^1$ denotes the one dimensional Hausdorff measure. We consider the following two cases:
- $[0,l]=A_1\cup A_2\cup A_3$,
- $[0,l]\neq A_1\cup A_2\cup A_3$.
We first consider the case (a). Since $$H^1([0,l]\setminus A_3)\leq 3 \delta^{1/10} l,$$ we have $$\begin{aligned}
\int_{[0,l]\setminus A_3}|\nabla^2 f|(s)\,d s
\leq&
\left(\int_{[0,l]\setminus A_3}|\nabla^2 f|^2(s)\,d s\right)^{1/2}H^1 ([0,l]\setminus A_3)^{1/2}\\
\leq &C \delta^{-1/100}\delta^{1/20} l=C\delta^{1/25}l.\end{aligned}$$ On the other hand, we have $$\int_{A_3} |\nabla^2 f|(s)\,d s\leq \delta^{1/250} l.$$ Therefore, we get $$\frac{1}{l}\int_0^l|\nabla^2 f|(s)\,d s\leq C\delta^{1/250}.$$ This implies (i). Moreover, since $|\Delta f|\leq \sqrt{n}|\nabla^2 f|$ and $\left\|\Delta f-(n-p)f\right\|_{\infty}\leq C\delta$, we get $$\frac{1}{l}\int_0^l|\nabla^2 f+f\sum_{i=1}^{n-p}E^i\otimes E^i|(s)\, d s\leq C\delta^{1/250},$$ where $\{E^1(s),\ldots,E^n(s)\}$ is any parallel orthonormal basis of $T_{\gamma(s)}^\ast M$ along $\gamma$.
We next consider the case (b). There exists $t\in[0,l]$ such that $$\begin{aligned}
\left|\sum_{i=1}^n e^i\otimes\iota(\nabla_{e_i}\nabla f)\omega\right|^2(t)&\leq\delta^{1/10},\\
\left||\nabla^2 f|^2-\frac{1}{n-p}(\Delta f)^2\right|(t)&\leq\delta^{1/10},\\
|\nabla^2 f|(t)&\geq\delta^{1/250}.\end{aligned}$$ Take an orthonormal basis $\{e_1,\ldots,e_n\}$ of $T_{\gamma(t)}M$ such that $$\nabla^2 f(e_i,e_j)=\mu_i\delta_{i j}\quad (\mu_i\in\mathbb{R})$$ for all $i,j=1,\ldots,n$. Let $\{e^1,\ldots,e^n\}$ be the dual basis of $T_{\gamma(t)}^\ast M$. Then, we have $$\delta^{1/10}\geq \left|\sum_{i=1}^n e^i\otimes\iota(\nabla_{e_i}\nabla f)\omega\right|^2(t)
=\sum_{i=1}^n\mu_i^2 |\iota(e_i)\omega|^2(t).$$ Thus, for each $i=1,\ldots,n$, we have at least one of the following:
- $|\mu_i|\leq \delta^{1/100}$,
- $|\iota(e_i)\omega|(t)\leq \delta^{1/25}$.
Since $|\omega|(t)\geq 1/2$, we have $${\mathop{\mathrm{Card}}\nolimits}\{i: |\iota(e_i)\omega|(t)\leq \delta^{1/25}\}\leq n-p,$$ and so $${\mathop{\mathrm{Card}}\nolimits}\{i: |\mu_i|\leq \delta^{1/100}\}\geq p.$$ Therefore, we can assume $|\mu_i|\leq \delta^{1/100}$ for all $i=n-p+1,\ldots, n$. Then, we get $$\begin{aligned}
\left|
\nabla^2 f+\frac{\Delta f}{n-p}\sum_{i=1}^{n-p} e^i\otimes e^i
\right|^2(t)
=&|\nabla^2 f|^2(t)+\frac{2}{n-p}(\Delta f)(t)\sum_{i=1}^{n-p}\mu_i
+\frac{(\Delta f)^2(t)}{n-p}\\
=&|\nabla^2 f|^2(t)-\frac{(\Delta f)^2(t)}{n-p}-\frac{2}{n-p}(\Delta f)(t)\sum_{i=n-p+1}^{n}\mu_i\\
\leq& C\delta^{1/100}. \end{aligned}$$ Putting $e_i\otimes e_i$ into the inside of the left hand side, we get $$\left|\mu_i+\frac{\Delta f(t)}{n-p}\right|^2\leq C\delta^{1/100}$$ for all $i=1,\ldots, n-p$, and so $$\begin{aligned}
|\mu_i|\geq \frac{|\Delta f(t)|}{n-p}-C\delta^{1/200}
\geq &\left(\frac{|\nabla^2 f|^2(t)-\delta^{1/10}}{n-p}\right)^{1/2}-C\delta^{1/200}\\
\geq &\left(\frac{\delta^{1/125}-\delta^{1/10}}{n-p}\right)^{1/2}-C\delta^{1/200}
>\delta^{1/100}.\end{aligned}$$ Thus, we have $$|\iota(e_i)\omega|(t)\leq \delta^{1/25}$$ for all $i=1,\ldots,n-p$. Thus, we get either $|\omega(t)-e^{n-p+1}\wedge\cdots\wedge e^n|\leq C\delta^{1/25}$ or $|\omega(t)+e^{n-p+1}\wedge\cdots\wedge e^n|\leq C\delta^{1/25}$ by $||\omega|^2(t)-1|\leq C\delta^{1/10}$. We can assume that $|\omega(t)-e^{n-p+1}\wedge\cdots\wedge e^n|\leq C\delta^{1/25}$.
Let $\{E_1,\ldots,E_n\}$ be the parallel orthonormal basis of $TM$ along $\gamma$ such that $E_i(t)=e_i$, and let $\{E^1,\ldots,E^n\}$ be its dual. Because $$\begin{aligned}
&\int_0^l \left|\frac{d}{d s}|\omega-E^{n-p+1}\wedge\cdots \wedge E^n|^2(s)\right|\,d s\\
\leq&
2 \int_0^l |\nabla \omega|(s)|\omega-E^{n-p+1}\wedge\cdots \wedge E^n|(s)\,d s
\leq C\delta^{1/10},\end{aligned}$$ we get $$|\omega-E^{n-p+1}\wedge\cdots\wedge E^n|(s)\leq C\delta^{1/25}$$ for all $s\in [0,l]$. Thus, we get $$|\langle\iota(E_i)\omega,\iota(E_j)\omega\rangle|\leq C\delta^{1/25}$$ for all $i=1,\cdots,n$ and $j=1,\ldots,n-p$, and $$|\langle\iota(E_i)\omega,\iota(E_j)\omega\rangle-\delta_{i j}|\leq C\delta^{1/25}$$ for all $i,j=n-p+1,\cdots,n$. Therefore, we get $$\begin{aligned}
&\left|
\left|\sum_{i=1}^n E^i\otimes\iota(\nabla_{E_i}\nabla f)\omega\right|^2-\sum_{i=1}^n\sum_{j=n-p+1}^n(\nabla^2 f(E_i,E_j))^2
\right|\\
=&\left|
\sum_{i,j,k=1}^n \nabla^2 f(E_i,E_j)\nabla^2 f(E_i,E_k)\langle\iota(E_j)\omega,\iota(E_k)\omega\rangle-\sum_{i=1}^n\sum_{j=n-p+1}^n(\nabla^2 f (E_i,E_j))^2
\right|\\
\leq&C |\nabla^2 f|^2 \delta^{1/25}.\end{aligned}$$ Thus, for all $i=1,\cdots,n$ and $j=1,\ldots,n-p$, we get $$|\nabla^2 f(E_i,E_j)|^2\leq \left|\sum_{k=1}^n E^k\otimes\iota(\nabla_{E_k}\nabla f)\omega\right|^2+C |\nabla^2 f|^2 \delta^{1/25},$$ and so $$\frac{1}{l}\int_0^l|\nabla^2 f (E_i,E_j)|^2(s)\,d s
\leq 2\delta^{1/5}+C\delta^{-1/50}\delta^{1/25}
\leq C\delta^{1/50}.$$ This gives $$\frac{1}{l}\int_0^l|\nabla^2 f (E_i,E_j)|(s)\,d s
\leq C\delta^{1/100}$$ for all $i=1,\cdots,n$ and $j=1,\ldots,n-p$. Because $$\begin{aligned}
\left|
\nabla^2 f+\frac{\Delta f}{n-p}\sum_{i=1}^{n-p}E^i\otimes E^i
\right|^2
=&|\nabla^2 f|^2+2\frac{\Delta f}{n-p}\sum_{i=1}^{n-p}\nabla^2 f (E_i,E_i)+\frac{(\Delta f)^2}{n-p}\\
=&|\nabla^2 f|^2-\frac{(\Delta f)^2}{n-p}-2\frac{\Delta f}{n-p}\sum_{i=n-p+1}^{n}\nabla^2 f (E_i,E_i),\end{aligned}$$ we have $$\frac{1}{l}\int_0^l\left|
\nabla^2 f+\frac{\Delta f}{n-p}\sum_{i=1}^{n-p}E^i\otimes E^i
\right|^2\,d s
\leq 2\delta^{1/5}+C\delta^{1/100}\leq C\delta^{1/100}.$$ Since $$\left\|\frac{\Delta f}{n-p}-f\right\|_{\infty}\leq C\delta$$ we get $$\frac{1}{l}\int_0^l|\nabla^2 f+f\sum_{i=1}^{n-p}E^i\otimes E^i|(s)\, d s\leq C\delta^{1/200}.$$ This implies (ii).
Let us show the final assertion. It is trivial that $|\dot{\gamma}^E|$ is constant along $\gamma$. Since we have $$\left(\nabla^2 f+f \sum_{i=1}^{n-p}E^i\otimes E^i\right)(\dot{\gamma},\dot{\gamma})
=\frac{d^2}{d s^2} f\circ \gamma + |\dot{\gamma}^E|^2 f\circ \gamma,$$ we get $$\int_0^l\left|\frac{d^2}{d s^2} f\circ \gamma(s) + |\dot{\gamma}^E|^2 f\circ \gamma(s)\right|\,d s\leq C\delta^{1/250}.$$ Thus, we get the lemma by Lemma \[trif\].
Almost Parallel $(n-p)$-form I
------------------------------
In this subsection, we consider the pinching condition on $\lambda_1(\Delta_{C,n-p})$ for $2\leq p< n/2$. If $M$ is oriented, then this is coincide with the pinching condition on $\lambda_1(\Delta_{C,p})$. Thus, we only consider the case when $M$ is not orientable.
\[anori\] In this subsection, we assume the following in addition to Assumption \[asu1\].
- $M$ is not orientable.
- $1\leq k\leq n-p+1$.
- $f_i\in C^\infty(M)$ ($i\in\{1,\ldots,k\}$) is an eigenfunction of the Laplacian acting on functions with $\|f_i\|_2^2=1/(n-p+1)$ corresponding to the eigenvalue $\lambda_i$ with $0<\lambda_i\leq n-p+\delta$ such that $$\int_M f_i f_j\,d\mu_g=0$$ holds for any $i\neq j$.
- $\xi\in\Gamma(\bigwedge^{n-p} T^\ast M)$ is an eigenform of the connection Laplacian $\Delta_{C,n-p}$ with $\|\xi\|_2=1$ corresponding to the eigenvalue $\lambda$ with $0\leq \lambda \leq \delta$.
Under these assumptions, we use the following notation.
\[np5f\] Take $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_k\}$ with $\|f\|_2^2=1/(n-p+1)$. Let $\pi\colon (\widetilde{M},\tilde{g})\to (M,g)$ be the two-sheeted oriented Riemannian covering. Put $
\tilde{f}:=f\circ \pi\in C^\infty(\widetilde{M})$, $\widetilde{\xi}:=\pi^\ast \xi\in\Gamma(\bigwedge^{n-p}T^\ast M)$ and $\omega:=\ast \widetilde{\xi}\in\Gamma(\bigwedge^{p}T^\ast \widetilde{M})$. Define $h_0,\ldots,h_6$, $Q_{\tilde{f}}$, $D_{\tilde{f}}(\tilde{y}_1)$, $R_{\tilde{f}}$ and $E_{\tilde{f}}(\tilde{y_1})$ as Notation \[np5d\] for $\tilde{f}$, $\omega$ and $\tilde{y}_1\in \widetilde{M}$. Put $$\begin{aligned}
Q_f:=&M\setminus \pi\left(\widetilde{M}\setminus Q_{\tilde{f}}\right),\quad D_f(y_1):=&&M\setminus \pi\left(\widetilde{M}\setminus\bigcap_{\tilde{y}\in\pi^{-1}(y_1)} D_{\tilde{f}}(\tilde{y})\right),\\
R_f:=&M\setminus \pi\left(\widetilde{M}\setminus R_{\tilde{f}}\right),\quad E_f(y_1):=&&U_{y_1} M\setminus \bigcup_{\tilde{y}\in\pi^{-1}(y_1)}\pi_\ast\left(U_{\tilde{y}}\widetilde{M}\setminus E_{\tilde{f}}(\tilde{y})\right)\end{aligned}$$ for each $y_1\in M$.
We immediately have the following lemmas by Lemma \[p5d\] and Lemma \[p5e\].
\[p5f\] We have the following:
- ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus Q_f)\leq C\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(M)$, and ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus D_f(y_1))\leq2\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(\widetilde{M})=4\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(M)$ for each $y_1\in Q_f$.
- ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus R_f)\leq C\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(M)$, and ${\mathop{\mathrm{Vol}}\nolimits}(U_{y_1} M\setminus E_f(y_1))\leq2\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(U_{y_1}M)$ for each $y_1\in R_f$.
- Take $y_1\in M$ and $y_2\in D_f(y_1)$ and one of the lift of $\gamma_{y_1,y_2}$: $$\tilde{\gamma}_{y_1,y_2}\colon[0,d(y_1,y_2)]\to \widetilde{M}.$$ Put $\tilde{y}_1:=\tilde{\gamma}_{y_1,y_2}(0)\in \widetilde{M}$ and $\tilde{y}_2:=\tilde{\gamma}_{y_1,y_2}(d(y_1,y_2))\in \widetilde{M}$. Then, we have $\tilde{y}_2\in D_{\tilde{f}}(\tilde{y}_1)$.
- Take $y_1\in M$ and $u\in E_f(y_1)$ and one of the lift of $\gamma_u$: $$\tilde{\gamma}_{u}\colon[0,\pi]\to \widetilde{M}.$$ Put $\tilde{y}_1:=\tilde{\gamma}_{u}(0)\in \widetilde{M}$ and $\tilde{u}:=\dot{\tilde{\gamma}}_{u}(0)\in U_{\tilde{y}_1}\widetilde{M}$. Then, we have $\tilde{u}\in E_{\tilde{f}}(\tilde{y}_1)$.
\[p5g\] Suppose that a geodesic $\gamma\colon [0,l]\to M$ satisfies one of the following:
- There exist $x\in M$ and $y\in D_f(x)$ such that $l=d(x,y)$ and $\gamma=\gamma_{x,y}$,
- There exist $x\in M$ and $u\in E_f(x)$ such that $l=\pi$ and $\gamma=\gamma_u$.
Let $\tilde{\gamma}\colon [0,l]\to\widetilde{M}$ be one of the lift of $\gamma$. Then, we have $$||\omega|^2(\tilde{\gamma}(s))-1|\leq C\delta^{1/10},\quad |\iota(\nabla \tilde{f})(\omega)|\circ\tilde{\gamma}(s)\leq C\delta^{1/10}$$ for all $s\in [0,l]$, and at least one of the following:
- $\frac{1}{l}\int_0^l|\nabla^2 f|\circ \gamma(s)\,d s\leq C\delta^{1/250}$,
- There exists a parallel orthonormal basis $\{E^1(s),\ldots,E^n(s)\}$ of $T_{\gamma(s)}^\ast M$ along $\gamma$ such that $$|\xi-E^{1}\wedge\cdots\wedge E^{n-p}|(s)\leq C\delta^{1/25}$$ for all $s\in[0,s]$, and $$\frac{1}{l}\int_0^l|\nabla^2 f+f\sum_{i=1}^{n-p}E^i\otimes E^i|(s)\, d s\leq C\delta^{1/200}.$$
In particular, for both cases, there exists a parallel orthonormal basis $\{E^1(s),\ldots,E^n(s)\}$ of $T_{\gamma(s)}^\ast M$ along $\gamma$ such that $$\frac{1}{l}\int_0^l|\nabla^2 f+f\sum_{i=1}^{n-p}E^i\otimes E^i|(s)\, d s\leq C\delta^{1/250}.$$ Moreover, if we put $$\dot{\gamma}^E:=\sum_{i=1}^{n-p} \langle\dot{\gamma},E_i\rangle E_i,$$ where $\{E_1,\ldots,E_n\}$ denotes the dual basis of $\{E^1,\ldots,E^n\}$, then $|\dot{\gamma}^E|$ is constant along $\gamma$, and $$\begin{split}
\left|f\circ \gamma(s)-f(\gamma(s_0))\cos (|\dot{\gamma}^E|(s-s_0))-\frac{1}{|\dot{\gamma}^E|}\langle\nabla f,\dot{\gamma}(s_0)\rangle\sin (|\dot{\gamma}^{E}|(s-s_0))\right|&\leq C\delta^{1/250},\\
\left| \langle \nabla f, \dot{\gamma}(s)\rangle+f(\gamma(s_0))|\dot{\gamma}^{E}|\sin (|\dot{\gamma}^{E}|(s-s_0))-\langle\nabla f,\dot{\gamma}(s_0)\rangle\cos (|\dot{\gamma}^{E}|(s-s_0))\right|&\leq C\delta^{1/250}
\end{split}$$ for all $s,s_0\in[0,l]$.
Eigenfunction and Distance
--------------------------
In this subsection, we show that the function is an almost cosine function in some sense under our pinching condition.
In this subsection, we assume the following in addition to Assumption \[asu1\].
- $1\leq k\leq n-p+1$.
- $f_i\in C^\infty(M)$ ($i\in\{1,\ldots,k\}$) is an eigenfunction of the Laplacian acting on functions with $\|f_i\|_2^2=1/(n-p+1)$ corresponding to the eigenvalue $\lambda_i$ with $0<\lambda_i\leq n-p+\delta$ such that $$\int_M f_i f_j\,d\mu_g=0$$ holds for any $i\neq j$.
- Either $\lambda_1(\Delta_{C,p})\leq \delta$ or $\lambda_1(\Delta_{C,n-p})\leq \delta.$
The following proposition is the goal of this subsection. See Notation \[np5d\] and Notation \[np5f\] for the definitions of $D_f$, $Q_f$, $E_f$ and $R_f$.
\[p53a\] Take $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots,f_{k}\}$ with $\|f\|_2^2=1/(n-p+1)$. There exists a point $p_f\in Q_f$ such that the following properties hold:
- $\sup_M f\leq f(p_f)+C\delta^{1/100n}$ and $|f(p_f)-1|\leq C\delta^{1/800n}$,
- For any $x\in D_f(p_f)$ with $|\nabla f|(x)\leq \delta^{1/800n}$, we have $$||f(x)|-1|\leq C\delta^{1/800n}.$$
- For any $x\in D_f(p_f)\cap Q_f\cap R_f$, we have $$|f(x)^2+|\nabla f|^2(x)-1|\leq C \delta^{1/800n}.$$
- Put $$A_f:=\{x\in M: |f(x)-1|\leq \delta^{1/900n}\}.$$ Then, we have $$|f(x)-\cos d(x,A_f)|\leq C\delta^{1/2000n}$$ for all $x\in M$, and $$\sup_{x\in M}d(x,A_f)\leq \pi+ C\delta^{1/100n}.$$
Take a maximum point $\tilde{p}\in M$ of $f$. Then, by the Bishop-Gromov theorem and Lemma \[p5d\] (or Lemma \[p5f\]), there exists a point $p_f\in Q_f$ with $d(\tilde{p},p_f)\leq C \delta^{1/100n}$. By Lemma \[chya\], we have $$|\nabla f|^2(p_f)\leq \frac{C}{\delta^{1/100n}}(f(\tilde{p})-f(p_f)+\delta^{1/100n}\|f\|_2)^2\leq C\delta^{1/100n},$$ and so $$\label{54b}
|\nabla f|(p_f)\leq C\delta^{1/200n}.$$
\[c0\] For any $x\in D_f(p_f)$ with $|\nabla f|(x)\leq C\delta^{1/800n}$, we have $$||f(x)|-|f(p_f)||\leq C\delta^{1/800n}.$$
Since $$|\nabla f|(p_f)\leq C\delta^{1/200n}, \quad |\nabla f|(x)\leq C\delta^{1/800n},$$ we get $$\begin{aligned}
|f\circ \gamma_{p_f,x}(s)-f(p_f)\cos ( |\dot{\gamma}_{p_f,x}^E| s)|&\leq C\delta^{1/200n},\\
|f\circ \gamma_{p_f,x}(d(p_f,x)-s)-f(x)\cos ( |\dot{\gamma}_{p_f,x}^E|s)|&\leq C\delta^{1/800n}\end{aligned}$$ for all $s\in[0,d(p_f,x)]$ by Lemma \[p5e\] or Lemma \[p5g\]. Thus, we have $$\begin{aligned}
|f(x)-f(p_f)\cos ( |\dot{\gamma}_{p_f,x}^E| d(p_f,x))|&\leq C\delta^{1/200n},\\
|f(p_f)-f(x)\cos ( |\dot{\gamma}_{p_f,x}^E|d(p_f,x))|&\leq C\delta^{1/800n},\end{aligned}$$ and so we get $||f(x)|-|f(p_f)||\leq C\delta^{1/800n}$.
Similarly to $p_f$, we take a point $q_f\in Q_{f}(x)$ with $d(\tilde{q},q_f)\leq C\delta^{1/100n}$, where $\tilde{q}\in M$ is minimum point of $f$. By $\|f\|_{\infty}\geq\|f\|_2=1/\sqrt{n-p+1}$, we have $\max\{|f(p_f)|,|f(q_f)|\}\geq 1/\sqrt{n-p+1}-C\delta^{1/100n}$. Since $|\nabla f|(q_f)\leq C\delta^{1/200n}$, we have $|f(p_f)|\geq |f(q_f)|-C\delta^{1/800n}$ by Claim \[c0\]. Therefore, we get $$\label{54c0}
f(p_f)\geq \frac{1}{\sqrt{n-p+1}}-C\delta^{1/800n}\geq\frac{1}{2\sqrt{n-p+1}}.$$
\[c1\] Take $x\in M$ and $y\in D_f(x)$. Let $\{E^1,\ldots,E^n\}$ be a parallel orthonormal basis along $\gamma_{x,y}$ in Lemma \[p5e\] or Lemma \[p5g\]. If $(i)$ holds in the lemmas, we can assume that $E_1=\dot{\gamma}_{x,y}$. Then, we have $$\begin{aligned}
\label{54ba}|\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle-\langle
\nabla f(x),\dot{\gamma}_{x,y}^{E}(0)\rangle|&\leq C\delta^{1/25},\\
\label{54c} |\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle|&\leq |\nabla f(x)||\dot{\gamma}_{x,y}^{E}|+C\delta^{1/25}\end{aligned}$$ and $$\begin{split}
\left|f\circ \gamma_{x,y}(s)-f(x)\cos (|\dot{\gamma}_{x,y}^{E}|s)-\frac{1}{|\dot{\gamma}_{x,y}^{E}|}\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle\sin (|\dot{\gamma}_{x,y}^{E}|s)\right|&\leq C\delta^{1/250},\\
\left| \langle \nabla f, \dot{\gamma}_{x,y}(s)\rangle+f(x)|\dot{\gamma}_{x,y}^{E}|\sin (|\dot{\gamma}_{x,y}^{E}|s)-\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle\cos (|\dot{\gamma}_{x,y}^{E}|s)\right|&\leq C\delta^{1/250}
\end{split}$$ for all $s\in[0,d(x,y)]$.
If (i) holds in the lemmas, $\dot{\gamma}_{x,y}=\dot{\gamma}_{x,y}^E$, and so (\[54ba\]) and (\[54c\]) are trivial. If (ii) in the lemma holds, we have $|\iota(\nabla f)(E^{n-p+1}\wedge\cdots\wedge E^n)|\leq C\delta^{1/25}$, and so $|\langle\nabla f(x),E_i\rangle|\leq C\delta^{1/25}$ for all $i=n-p+1,\ldots,n$. This gives (\[54ba\]) and (\[54c\]). We get the remaining part of the claim by Lemma \[p5e\] or Lemma \[p5g\] putting $s_0=0$.
\[c2\] For any $x\in D_f(p_f)\cap Q_f\cap R_f$ with $|\nabla f|(x)\geq \delta^{1/800n}$, we have $$|f(x)^2+|\nabla f|^2(x)-f(p_f)^2|\leq C\delta^{1/800n}.$$ Moreover, there exists a point $y\in D_f(p_f)\cap D_f(x)$ such that the following properties hold.
- $d(x,y)< \pi$,
- $|f(p_f)-f(y)|\leq C \delta^{1/800n}$,
- $|f(x)-f(p_f)\cos d(x,y)|\leq C \delta^{1/800n},$
- For any $z\in M$ with $d(x,z)\leq d(x,y)-\delta^{1/2000n}$, we have $$f(p_f)-f(z)\geq \frac{1}{C}\delta^{1/1000n}.$$
Take $x\in D_f(p_f)\cap Q_f\cap R_f$ with $|\nabla f|(x)\geq \delta^{1/800n}$. By the definition of $R_f$, there exists a vector $u\in E_f(x)$ with $$\left|
\frac{\nabla f}{|\nabla f|}(x)-u
\right|\leq C \delta^{1/100n}.$$ Thus, we have $$\label{54d}
\Big|\langle\nabla f(x),\dot{\gamma}_u(0)\rangle-|\nabla f|(x)\Big|=|\nabla f|(x)-\langle\nabla f(x), u\rangle\leq C\delta^{1/100n}.$$ Let $\{E^1,\ldots,E^n\}$ be a parallel orthonormal basis along $\gamma_{u}$ in Lemma \[p5e\] or Lemma \[p5g\]. We first suppose that (ii) holds in the lemmas. Then, for all $i=n-p+1,\ldots, n$, we have $|\langle\nabla f, E_i\rangle|\leq C\delta^{1/25}$, and so $$|\langle u,E_i\rangle|\leq \left| u-\frac{\nabla f}{|\nabla f|}(x)\right|+\left|\langle \frac{\nabla f}{|\nabla f|}(x), E_i\rangle\right|\leq C\delta^{1/100n}+C\delta^{1/25}\delta^{-1/800n}\leq C\delta^{1/100n}.$$ Thus, we get $$|\dot{\gamma}_u^E|^2=|u^E|^2=1-\sum_{i=n-p+1}^n\langle u, E_i\rangle^2\geq 1-C\delta^{1/100n}.$$ If (i) holds in the lemmas, we can assume $u=E_1$, and so $|\dot{\gamma}_u^E|=|u^E|=1$. For both cases, we get we get $$\label{54e}
\begin{split}
|f\circ \gamma_u(s)-f(x)\cos s-|\nabla f|(x)\sin s|\leq& C\delta^{1/100n}\\
|\langle\nabla f,\dot{\gamma}_u(s)\rangle+f(x)\sin s-|\nabla f|(x)\cos s|\leq& C\delta^{1/100n}
\end{split}$$ for all $s\in [0,\pi]$ by (\[54d\]). Take $s_0\in[0,\pi]$ such that $$\begin{aligned}
\frac{f(x)}{(f(x)^2+|\nabla f|^2(x))^{1/2}}=&\cos s_0,\\
\frac{|\nabla f|(x)}{(f(x)^2+|\nabla f|^2(x))^{1/2}}=&\sin s_0.\end{aligned}$$ Since $\sin s_0\geq \frac{1}{C}\delta^{1/800n}$ by the assumption, we have $$\label{54ea}
\frac{1}{C}\delta^{1/800n}\leq s_0\leq \pi-\frac{1}{C}\delta^{1/800n}.$$ By the addition theorem, we have $$\begin{split}
(f(x)^2+|\nabla f|^2(x))^{1/2}\cos (s-s_0)=&f(x)\cos s+|\nabla f|(x)\sin s,\\
(f(x)^2+|\nabla f|^2(x))^{1/2}\sin (s-s_0)=&f(x)\sin s-|\nabla f|(x)\cos s,
\end{split}$$ and so we get $$\begin{split}
|f\circ \gamma_u(s)-(f(x)^2+|\nabla f|^2(x))^{1/2}\cos (s-s_0)|\leq& C\delta^{1/100n},\\
|\langle\nabla f,\dot{\gamma}_u(s)\rangle+(f(x)^2+|\nabla f|^2(x))^{1/2}\sin (s-s_0)|\leq& C\delta^{1/100n}
\end{split}$$ for all $s\in [0,\pi]$ by (\[54e\]). In particular, we get $$\label{54f}
\begin{split}
|f\circ \gamma_u(s_0)-(f(x)^2+|\nabla f|^2(x))^{1/2}|\leq& C\delta^{1/100n},\\
|\langle\nabla f,\dot{\gamma}_u(s_0)\rangle|\leq& C\delta^{1/100n}.
\end{split}$$ Take $y\in D_f(p_f)\cap D_f(x)$ with $d(\gamma_u(s_0),y)\leq C\delta^{1/100n}$. We have $$\label{54fa}
d(x,y)\leq d(x,\gamma_u(s_0))+d(\gamma_u(s_0),y)\leq s_0+C\delta^{1/100n}.$$ By (\[54f\]), we get $$\label{54fb}
|f(y)-(f(x)^2+|\nabla f|^2(x))^{1/2}|\leq C\delta^{1/100n}$$ Take a parallel orthonormal basis $\{\widetilde{E^1},\ldots,\widetilde{E^n}\}$ of $T^\ast M$ along $\gamma_{x,y}$ in Lemma \[p5e\] or Lemma \[p5g\]. By (\[54ea\]) and (\[54fa\]), we get (a) and $$\frac{1}{C}\delta^{1/800n}\leq
|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)+s_0
\leq 2\pi-\frac{1}{C}\delta^{1/800n},$$ and so $$\label{54g}
\cos (|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)+s_0)\leq 1-\frac{1}{C}\delta^{1/400n}.$$ If $|\dot{\gamma}_{x,y}^{\widetilde{E}}|\leq \delta^{1/100}$, we have $|f(y)-f(x)|\leq C\delta^{1/250}$ by Claim \[c1\], and so $$(f(x)^2+|\nabla f|^2(x))^{1/2}-f(x)\leq C\delta^{1/100n}$$ by (\[54f\]). This contradicts to $
|\nabla f|(x)\geq \delta^{1/800n}.
$ Thus, we get $|\dot{\gamma}_{x,y}^{\widetilde{E}}|\geq \delta^{1/100}$. Then, we have $$\label{54g1}
\frac{1}{|\dot{\gamma}_{x,y}^{\widetilde{E}}|}\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle\leq |\nabla f|(x)+C\delta^{3/100}$$ and $$\begin{aligned}
&(f(x)^2+|\nabla f|^2(x))^{1/2}\\
\leq& f(y)+C\delta^{1/100n}\\
\leq& f(x)\cos (|\dot{\gamma}_{x,y}^{E}|d(x,y))+\frac{1}{|\dot{\gamma}_{x,y}^{E}|}\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle\sin (|\dot{\gamma}_{x,y}^{E}|d(x,y))+C\delta^{1/100n}\\
\leq &\left(f(x)^2+\frac{1}{|\dot{\gamma}_{x,y}^{\widetilde{E}}|^2}\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle^2\right)^{1/2}+C\delta^{1/100n}\end{aligned}$$ by Claim \[c1\] and (\[54fb\]). Thus, $$\label{54g2}
|\nabla f|^2(x)
\leq
\frac{1}{|\dot{\gamma}_{x,y}^{\widetilde{E}}|^2}\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle^2 +C\delta^{1/100n}.$$ By (\[54g1\]) and (\[54g2\]), we get $$\label{54h0}
\left|\frac{1}{|\dot{\gamma}_{x,y}^{\widetilde{E}}|^2}\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle^2-|\nabla f|^2(x)\right|\leq C\delta^{1/100n}.$$ This gives $$\label{54h}
\begin{split}
&\left|\frac{1}{|\dot{\gamma}_{x,y}^{\widetilde{E}}|}|\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle|-|\nabla f|(x)\right|\\
\leq& \left|\frac{1}{|\dot{\gamma}_{x,y}^{\widetilde{E}}|^2}\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle^2-|\nabla f|^2(x)\right|\delta^{-1/800n}\leq
C\delta^{7/800n}.
\end{split}$$ We show that $\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle> 0$. If $\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle\leq 0$, we get $$\left|f(y)-f(x)\cos (|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y))+ |\nabla f|\sin (|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y))\right|\leq C\delta^{7/800n}$$ by (\[54h\]) and Claim \[c1\], and so $$\left|f(y)-(f(x)^2+|\nabla f|^2(x))^{1/2}\cos (|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)+s_0)\right|\leq C\delta^{7/800n}.$$ Thus, we get $$\begin{split}
&(f(x)^2+|\nabla f|^2(x))^{1/2}\\
\leq &f(y)+C\delta^{1/100n}\\
\leq &(f(x)^2+|\nabla f|^2(x))^{1/2} \cos (|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)+s_0)+C\delta^{7/800n}\\
\leq &(f(x)^2+|\nabla f|^2(x))^{1/2} -\frac{1}{C}\delta^{3/800n}
\end{split}$$ by (\[54fb\]), (\[54g\]) and $|\nabla f|(x)\geq \delta^{1/800n}$. This is a contradiction. Therefore, we get $\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle>0$. Thus, $$\label{54ha}
\begin{split}
\left|f(y)-(f(x)^2+|\nabla f|^2(x))^{1/2}\cos (|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)-s_0)\right|&\leq C\delta^{7/800n},\\
\left| \langle \nabla f(y), \dot{\gamma}_{x,y}\rangle+|\dot{\gamma}_{x,y}^{\widetilde{E}}|(f(x)^2+|\nabla f|^2(x))^{1/2}\sin (|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)-s_0)\right|&\leq C\delta^{7/800n}
\end{split}$$ by (\[54h\]) and Claim \[c1\]. Then, we have $$\begin{aligned}
(f(x)^2+|\nabla f|^2(x))^{1/2} (1-\cos(|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)-s_0))
\leq C\delta^{7/800n}\end{aligned}$$ by (\[54fb\]), and so $$1-\cos(|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)-s_0)\leq C\delta^{3/400n}.$$ by $|\nabla f|(x)\geq\delta^{1/800n}$. Since $$-\pi<|\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)-s_0<\pi,$$ we get $$\label{54i}
\left||\dot{\gamma}_{x,y}^{\widetilde{E}}|d(x,y)-s_0\right|\leq C\delta^{3/800n}.$$ Thus, we have $$s_0\leq |\dot{\gamma}_{x,y}^{\widetilde{E}}|s_0+ C\delta^{3/800n}$$ by (\[54fa\]), and so $$\label{54j}
1-|\dot{\gamma}_{x,y}^{\widetilde{E}}| \leq C\delta^{1/400n}$$ by (\[54ea\]). Thus, we get $$\label{54k}
|d(x,y)-s_0|\leq C\delta^{1/400n}.$$ By (\[54ha\]) and (\[54i\]), we have $$\label{54l}
|\langle\nabla f(y), \dot{\gamma}_{x,y}(d(x,y))\rangle|\leq C\delta^{3/800n}.$$
We have $$\label{54m}
\begin{split}
&\frac{d}{d s}\left(|\nabla f|^2(s)-\langle\nabla f,\dot{\gamma}_{x,y}(s)\rangle^2\right)\\
=&2\left(\langle\nabla_{\dot{\gamma}_{x,y}}\nabla f,\nabla f\rangle(s)-\langle\nabla_{\dot{\gamma}_{x,y}}\nabla f,\dot{\gamma}_{x,y}(s)\rangle\langle\nabla f,\dot{\gamma}_{x,y}(s)\rangle\right)\\
=&2\langle \nabla^2 f+ f\sum_{i=1}^{n-p}\widetilde{E}^i\otimes \widetilde{E}^i,\dot{\gamma}_{x,y}\otimes\nabla f\rangle(s)-2f\langle\nabla f,\dot{\gamma}_{x,y}^{\widetilde{E}}\rangle\\
&-2\langle \nabla^2 f+ f\sum_{i=1}^{n-p}\widetilde{E}^i\otimes \widetilde{E}^i,\dot{\gamma}_{x,y}\otimes\dot{\gamma}_{x,y}
\rangle(s)\langle\nabla f,\dot{\gamma}_{x,y}(s)\rangle\\
&+2f|\dot{\gamma}_{x,y}^{\widetilde{E}}|^2\langle\nabla f,\dot{\gamma}_{x,y}(s)\rangle.
\end{split}$$ Thus, we get $$\label{54n}
\begin{split}
&\left|\frac{d}{d s}\left(|\nabla f|^2(s)-\langle\nabla f,\dot{\gamma}_{x,y}(s)\rangle^2\right)\right|\\
\leq&C\left|\nabla^2 f+ f\sum_{i=1}^{n-p}\widetilde{E}^i\otimes \widetilde{E}^i\right|
+C\left|\langle\nabla f,\dot{\gamma}_{x,y}^{\widetilde{E}}\rangle-|\dot{\gamma}_{x,y}^{\widetilde{E}}|^2\langle\nabla f,\dot{\gamma}_{x,y}(s)\rangle\right|.\\
\leq &C\left|\nabla^2 f+ f\sum_{i=1}^{n-p}\widetilde{E}^i\otimes \widetilde{E}^i\right|+
C\delta^{1/400n}
\end{split}$$ by (\[54ba\]) and (\[54j\]). By integration, we get $$\int_0^{d(x,y)} \left|\frac{d}{d s}\left(|\nabla f|^2(s)-\langle\nabla f,\dot{\gamma}_{x,y}(s)\rangle^2\right)\right|\,d s
\leq C\delta^{1/400n},$$ and so $$\Big||\nabla f|^2(y)-\langle\nabla f(y),\dot{\gamma}_{x,y}(d(x,y))\rangle^2
-|\nabla f|^2(x)+\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle^2
\Big|\leq C\delta^{1/400n}.$$ Thus, we get $$|\nabla f|(y)\leq C\delta^{1/800n}.$$ by (\[54h0\]), (\[54j\]) and (\[54l\]). By Claim \[c0\] and (\[54c0\]), we get $$\left||f(y)|-f(p_f)\right|\leq C\delta^{1/800n}.$$ Since $$f(y)\geq (f(x)^2+|\nabla f|^2(x))^{1/2}-C\delta^{1/100n}\geq \delta^{1/800n}-C\delta^{1/100n}>0$$ by (\[54fb\]), we get $$\left|f(y)-f(p_f)\right|\leq C\delta^{1/800n}.$$ This gives (b). We get $$\label{54o}
|(f(x)^2+|\nabla f|^2(x))^{1/2}-f(p_f)|
\leq C\delta^{1/800n}$$ by (\[54ha\]), (\[54i\]) and (b), and so we get (c) by the definition of $s_0$ and (\[54k\]). (\[54o\]) implies the first assertion.
Finally, we show (d). Suppose that a point $z\in M$ satisfies $d(x,z)\leq d(x,y)-\delta^{1/2000n}$. Then, $d(x,y)\geq \delta^{1/2000n}$, and so $$f(x)\leq f(p_f)\cos d(x,y)+C\delta^{1/800n}\leq f(p_f)-\frac{1}{C}\delta^{1/1000n}$$ by (\[54c0\]). There exists $w\in D_f(x)$ with $d(z,w)\leq C\delta^{1/100n}$. Let $\{\overline{E}^1,\ldots,\overline{E}^n\}$ be a parallel orthonormal basis along $\gamma_{x,w}$ in Lemma \[p5e\] or Lemma \[p5g\]. If (i) holds in the lemmas, we assume that $\overline{E}_1=\dot{\gamma}_{x,w}$. If $|\dot{\gamma}_{x,w}^{\overline{E}}|\leq \delta^{1/100}$, we have $$f(z)\leq f(w)+C\delta^{1/100n}
\leq f(x)+ C\delta^{1/100n}
\leq f(p_f)-\frac{1}{C}\delta^{1/1000n}$$ by Claim \[c1\]. If $|\dot{\gamma}_{x,w}^{\overline{E}}|\geq \delta^{1/100}$, we have $$\begin{split}
f(z)\leq& f(w)+C\delta^{1/100n}\\
\leq& f(x)\cos (|\dot{\gamma}_{x,w}^{\overline{E}}|d(x,z))+|\nabla f|(x)\sin (|\dot{\gamma}_{x,w}^{\overline{E}}|d(x,z))+C\delta^{1/100n}\\
\leq& f(p_f)\cos (|\dot{\gamma}_{x,w}^{\overline{E}}|d(x,z)-d(x,y))+\delta^{1/800n}
\leq f(p_f)-\frac{1}{C}\delta^{1/1000n}
\end{split}$$ by Claim \[c1\], (\[54k\]), (\[54o\]) and $-\pi\leq|\dot{\gamma}_{x,w}^{\overline{E}}|d(x,z)-d(x,y)\leq -\delta^{1/2000n}$. For both cases, we get (d).
By Claim \[c0\] and Claim \[c2\], we get $$\label{54p}
|f(x)^2+|\nabla f|^2(x)-f(p_f)^2|\leq C\delta^{1/800n}$$ for all $x\in D_f(p_f)\cap Q_f\cap R_f$.
\[c3\] We have $$|f(p_f)-1|\leq C\delta^{1/800n}.$$
Since $$\|f^2+|\nabla f|^2-f(p_f)^2\|_{\infty}\leq C$$ and $${\mathop{\mathrm{Vol}}\nolimits}(M\setminus (D_f(p_f)\cap Q_f\cap R_f) )\leq C\delta^{1/100},$$ we get $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M|f(x)^2+|\nabla f|^2(x)-f(p_f)^2| \,d\mu_g\leq C \delta^{1/800n}$$ by (\[54p\]). By the assumption, we have $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\left|\int_M (f(x)^2+|\nabla f|^2(x)-1) \,d\mu_g\right|\leq C \delta^{1/2}$$ Thus, we get $$|f(p_f)^2-1|\leq C\delta^{1/800n}.$$ Since $f(p_f)>0$, we get the claim.
By Claim \[c0\], Claim \[c3\] and (\[54p\]), we get (i), (ii) and (iii).
Finally, we prove (iv). Put $$A_f:=\{x\in M: |f(x)-1|\leq \delta^{1/900n}\}.$$ Since we have $$|f(w)-\cos d(w,A_f)|\leq \delta^{1/900n}$$ for all $w\in A_f$, we get (iv) on $A_f$.
Let us show (iv) on $M\setminus A_f$. Take $w\notin A_f$ and $x\in D_f(p_f)\cap Q_f\cap R_f $ with $d(w,x)\leq C\delta^{1/100n}$.
We first suppose that $|\nabla f|(x)\geq \delta^{1/800n}$. Take $y\in D_f(p_f)\cap D_f(x)$ of Claim \[c2\]. Then, $|f(y)-1|\leq C\delta^{1/800n}$, and so $y\in A_f$. Thus, $$\label{54q}
d(x, A_f)\leq d(x,y)<\pi.$$ For all $z\in A_f$, we have $|f(p_f)-f(z)|\leq C\delta^{1/900n}$, and so $d(x,z)> d(x,y)-\delta^{1/2000n}$ by Claim \[c2\] (d). Thus, $$\label{54r}
d(x,A_f)\geq d(x,y)-\delta^{1/2000n}.$$ By (\[54q\]) and (\[54r\]), we get $$|d(x,A_f)- d(x,y)|\leq \delta^{1/2000n}.$$ Therefore, we have $|f(x)-\cos d(x,A_f)|\leq C\delta^{1/2000n}$ by Claim \[c2\] (c), and so $|f(w)-\cos d(w,A_f)|\leq C\delta^{1/2000n}$. By (\[54q\]), we have $d(w,A_f)\leq \pi+C\delta^{1/100n}$.
We next suppose that $|\nabla f|(x)\leq \delta^{1/800n}$. Then, $||f|(x)-1|\leq C\delta^{1/800n}$ by Claim \[c0\]. If $f(x)\geq 0$, then $w \in A_f$. This contradicts to $w\notin A_f$. Thus, we have $|f(x)+1|\leq C\delta^{1/800n}$. We see that (i) in Lemma \[p5e\] or Lemma \[p5g\] cannot occur for $\gamma_{p_f,x}$ because we have $$|\nabla^2 f|\geq \frac{1}{\sqrt{n}}|\Delta f|\geq\frac{n-p}{\sqrt{n}}|f|-C\delta^{1/2}.$$ Thus, there exists an orthonormal basis $\{e^1,\ldots,e^n\}$ of $T_x^\ast M$ such that $|\omega(x)-e^{n-p+1}\wedge\cdots\wedge e^n|\leq C\delta^{1/25}$ if $\lambda_1(\Delta_{C,p})\leq \delta$, where $\omega$ denotes the first eigenform of $\Delta_{C,p}$ with $\|\omega\|_2=1$, and $|\xi(x)-e^{1}\wedge\cdots\wedge e^{n-p}|\leq C\delta^{1/25}$ if $\lambda_1(\Delta_{C,n-p})\leq \delta$, where $\xi$ denotes the first eigenform of $\Delta_{C,n-p}$ with $\|\xi\|_2=1$. Take $u\in E_f(x)$ with $|u-e_1|\leq C\delta^{1/100n}$. Then, we get $$|f\circ\gamma_u(s)+\cos s|\leq C\delta^{1/800n}$$ for all $s\in [0,\pi]$ by Lemma \[p5e\] or Lemma \[p5g\]. Thus, we get $\gamma_u(\pi)\in A_f$, and so $$\label{54s}
d(w,A_f)\leq \pi+C\delta^{1/100n}.$$ For any $y\in A_f$, there exists $z\in D_f(x)$ with $d(y,z)\leq C\delta^{1/100n}$. Let $\{E^1,\ldots,E^n\}$ be a parallel orthonormal basis of $T^\ast M$ along $\gamma_{x,z}$ of Claim\[c1\]. Then, $$|1+\cos ( |\dot{\gamma}_{x,z}^{E}| d(x,z))|\leq C\delta^{1/900n}$$ by Claim \[c1\]. Thus, we get $d(x,z)\geq\pi-C\delta^{1/1800n}$, and so $$\label{54t}
d(w,A_f)\geq \pi- C\delta^{1/1800n}.$$ By (\[54s\]) and (\[54t\]), we get $|d(w,A_f)-\pi|\leq C\delta^{1/1800n}$, and so $|f(w)-\cos d(w,A_f)|\leq C\delta^{1/1800n}$.
For both cases, we get (iv).
Gromov-Hausdorff Approximation
------------------------------
In this subsection, we construct a Hausdorff approximation map, and show that the Riemannian manifold is close to the product metric space $S^{n-p}\times X$ in the Gromov-Hausdorff topology under our pinching condition.
In this subsection, we assume the following in addition to Assumption \[asu1\].
- $\lambda_{n-p+1}(g)\leq n-p+\delta$.
- $f_i$ is the $i$-th eigenfunction of the Laplacian acting on functions with $\|f_i\|_2^2=1/(n-p+1)$.
- Either $\lambda_1(\Delta_{C,p})\leq \delta$ or $\lambda_1(\Delta_{C,n-p})\leq \delta.$
The following proposition is based on [@Pe1 Lemma 5.2].
\[p54a\] Define $\widetilde{\Psi}:=(f_1,\dots,f_{n-p+1})\colon M\to \mathbb{R}^{n-p+1}$. Then, we have $$\||\widetilde{\Psi}|^2-1\|_{\infty}\leq C\delta^{1/800n^2}.$$
We first prove the following claim:
\[p54b\] For any $x\in M$, we have $|\widetilde{\Psi}|(x)\leq 1+C\delta^{1/800n}$
If $|\widetilde{\Psi}|(x)=0$, the claim is trivial. Thus, we assume that $|\widetilde{\Psi}|(x)\neq 0$. Put $$f_x:=\frac{1}{|\widetilde{\Psi}|(x)}\sum_{i=1}^{n-p+1} f_i(x)f_i.$$ Then, we have $$\|f_x\|_2^2=\frac{1}{n-p+1}.$$ Thus, we get $$|\widetilde{\Psi}|(x)=f_x(x)\leq 1+ C\delta^{1/800n}$$ by Proposition \[p53a\] (i). This gives the claim.
We need the following claim [@Pe1 Theorem 7.1]. Note that our sign convention of the Laplacian is different from [@Pe3].
\[claimpe\] For a smooth functions $u\in C^\infty(M)$ and a non-negative continuous function $F$ with $\Delta u\leq F$, we have $$u\leq C\|F\|_n+\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}M}\int_M u\,d\mu_g.$$
To apply Claim \[claimpe\] to $-|\widetilde{\Psi}|^2$, we compute $\Delta|\widetilde{\Psi}|^2$. $$\label{55a}
\begin{split}
\Delta|\widetilde{\Psi}|^2
=&\Delta \sum_{i=1}^{n-p+1} f_i^2\\
=&2\sum_{i=1}^{n-p+1} f_i \Delta f_i-2\sum_{i=1}^{n-p+1} |\nabla f_i|^2\\
=&2\sum_{i=1}^{n-p+1} (\Delta f_i-(n-p) f_i)f_i+2(n-p+1)(|\widetilde{\Psi}|^2-1)\\
&\quad-2\sum_{i=1}^{n-p+1} (f_i^2+|\nabla f_i|^2-1).
\end{split}$$ We estimate each component.
By the assumption, we have $$\label{55b}
\|(\Delta f_i-(n-p) f_i)f_i\|_{\infty}\leq C \delta^{1/2}$$ for each $i$.
We next estimate $\||\widetilde{\Psi}|^2-1\|_n$. For $x\in M$ with $|\widetilde{\Psi}(x)|^2-1< 0$, we have $||\widetilde{\Psi}(x)|^2-1|=1-|\widetilde{\Psi}(x)|^2$. For $x\in M$ with $|\widetilde{\Psi}(x)|^2-1\geq 0$, we have $||\widetilde{\Psi}(x)|^2-1|=|\widetilde{\Psi}(x)|^2-1 \leq 1-|\widetilde{\Psi}(x)|^2+C\delta^{1/800n}$ by Claim \[p54b\]. For both cases, we have $||\widetilde{\Psi}(x)|^2-1|\leq 1-|\widetilde{\Psi}(x)|^2+C\delta^{1/800n}$. Combining this and $\|\widetilde{\Psi}\|_2=1$, we get $$\||\widetilde{\Psi}|^2-1\|_1 \leq C \delta^{1/800n}.$$ Since $||\widetilde{\Psi}(x)|^2-1|\leq 1$ for all $x\in M$, we get $$\||\widetilde{\Psi}|^2-1\|_n^n\leq \||\widetilde{\Psi}|^2-1\|_1\leq C \delta^{1/800n}.$$ Thus, we get $$\label{55c}
\||\widetilde{\Psi}|^2-1\|_n\leq C\delta^{1/800n^2}.$$
Finally, we estimate $\|f_i^2+|\nabla f_i|^2-1\|_n$. Since we have $$\|f_i^2 +|\nabla f_i|^2-1\|_\infty\leq C,$$ we get $$\begin{split}
\|f_i^2+|\nabla f_i|^2 -1\|_n^n
\leq &C \|f_i^2 +|\nabla f_i|^2-1\|_1
\leq C\delta^{1/800n}
\end{split}$$ by Proposition \[p53a\] (iii). Therefore, we get $$\label{55d}
\|f_i^2+|\nabla f_i|^2 -1\|_n\leq C\delta^{1/800n^2}.$$
By (\[55a\]), (\[55b\]), (\[55c\]) and (\[55d\]), we get $$\|\Delta|\widetilde{\Psi}|^2\|_n\leq C \delta^{1/800n^2}.$$ Since we have $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M |\widetilde{\Psi}|^2\,d\mu_g=1,$$ we get the lemma by Claim \[claimpe\].
In the remaining part of this subsection, we use the following notation.
- Let $d_S$ denotes the intrinsic distance function on $S^{n-p}(1)$. Note that we have $\cos d_S(x,y)=x\cdot y$ and $$d_{\mathbb{R}^{n-p+1}}(x,y)\leq d_{S}(x,y)\leq 3 d_{\mathbb{R}^{n-p+1}}(x,y)$$ for all $x,y\in S^{n-p}\subset\mathbb{R}^{n-p+1}$.
- For each $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$, we use the notation $p_f$ and $A_f$ of Proposition \[p53a\]. Recall that we defined $$A_f:=\{x\in M: |f(x)-1|\leq \delta^{1/900n}\}.$$
- Define $\widetilde{\Psi}:=(f_1,\dots,f_{n-p+1})\colon M\to \mathbb{R}^{n-p+1}$ and $$\Psi:=\frac{\widetilde{\Psi}}{|\widetilde{\Psi}|}\colon M\to S^{n-p}.$$
- For each $x\in M$, put $$f_x:=\frac{1}{|\widetilde{\Psi}|(x)}\sum_{i=1}^{n-p+1} f_i(x)f_i=\sum_{i=1}^{n-p+1} \Psi_i(x)f_i,$$ $p_x:=p_{f_x}$ and $A_x:=A_{f_x}$.
- For each $x\in M$ and $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$, choose $a_f(x)\in A_f$ such that $$d(x,A_f)=d(x,a_f(x)).$$
The goal of this subsection is to show that $$\Phi_f \colon M\to S^{n-p}\times A_f,\,x\mapsto (\Psi(x),a_f(x))$$ is an approximation map.
\[p54c00\] For all $x,y\in M$, we have $$|\Psi(x)-\Psi(y)|\leq Cd(x,y).$$
Since we have $\|\nabla f_i\|_{\infty}\leq C$ for all $i\in\{1,\ldots,n-p+1\}$, we get $|\widetilde{\Psi}(x)-\widetilde{\Psi}(y)|\leq Cd(x,y)$ for all $x,y\in M$. Thus, we get the lemma by Lemma \[p54a\].
\[p54c0\] Take $u\in S^{n-p}$ and put $f=\sum_{i=1}^{n-p+1}u_i f_i$. Then, we have $$|d_S(\Psi(y),u)-d(y,A_{f})|\leq C\delta^{1/1600n^2}$$ for all $y\in M$.
Since $$f(y)=u\cdot\widetilde{\Psi}(y),$$ we have $$|u \cdot\widetilde{\Psi}(y)-\cos d(y,A_{f})|\leq C\delta^{1/2000n}$$ by Proposition \[p53a\], and so $$|u\cdot \Psi(y)-\cos d(y,A_{f})|\leq C\delta^{1/800n^2}$$ by Lemma \[p54a\]. Thus, we get $$|\cos d_S(\Psi(y),u)-\cos d(y,A_{f})|\leq C\delta^{1/800n^2}.$$ This and $d(y,A_{f})\leq \pi+C\delta^{1/100n}$ give the lemma.
By the definition of $A_{y}$, we immediately get the following corollaries:
\[p54c01\] Take $u\in S^{n-p}$ and put $f=\sum_{i=1}^{n-p+1}u_i f_i$. Then, we have $$d_S(\Psi(p_f),u)\leq C\delta^{1/1600n^2}.$$
\[p54c\] For each $y_1,y_2\in M$, we have $$|d_S(\Psi(y_1),\Psi(y_2))-d(y_2,A_{y_1})|\leq C\delta^{1/1600n^2}.$$
\[p54c1\] For each $y\in M$, we have $$d(y,A_{y})\leq C\delta^{1/1600n^2}.$$
We need to show the almost Pythagorean theorem for our purpose. To do this, we regard $|\dot{\gamma}^E| s$ in Lemma \[p5e\] or Lemma \[p5g\] as a moving distance in $S^{n-p}$. We first approximate their cosine.
\[p54d\] Take $y_1\in M$, $\tilde{y}_1\in D_{f_{y_1}}(p_{y_1})\cap R_{f_{y_1}}\cap Q_{f_{y_1}}$ with $d(y_1,\tilde{y}_1)\leq C\delta^{1/100n}$ and $y_2\in D_{f_{y_1}}(\tilde{y}_1)$ $($note that we can take such $\tilde{y}_1$ for any $y_1$ by the Bishop-Gromov theorem$)$. Let $\{E^1,\ldots,E^n\}$ be a parallel orthonormal basis of $T^\ast M$ along $\gamma_{\tilde{y}_1,y_2}$ in Lemma \[p5e\] or Lemma \[p5g\] for $f_{y_1}$. Then, $(ii)$ holds in the lemmas, and $$|\cos(|\dot{\gamma}_{\tilde{y}_1,y_2}^E|s)-\cos d_S(\Psi(y_1),\Psi(\gamma_{\tilde{y}_1,y_2}(s)))|\leq C\delta^{1/1600n^2}$$ for all $s\in[0,d(\tilde{y}_1,y_2)]$. In particular, we have $$|\cos(|\dot{\gamma}_{\tilde{y}_1,y_2}^E|d(\tilde{y}_1,y_2))-\cos d_S(\Psi(y_1),\Psi(y_2))|\leq C\delta^{1/1600n^2}.$$
By Corollary \[p54c1\], we have $$d(\tilde{y}_1,A_{y_1})\leq C\delta^{1/1600n^2},$$ and so we get $$f\circ \gamma_{\tilde{y}_1,y_2}(s)
\geq \cos d(\gamma_{\tilde{y}_1,y_2}(s),A_{y_1})- C\delta^{1/2000n}
\geq \cos s- C\delta^{1/1600n^2}
\geq \frac{1}{\sqrt{2}}- C\delta^{1/1600n^2}$$ for all $s\leq\min\{\pi/4,d(\tilde{y}_1,y_2)\}$. Therefore, we have $$|\nabla^2 f|(\gamma_{\tilde{y}_1,y_2}(s))
\geq \frac{1}{\sqrt{n}}|\Delta f|(\gamma_{\tilde{y}_1,y_2}(s))
\geq \frac{n-p}{\sqrt{2n}}- C\delta^{1/1600n^2}$$ for all $s\leq\min\{\pi/4,d(\tilde{y}_1,y_2)\}$. Thus, (i) in Lemma \[p5e\] or Lemma \[p5g\] cannot occur, and so (ii) holds in the lemmas.
Since we have $f_{y_1}(y_1)=|\widetilde{\Psi}(y_1)|$, we get $$\label{ad1}
|f_{y_1}(\tilde{y}_1)-1|\leq C\delta^{1/800n^2}$$ by Lemma \[p54a\] and $d(y_1,\tilde{y}_1)\leq C\delta^{1/100n}$. By (\[ad1\]) and Proposition \[p53a\] (iii), we have $$|\nabla f_{y_1}|(\tilde{y}_1)\leq C\delta^{1/1600n^2}.$$ Thus, we get $$|f_{y_1}(\gamma_{\tilde{y}_1,y_2}(s))-\cos(|\dot{\gamma}_{\tilde{y}_1,y_2}^E|s)|\leq C\delta^{1/1600n^2}$$ for all $s\in[0,d(\tilde{y}_1,y_2)]$ by Lemma \[p5e\] or Lemma \[p5g\]. On the other hand, we have $$|f_{y_1}(\gamma_{\tilde{y}_1,y_2}(s))-\cos d_S(\Psi(y_1),\Psi(\gamma_{\tilde{y}_1,y_2}(s)))|\leq C\delta^{1/1600n^2}$$ for all $s\in[0,d(\tilde{y}_1,y_2)]$ by Proposition \[p53a\] (iv) and Corollary \[p54c\]. Thus, we get the lemma.
We use the following notation:
- For any $y_1,y_2\in M$ and $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$, define $$\begin{aligned}
&G_f^{y_1}(y_2)\\
:=&\langle\dot{\gamma}_{y_2,y_1}(0),\nabla f(y_2)\rangle d(y_1,y_2)\sin d_S(\Psi(y_1),\Psi(y_2))\\
&\quad +\Big(\cos d(y_2, A_f)\cos d_S(\Psi(y_1),\Psi(y_2))-\cos d(y_1,A_f)\Big)
d_S(\Psi(y_1),\Psi(y_2)).\end{aligned}$$
- For any $y_1,y_2\in M$, define
[align\*]{} &1 d(y\_1,y\_2),\
&0 d(y\_1,y\_2)>.
- For any $y_1,y_2\in M$ and $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$, define $$\begin{aligned}
C_f^{y_1}(y_2):=&\Big\{y_3\in M : \gamma_{y_2,y_3}(s)\in I_{y_1}\setminus\{y_1\} \text{ for almost all $s\in[0,d(y_2,y_3)]$, and}\\
&\qquad \qquad\qquad \qquad \int_{0}^{d(y_2,y_3)} |G_f^{y_1}H^{y_1}|(\gamma_{y_2,y_3}(s))\,d s\leq \delta^{1/9600n^2}\Big\},\\
P_f^{y_1}:=&\{y_2\in M: {\mathop{\mathrm{Vol}}\nolimits}(M\setminus C_f^{y_1}(y_2))\leq\delta^{1/9600n^2}{\mathop{\mathrm{Vol}}\nolimits}(M)\}.\end{aligned}$$
Pinching condition on $G_f^{y_1}$ plays a crucial role for our purpose. Let us estimate $G_f^{y_1}$.
\[p54e\] Take $\eta>0$ with $\eta\geq \delta^{1/2000n}$, $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$, $y_1\in Q_f$ and $y_2\in D_f(y_1)$. Let $\{E^1,\ldots,E^n\}$ be a parallel orthonormal basis of $T^\ast M$ along $\gamma_{y_1,y_2}$ in Lemma \[p5e\] or Lemma \[p5g\] for $f$. If $$||\dot{\gamma}_{y_1,y_2}^E|d(y_1,y_2)-d_S(\Psi(y_1),\Psi(y_2))|\leq \eta,$$ then $$|G_f^{y_1}(y_2)|\leq C\eta.$$
We have $$\begin{aligned}
\Big|f(y_1)-f(y_2)\cos &(|\dot{\gamma}_{y_1,y_2}^E|d(y_1,y_2))\\
&-\frac{1}{|\dot{\gamma}_{y_1,y_2}^E|}\langle\nabla f(y_2),\dot{\gamma}_{y_2,y_1}(0)\rangle\sin (|\dot{\gamma}_{y_1,y_2}^E|d(y_1,y_2))
\Big|
\leq C\delta^{1/250}\end{aligned}$$ by Lemma \[p5e\] or Lemma \[p5g\]. Thus, by Proposition \[p53a\] (iv), we get $$\begin{aligned}
\Big||\dot{\gamma}_{y_1,y_2}^E|\cos d(y_1,A_f)&-|\dot{\gamma}_{y_1,y_2}^E|\cos d(y_2, A_f)\cos (|\dot{\gamma}_{y_1,y_2}^E|d(y_1,y_2))\\
&-\langle\nabla f(y_2),\dot{\gamma}_{y_2,y_1}(0)\rangle\sin (|\dot{\gamma}_{y_1,y_2}^E|d(y_1,y_2))
\Big|
\leq C\delta^{1/2000n},\end{aligned}$$ and so we get the lemma.
The quantity $|\dot{\gamma}_{y_1,y_2}^E|$ in the above lemma is slightly different from that of Lemma \[p54d\]. Comparing these two quantity, we get the following:
\[p54f0\] Take $\eta>0$ with $\eta\geq \delta^{1/2000n}$, $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$, $y_1\in M$, $\tilde{y}_1\in D_{f_{y_1}}(p_{y_1})\cap R_{f_{y_1}}\cap Q_{f_{y_1}}\cap Q_f$ with $d(y_1,\tilde{y}_1)\leq C\delta^{1/100n}$ and $y_2\in D_{f_{y_1}}(\tilde{y}_1)\cap D_f(\tilde{y}_1)$. Let $\{E^1,\ldots,E^n\}$ be a parallel orthonormal basis of $T^\ast M$ along $\gamma_{\tilde{y}_1,y_2}$ in Lemma \[p5e\] or Lemma \[p5g\] for $f_{y_1}$. If $$||\dot{\gamma}_{\tilde{y}_1,y_2}^E|d(\tilde{y}_1,y_2)-d_S(\Psi(\tilde{y}_1),\Psi(y_2))|\leq \eta,$$ then $$|G_f^{\tilde{y}_1}(y_2)|\leq C\eta.$$
Let $\{\widetilde{E}^1,\ldots,\widetilde{E}^n\}$ be a parallel orthonormal basis of $T^\ast M$ along $\gamma_{\tilde{y}_1,y_2}$ in Lemma \[p5e\] or Lemma \[p5g\] for $f$ (if (i) holds, then we can assume that $\widetilde{E}^i=E^i$ for all $i$). We show that $$\left||\dot{\gamma}_{\tilde{y}_1,y_2}^E|-|\dot{\gamma}_{\tilde{y}_1,y_2}^{\widetilde{E}}|\right|\leq C\delta^{1/50}.$$ Then, we immediately get the corollary by Lemma \[p54e\].
We first suppose that $\lambda_1(\Delta_{C,p})\leq \delta$. Let $\omega$ be the first eigenform of $\Delta_{C,p}$ with $\|\omega\|_2=1$. We have $|\omega(y_2)-E^{n-p+1}\wedge \cdots\wedge E^n|\leq C\delta^{1/25}$ by Lemma \[p5e\] and Lemma \[p54d\]. Since $|\dot{\gamma}_{\tilde{y}_1,y_2}^E|^2=1-|\iota(\dot{\gamma}_{\tilde{y}_1,y_2})(E^{n-p+1}\wedge \cdots\wedge E^n)|^2$, we get $$\label{55e}
\left||\dot{\gamma}_{\tilde{y}_1,y_2}^E|^2-\left(1-|\iota(\dot{\gamma}_{\tilde{y}_1,y_2})\omega|^2(y_2)\right)\right|\leq C\delta^{1/25}.$$ Similarly, we get $$\label{55f}
\left||\dot{\gamma}_{\tilde{y}_1,y_2}^{\widetilde{E}}|^2-\left(1-|\iota(\dot{\gamma}_{\tilde{y}_1,y_2})\omega|^2(y_2)\right)\right|\leq C\delta^{1/25}.$$ By (\[55e\]) and (\[55f\]), we get $$\left||\dot{\gamma}_{\tilde{y}_1,y_2}^E|-|\dot{\gamma}_{\tilde{y}_1,y_2}^{\widetilde{E}}|\right|\leq C\delta^{1/50}.$$
We next suppose that $\lambda_1(\Delta_{C,n-p})\leq \delta$. Let $\xi$ be the first eigenform of $\Delta_{C,n-p}$ with $\|\xi\|_2=1$. Similarly to the case when $\lambda_1(\Delta_{C,p})\leq \delta$, we have $$\begin{aligned}
\left||\dot{\gamma}_{\tilde{y}_1,y_2}^E|^2-|\iota(\dot{\gamma}_{\tilde{y}_1,y_2})\xi|^2(y_2)\right|\leq& C\delta^{1/25},\\
\left||\dot{\gamma}_{\tilde{y}_1,y_2}^{\widetilde{E}}|^2-|\iota(\dot{\gamma}_{\tilde{y}_1,y_2})\xi|^2(y_2)\right|\leq& C\delta^{1/25},\end{aligned}$$ and so $$\left||\dot{\gamma}_{\tilde{y}_1,y_2}^E|-|\dot{\gamma}_{\tilde{y}_1,y_2}^{\widetilde{E}}|\right|\leq C\delta^{1/50}.$$
By the above two cases, we get the corollary.
Let us show the integral pinching condition.
\[p54f\] Take $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$, $y_1\in M$ and $\tilde{y}_1\in D_{f_{y_1}}(p_{y_1})\cap R_{f_{y_1}}\cap Q_{f_{y_1}}\cap Q_f$ with $d(y_1,\tilde{y}_1)\leq C\delta^{1/100n}$. Then, $$\|G_f^{\tilde{y}_1} H_{\tilde{y}_1}\|_1\leq C\delta^{1/3200n^2},$$ and $${\mathop{\mathrm{Vol}}\nolimits}(M\setminus P_f^{\tilde{y}_1})\leq C\delta^{1/9600n^2}.$$
Take arbitrary $y_2\in D_f(\tilde{y_1})\cap D_{f_{y_1}}(\tilde{y}_1)$. Let $\{E^1,\ldots,E^n\}$ be a parallel orthonormal basis of $T^\ast M$ along $\gamma_{y_1,y_2}$ in Lemma \[p5e\] or Lemma \[p5g\] for $f_{y_1}$. Then, we have $$||\dot{\gamma}_{\tilde{y}_1,y_2}^E|d(\tilde{y}_1,y_2)-d_S(\Psi(\tilde{y}_1),\Psi(y_2))|\leq C\delta^{1/3200n^2},$$ if $d(\tilde{y}_1,y_2)\leq \pi$ by Lemma \[p54c00\] and Lemma \[p54d\]. Thus, by Corollary \[p54f0\], we have $$\sup_{D_f(\tilde{y_1})\cap D_{f_{y_1}}(\tilde{y}_1)}|G_f^{\tilde{y}_1} H^{\tilde{y}_1}|\leq C\delta^{1/3200n^2}.$$ Since ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus (D_f(\tilde{y}_1)\cap D_{f_{y_1}}(\tilde{y}_1)))\leq C\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(M)$ and $\|G_f^{\tilde{y}_1} H^{\tilde{y}_1}\|_\infty\leq C$, we get $$\|G_f^{\tilde{y}_1} H^{\tilde{y}_1}\|_1\leq C\delta^{1/3200n^2}.$$ By the segment inequality (Theorem \[seg\]), we get the lemma.
\[order\] We use the following notation.
- $\eta_0=\delta^{1/9600n^3}$.
- $\eta_1=\eta_0^{1/26}$.
- $\eta_2=\eta_1^{1/78}$.
- $L=\eta_2^{1/150}$.
We use Lemma \[p54f\] to give the almost Pythagorean theorem for the special case (see Lemma \[p54l\]). For the general case, we need to estimate $\|G_f^{\tilde{y}_1}\|_1$. To do this, we show that $|\dot{\gamma}_{\tilde{y}_1,y_2}^E|d(\tilde{y}_1,y_2)\leq \pi+L$ under the assumption of Lemma \[p54d\] in Lemma \[p54n\]. Then, we can estimate $\|G_f^{\tilde{y}_1}\|_1$ similarly to Lemma \[p54f\]. After proving that, we use Lemma \[p54i\] again to give the almost Pythagorean theorem for the general case. The following lemma, which guarantees that an almost shortest pass from a point in $M$ to $A_f$ almost corresponds to a geodesic in $S^{n-p}$ through $\Psi$ under some assumptions, is the first step to achieve these objectives.
\[p54g\] Take
- $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$,
- $u\in S^{n-p}$ with $f=\sum_{i=1}^{n-p+1}u_i f_i$,
- $x,y\in M$,
- $\eta>0$ with $\eta_0\leq\eta\leq L^{1/3n}$.
Suppose
- $d(y,A_f)\leq C \eta$,
- $|d(x,A_f)-d(x,y)|\leq C\eta$.
Then, we have the following for all $s,s'\in[0,d(x,y)]$:
- $|d(\gamma_{y,x}(s),A_f)-s|\leq C\eta$,
- $\left||s-s'|-d_S\left(\Psi(\gamma_{y,x}(s)),\Psi(\gamma_{y,x}(s'))\right)\right|\leq C\eta$,
- If in addition $d(x,A_f)\geq \frac{1}{C}\eta^{1/26}$, there exists $v\in S^{n-p}$ such that $u\cdot v=0$ and $$d_S(\Psi(\gamma_{y,x}(s)),\gamma_v(s))\leq C\eta^{3/13}$$ for all $s\in[0,d(x,y)]$, where we define $\gamma_v(s):=\cos s u+\sin s v\in S^{n-p}$.
We first prove (i). We have $$d(\gamma_{y,x}(s),A_f)\leq d(\gamma_{y,x}(s),y)+d(y,A_f)\leq s+ C\eta,$$ and $$\begin{aligned}
d(x,y)-C\eta\leq d(x,A_f)\leq &d(\gamma_{y,x}(s),A_f)+d(\gamma_{y,x}(s),x)\\
=&d(\gamma_{y,x}(s),A_f)+d(x,y)-s.\end{aligned}$$ Thus, we get (i).
We next prove (ii). Since $d(y,A_f)\leq C \eta$, we have $\cos d(y,A_f)\geq 1-C\eta^2$, and so $$\label{55g}
|f(y)-1|\leq C\eta^2$$ by Proposition \[p53a\] (iv). On the other hand, we have $f(y)=u\cdot \widetilde{\Psi}(y)$, and so $$\label{55h}
|f(y)-\cos d_S(u,\Psi(y))|\leq C\delta^{1/800n^2}$$ by Lemma \[p54a\]. By (\[55g\]) and (\[55h\]), we get $|1-\cos d_S(u,\Psi(y))|\leq C\eta^2$. This gives $d_S(u,\Psi(y))\leq C\eta$. Thus, we get $$\label{55i}
|s-d_S(\Psi(\gamma_{y,x}(s)),\Psi(y))|\leq C \eta$$ for all $s\in[0,d(x,y)]$ by (i) and Lemma \[p54c0\]. Take arbitrary $s,s'\in[0,d(x,y)]$ with $s<s'$. Then, $$\label{55j}
\begin{split}
s'-s=d(\gamma_{y,x}(s),\gamma_{y,x}(s'))&\geq d(\gamma_{y,x}(s),A_{\gamma_{y,x}(s')})-d(\gamma_{y,x}(s'),A_{\gamma_{y,x}(s')})\\
&\geq d_S(\Psi(\gamma_{y,x}(s)),\Psi(\gamma_{y,x}(s')))-C\delta^{1/1600n^2}
\end{split}$$ by Corollary \[p54c\] and Corollary \[p54c1\]. On the other hand, we have $$\begin{split}
s'-C\eta\leq& d_S(\Psi(\gamma_{y,x}(s')),\Psi(y))\\
\leq &d_S(\Psi(\gamma_{y,x}(s)),\Psi(\gamma_{y,x}(s')))+d_S(\Psi(\gamma_{y,x}(s)),\Psi(y))\\
\leq &d_S(\Psi(\gamma_{y,x}(s)),\Psi(\gamma_{y,x}(s'))) +s+C\eta
\end{split}$$ by (\[55i\]), and so $$\label{55k}
s'-s\leq d_S(\Psi(\gamma_{y,x}(s)),\Psi(\gamma_{y,x}(s'))) +C\eta.$$ By (\[55j\]) and (\[55k\]), we get (ii).
Finally, we prove (iii). Since $d(x,A_f)\geq\frac{1}{C}\eta^{1/26}$, there exists $s_0\in[0,d(x,y)]$ such that $\frac{1}{C}\eta^{1/26}\leq d(z,y)\leq \pi- \frac{1}{C}\eta^{1/26}$, where we put $z=\gamma_{y,x}(s_0)$. Then, there exists $v\in S^{n-p}$ with $u\cdot v=0$ and $t_1\in[0,\pi]$ such that $$\Psi(z)=\cos t_1 u+\sin t_1 v.$$ We have $$\begin{aligned}
|\cos t_1-\cos d(z,y)|=&|\cos d_S(\Psi(z),u)-\cos s_0|\\
\leq& |\cos d(z,A_f)-\cos s_0|+C\delta^{1/1600n^2}
\leq C\eta\end{aligned}$$ by Lemma \[p54c0\] and (i). This gives $$\label{55l}
|t_1-d(z,y)|\leq C\eta^{1/2}.$$ Take arbitrary $s\in [0,d(x,y)]$. Then, there exist $w\in S^{n-p}$ and $x_1,x_2,x_3\in \mathbb{R}$ such that $w\perp {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{u,v\}$, $x_1^2+x_2^2+x_3^2=1$ and $$\Psi(\gamma_{y,x}(s))=x_1 u+x_2 v+ x_3 w.$$ Since we have $$|s-d_S(\Psi(\gamma_{y,x}(s)),u)|\leq C\eta$$ by (i) and Lemma \[p54c0\], and $\cos d_S(\Psi(\gamma_{y,x}(s)),u)=x_1$, we get $$\label{55m}
|\cos s- x_1|\leq C\eta.$$ We have $$\left||d(z,y)-s|-d_S(\Psi(\gamma_{y,x}(s)),\Psi(z))\right|\leq C\eta$$ by (ii). Since $\cos d_S(\Psi(\gamma_{y,x}(s)),\Psi(z))=x_1 \cos t_1+x_2\sin t_1$, we get $$\label{55n}
|\cos(d(z,y)-s)- x_1 \cos d(z,y)-x_2\sin d(z,y)|\leq C\eta^{1/2}$$ by (\[55l\]). By (\[55m\]) and (\[55n\]), we have $$\sin d(z,y)|\sin s- x_2|\leq C\eta^{1/2}.$$ By the assumption, we have $$\sin d(z,y)\geq \frac{1}{C}\eta^{1/26},$$ and so we get $$\label{55o}
|\sin s- x_2|\leq C\eta^{6/13}.$$ By (\[55m\]) and (\[55o\]), we get $$|\cos d_S(\Psi(\gamma_{y,x}(s)),\gamma_v(s))-1|
=|x_1 \cos s+x_2\sin s-1|\leq C\eta^{6/13}.$$ Thus, we get (iii).
The following lemma asserts that the differential of an almost shortest pass from a point in $M$ to $A_f$ is in the direction of $\nabla f$ under some assumptions.
\[p54h\] Take
- $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$,
- $x\in D_f(p_f)\cap Q_f \cap R_f$,
- $y\in D_f(x)\cap D_f(p_f)\cap Q_f\cap R_f$,
- $\eta>0$ with $\eta_0\leq\eta\leq L^{1/3n}$.
Suppose
- $d(x,A_f)\geq\frac{1}{C}\eta^{1/26}$,
- $d(y,A_f)\leq C \eta$,
- $|d(x,A_f)-d(x,y)|\leq C\eta$.
Let $\{E^1,\ldots,E^n\}$ be a parallel orthonormal basis of $T^\ast M$ along $\gamma_{y,x}$ in Lemma \[p5e\] or Lemma \[p5g\] for $f$. Then, we have the following for all $s\in[0,d(x,y)]$:
- $||\dot{\gamma}^E_{y,x}|-1|\leq C \eta^{6/13}$,
- $|\nabla f (\gamma_{y,x}(s))+\sin s \dot{\gamma}_{y,x}(s)|\leq C\eta^{3/26}$.
We first note that we have $$\label{55p}
d(x,y)\leq \pi+C\eta$$ by the assumption and Proposition \[p53a\] (iv).
Let us prove (i). By $d(y,A_f)\leq C \eta$, we have $\cos d(y,A_f)\geq 1- C\eta^2$. Thus, we have $$\label{55q}
|1-f(y)|\leq C\eta^2$$ by Proposition \[p53a\] (iv). By Proposition \[p53a\] (iii), we get $$|\nabla f|(y)\leq C\eta.$$ Thus, we have $$\label{55r}
|f(x)-\cos(|\dot{\gamma}_{y,x}^E|d(x,y))|\leq C\eta$$ by Lemma \[p5e\] or Lemma \[p5g\], and so $$||\dot{\gamma}_{y,x}^E|d(x,y)-d(x,A_f)|\leq C\eta^{1/2}$$ by Proposition \[p53a\] (iv) and (\[55p\]). By the assumption, we get $$||\dot{\gamma}_{y,x}^E|-1|d(x,A_f)\leq C\eta^{1/2}.$$ This gives (i).
We next prove (ii). By Proposition \[p53a\], we have $$||\nabla f|^2(x)-\sin^2 d(x,A_f)|\leq C\delta^{1/2000n}.$$ Thus, we get $$||\nabla f|(x)-|\sin d(x,A_f)||\leq C\delta^{1/4000n}.$$ Since $\sin d(x, A_f)\geq -C\delta^{1/100n}$ by Proposition \[p53a\] (iv), we have $$||\nabla f|(x)-\sin d(x,A_f)|\leq C\delta^{1/4000n}.$$ Thus, we get $$\label{55s}
||\nabla f|(x)-\sin d(x,y)|\leq C\eta$$ by the assumption. On the other hand, by (i) and Lemma \[p5e\] or Lemma \[p5g\], we have $$|f(y)-f(x)\cos d(x,y)-\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle\sin d(x,y)|\leq C\eta^{6/13},$$ and so $$\label{55t}
|\sin^2 d(x,y)-\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle\sin d(x,y)|\leq C\eta^{6/13}$$ by (\[55q\]) and (\[55r\]).
We consider the following two cases:
- $d(x,y)\leq \pi-\eta^{3/13}$,
- $d(x,y)> \pi-\eta^{3/13}$.
We first suppose that $d(x,y)\leq \pi-\eta^{3/13}$. We get $$|\sin d(x,y)-\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle|\leq C\eta^{3/13}$$ by the assumption and (\[55t\]). By (\[55s\]), we get $$\label{55u}
|\nabla f|(x)-\langle\nabla f(x),\dot{\gamma}_{x,y}(0)\rangle \leq C\eta^{3/13}.$$
We next suppose that $d(x,y)> \pi-\eta^{3/13}$. Then, we have $\cos d(x,A_f)\leq -1+C\eta^{6/13}$, and so $|\nabla f|(x)\leq C\eta^{3/13}$ by Proposition \[p53a\] (iii) and (iv). Thus, we also get (\[55u\]) for this case.
By (i), (\[54m\]) and Lemma \[p5e\] or \[p5g\], we have $$\int_0^{d(x,y)} \left|\frac{d}{d s}\left(|\nabla f|^2(\gamma_{x,y}(s))-\langle\nabla f(\gamma_{x,y}(s)), \dot{\gamma}_{x,y}(s)\rangle^2\right)\right|\,d s\leq C\eta^{6/13}.$$ Thus, we get $$\label{55v}
|\nabla f|^2(\gamma_{x,y}(s))-\langle\nabla f(\gamma_{x,y}(s)), \dot{\gamma}_{x,y}(s)\rangle^2\leq C\eta^{3/13}$$ for all $s\in[0,d(x,y)]$ by (\[55u\]). Since $$|\nabla f (\gamma_{x,y}(s))-\langle\nabla f(\gamma_{x,y}(s)), \dot{\gamma}_{x,y}(s)\rangle\dot{\gamma}_{x,y}(s)|^2=|\nabla f|^2(\gamma_{x,y}(s))-\langle\nabla f(\gamma_{x,y}(s)), \dot{\gamma}_{x,y}(s)\rangle^2,$$ we get $$|\nabla f (\gamma_{x,y}(s))-\langle\nabla f(\gamma_{x,y}(s)), \dot{\gamma}_{x,y}(s)\rangle\dot{\gamma}_{x,y}(s)|\leq
C\eta^{3/26}$$ by (\[55v\]). Since we have $$|\langle\nabla f(\gamma_{x,y}(s)), \dot{\gamma}_{x,y}(s)\rangle+\cos d(x,y)\sin s-\sin d(x,y) \cos s|\leq C\eta^{3/13}$$ by (\[55r\]), (\[55s\]), (\[55u\]), (i) and Lemma \[p5e\] or Lemma \[p5g\], we get $$|\nabla f (\gamma_{x,y}(s))-\sin (d(x,y)-s)\dot{\gamma}_{x,y}(s)|\leq
C\eta^{3/26}$$ This gives (ii).
The following lemma is crucial to show the almost Pythagorean theorem.
\[p54i\] Take
- $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$,
- $x\in D_f(p_f)\cap Q_f \cap R_f$,
- $y\in D_f(x)\cap D_f(p_f)\cap Q_f\cap R_f$,
- $z\in M$,
- $\eta>0$ with $\eta_0\leq\eta\leq L^{1/3n}$ and $T\in [0, d(x,y)]$.
Suppose
- $d(y,A_f)\leq C\eta$,
- $|d(x,A_f)-d(x,y)|\leq C\eta$,
- $\gamma_{y,x}(s)\in I_z\setminus\{z\}$ for almost all $s\in [T,d(x,y)]$,
- $\int_T^{d(x,y)} |G_f^z(\gamma_{y,x}(s))|\,d s\leq C\eta^{3/26}$.
Then, we have $$\left| d(z,x)^2-d_S(\Psi(z),\Psi(x))^2- d(z,\gamma_{y,x}(T))^2+d_S(\Psi(z),\Psi(\gamma_{y,x}(T)))^2
\right|\leq C\eta^{1/26}.$$
If $d(x,A_f)\leq\eta^{1/26}$, then $d(x,y)\leq C\eta^{1/26}$, and so $d(x,\gamma_{y,x}(T))\leq C\eta^{1/26}$. Thus, we immediately get the lemma by Lemma \[p54c00\] if $d(x,A_f)\leq\eta^{1/26}$. In the following, we assume that $d(x,A_f)\geq\eta^{1/26}$. Take $u\in S^{n-p}$ with $f=\sum_{i=1}^{n-p}u_i f_i$, and $v\in S^{n-p}$ of Lemma \[p54g\] (iii). Define $$r(s):=d_S(\Psi (z),\gamma_v(s)).$$ Then, by the triangle inequality and Lemma \[p54g\] (iii), we have $$\label{55w}
|r(s)-d_S(\Psi (z),\Psi(\gamma_{y,x}(s)))|\leq C\eta^{3/13}.$$
There exist $w\in S^{n-p}$ and $x_1,x_2,x_3\in \mathbb{R}$ such that $w\perp {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{u,v\}$, $x_1^2+x_2^2+x_3^2=1$ and $$\Psi(z)=x_1 u+x_2 v+ x_3 w.$$ Then, $$\label{55x}
\cos r(s)=x_1\cos s+x_2\sin s$$ by the definition of $\gamma_v$ in Lemma \[p54g\] (iii), and so $$-x_1\sin s+x_2\cos s
=\frac{d}{d s} \cos r(s)
=-r'(s)\sin r(s)$$ Thus, we get $$\label{55y}
\begin{split}
-r'(s)\sin r(s) \sin s=-x_1\sin^2 s+x_2\sin s\cos s=\cos r(s)\cos s-x_1
\end{split}$$ by (\[55x\]). Since $x_1=\Psi(z)\cdot u$ and $f(z)=\widetilde{\Psi}(z)\cdot u$, we have $$\label{55z}
|x_1-\cos d(z,A_f)|\leq C\delta^{1/800n^2}$$ by Proposition \[p53a\] (iv) and Lemma \[p54a\]. By Lemma \[p54g\], (\[55w\]), (\[55y\]) and (\[55z\]), we get $$\label{56a}
\begin{split}
&\Big|\Big(\cos d(\gamma_{y,x}(s),A_f)\cos d_S(\Psi(z),\Psi(\gamma_{y,x}(s)))-\cos d(z,A_f)\Big)d_S(\Psi(z),\Psi(\gamma_{y,x}(s)))\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad +r'(s)r(s)\sin r(s) \sin s\Big|\leq C\eta^{3/13}.
\end{split}$$
Define $$l(s):=d(z,\gamma_{y,x}(s)).$$ Then, for all $s\in [0,d(x,y)]$ with $\gamma_{y,x}(s)\in I_z\setminus\{z\}$, we have $$l'(s)=\langle\dot{\gamma}_{z,\gamma_{y,x}(s)}(l(s)),\dot{\gamma}_{y,x}(s)\rangle$$ by the first variation formula, and so $$|l'(s)\sin s+\langle\dot{\gamma}_{z,\gamma_{y,x}(s)}(l(s)),\nabla f(\gamma_{y,x}(s))\rangle|\leq C\eta^{3/26}$$ by Lemma \[p54h\] (ii). Thus, for almost all $s\in [0,d(x,y)]$, we have $$\label{56b}
\begin{split}
\Big|\langle\dot{\gamma}_{\gamma_{y,x}(s),z}(0),&\nabla f(\gamma_{y,x}(s))\rangle l(s)\sin d_S(\Psi(z),\Psi(\gamma_{y,x}(s))\\
&-l'(s)l(s)\sin r(s)\sin s \Big|\leq C\eta^{3/26}
\end{split}$$ by (\[55w\]). By the definition of $G_f^z$, (\[56a\]) and (\[56b\]), for almost all $s\in [0,d(x,y)]$, we have $$\begin{aligned}
\Big|
G_f^z(\gamma_{y,x}(s))-l'(s)l(s)\sin r(s)\sin s+r'(s)r(s)\sin r(s) \sin s
\Big|\leq C\eta^{3/26}.\end{aligned}$$ Thus, by the assumption, we get $$\label{56c}
\int_T^{d(x,y)}\left|\left(\frac{d}{d s}(l(s)^2-r(s)^2)\right)\sin r(s)\sin s\right|\,d s
\leq C\eta^{3/26}.$$ Define $$\begin{aligned}
I&:=\{s\in [T,d(x,y)]: \eta^{1/26}\leq s\leq \pi -\eta^{1/26}\text{ and }\eta^{1/26}\leq r(s) \leq\pi -\eta^{1/26}
\}\\
II&:=[T,d(x,y)]\setminus I.\end{aligned}$$ Then, we have $$\label{56ca}
\int_I \left|\frac{d}{d s}(l(s)^2-r(s)^2)\right|\,d s
\leq C\eta^{1/26}$$ by (\[56c\]). Let us estimate $H^1(II)$, where $H^1$ denotes the $1$-dimensional Hausdorff measure. Suppose that $$\{s\in [T,d(x,y)]: r(s)<\eta^{1/26} \text{ or } r(s)>\pi-\eta^{1/26}\}\neq \emptyset,$$ and take arbitrary $s\in[T,d(x,y)]$ such that $r(s)<\eta^{1/26}$ or $r(s)>\pi-\eta^{1/26}$. Then, we have $$\label{56d}
||\cos r(s)|-1|\leq C\eta^{1/13}.$$ Note that we have $r(s)\leq \pi$ by ${\mathop{\mathrm{diam}}\nolimits}(S^{n-p})=\pi$. By (\[55x\]), we get $$\label{56e}
1-C\eta^{1/13}\leq (x_1^2+x_2^2)^{1/2}\leq 1.$$ Take $s_1\in[0,2\pi]$ such that $$\begin{aligned}
\cos s_1=&\frac{x_1}{(x_1^2+x_2^2)^{1/2}},\\
\sin s_1=&\frac{x_2}{(x_1^2+x_2^2)^{1/2}}.\end{aligned}$$ Then, we get $||\cos (s-s_1)|-1|\leq C\eta^{1/13}$ by (\[55x\]), (\[56d\]) and (\[56e\]). Thus, there exists $n\in \mathbb{Z}$ such that $$|s-s_1-n\pi|\leq C\eta^{1/26}.$$ Then, we have $|n|\leq 2$, and so $$H^1\left(\{s\in [T,d(x,y)]: r(s)<\eta^{1/26} \text{ or } r(s)>\pi-\eta^{1/26}\}\right)\leq C\eta^{1/26}.$$ Note that we have $d(x,y)\leq d(x,A_f)+C\eta\leq \pi+C\eta$ by the assumption and Proposition \[p53a\] (iv). Since we have $$H^1\left(\{s\in [T,d(x,y)]: s<\eta^{1/26} \text{ or } s>\pi-\eta^{1/26}\}\right)\leq C\eta^{1/26},$$ we get $H^1(II)\leq C\eta^{1/26}$. Since $\left|\frac{d}{d s}(l(s)^2-r(s)^2)\right|\leq C$ for almost all $s\in[L,d(x,y)]$, we get $$\label{56f}
\int_{II} \left|\frac{d}{d s}(l(s)^2-r(s)^2)\right|\,d s
\leq C\eta^{1/26}.$$
By (\[56ca\]) and (\[56f\]), we get $$\int_T^{d(x,y)}\left|\frac{d}{d s}(l(s)^2-r(s)^2)\right|\,d s
\leq C\eta^{1/26}.$$ Thus, we have $$|l(d(x,y))^2-r(d(x,y))^2-l(T)^2+r(T)^2|\leq C\eta^{1/26}.$$ By (\[55w\]) and the definition of $l$, we get the lemma.
In the following, the term “$\eta^{1/26}$” frequently appears. Since it appears only due to technical reasons, we put $\tau:=1/26$.
Take $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$. By Lemma \[p54f\] and the Bishop-Gromov inequality, for any triple $(x_1,x_2,x_3)\in M\times M\times M$, we can take points $\tilde{x}_1\in D_{f_{x_1}}(p_{x_1})\cap Q_{f_{x_1}} \cap R_{f_{x_1}}\cap Q_f$, $\tilde{x}_2\in D_f(p_f)\cap Q_f \cap R_f\cap P_f^{\tilde{x}_1}$ and $\tilde{x}_3\in D_f(\tilde{x}_2)\cap D_f(p_f)\cap Q_f\cap R_f\cap C_f^{\tilde{x}_1}(\tilde{x}_2)$ such that $d(x_1,\tilde{x}_1)\leq C\delta^{1/100n}$, $d(x_2,\tilde{x}_2)\leq C\eta_0$, $d(x_3,\tilde{x}_3)\leq C\eta_0$. We call the triple $(\tilde{x}_1,\tilde{x}_2,\tilde{x}_3)$ a “[*$\Pi$-triple for $(x_1,x_2,x_3,f)$*]{}”.
\[ptrp\] Take
- $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$,
- $x,y,z\in M$,
- $\eta>0$ with $\eta_0\leq\eta\leq L^{1/3n}$ and $T\in [0, d(x,y)]$.
Take a $\Pi$-triple $(\tilde{z},\tilde{x},\tilde{y})$ for $(z,x,y,f)$. Suppose
- $d(y,A_f)\leq C\eta$,
- $|d(x,A_f)-d(x,y)|\leq C\eta$,
- $d(\tilde{z},\gamma_{\tilde{y},\tilde{x}}(s))\leq \pi$ for all $s\in[T,d(\tilde{x},\tilde{y})]$.
Then, we have $$\left| d(\tilde{z},\tilde{x})^2-d_S(\Psi(\tilde{z}),\Psi(\tilde{x}))^2- d(\tilde{z},\gamma_{\tilde{y},\tilde{x}}(T))^2+d_S(\Psi(\tilde{z}),\Psi(\gamma_{\tilde{y},\tilde{x}}(T)))^2
\right|\leq C\eta^{\tau}.$$
We have $(G^{\tilde{z}}_f H^{\tilde{z}})(\gamma_{\tilde{y},\tilde{x}}(s))=G^{\tilde{z}}_f(\gamma_{\tilde{y},\tilde{x}}(s))$ for all $s\in[T,d(\tilde{x},\tilde{y})]$. Thus, we get the lemma immediately by the definition of $C_f^{\tilde{z}}(\tilde{x})$ and Lemma \[p54i\].
The following lemma guarantees that if the images of two points in $M$ under $\Phi_f$ are close to each other in $S^{n-p}\times A_f$, then their distance in $M$ are close to each other under some assumptions.
\[p54j\] Take
- $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$,
- $x,y,z,w\in M$,
- $\eta>0$ with $\eta_0\leq\eta\leq L^{1/3n}$.
Suppose
- $d(x,A_f)\leq \pi- \frac{1}{C}\eta^{\tau/3}$ and $d(z,A_f)\leq \pi- \frac{1}{C}\eta^{\tau/3}$,
- $d(y,A_f)\leq C\eta$ and $d(w,A_f)\leq C\eta$,
- $|d(x,A_f)-d(x,y)|\leq C\eta$ and $|d(z,A_f)-d(z,w)|\leq C\eta$
- $d(y,w)\leq C\eta$,
- $d_S(\Psi(x),\Psi(z))\leq C\eta$.
Then, we have $$d(x,z)\leq C\eta^{\tau/2}.$$
We first show the following claim.
\[p54k\] If $x,y,z,w\in M$ satisfies:
- $d(x,A_f)\leq \frac{1}{2}\pi- \frac{1}{C}\eta^{1/2}$ and $d(z,A_f)\leq \frac{1}{2}\pi- \frac{1}{C}\eta^{1/2}$,
- $d(y,A_f)\leq C\eta$ and $d(w,A_f)\leq C\eta$,
- $|d(x,A_f)-d(x,y)|\leq C\eta$ and $|d(z,A_f)-d(z,w)|\leq C\eta$
- $d(y,w)\leq C\eta$,
- $d_S(\Psi(x),\Psi(z))\leq C\eta^{\tau/2}$.
Then, we have $$d(x,z)\leq C\eta^{\tau/2}.$$
Take $u\in S^{n-p}$ with $f=\sum_{i=1}^{n-p+1} u_i f_i$. By the assumptions and Lemma \[p54c0\], we have $$\begin{aligned}
d_S(u,\Psi(y))\leq& C\eta,\\
|d_S(\Psi(z),u)-d(z,A_f)|\leq &C\delta^{1/1600n^2}.\end{aligned}$$ Since we have $|d(z,A_f)-d(z,y)|\leq C\eta$ by the assumptions, we get $$\label{57a0}
|d_S(\Psi(z),\Psi(y))-d(z,y)|\leq C\eta.$$ Take a $\Pi$-triple $(\tilde{z},\tilde{x},\tilde{y})$ for $(z,x,y,f)$. Then, we have $$\begin{aligned}
d(\tilde{z},\gamma_{\tilde{y},\tilde{x}}(s))\leq d(\tilde{z},\tilde{y})+d(\tilde{y},\tilde{x})
\leq &d(z,w)+d(y,w)+d(x,y)+C\eta_0\\
\leq &\pi-\frac{1}{C}\eta^{1/2}+C\eta\leq \pi\end{aligned}$$ for all $s\in[0,d(\tilde{x},\tilde{y})]$, and so $$\left| d(z,x)^2-d_S(\Psi(z),\Psi(x))^2- d(z,y)^2+d_S(\Psi(z),\Psi(y))^2
\right|\leq C\eta^{\tau}$$ by Lemma \[p54c00\] and Lemma \[ptrp\]. Thus, we get $d(x,z)\leq C\eta^{\tau/2}$ by (\[57a0\]).
Let us suppose that $x,y,z,w\in M$ satisfies the assumptions of the lemma. Take $u\in S^{n-p}$ with $f=\sum_{i=1}^{n-p+1}u_i f_i$. By the assumptions and Lemma \[p54c0\], we have $$\label{57a1}
|d(x,A_f)-d(z,A_f)|
\leq |d_S(\Psi(x),u)-d(\Psi(z),u)|+C\delta^{1/1600n^2}
\leq C\eta$$ Thus, if either $d(x,A_f)\leq \eta^{\tau}$ or $d(z,A_f)\leq \eta^{\tau}$ holds, then the lemma is trivial. In the following, we assume $d(x,A_f)\geq \eta^{\tau}$ and $d(z,A_f)\geq \eta^{\tau}$. Take a $\Pi$-triple $(\tilde{z},\tilde{x},\tilde{y})$ for $(z,x,y,f)$. By Lemma \[p54g\] (iii), we can take $v_1,v_2\in S^{n-p}$ such that $u\cdot v_i=0$ ($i=1,2$), $$\label{57a}
d_S(\Psi(\gamma_{\tilde{y},\tilde{x}}(s)),\gamma_{v_1}(s))\leq C\eta^{3/13}$$ for all $s\in [0,d(\tilde{y},\tilde{x})]$ and $$\label{57b}
d_S(\Psi(\gamma_{w,\tilde{z}}(s)),\gamma_{v_2}(s))\leq C\eta^{3/13}$$ for all $s\in [0,d(w,\tilde{z})]$, where $\gamma_{v_i}(s):=\cos s u+\sin s v_i \in S^{n-p}$ ($i=1,2$). By the assumptions and (\[57a1\]), we get $$\label{57b1}
|d(\tilde{y},\tilde{x})-d(w,\tilde{z})|\leq C\eta,$$ and so $$\begin{aligned}
\sin d(\tilde{y},\tilde{x}) |v_1- v_2|
\leq &C d_S(\gamma_{v_1}(d(\tilde{y},\tilde{x})),\gamma_{v_2}(d(\tilde{y},\tilde{x})))\\
\leq &Cd_S(\Psi(\tilde{x}),\Psi(\tilde{z}))+C\eta^{3/13}
\leq C\eta^{3/13}\end{aligned}$$ by (\[57a\]) and (\[57b\]). By $\eta^{\tau}\leq d(x,A_f)\leq \pi-\frac{1}{C}\eta^{\tau/3}$, we have $\sin d(\tilde{y},\tilde{x})\geq \frac{1}{C}\eta^{\tau}$. Thus, we get $$|v_1-v_2|\leq C\eta^{\tau}.$$ This gives $$\label{57c}
d_S(\gamma_{v_1}(s),\gamma_{v_2}(s))\leq C\eta^{\tau}.$$ for all $s\in \mathbb{R}$.
Put $$a=\gamma_{\tilde{y},\tilde{x}}\left(\frac{1}{2}d(\tilde{y},\tilde{x})\right),
\quad b=\gamma_{w,\tilde{z}}\left(\frac{1}{2}d(w,\tilde{z})\right).$$ By (\[57a\]), (\[57b\]), (\[57b1\]) and (\[57c\]), we have $$d_S(\Psi(a),\Psi(b))\leq C\eta^{\tau}.$$ Moreover, other assumptions of Claim \[p54k\] hold for the pair $(a,y,b,w)$ by Lemma \[p54g\] (i). Thus, we get $$d(a,b)\leq C\eta^{\tau/2}$$ by Claim \[p54k\]. Therefore, we have $$d(\tilde{z}, \gamma_{\tilde{y},\tilde{x}}(s))\leq d(\tilde{z},b)+d(a,b)+d(\gamma_{\tilde{y},\tilde{x}}(s),a)\leq \frac{1}{2}d(\tilde{x},\tilde{y})+\frac{1}{2}d(\tilde{z},w)+C\eta^{\tau/2}\leq \pi$$ for all $s\in[0,d(\tilde{y},\tilde{x})]$, and so $d(\tilde{x},\tilde{z})\leq C\eta^{\tau/2}$ similarly to Claim \[p54k\]. Thus, we get the lemma.
Let us show the almost Pythagorean theorem for the special case. Recall that we defined $\eta_1:=\eta_0^\tau$.
\[p54l\] Take
- $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$,
- $x,y,z,w\in M$,
- $\eta>0$ with $\eta_1\leq \eta\leq L^{1/3n}$.
Suppose
- $d(x,z)\leq C\eta$,
- $d(x,A_f)\leq \pi- \frac{1}{C}\eta^{1/2}$ and $d(z,A_f)\leq \pi- \frac{1}{C}\eta^{1/2}$,
- $d(y,A_f)\leq C\eta_0$ and $d(w,A_f)\leq C\eta_0$,
- $|d(x,A_f)-d(x,y)|\leq C\eta_0$ and $|d(z,A_f)-d(z,w)|\leq C\eta_0$.
Then, we have $$|d(x,z)^2-d_S(\Psi(x),\Psi(z))^2-d(y,w)^2|\leq C\eta_1.$$
By Lemma \[p54c0\], we have $$\label{57d0}
d_S(\Psi(y),\Psi(w))
\leq d(y,A_f)+d(w,A_f)+C\delta^{1/1600n^2}
\leq C\eta_0.$$
Put $a_0:=x$ and $b_0:=z$. In the following, we define $a_{i},b_{i}\in M$ ($i=1,2,3$) so that
- $d(a_{i},b_{i})\leq C\eta^{1/2}$,
- $|d(a_{i},A_f)-d(a_{i},y)|\leq C\eta_0$ and $|d(b_{i},A_f)-d(b_{i},w)|\leq C\eta_0$,
- $d(a_{i},A_f)\leq \frac{3-i}{3}\pi+C\eta_0$ and $d(b_{i},A_f)\leq \frac{3-i}{3}\pi+C\eta_0$,
- $|d(a_{i+1},b_{i+1})^2-d_S(\Psi(a_{i+1}),\Psi(b_{i+1}))^2-d(a_{i},b_{i})^2+d_S(\Psi(a_{i}),\Psi(b_{i}))^2|\leq C\eta_0^{\tau}$ ($i=0,1,2$),
- $d(y,a_3)\leq C\eta_0$ and $d(w,b_3)\leq C\eta_0$.
If we succeed in defining such $a_i$ and $b_i$, we have $$|d(x,z)^2-d_S(\Psi(x),\Psi(z))^2-d(y,w)^2+d_S(\Psi(y),\Psi(w))^2|\leq C\eta_0^{\tau}=C\eta_1$$ by (iv) and (v), and so we get the lemma by (\[57d0\]).
Take arbitrary $i\in\{0,1,2\}$ and suppose that we have chosen $a_i,b_i\in M$ such that (i), (ii) and (iii) hold if $i\geq 1$. Let us define $a_{i+1},b_{i+1}\in M$ that satisfy our properties. Take a $\Pi$-triple $(\tilde{b}_i,\tilde{a}_i, \tilde{y}_i)$ for $(b_i,a_i,y,f)$. Define $$a_{i+1}:=\gamma_{\tilde{y}_i,\tilde{a}_i}\left(\frac{2-i}{3-i}d(\tilde{y}_i,\tilde{a}_i)\right).$$ Since $$d(\tilde{b}_i, \gamma_{\tilde{y}_i,\tilde{a}_i}(s))\leq d(\tilde{a}_i,\tilde{b}_i)
+d(\tilde{a}_i,\gamma_{\tilde{y}_i,\tilde{a}_i}(s))\leq \frac{\pi}{3}+C\eta^{1/2}$$ for all $s\in\left[\frac{2-i}{3-i}d(\tilde{y}_i,\tilde{a}_i),d(\tilde{y}_i,\tilde{a}_i)\right]$ by the assumptions, we get $$\label{57d}
|d(a_{i+1},b_{i})^2-d_S(\Psi(a_{i+1}),\Psi(b_{i}))^2-d(a_{i},b_{i})^2+d_S(\Psi(a_{i}),\Psi(b_{i}))^2|\leq C\eta_0^{\tau}$$ by Lemma \[p54c00\] and Lemma \[ptrp\]. Take a $\Pi$-triple $(\overline{a}_{i+1},\overline{b}_i,\overline{w}_i)$ for $(a_{i+1},b_i,w,f)$. Define $$b_{i+1}:=\gamma_{\overline{w}_i,\overline{b}_i}\left(\frac{2-i}{3-i}d(\overline{w}_i,\overline{b}_i)\right).$$ Since $$d(\overline{a}_{i+1},\gamma_{\overline{w}_i,\overline{b}_i}(s))\leq d(\overline{a}_{i+1},a_i)+d(a_i,\overline{b}_i)+d(\overline{b_i},\gamma_{\overline{w}_i,\overline{b}_i}(s))\leq \frac{2}{3}\pi +C\eta^{1/2}$$ for all $s\in\left[\frac{2-i}{3-i}d(\overline{w}_i,\overline{b}_i),d(\overline{w}_i,\overline{b}_i)\right]$ by the assumptions, we get $$\label{57e}
|d(a_{i+1},b_{i+1})^2-d_S(\Psi(a_{i+1}),\Psi(b_{i+1}))^2-d(a_{i+1},b_{i})^2+d_S(\Psi(a_{i+1}),\Psi(b_{i}))^2|\leq C\eta_0^{\tau}$$ by Lemma \[p54c00\] and Lemma \[ptrp\]. By (\[57d\]) and (\[57e\]), we get (iv).
By the assumptions and Lemma \[p54g\], we get (ii) for $a_{i+1}$ and $b_{i+1}$.
By the assumptions, we have $$\begin{aligned}
d(a_{i+1},A_f)
\leq &d(a_{i+1},\tilde{y}_i)+d(y,A_f)+C\eta_0\\
=&\frac{2-i}{3-i}d(\tilde{a}_{i},\tilde{y}_i)+C\eta_0
\leq \frac{2-i}{3}\pi+C\eta_0.\end{aligned}$$ Similarly, we have $d(b_{i+1},A_f)\leq \frac{2-i}{3}\pi+C\eta_0$. Thus, we get (iii) for $a_{i+1}$ and $b_{i+1}$.
By definition, we have $$a_3=\tilde{y}_3,\quad b_3=\overline{w}_3.$$ Thus, we get (v).
In the following, we prove (i) for $a_{i+1}$ and $b_{i+1}$. If $d(a_i,y)\leq \eta_0^{\tau}$, then we have $$\begin{aligned}
d(b_i,w)\leq d(b_i,A_f)+C\eta_0
\leq d(a_i,A_f)+C\eta^{1/2}
\leq C\eta^{1/2},\end{aligned}$$ and so $$\begin{aligned}
d(y,w)\leq&C\eta^{1/2},\\
d(a_{i+1},y)\leq& C\eta^{1/2},\\
d(b_{i+1},w)\leq & C\eta^{1/2}.\end{aligned}$$ Then, we have $d(a_{i+1},b_{i+1})\leq C\eta^{1/2}$. Similarly, if $d(b_i,w)\leq \eta_0^{\tau}$, then $d(a_{i+1},b_{i+1})\leq C\eta^{1/2}$. Thus, in the following, we assume that $d(a_i,y)\geq \eta_0^{\tau}$ and $d(b_i,w)\geq \eta_0^{\tau}$. By Lemma \[p54g\], we can take $u,v_1,v_2\in S^{n-p}$ such that $f=\sum_{j=1}^{n-p+1}u_j f_j$, $ u\cdot v_k=0$ ($k=1,2$), $$\label{57f}
d_S(\Psi(\gamma_{\tilde{y}_i,\tilde{a}_i}(s)),\gamma_{v_1}(s))\leq C\eta_0^{3/13}$$ for all $s\in [0,d(\tilde{a}_i,\tilde{y}_i)]$ and $$\label{57g}
d_S(\Psi(\gamma_{\overline{w}_i,\overline{b}_i}(s)),\gamma_{v_2}(s))\leq C\eta_0^{3/13}$$ for all $s\in [0,d(\overline{b}_i,\overline{w}_i)]$, where $\gamma_{v_k}(s):=\cos s u+\sin s v_k\in S^{n-p}$ ($k=1,2$). Since $$|d(\tilde{a}_i,\tilde{y}_i)-d(\overline{b}_i,\overline{w}_i)|\leq
|d(a_i,A_f)-d(b_i,A_f)|+C\eta_0\leq d(a_i,b_i)+C\eta_0,$$ we have $$\label{57h}
\left|d_S(\Psi(\tilde{a}_i),\Psi(\overline{b}_i))
-d_S\left(\gamma_{v_1}(l_i),\gamma_{v_2} (l_i)\right)
\right|\leq d(a_i,b_i)+C\eta_0^{3/13}$$ and $$\label{57i}
\left|d_S(\Psi(a_{i+1}),\Psi(b_{i+1}))
-d_S\left(\gamma_{v_1}\left(\frac{2-i}{3-i}l_i\right),\gamma_{v_2} \left(\frac{2-i}{3-i}l_i\right)\right)
\right|\leq d(a_i,b_i)+C\eta_0^{3/13}$$ by (\[57f\]) and (\[57g\]), where we put $l_i:=d(\tilde{a}_i,\tilde{y}_i)$. By (\[57h\]) and Lemma \[p54c00\], we get $$\label{57i1}
|v_1-v_2|\sin l_i
\leq C d_S\left(\gamma_{v_1}(l_i),\gamma_{v_2} (l_i)\right)
\leq Cd(a_i,b_i)+C\eta_0^{3/13}.$$ Note that we assumed $$\label{57j}
d(a_i,b_i)\leq C\eta^{1/2}$$ and $$\label{57k}
d(a_0,b_0)\leq C\eta.$$
We first suppose that $d(a_i,y)\leq \pi/6$. By (\[57h\]), (\[57i\]) and (\[57j\]), we get $$\label{57l}
\left|d_S(\Psi(\tilde{a}_i),\Psi(\overline{b}_i))
-d_S\left(\gamma_{v_1}(l_i),\gamma_{v_2} (l_i)\right)
\right|\leq C\eta^{1/2}$$ and $$\label{57m}
\left|d_S(\Psi(a_{i+1}),\Psi(b_{i+1}))
-d_S\left(\gamma_{v_1}\left(\frac{2-i}{3-i}l_i\right),\gamma_{v_2} \left(\frac{2-i}{3-i}l_i\right)\right)
\right|\leq C\eta^{1/2},$$ Since $l_i\leq \pi/2$, we have $$\sin \left(\frac{2-i}{3-i}l_i\right)
\leq \sin l_i,$$ and so $$\begin{aligned}
d_S(\Psi(a_{i+1}),\Psi(b_{i+1}))
\leq& d_S\left(\gamma_{v_1}\left(\frac{2-i}{3-i}l_i\right),\gamma_{v_2} \left(\frac{2-i}{3-i}l_i\right)\right)+C\eta^{1/2}\\
\leq& C|v_1-v_2|\sin \left(\frac{2-i}{3-i}l_i\right)+C\eta^{1/2}\\
\leq& C|v_1-v_2|\sin l_i+C\eta^{1/2}\\
\leq& C d_S(\Psi(\tilde{a}_i),\Psi(\overline{b}_i))+C\eta^{1/2}
\leq C\eta^{1/2}\end{aligned}$$ by (\[57l\]) and (\[57m\]). Thus, we get $d(a_{i+1},b_{i+1})\leq C\eta^{1/2}$ by (iv).
We next suppose that $\pi/6\leq d(a_i,y)\leq 5\pi/6$. By (\[57i1\]) and (\[57j\]), we have $|v_1-v_2|\leq C\eta^{1/2}$. Thus, we get $$\begin{aligned}
d_S(\Psi(a_{i+1}),\Psi(b_{i+1}))
\leq C\eta^{1/2}\end{aligned}$$ by (\[57i\]). Thus, we get $d(a_{i+1},b_{i+1})\leq C\eta^{1/2}$ by (iv).
If $i\geq 1$, we have $d(a_i,y)\leq 5\pi/6$, and so we get $d(a_{i+1},b_{i+1})\leq C\eta^{1/2}$ by the above two cases.
Finally, we suppose that $i=0$ and $d(x,y)\geq 5\pi/6$. By (\[57i1\]) and (\[57k\]), we have $|v_1-v_2|\sin l_0\leq C\eta$. By the definition of $l_0$, we have $|l_0-d(x,y)|\leq C\eta_0.$ Thus, we have $\sin l_0\geq \frac{1}{C}(\pi- l_0)\geq \frac{1}{C}\eta^{1/2}$, and so we get $|v_1-v_2|\leq C\eta^{1/2}$, and so we have $$\begin{aligned}
d_S(\Psi(a_{i+1}),\Psi(b_{i+1}))
\leq C\eta^{1/2}\end{aligned}$$ by (\[57i\]). Thus, $d(a_{i+1},b_{i+1})\leq C\eta^{1/2}$ by (iv).
Therefore, we have (i) for all cases, and we get the lemma.
Let us show that the map $\Phi_f\colon M\to S^{n-p}\times A_f,\,x\mapsto (\Psi(x), a_f(x))$ is almost surjective.
\[p54m\] Take $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$. For any $(v,a)\in S^{n-p}\times A_f$, there exists $x\in M$ such that $d(\Phi_f(x),(v,a))\leq C\eta_1^{1/2}$ holds.
Take arbitrary $(v,a)\in S^{n-p}\times A_f$. Take $u\in S^{n-p}$ with $f=\sum_{i=1}^{n-p+1} u_i f_i$. Since there exists $\tilde{v}\in S^{n-p}$ such that $d_S(u,\tilde{v})\leq \pi-\eta_1^{1/2}$ and $d_S(v,\tilde{v})\leq \eta_1^{1/2}$, it is enough to prove the proposition assuming $d_S(u,v)\leq \pi-\eta_1^{1/2}$.
Put $F_v:=\sum_{i=1}^{n-p+1}v_i f_i$. Then, $|F_v(p_{F_v})-1|\leq C\delta^{1/800n}$ and $A_{F_v}=\{x\in M:|F_v(x)-1|\leq \delta^{1/900n}\}$ by Proposition \[p53a\]. In the following, we show that $a_v:=a_{F_v}(a)\in A_{F_v}$ has the desired property. By Lemma \[p54c0\], we get $$\begin{aligned}
\notag d_S(\Psi(a),u)\leq &C\delta^{1/1600n^2},\\
\label{57n}
d_S(\Psi(a_v),v)\leq &C\delta^{1/1600n^2}.\end{aligned}$$ Thus, by Lemma \[p54c0\], we get $$\begin{aligned}
|d(a,a_v)-d(a_f(a_v),a_v)|=&|d(a,A_{F_v})-d(a_v,A_f)|\\
\leq& |d_S(\Psi(a),v)-d_S(\Psi(a_v),u)|+C\delta^{1/1600n^2}\\
\leq& C\delta^{1/1600n^2}\leq \eta_0\end{aligned}$$ and $$d(a_v,A_f)\leq d_S(\Psi(a_v),u)+C\delta^{1/1600n^2}
\leq d_S(u,v)+C\delta^{1/1600n^2}\leq \pi-\frac{1}{2}\eta_1^{1/2}.$$ Since we have $d(a_v,A_f)=d(a_v,a_f(a_v))$, we get $$\begin{aligned}
|d(a_v,A_f)-d(a_v,a)|\leq |d(a_v,A_f)-d(a_v,a_f(a_v))|+\eta_0=\eta_0,\end{aligned}$$ and so we get $$\label{57o}
d(a,a_f(a_v))\leq C\eta_0^{\tau/2}=C\eta_1^{1/2}$$ by Lemma \[p54l\] putting $x=z=a_v$, $y=a$ and $w=a_f(a_v)$.
By (\[57n\]) and (\[57o\]), putting $x=a_v$, we get the proposition.
Now, we are in position to show $|\dot{\gamma}_{\tilde{y}_1,y_2}^E|d(\tilde{y}_1,y_2)\leq \pi+L$ under the assumption of Lemma \[p54d\]. Note that we defined $\tau=1/26$, $\eta_2=\eta_1^{\tau/3}$ and $L=\eta_2^{1/150}$.
\[p54n\] Take $y_1\in M$, $\tilde{y}_1\in D_{f_{y_1}}(p_{y_1})\cap R_{f_{y_1}}\cap Q_{f_{y_1}}$ with $d(y_1,\tilde{y}_1)\leq C\delta^{1/100n}$ and $y_2\in D_{f_{y_1}}(\tilde{y}_1)$. Let $\{E_1,\ldots,E_n\}$ be a parallel orthonormal basis of $TM$ along $\gamma_{\tilde{y}_1,y_2}$ in Lemma \[p5e\] or Lemma \[p5g\] for $f_{y_1}$. Then, $$|\dot{\gamma}_{\tilde{y}_1,y_2}^E|d(\tilde{y}_1,y_2)\leq \pi+ L$$ and $$||\dot{\gamma}_{\tilde{y}_1,y_2}^E|d(\tilde{y}_1,y_2)-d_S(\Psi(y_1),\Psi(y_2))|\leq CL.$$
We immediately get the second assertion by the first assertion and Lemma \[p54d\].
Let us show the first assertion. Suppose that $$|\dot{\gamma}_{\tilde{y}_1,y_2}^E|d(\tilde{y}_1,y_2)>\pi+ L.$$ Put $$\begin{aligned}
f:=&-f_{y_1},\\
\gamma:=&\gamma_{\tilde{y}_1,y_2},\\
s_0:=&\frac{1}{|\dot{\gamma}^E|}\eta_2^{\tau/4},\\
s_1:=&\frac{1}{|\dot{\gamma}^E|}(\pi+L).\end{aligned}$$ Take $k\in \mathbb{N}$ to be $$\frac{1}{\eta_2}(s_1-s_0)<k\leq \frac{1}{\eta_2}(s_1-s_0)+1,$$ and put $$t_j:=\frac{j}{k}(s_1-s_0)+s_0$$ for each $j\in\{0,\ldots,k\}$. Note that we have $t_0=s_0$, $t_k=s_1$ and $$\begin{aligned}
\label{57p0}
k\leq& C\eta_2^{-1},\\
\label{57p1}\frac{1}{k}\leq& C\eta_2.\end{aligned}$$
For all $s\in[s_0,s_1]$, we have $$\begin{aligned}
\cos d_S(\Psi(y_1),\Psi(\gamma(s)))
\leq \cos (|\dot{\gamma}^E| s)+C\delta^{1/1600n^2}
\leq 1-\frac{1}{C}\eta_2^{\tau/2}\end{aligned}$$ for all $s\in[s_0,s_1]$ by Lemma \[p54d\]. Since $$f(\gamma(s))=-|\widetilde{\Psi}|(\gamma(s))\cos d_S(\Psi(y_1),\Psi(\gamma(s)))$$ by the definitions of $f_{y_1}$ and $f$, we get $$f(\gamma(s))\geq -1+\frac{1}{C}\eta_2^{\tau/2}$$ for all $s\in[s_0,s_1]$ by Lemma \[p54a\]. This gives $$\label{57p}
d(\gamma(s),A_f)\leq \pi-\frac{1}{C}\eta_2^{\tau/4}$$ $s\in[s_0,s_1]$ by Proposition \[p53a\]. By the definition of $t_j$ and (\[57p\]), we have $$\begin{aligned}
\label{57p11} d(\gamma(t_j),\gamma(t_{j+1}))\leq& \eta_2,\\
\notag d(\gamma(t_j),A_f)\leq & \pi-\frac{1}{C}\eta_2^{\tau/4}\leq \pi-\eta_2^{1/2},\\
\notag d(\gamma(t_{j+1}),A_f)\leq &\pi-\frac{1}{C}\eta_2^{\tau/4}\leq\pi-\eta_2^{1/2}\end{aligned}$$ for all $j\in\{0,\ldots,k-1\}$, and so we get $$\label{57q}
|d(\gamma(t_j),\gamma(t_{j+1}))^2-d_S(\Psi(\gamma(t_j)),\Psi(\gamma(t_{j+1})))^2-d(a_f(\gamma(t_j)),a_f(\gamma(t_{j+1})))^2|\leq C\eta_1$$ by Lemma \[p54l\]. In particular, we get $$\label{57r}
d(a_f(\gamma(t_j)),a_f(\gamma(t_{j+1})))\leq C\eta_2$$ by (\[57p11\]).
Take $j_0\in\{1,\ldots, k-1\}$ to be $$|\dot{\gamma}^E|t_{j_0}< \pi \leq |\dot{\gamma}^E|t_{j_0+1}.$$ Since $$||\dot{\gamma}^E|s-d_S(\Psi(y_1),\Psi(\gamma(s)))|\leq C\delta^{1/3200n^2}$$ for all $s\in\left[0,\frac{1}{|\dot{\gamma}^E|}\pi\right]$ by Lemma \[p54d\], we get $$\label{57s}
\begin{split}
d_S(\Psi(\gamma(t_j)),\Psi(\gamma(t_{j+1})))
\geq &d_S(\Psi(y_1),\Psi(\gamma(t_{j+1})))- d_S(\Psi(y_1),\Psi(\gamma(t_{j})))\\
\geq &|\dot{\gamma}^E|(t_{j+1}-t_j)-C\delta^{1/3200n^2}
\end{split}$$ for all $j\in \{0,\ldots,j_0-1\}$. Since $$|2\pi-|\dot{\gamma}^E|s-d_S(\Psi(y_1),\Psi(\gamma(s)))|\leq C\delta^{1/3200n^2}$$ for all $s\in\left[\frac{1}{|\dot{\gamma}^E|}\pi,s_1\right]$ by Lemma \[p54d\], we get $$\label{57t}
d_S(\Psi(\gamma(t_j)),\Psi(\gamma(t_{j+1})))
\geq |\dot{\gamma}^E|(t_{j+1}-t_j)-C\delta^{1/3200n^2}$$ for all $j\in \{j_0+1,\ldots,k-1\}$. By (\[57q\]), (\[57s\]) and (\[57t\]), we get $$\label{57u}
d(a_f(\gamma(t_j)),a_f(\gamma(t_{j+1})))^2
\leq d(\gamma(t_j),\gamma(t_{j+1}))^2-|\dot{\gamma}^E|^2(t_{j+1}-t_j)^2+C\eta_1$$ for all $j\in\{0,\ldots,k-1\}\setminus \{j_0\}$.
Since we have $$\begin{aligned}
d_S(\Psi(\gamma(s_l)),\Psi(p_f))
\leq d(\gamma(s_l),A_f)+C\delta^{1/1600n^2}
\leq \pi-\frac{1}{C}\eta_2^{\tau/4}\end{aligned}$$ for each $l=0,1$ by Lemma \[p54c0\], Corollary \[p54c01\] and (\[57p\]), we can take a curve $\beta\colon[0,K]\to S^{n-p}$ in $S^{n-p}$ with unit speed ($K$ is some constant) such that $$\begin{aligned}
\beta(0)=&\Psi(\gamma(s_0)),\\
\beta(K)=&\Psi(\gamma(s_1)),\\
|d_S(\Psi(\gamma(s_0)),\Psi(\gamma(s_1)))-K|\leq &C\eta_2^{\tau/4},\\
d_S(\beta(s),\Psi(p_f))\leq &\pi-\frac{1}{C}\eta_2^{\tau/4}\end{aligned}$$ for all $s\in[0,K]$. Note that we can find such $\beta$ by taking an almost shortest pass in $$\{u\in S^{n-p}: d(u,\Psi(p_f))\leq \pi-\frac{1}{C}\eta_2^{\tau/4}\}.$$ By Proposition \[p54m\], there exists $x_j\in M$ such that $$\label{57v}
d\left(\Phi_f(x_j),\left(\beta\left(\frac{j}{k}K\right),a_f(\gamma(t_j))\right)\right)\leq C\eta_1^{1/2}$$ for each $j\in\{0,\ldots,k\}$. By (\[57p1\]), (\[57r\]), (\[57v\]), Lemma \[p54c0\] and Corollary \[p54c01\], we have $$\begin{aligned}
\notag d(a_f(x_j),a_f(x_{j+1}))\leq &C\eta_2,\\
\label{57v1}d_S(\Psi(x_j),\Psi(x_{j+1}))\leq &\frac{1}{k}K+C\eta_1^{1/2}\leq C\eta_2,\\
\label{57v2}d(x_j,A_f)\leq &d_S(\Psi(x_j),\Psi(p_f))+C\delta^{1/1600n^2}\\
\notag \leq &d_S\left(\beta\left(\frac{j}{k}K\right),\Psi(p_f)\right)+C\eta_1^{1/2}
\leq\pi-\frac{1}{C}\eta_2^{\tau/4}\end{aligned}$$ for all $j$, and so $$\label{57v3}
d(x_j,x_{j+1})\leq C\eta_2^{\tau/2}$$ by Lemma \[p54j\] putting $x=x_j, y=a_f(x_j), z=x_{j+1}, w=a_f(x_{j+1})$ and $\eta=\eta_2$. By (\[57v2\]), (\[57v3\]) and Lemma \[p54l\] putting $x=x_j, y=a_f(x_j), z=x_{j+1}, w=a_f(x_{j+1})$ and $\eta=\eta_2^{\tau/2}$, we get $$\label{57w}
|d(x_j,x_{j+1})^2-d_S(\Psi(x_j),\Psi(x_{j+1}))^2-d(a_f(x_j),a_f(x_{j+1}))^2|\leq C\eta_1$$ for all $j\in\{0,\ldots,k-1\}$. By (\[57u\]), (\[57v1\]) and (\[57w\]), we have $$\label{57x}
\begin{split}
d(x_j,x_{j+1})^2
\leq &\frac{1}{k^2}K^2+d(a_f(x_j),a_f(x_{j+1}))^2+C\eta_1^{1/2}\\
\leq &\frac{1}{k^2}K^2+d(\gamma(t_j),\gamma(t_{j+1}))^2-|\dot{\gamma}^E|^2(t_{j+1}-t_j)^2+C\eta_1^{1/2}
\end{split}$$ for all $j\in\{0,\ldots,k-1\}\setminus\{j_0\}$. Since $K\leq \pi+C\eta_2^{\tau/4}$, we have $$\label{57y}
\frac{1}{k^2}K^2\leq \frac{\pi^2}{k^2}+\frac{C}{k^2}\eta_2^{\tau/4}.$$ Since $$|\dot{\gamma}^E|(t_{j+1}-t_j)=\frac{|\dot{\gamma}^E|}{k}(s_1-s_0)
=\frac{1}{k}(\pi+L-\eta_2^{\tau/4})
\geq\frac{1}{k}\left(\pi+\frac{1}{2}L\right),$$ we have $$\label{57z}
|\dot{\gamma}^E|^2(t_{j+1}-t_j)^2\geq \frac{\pi^2}{k^2}+\frac{1}{k^2}L$$ for all $j\in\{0,\ldots,k-1\}$. By (\[57y\]) and (\[57z\]), we get $$|\dot{\gamma}^E|^2(t_{j+1}-t_j)^2-\frac{1}{k^2}K^2\geq \frac{1}{k^2}L-\frac{C}{k^2}\eta_2^{\tau/4} \geq
\frac{1}{2k^2}L$$ for all $j\in\{0,\ldots,k-1\}$. Thus, by (\[57x\]), we have $$d(x_j,x_{j+1})^2
\leq d(\gamma(t_j),\gamma(t_{j+1}))^2-\frac{1}{2k^2}L+C\eta_1^{1/2}
\leq d(\gamma(t_j),\gamma(t_{j+1}))^2-\frac{1}{4k^2}L$$ for all $j\in\{0,\ldots,k-1\}\setminus\{j_0\}$. Since $d(\gamma(t_j),\gamma(t_{j+1}))+d(x_j,x_{j+1})\leq 1$, we get $$\label{58a}
\frac{1}{4k^2}L\leq d(\gamma(t_j),\gamma(t_{j+1}))^2-d(x_j,x_{j+1})^2
\leq d(\gamma(t_j),\gamma(t_{j+1}))-d(x_j,x_{j+1})$$ $j\in\{0,\ldots,k-1\}\setminus\{j_0\}$. By (\[57p0\]), (\[57v3\]) and (\[58a\]), we get $$\label{58b}
\begin{split}
d(x_0,x_k)\leq &\sum_{i=0}^{k-1}d(x_j,x_{j+1})
\leq \sum_{i=0}^{k-1}d(\gamma(t_j),\gamma(t_{j+1}))-\frac{k-1}{4k^2}L+d(x_{j_0},x_{j_0+1})\\
\leq& d(\gamma(s_0),\gamma(s_1))-\frac{1}{8k}L\leq d(\gamma(s_0),\gamma(s_1))-\frac{1}{C}\eta_2 L.
\end{split}$$ By (\[57v\]), we have $$\begin{aligned}
d_S(\Psi(x_0),\Psi(\gamma(s_0)))\leq C\eta_1,\quad
d(a_f(x_0),a_f(\gamma(s_0)))\leq C\eta_1.\end{aligned}$$ Thus, by (\[57p\]), (\[57v2\]) and Lemma \[p54j\], we get $$\label{58c}
d(x_0,\gamma(s_0))\leq C\eta_1^{\tau/2}.$$ Similarly, we get $$\label{58d}
d(x_k,\gamma(s_1))\leq C\eta_1^{\tau/2}.$$ By (\[58b\]), (\[58c\]) and (\[58d\]), we get $$\eta_2 L\leq C\eta_1^{\tau/2}.$$ This contradicts to the definitions of $\eta_2$ and $L$. Thus, we get the lemma.
For all $y_1,y_2\in M$, define $$\begin{aligned}
\overline{C}_f^{y_1}(y_2)=&\Big\{y_3\in M : \gamma_{y_2,y_3}(s)\in I_{y_1}\setminus\{y_1\} \text{ for almost all $s\in[0,d(y_2,y_3)]$, and}\\
&\qquad \qquad\qquad \qquad \int_{0}^{d(y_2,y_3)} |G_f^{y_1}|(\gamma_{y_2,y_3}(s))\,d s\leq L^{1/3}\Big\},\\
\overline{P}_f^{y_1}=&\{y_2\in M: {\mathop{\mathrm{Vol}}\nolimits}(M\setminus C_f^{y_1}(y_2))\leq L^{1/3}{\mathop{\mathrm{Vol}}\nolimits}(M)\}.\end{aligned}$$
Let us complete the Gromov-Hausdorff approximation.
\[MT2\] Take $f\in{\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p+1}\}$ with $\|f\|_2^2=1/(n-p+1)$. Then, the map $\Phi_f\colon M\to S^{n-p}\times A_f$ is a $CL^{1/156n}$-Hausdorff approximation map. In particular, we have $d_{GH}(M, S^{n-p}\times A_f)\leq CL^{1/156n}$.
Take arbitrary $y_1\in M$ and $\tilde{y}_1\in D_{f_{y_1}}(p_{y_1})\cap R_{f_{y_1}}\cap Q_{f_{y_1}}\cap Q_f$ with $d(y_1,\tilde{y}_1)\leq C\delta^{1/100n}$. By Lemma \[p54c00\], Corollary \[p54f0\] and Lemma \[p54n\], we have $$|G_f^{\tilde{y}_1}|(y_2)\leq CL$$ for all $y\in D_f(\tilde{y_1})\cap D_{f_{y_1}}(\tilde{y}_1)$. Since ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus (D_f(\tilde{y_1})\cap D_{f_{y_1}}(\tilde{y}_1)))\leq C\delta^{1/100}{\mathop{\mathrm{Vol}}\nolimits}(M)$ and $\|G_f^{\tilde{y}_1}\|_\infty\leq C$, we get $$\|G_f^{\tilde{y}_1}\|_1\leq CL.$$ Thus, by the segment inequality, we get $${\mathop{\mathrm{Vol}}\nolimits}(M\setminus \overline{P}^{\tilde{y}_1}_f)\leq CL^{1/3}.$$
Take arbitrary $x,z\in M$. By the Bishop-Gromov inequality, there exist $\tilde{z}\in D_{f_{z}}(p_{z})\cap Q_{f_{z}} \cap R_{f_{z}}\cap Q_f$, $\tilde{x}\in D_f(p_f)\cap Q_f \cap R_f\cap\overline{P}_f^{\tilde{z}}$ and $\tilde{y}\in D_f(\tilde{x})\cap D_f(p_f)\cap Q_f\cap R_f\cap \overline{C}_f^{\tilde{x}}(\tilde{z})$ such that $d(z,\tilde{z})\leq C \delta^{1/100n}$, $d(x,\tilde{x})\leq CL^{1/3n}$ and $d(a_f(x),\tilde{y})\leq CL^{1/3n}$. Here, we used the estimate ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus \overline{P}^{\tilde{z}}_f)\leq CL^{1/3}$. Then, we get $$\left| d(\tilde{z},\tilde{x})^2-d_S(\Psi(\tilde{z}),\Psi(\tilde{x}))^2- d(\tilde{z},\tilde{y})^2+d_S(\Psi(\tilde{z}),\Psi(\tilde{y}))^2
\right|\leq CL^{1/78n}$$ by Lemma \[p54i\]. Thus, we get $$\label{59a}
\left| d(z,x)^2-d_S(\Psi(z),\Psi(x))^2- d(z,a_f(x))^2+d_S(\Psi(z),\Psi(a_f(x)))^2
\right|\leq CL^{1/78n}$$ by Lemma \[p54c00\]. Similarly, we have $$\label{59b}
\begin{split}
&\left| d(a_f(x),z)^2-d_S(\Psi(a_f(x)),\Psi(z))^2- d(a_f(x),a_f(z))^2+ d_S(\Psi(a_f(x)),\Psi(a_f(z)))^2\right|\\
\leq &CL^{1/78n}.
\end{split}$$ Since we have $d_S(\Psi(a_f(x)),\Psi(a_f(z)))\leq C\delta^{1/1600n^2}$ by Lemma \[p54c0\], we get $$\left| d(z,x)^2-d_S(\Psi(z),\Psi(x))^2- d(a_f(x),a_f(z))^2\right|\leq CL^{1/78n}.$$ by (\[59a\]) and (\[59b\]). This gives $$\begin{split}
&\left| d(z,x)-d(\Phi_f(z),\Phi_f(x))\right|\\
=&\left| d(z,x)-\left(d_S(\Psi(z),\Psi(x))^2+ d(a_f(x),a_f(z))^2\right)^{1/2}\right|\leq CL^{1/156n}.
\end{split}$$ Combining this and Proposition \[p54l\], we get the theorem.
By the above theorem, we get Main Theorem 2 except for the orientability, which is proved in subsection 5.7.
Further Inequalities
--------------------
In this subsection, we show two lemmas to prove the remaining part of main theorems.
In this subsection, we assume the following in addition to Assumption \[asu1\].
- $1\leq k\leq n-p+1$.
- $f_i\in C^\infty(M)$ ($i\in\{1,\ldots,k\}$) is an eigenfunction of the Laplacian acting on functions with $\|f_i\|_2^2=1/(n-p+1)$ corresponding to the eigenvalue $\lambda_i$ with $0<\lambda_i\leq n-p+\delta$ such that $$\int_M f_i f_j\,d\mu_g=0$$ holds for any $i\neq j$.
- $\omega\in\Gamma(\bigwedge^p T^\ast M)$ is an eigenform of the connection Laplacian $\Delta_{C,p}$ with $\|\omega\|_2=1$ corresponding to the eigenvalue $\lambda$ with $0\leq \lambda \leq \delta$.
Note that we have $\|\omega\|_\infty\leq C$, $\|f_i\|_\infty \leq C $ and $\|\nabla f_i\|_\infty \leq C$ for all $i\in\{1,\ldots,k\}$ by Lemma \[Linfes\], and $\lambda_i\geq n-p-C\delta^{1/2}$ by Main Theorem 1.
\[pfua\] For any $f\in {\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots,f_{k}\}$, we have $$\left\|\sum_{i=1}^n e^i\otimes (\nabla_{e_i}d f+f e^i)\wedge \omega \right\|_2\leq C\delta^{1/8}\|f\|_2.$$
We have $$\label{fua}
\begin{split}
&\left|\sum_{i=1}^n e^i\otimes (\nabla_{e_i}d f+f e^i)\wedge \omega\right|^2\\
=&\sum_{i=1}^n \langle (\nabla_{e_i}d f) \wedge \omega, (\nabla_{e_i}d f) \wedge \omega\rangle+ 2\sum_{i=1}^n \langle(\nabla_{e_i}d f) \wedge \omega, f e^i \wedge \omega\rangle\\
&\qquad\qquad\qquad\qquad \qquad\qquad\qquad \qquad+\sum_{i=1}^n f^2 \langle e^i \wedge \omega ,e^i \wedge \omega\rangle\\
=&|\nabla^2 f|^2|\omega|^2 -\sum_{i=1}^n|\iota(\nabla_{e_i}\nabla f)\omega|^2
-2 f\Delta f|\omega|^2\\
&\qquad \qquad\qquad \qquad\qquad -2\sum_{i=1}^n f \langle\omega , e^i\wedge \iota(\nabla_{e_i}\nabla f)\omega\rangle
+(n-p)f^2|\omega|^2\\
=& |\nabla^2 f|^2|\omega|^2-\frac{1}{n-p}(\Delta f)^2|\omega|^2
+2\Delta f \left(\frac{1}{n-p}\Delta f-f\right)|\omega|^2\\
-&(n-p)\left(\left(\frac{\Delta f}{n-p}\right)^2-f^2\right)|\omega|^2
-\left|\sum_{i=1}^n e^i \otimes\iota(\nabla_{e_i}\nabla f)\omega\right|^2-2\sum_{i=1}^n f \langle\omega , e^i\wedge \iota(\nabla_{e_i}\nabla f)\omega\rangle.
\end{split}$$ By the assumption, we have $$\begin{aligned}
\label{fub}\left\|\Delta f \left(\frac{1}{n-p}\Delta f-f\right)|\omega|^2\right\|_1\leq &C\delta^{1/2}\|f\|_2^2,\\
\label{fuc}\left\|\left(\left(\frac{\Delta f}{n-p}\right)^2-f^2\right)|\omega|^2\right\|_1\leq &C\delta^{1/2}\|f\|_2^2.\end{aligned}$$ By Lemma \[p4d\] (iv) and Lemma \[p5c\] (ii), we have $$\label{fud}
\left\|\sum_{i=1}^n e^i \otimes\iota(\nabla_{e_i}\nabla f)\omega\right\|_2
\leq \|\nabla (\iota(\nabla f)\omega)\|_2+ C\delta^{1/2}\|f\|_2\leq C\delta^{1/4}\|f\|_2,$$ and so $$\label{fue}
\left\|\sum_{i=1}^n f \langle\omega , e^i\wedge \iota(\nabla_{e_i}\nabla f)\omega\rangle\right\|_1\leq C\|f\|_2\left\|\sum_{i=1}^n e^i \otimes\iota(\nabla_{e_i}\nabla f)\omega\right\|_2
\leq C\delta^{1/4}\|f\|_2^2.$$ By Lemma \[p5c\], (\[fua\]), (\[fub\]), (\[fuc\]), (\[fud\]) and (\[fue\]), we get the lemma.
\[pfub\] Define $G=G(f_1,\ldots,f_k)$ by $$\begin{split}
G:=\Big\{x\in M: & |f_i^2+|\nabla f_i|^2-1|(x)\leq\delta^{1/1600n}\text{ for all $i=1,\ldots,k$, and}\\
&\left|\frac{1}{2}(f_i+f_j)^2+\frac{1}{2}|\nabla f_i+\nabla f_j|^2-1\right|(x)\leq \delta^{1/1600n},\\
&\left|\frac{1}{2}(f_i-f_j)^2+\frac{1}{2}|\nabla f_i-\nabla f_j|^2-1\right|(x)\leq \delta^{1/1600n}\text{ for all $i\neq j$}
\Big\}.
\end{split}$$ Then, we have the following properties.
- We have ${\mathop{\mathrm{Vol}}\nolimits}(M\setminus G)\leq C\delta^{1/1600n}{\mathop{\mathrm{Vol}}\nolimits}(M)$.
- For all $x\in G$ and $i,j$ with $i\neq j$, we have $\left|f_i f_j+\langle\nabla f_i,\nabla f_j\rangle\right|(x)\leq\delta^{1/1600n}$.
By Proposition \[p53a\] (iii), we have $$\begin{split}
\|f_i^2+|\nabla f_i|^2-1\|_1\leq&C\delta^{1/800n},\\
\left\|\frac{1}{2}(f_i+f_j)^2+\frac{1}{2}|\nabla f_i+\nabla f_j|^2-1\right\|_1\leq &C\delta^{1/800n},\\
\left\|\frac{1}{2}(f_i-f_j)^2+\frac{1}{2}|\nabla f_i-\nabla f_j|^2-1\right\|_1\leq &C\delta^{1/800n}
\end{split}$$ for all $i\neq j$. Therefore, we get $$\begin{aligned}
&{\mathop{\mathrm{Vol}}\nolimits}\left(\left\{x\in M: \left|f_i^2+|\nabla f_i|^2-1\right|(x)>\delta^{1/1600n}\right\}\right)\\
\leq& \delta^{-1/1600n}\int_M \left|f_i^2+|\nabla f_i|^2-1\right|\,d\mu_g\\
=& \delta^{-1/1600n}\|f_i^2+|\nabla f_i|^2-1\|_1{\mathop{\mathrm{Vol}}\nolimits}(M)\leq C\delta^{1/1600n}{\mathop{\mathrm{Vol}}\nolimits}(M)\end{aligned}$$ for all $i$. Similarly, we have $$\begin{aligned}
{\mathop{\mathrm{Vol}}\nolimits}\left(\left\{x\in M: \left|\frac{1}{2}(f_i+f_j)^2+\frac{1}{2}|\nabla f_i+\nabla f_j|^2-1\right|(x)>\delta^{1/1600n}\right\}\right)&\leq C\delta^{1/1600n}{\mathop{\mathrm{Vol}}\nolimits}(M),\\
{\mathop{\mathrm{Vol}}\nolimits}\left(\left\{x\in M: \left|\frac{1}{2}(f_i-f_j)^2+\frac{1}{2}|\nabla f_i-\nabla f_j|^2-1\right|(x)>\delta^{1/1600n}\right\}\right)&\leq C\delta^{1/1600n}{\mathop{\mathrm{Vol}}\nolimits}(M)\end{aligned}$$ for all $i\neq j$. Thus, we get (i).
For all $x\in G$ and $i,j$ with $i\neq j$, we have $$\begin{aligned}
&\left|f_i f_j+\langle\nabla f_i,\nabla f_j\rangle\right|(x)\\
=&\frac{1}{2}\left|
\frac{1}{2}(f_i+f_j)^2+\frac{1}{2}|\nabla f_i+\nabla f_j|^2
-\frac{1}{2}(f_i-f_j)^2-\frac{1}{2}|\nabla f_i-\nabla f_j|^2\right|(x)
\leq\delta^{1/1600n}.\end{aligned}$$ Thus, we get (ii).
Orientability
-------------
The goal of this subsection is to show the orientability of the manifold under the assumption of Main Theorem 2.
Note that our assumptions are Assumption \[asu1\].
\[pora\] If $$\lambda_{n-p+1}(g)\leq n-p+\delta$$ and $$\lambda_1(\Delta_{C,p})\leq \delta$$ hold, then $M$ is orientable.
To prove the theorem, we use the following claim:
\[porb\] Define $$\lambda_1(\Delta_{C,n}):=\inf \left\{\frac{\|\nabla \eta\|_2^2}{\|\eta\|_2^2}: \eta\in \Gamma(\bigwedge^n T^\ast M)\text{ with } \eta\neq 0\right\}.$$ If $$\lambda_1(\Delta_{C,n})< \frac{n}{n-1}(n-p-1)$$ holds, then $M$ is orientable.
See Corollary \[papeb\] and Remark \[papec\] in the appendix for the proof of Claim \[porb\].
Let $f_i$ be the $i$-th eigenfunction of the Laplacian with $\|f_i\|_2^2=1/(n-p+1)$ for each $i$, and $\omega$ be the first eigenform of the connection Laplacian $\Delta_{C,p}$ acting on $\Gamma(\bigwedge^p T^\ast M)$ with $\|\omega\|_2=1$. Put $$V:=\sum_{i=1}^{n-p+1} (-1)^{i-1} f_i d f_1\wedge\cdots \wedge \widehat{d f_i}\wedge\cdots \wedge d f_{n-p+1}\wedge \omega\in \Gamma(\bigwedge^n T^\ast M).$$ In the following, we show that $\|\nabla V\|_2^2/\|V\|_2^2< n(n-p+1)/(n-1)$.
Define a vector bundle $E:=T^\ast M\oplus \mathbb{R}e$, where $\mathbb{R}e$ denotes the trivial bundle of rank $1$ with a global non-vanishing section $e$. We consider an inner product $\langle\cdot,\cdot\rangle$ on $E$ defined by $$\langle \alpha+ f e,\beta +h e\rangle:=\langle\alpha,\beta\rangle+ fh$$ for all $\alpha,\beta\in\Gamma(T^\ast M)$ and $f,h\in C^\infty(M)$. Put $$S_i=d f_i +f_i e\in \Gamma(E)$$ for each $i$, and $$\alpha:=S_1\wedge\cdots \wedge S_{n-p+1}\in \Gamma(\bigwedge^{n-p+1} E).$$ Then, we have $\alpha\wedge \omega=e\wedge V$, and so $$\label{ora}
|\alpha\wedge \omega|=|V|.$$
For each $k=1,\ldots,n-p+1$, we have $$\begin{split}
&\Big\|
\big\langle
S_k\wedge\cdots \wedge S_{n-p+1}\wedge \omega,
\left(\iota(S_{k-1})\cdots\iota(S_1)\alpha\right)\wedge \omega
\big\rangle\\
&\qquad-\big\langle S_{k+1} \wedge\cdots \wedge S_{n-p+1}\wedge \omega,
\left(\iota(S_k)\cdots\iota(S_1)\alpha\right)\wedge \omega
\big\rangle
\Big\|_1\\
=&\left\|
\big\langle S_{k+1} \wedge\cdots \wedge S_{n-p+1}\wedge \omega,
\left(\iota(S_{k-1})\cdots\iota(S_1)\alpha\right)\wedge \iota(d f_k)\omega
\big\rangle
\right\|_1\\
\leq& C\|\iota(d f_k)\omega\|_2\leq C\delta^{1/4}
\end{split}$$ by Lemma \[p5c\] (i). By induction, we get $$\label{orb}
\||\alpha\wedge \omega|^2-|\alpha|^2|\omega|^2\|_1\leq C\delta^{1/4}.$$ In particular, we have $$\label{orc}
\left|\|\alpha\wedge \omega\|_2^2-\||\alpha|^2|\omega|^2\|_1\right|
\leq C\delta^{1/4}.$$
Since we have $$\left|\langle S_i(x), S_j(x)\rangle -\delta_{i j}\right|\leq \delta^{1/1600n}$$ for all $x\in G=G(f_1,\ldots,f_{n-p+1})$ and $i,j$ by Lemma \[pfub\] (ii), we get $$||\alpha|^2(x)-1|\leq C\delta^{1/1600n}$$ for all $x\in G$. Thus, we get $$\label{ord}
\begin{split}
&\left|
\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M(|\alpha|^2|\omega|^2-1) \,d\mu_g
\right|\\
=&\Bigg|
\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_G(|\alpha|^2-1)|\omega|^2 \,d\mu_g\\
&\qquad+\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_{M\setminus G}(|\alpha|^2-1)|\omega|^2 \,d\mu_g+\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M(|\omega|^2-1) \,d\mu_g
\Bigg|\\
\leq &C\delta^{1/1600n}
\end{split}$$ by Lemma \[p4c\] and Lemma \[pfub\] (i). By (\[ora\]), (\[orc\]) and (\[ord\]), we get $$\label{ore}
|\|V\|_2^2-1|\leq C\delta^{1/1600n}.$$
We next estimate $\|\nabla V\|_2^2$. We have $$\begin{split}
&\nabla V\\
=& \sum_{i=1}^{n-p+1} (-1)^{i-1} d f_i\otimes d f_1\wedge\cdots \wedge \widehat{d f_i}\wedge\cdots \wedge d f_{n-p+1}\wedge \omega\\
+&\sum_{j<i}\sum_{k=1}^n (-1)^{i-1}(-1)^{j-1} f_i e^k\otimes (\nabla_{e_k} d f_j)\wedge d f_1\wedge\cdots\wedge\widehat{d f_j} \wedge\cdots\wedge \widehat{d f_i}\wedge\cdots \wedge d f_{n-p+1}\wedge \omega\\
+&\sum_{i<j}\sum_{k=1}^n (-1)^{i-1}(-1)^{j} f_i e^k\otimes (\nabla_{e_k} d f_j)\wedge d f_1\wedge\cdots\wedge\widehat{d f_i} \wedge\cdots\wedge\widehat{d f_j}\wedge\cdots \wedge d f_{n-p+1}\wedge \omega\\
+&\sum_{i=1}^{n-p+1} \sum_{k=1}^n (-1)^{i-1} f_i e^k \otimes d f_1\wedge\cdots \wedge \widehat{d f_i}\wedge\cdots \wedge d f_{n-p+1}\wedge \nabla_{e_k}\omega.
\end{split}$$ Thus, we get $$\label{orf}
\begin{split}
&\left\|
\nabla V
- \sum_{i=1}^{n-p+1} (-1)^{i-1} d f_i\otimes d f_1\wedge\cdots \wedge \widehat{d f_i}\wedge\cdots \wedge d f_{n-p+1}\wedge \omega
\right\|_2\\
\leq&
\Bigg\|\sum_{j<i}\sum_{k=1}^n (-1)^{i-1}(-1)^{j-1} f_i f_j e^k\otimes e^k\wedge d f_1\wedge\cdots\wedge\widehat{d f_j} \wedge\cdots\wedge \widehat{d f_i}\wedge\cdots \wedge d f_{n-p+1}\wedge \omega\\
&+\sum_{i<j}\sum_{k=1}^n (-1)^{i-1}(-1)^{j} f_i f_j e^k\otimes e^k\wedge d f_1\wedge\cdots\wedge\widehat{d f_i} \wedge\cdots\wedge\widehat{d f_j}\wedge\cdots \wedge d f_{n-p+1}\wedge \omega\Bigg\|_2\\
&+C\sum_{i=1}^{n-p+1}\left\|\sum_{k=1}^n e^k\otimes (\nabla_{e_k}d f_i+f_i e^k)\wedge\omega\right\|_2
+ C\|\nabla\omega\|_2\\
\leq &C\delta^{1/8}
\end{split}$$ by Lemma \[pfua\].
Similarly to (\[orb\]), we have $$\label{org}
\begin{split}
&\Bigg\|\left|\sum_{i=1}^{n-p+1} (-1)^{i-1} d f_i\otimes d f_1\wedge\cdots \wedge \widehat{d f_i}\wedge\cdots \wedge d f_{n-p+1}\wedge \omega\right|^2\\
&\qquad-\left|\sum_{i=1}^{n-p+1} (-1)^{i-1} d f_i\otimes d f_1\wedge\cdots \wedge \widehat{d f_i}\wedge\cdots \wedge d f_{n-p+1}\right|^2|\omega|^2\Bigg\|_1\\
&\leq C\delta^{1/4}.
\end{split}$$
Since we have $$d f_1\wedge\cdots\wedge d f_{n-p+1}\wedge\omega=0,$$ we get $$\label{orh}
\begin{split}
&\|
|d f_1\wedge\cdots\wedge d f_{n-p+1}|^2|\omega|^2
\|_1\\
=&
\||d f_1\wedge\cdots\wedge d f_{n-p+1}|^2|\omega|^2-
|d f_1\wedge\cdots\wedge d f_{n-p+1}\wedge\omega|^2
\|_1
\leq C\delta^{1/4}
\end{split}$$ similarly to (\[orb\]). By (\[q1k\]), we get $$\label{ori}
\begin{split}
&\left|
\sum_{i=1}^{n-p+1}(-1)^{i-1}d f_i\otimes d f_1\wedge\cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p+1}
\right|^2\\
=&(n-p+1) |d f_1\wedge \cdots\wedge d f_{n-p+1}|^2.
\end{split}$$ By (\[orh\]) and (\[ori\]), we get $$\label{orj}
\left\|
\left|\sum_{i=1}^{n-p+1}(-1)^{i-1}d f_i\otimes d f_1\wedge\cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p+1}\right|^2|\omega|^2
\right\|_1\leq C\delta^{1/4}.$$ By (\[org\]) and (\[orj\]), we have $$\label{ork}
\begin{split}
&\left\|
\sum_{i=1}^{n-p+1}(-1)^{i-1}d f_i\otimes d f_1\wedge\cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p+1}\wedge\omega
\right\|_2^2\\
=&\left\|
\left|\sum_{i=1}^{n-p+1}(-1)^{i-1}d f_i\otimes d f_1\wedge\cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p+1}\wedge\omega\right|^2
\right\|_1\leq C\delta^{1/4}.
\end{split}$$ By (\[orf\]) and (\[ork\]), we get $$\label{orl}
\|\nabla V\|_2\leq C\delta^{1/8}.$$ By (\[ore\]) and (\[orl\]), we get $$\lambda_1(\Delta_{C,n})\leq C\delta^{1/4},$$ and so we get the theorem by Claim \[porb\].
Combining Theorem \[MT2\] and Theorem \[pora\], we get Main Theorem 2.
Almost Parallel $(n-p)$-form II
-------------------------------
In this subsection, we show that the assumption “$\lambda_{n-p}(g)$ is close to $n-p$” implies the condition “$\lambda_{n-p+1}(g)$ is close to $n-p$” under the assumption $\lambda_1(\Delta_{C,n-p})\leq \delta$.
Note that our assumptions are Assumption \[asu1\].
\[pala\] Suppose that
- $f_i\in C^\infty(M)$ $(i\in\{1,\ldots,n-p\})$ is an eigenfunction of the Laplacian acting on functions with $\|f_i\|_2^2=1/(n-p+1)$ corresponding to the eigenvalue $\lambda_i$ with $0<\lambda_i\leq n-p+\delta$ such that $$\int_M f_i f_j\,d\mu_g=0$$ holds for any $i\neq j$,
- $\xi\in\Gamma(\bigwedge^{n-p} T^\ast M)$ is an eigenform of the connection Laplacian $\Delta_{C,n-p}$ with $\|\xi\|_2=1$ corresponding to the eigenvalue $\lambda$ with $0\leq \lambda \leq \delta$.
Put $$F:= \langle d f_1\wedge\ldots\wedge d f_{n-p}, \xi \rangle\in C^\infty(M).$$ Then, we have $$\left|\|F\|_2^2-\frac{1}{n-p+1}\right|\leq C\delta^{1/1600n},\quad \left|\|\nabla F\|_2^2-\frac{n-p}{n-p+1}\right|\leq C\delta^{1/1600n}$$ and $$\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M f_i F\,d\mu_g\right|\leq C\delta^{1/2}$$ for all $i=1,\ldots, n-p$.
If $M$ is not orientable, we take the two-sheeted oriented Riemannian covering $\pi\colon (\widetilde{M},\tilde{g})\to (M,g)$, and put $$\widetilde{F}:=F\circ \pi,\quad \tilde{f}_i:=f_i\circ \pi.$$ Then, we have $
\|F\|_2=\|\widetilde{F}\|_2$, $\|\nabla F\|_2=\|\nabla \widetilde{F}\|_2,$ $$\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(\widetilde{M})}\int_{\widetilde{M}} \tilde{f}_i \widetilde{F} \,d\mu_{\tilde{g}}=
\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M f_i F \,d\mu_g$$ and $$\widetilde{F}=\langle d \tilde{f}_1\wedge\ldots\wedge d \tilde{f}_{n-p}, \pi^\ast \xi\rangle.$$ Thus, it is enough to consider the case when $M$ is orientable. In the following, we assume that $M$ is orientable, and we fix an orientation of $M$.
Put $$\omega:=\ast \xi\in \Gamma(\bigwedge^p T^\ast M).$$ Let $V_g\in \Gamma(\bigwedge^n T^\ast M)$ be the volume form of $(M,g)$. Then, we have $$\label{ala}
F V_g= d f_1\wedge\cdots \wedge d f_{n-p}\wedge \omega.$$ Define a vector bundle $E:=T^\ast M\oplus \mathbb{R}e$, where $\mathbb{R}e$ denotes the trivial bundle of rank $1$ with a global non-vanishing section $e$. We consider an inner product $\langle\cdot,\cdot\rangle$ on $E$ defined by $$\langle \alpha+ f e,\beta +h e\rangle:=\langle\alpha,\beta\rangle+ fh$$ for all $\alpha,\beta\in\Gamma(T^\ast M)$ and $f,h\in C^\infty(M)$. Put $$S_i:=d f_i +f_i e\in \Gamma(E)$$ for each $i$, and $$\beta:=S_1\wedge\cdots \wedge S_{n-p}\in \Gamma(\bigwedge^{n-p} E).$$
Since we have $|F|=|F V_g|$, we get $$\||F|^2-|d f_1\wedge\cdots \wedge d f_{n-p} |^2|\omega|^2\|_1\leq C\delta^{1/4}$$ similarly to (\[orb\]) by (\[ala\]), and so $$\label{alb}
\left|\|F\|_2^2-\left\||d f_1\wedge\cdots \wedge d f_{n-p} |^2|\omega|^2\right\|_1\right|\leq C\delta^{1/4}$$
By Lemma \[pfua\] and (\[ala\]), we have $$\left\|
\nabla (F V_g)+\sum_{i=1}^{n-p} \sum_{k=1}^n(-1)^{i-1} f_i e^k\otimes e^k\wedge d f_1\wedge \cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p}\wedge \omega
\right\|_2\leq C\delta^{1/8}.$$ Since $|\nabla(F V_g)|=|\nabla F|$, we get $$\label{alc}
\left|\|\nabla F\|_2^2-\left\|\left|\sum_{i=1}^{n-p} \sum_{k=1}^n(-1)^{i-1} f_i e^k\otimes e^k\wedge d f_1\wedge \cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p}\wedge \omega\right|^2\right\|_1\right|\leq C\delta^{1/8}.$$ We have $$\label{ald}
\begin{split}
&\left|\sum_{i=1}^{n-p} \sum_{k=1}^n(-1)^{i-1} f_i e^k\otimes e^k\wedge d f_1\wedge \cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p}\wedge \omega
\right|^2\\
=&\sum_{k=1}^n \left|e^k\wedge\left(\sum_{i=1}^{n-p} (-1)^{i-1} f_i d f_1\wedge \cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p}\right)\wedge \omega\right|^2\\
=&\left|\sum_{i=1}^{n-p} (-1)^{i-1} f_i d f_1\wedge \cdots \wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p}\wedge \omega\right|^2.
\end{split}$$ Similarly to (\[orb\]), we have $$\begin{split}
&\Bigg\|\left|\sum_{i=1}^{n-p} (-1)^{i-1} f_i d f_1\wedge \cdots \wedge\widehat{d f_i}\wedge \cdots\wedge d f_{n-p}\wedge \omega\right|^2\\
&\qquad -\left|\sum_{i=1}^{n-p} (-1)^{i-1} f_i d f_1\wedge \cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p}\right|^2|\omega|^2\Bigg\|_1\leq C\delta^{1/4}.
\end{split}$$ Since we have $$\iota(e)\beta=\sum_{i=1}^{n-p} (-1)^{i-1} f_i d f_1\wedge \cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p},$$ we get $$\label{alf}
\left\|\left|\sum_{i=1}^{n-p} (-1)^{i-1} f_i d f_1\wedge \cdots\wedge \widehat{d f_i}\wedge \cdots\wedge d f_{n-p}\wedge \omega\right|^2 -|\iota(e)\beta|^2|\omega|^2\right\|_1\leq C\delta^{1/4}.$$ By (\[alc\]), (\[ald\]) and (\[alf\]), we get $$\label{alf1}
\left|\|\nabla F\|_2^2-\left\||\iota(e)\beta|^2|\omega|^2\right\|_1\right|\leq C\delta^{1/8}.$$
We have $$\label{alg}
|\beta|^2=|d f_1\wedge\cdots\wedge d f_{n-p}|^2+|\iota(e)\beta|^2.$$ We calculate $\sum_{k=1}^n\left|e^k\wedge \beta\right|^2$ in two ways. We have $$\label{alh}
\begin{split}
\sum_{k=1}^n|e^k\wedge \beta|^2=&(p+1)|\beta|^2-|e\wedge\beta|^2\\
=&(p+1)|\beta|^2-|d f_1\wedge\cdots\wedge d f_{n-p}|^2= p|\beta|^2+|\iota(e)\beta|^2
\end{split}$$ by (\[alg\]). For all $\eta\in \Gamma(T^\ast M)$, we have $$\begin{aligned}
&|\eta\wedge\beta|^2\\
=&|\eta|^2|\beta|^2-\langle\iota(\eta)\beta,\iota(\eta)\beta\rangle\\
=&|\eta|^2|\beta|^2-\sum_{i,j=1}^{n-p}(-1)^{i+j}\langle \eta, d f_i\rangle\langle\eta, d f_j\rangle
\langle S_1\wedge\cdots\wedge \widehat{S_i}\wedge\cdots \wedge S_{n-p},S_1\wedge\cdots\wedge \widehat{S_j}\wedge\cdots \wedge S_{n-p}
\rangle,\end{aligned}$$ and so we get $$\label{ali}
\begin{split}
&\sum_{k=1}^n|e^k\wedge \beta|^2\\
=&n|\beta|^2-\sum_{i,j=1}^{n-p}(-1)^{i+j}\langle d f_i,d f_j\rangle
\langle S_1\wedge\cdots\wedge \widehat{S_i}\wedge\cdots \wedge S_{n-p},S_1\wedge\cdots\wedge \widehat{S_j}\wedge\cdots \wedge S_{n-p}\rangle.
\end{split}$$ By (\[alh\]) and (\[ali\]), we get $$\label{alj}
\begin{split}
&|\iota(e)\beta|^2\\
=&(n-p)|\beta|^2-\sum_{i,j=1}^{n-p}(-1)^{i+j}\langle d f_i,d f_j\rangle
\langle S_1\wedge\cdots\wedge \widehat{S_i}\wedge\cdots \wedge S_{n-p},S_1\wedge\cdots\wedge \widehat{S_j}\wedge\cdots \wedge S_{n-p}\rangle
\end{split}$$
Since we have $|\langle S_i,S_j\rangle(x)-\delta_{i j}|\leq C\delta^{1/1600n}$ for all $x\in G=G(f_1,\ldots, f_{n-p})$ by Lemma \[pfub\] (ii), we have $$\label{alk}
\begin{split}
&\Bigg\|\sum_{i=1}^{n-p}|d f_i|^2\\
&-\sum_{i,j=1}^{n-p}(-1)^{i+j}\langle d f_i,d f_j\rangle
\langle S_1\wedge\cdots\wedge \widehat{S_i}\wedge\cdots \wedge S_{n-p},S_1\wedge\cdots\wedge \widehat{S_j}\wedge\cdots \wedge S_{n-p}\rangle|\omega|^2
\Bigg\|_1
\leq C\delta^{1/1600n}
\end{split}$$ and $$\label{all}
\left|\left\||\beta|^2|\omega|^2\right\|_1-1\right|\leq C\delta^{1/1600n}$$ by Lemma \[p4c\] and Lemma \[pfub\] (i). By the assumption, we have $$\label{alm}
\left|\sum_{i=1}^{n-p}\|d f_i\|_2^2-\frac{(n-p)^2}{n-p+1}\right|\leq C\delta^{1/2}.$$ By (\[alj\]), (\[alk\]), (\[all\]) and (\[alm\]), we get $$\label{aln}
\left|\left\||\iota(e)\beta|^2|\omega|^2\right\|_1-\frac{n-p}{n-p+1}\right|\leq C\delta^{1/1600n},$$ and so $$\label{alo}
\left|\left\||d f_1\wedge\cdots\wedge d f_{n-p}|^2|\omega|^2\right\|_1-\frac{1}{n-p+1}\right|\leq C\delta^{1/1600n}$$ by (\[alg\]) and (\[all\]). By (\[alb\]) and (\[alo\]), we get $$\left|\|F\|_2^2-\frac{1}{n-p+1}\right|\leq C\delta^{1/1600n}.$$ By (\[alf1\]) and (\[aln\]), we get $$\left|\|\nabla F\|_2^2- \frac{n-p}{n-p+1}\right| \leq C\delta^{1/1600n}.$$
Let us show the remaining assertion. Since we have $$\begin{aligned}
f_i F V_g=&\frac{1}{2}(-1)^{i-1} d \left(f_i^2 d f_1\wedge\cdots\wedge\widehat{d f_i}\wedge \cdots \wedge d f_{n-p}\wedge\omega\right)\\
-&\frac{1}{2}(-1)^{i-1} (-1)^{n-p-1}f_i^2 d f_1\wedge\cdots\wedge\widehat{d f_i}\wedge\cdots \wedge d f_{n-p}\wedge d \omega,\end{aligned}$$ we get $$\left|\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M f_i F\,d\mu_g\right|\leq
C\|\nabla \omega\|_2 \leq C\delta^{1/2}$$ by the Stokes theorem.
By applying the Rayleigh principle $$\begin{aligned}
&\lambda_{n-p+1}(g)\\
=&\inf\left\{\sup_{f\in V\setminus\{0\}}\frac{\|\nabla f\|_2^2}{\|f\|_2^2}: V\text{ is an $(n-p+1)$-dimensional subspace of } C^\infty (M)
\right\}\end{aligned}$$ to the subspace ${\mathop{\mathrm{Span}}\nolimits}_{\mathbb{R}}\{f_1,\ldots, f_{n-p}, F\}$, we immediately get the following corollary:
\[palb\] If $$\lambda_{n-p}(g)\leq n-p+\delta$$ and $$\lambda_1(\Delta_{C,n-p})\leq \delta$$ hold, then we have $$\lambda_{n-p+1}(g)\leq n-p+C\delta^{1/1600n}.$$
Combining Theorem \[MT2\] and Corollary \[palb\], we get Main Theorem 4.
Finally, we investigate the Gromov-Hausdorff limit of the sequence of the Riemannian manifolds that satisfy our pinching condition.
Take $n\geq 5$ and $2\leq p < n/2$. Let $\{(M_i,g_i)\}_{i\in\mathbb{N}}$ be a sequence of $n$-dimensional closed Riemannian manifolds with ${\mathop{\mathrm{Ric}}\nolimits}_{g_i}\geq (n-p-1)g_i$ that satisfies one of the following:
- $\lim_{i\to\infty}\lambda_{n-p+1}(g_i)=n-p$ and $\lim_{i\to \infty}\lambda_1(\Delta_{C,p},g_i)=0$,
- $M_i$ is orientable for each $i$, $\lim_{i\to\infty}\lambda_{n-p}(g_i)=n-p$ and $\lim_{i\to \infty}\lambda_1(\Delta_{C,p},g_i)=0$,
- $\lim_{i\to\infty}\lambda_{n-p}(g_i)=n-p$ and $\lim_{i\to \infty}\lambda_1(\Delta_{C,n-p},g_i)=0$.
If $\{(M_i,g_i)\}_{i\in\mathbb{N}}$ converges to a geodesic space $X$, then there exists a geodesic space $Y$ such that $X$ is isometric to $S^{n-p}\times Y$.
By Main Theorem 2 and Main Theorem 4, we get that there exist a sequence of positive real numbers $\{\epsilon_i\}$ and compact metric spaces $\{Y_i\}$ such that $\lim_{i\to \infty}\epsilon_i=0$ and $d_{GH}(M_i,S^{n-p}\times Y_i)\leq \epsilon_i$. Then, $\{S^{n-p}\times Y_i\}$ converges to $X$ in the Gromov-Hausdorff topology, and so $\{Y_i\}$ is pre-compact in the Gromov-Hausdorff topology by [@Pe3 Theorem 11.1.10]. Thus, there exists a subsequence that converges to some compact metric space $Y$. Therefore, we get that $X$ is isometric to $S^{n-p}\times Y$. Since $X$ is a geodesic space, $Y$ is also a geodesic space.
Limit Spaces and Unorientability
================================
In this appendix, we generalize our main theorems to certain limit spaces. We first give the proof of Claim \[porb\] in the following form.
\[papea\] Let $(M,g)$ be an $n$-dimensional unorientable closed Riemannian manifold. Define $$\lambda_1(\Delta_{C,n},g):=\inf \left\{\frac{\|\nabla \eta\|_2^2}{\|\eta\|_2^2}: \eta\in \Gamma(\bigwedge^n T^\ast M)\text{ with } \eta\neq 0\right\}.$$ Then, we have the following:
- If ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-1)g$, then we have $$\lambda_1(\Delta_{C,n},g)\geq n.$$
- If ${\mathop{\mathrm{Ric}}\nolimits}\geq -K g$ and ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$ $(K,D>0)$, then we have $$\lambda_1(\Delta_{C,n},g)\geq C_1(n,K,2D),$$ where $C_1(n,K,D)$ is defined by $$C_1(n,K,D):=\frac{1}{(n-1)D^2\exp\left(1+\sqrt{1+4(n-1)KD^2}\right)}.$$
Take the two-sheeted oriented Riemannian covering $\pi\colon (\widetilde{M},\tilde{g})\to (M,g)$. Let us show that $$\lambda_1(\Delta_{C,n},g)\geq \lambda_1(\tilde{g}).$$ If we succeed in proving this inequality, we get (i) by the Lichnerowicz inequality, and we get (ii) by ${\mathop{\mathrm{diam}}\nolimits}(\widetilde{M})\leq 2 {\mathop{\mathrm{diam}}\nolimits}(M)$ and the Li-Yau estimate [@SY p.116], which asserts that $$\lambda_1(g_1)\geq C_1(n,K,D)$$ holds for any $n$-dimensional closed Riemannian manifold $(N_1,g_1)$ with ${\mathop{\mathrm{Ric}}\nolimits}_{g_1}\geq -K g_1$ and ${\mathop{\mathrm{diam}}\nolimits}(N_1)\leq D$.
Take the first eigenform $\omega\in\Gamma(\bigwedge^n T^\ast M)$ of the connection Laplacian $\Delta_{C,n}$. Put $$\widetilde{\omega}:=\pi^\ast \omega \in\Gamma(\bigwedge^n T^\ast \widetilde{M}).$$ Then, we have $$\Delta_{C,n}\widetilde{\omega}=\lambda_1(\Delta_{C,n},g)\widetilde{\omega}.$$ By the correspondence through the Hodge star operator, we have that $\lambda_1(\Delta_{C,n},g)$ is the eigenvalue of the Laplacian acting on functions on $\widetilde{M}$. If $\lambda_1(\Delta_{C,n},g)=0$, then $\omega$ defines an orientation of $M$, and so we have $$\lambda_1(\Delta_{C,n},g)\neq 0=\lambda_0(\tilde{g}).$$ Thus, we get $$\lambda_1(\Delta_{C,n},g)\geq\lambda_1(\tilde{g}).$$ Therefore, we get the lemma.
We immediately get the following corollary.
\[papeb\] Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold. If one of the following properties holds, then $M$ is orientable.
- ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-1)g$ and $$\lambda_1(\Delta_{C,n},g)< n.$$
- ${\mathop{\mathrm{Ric}}\nolimits}\geq -K g$, ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$ $(K,D>0)$ and $$\lambda_1(\Delta_{C,n},g)< C_1(n,K,2D).$$
\[papec\] For any closed Riemannian manifold $(M,g)$ and a positive real number $a>0$, we have ${\mathop{\mathrm{Ric}}\nolimits}_{g}={\mathop{\mathrm{Ric}}\nolimits}_{a g}$ and $$\lambda_k(a g)=\frac{1}{a}\lambda_k(g)$$ for all $k$. Thus, we get Claim \[porb\] by Corollary \[papeb\].
As an application of Lemma \[papea\], we show the stability of unorientability under the non-collapsing Gromov-Hausdorff convergence assuming the two-sided bound on the Ricci curvature.
\[paped\] Take real numbers $K_1,K_2\in\mathbb{R}$ and positive real numbers $D>0$ and $v>0$. Let $\{(M_i,g_i)\}$ be a sequence of $n$-dimensional unorientable closed Riemannian manifolds with $K_1 g_i\leq {\mathop{\mathrm{Ric}}\nolimits}_{g_i}\leq K_2 g_i$, ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$ and ${\mathop{\mathrm{Vol}}\nolimits}(M)\geq v$. Suppose that $\{(M_i,g_i)\}$ converges to a limit space $X$ in the Gromov-Hausdorff sense. Then, $X$ is not orientable in the sense of Honda [@Hoor] $($see also the definition below$)$.
Note that Honda [@Hoor Theorem 1.3] showed the stability of orientability without assuming the upper bound on the Ricci curvature.
Before proving Theorem \[paped\], we fix our notation and recall definitions about limit spaces.
Take real numbers $K_1,K_2\in\mathbb{R}$ and positive real numbers $D>0$ and $v>0$.
- Let $\mathcal{M}_1=\mathcal{M}_1(n,K_1,K_2,v)$ be the set of isometry classes of $n$-dimensional closed Riemannian manifolds $(M,g)$ with $K_1g\leq {\mathop{\mathrm{Ric}}\nolimits}_g \leq K_2 g$ and ${\mathop{\mathrm{Vol}}\nolimits}(M)\geq v$. Let $\overline{\mathcal{M}}_1=\overline{\mathcal{M}}_1(n,K_1,K_2,v)$ be the closure of $\mathcal{M}$ in the Gromov-Hausdorff topology.
- Let $\mathcal{M}_2=\mathcal{M}_2(n,K_1,K_2,D,v)$ be the set of isometry classes of $n$-dimensional closed Riemannian manifolds $(M,g)$ with $K_1g\leq {\mathop{\mathrm{Ric}}\nolimits}_g \leq K_2 g$, ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$ and ${\mathop{\mathrm{Vol}}\nolimits}(M)\geq v$. Let $\overline{\mathcal{M}}_2=\overline{\mathcal{M}}_2(n,K_1,K_2,D,v)$ be the closure of $\mathcal{M}$ in the Gromov-Hausdorff topology.
By the Myers theorem, we have $$\begin{aligned}
\mathcal{M}_1(n,K_1,K_2,v)\subset &\mathcal{M}_2(n,K_1,K_2,\sqrt{(n-1)/K_1}\pi,v),\\
\overline{\mathcal{M}}_1(n,K_1,K_2,v)\subset&\overline{\mathcal{M}}_2(n,K_1,K_2,\sqrt{(n-1)/K_1}\pi,v)\end{aligned}$$ if $K_1>0$.
If $X_i\in\overline{\mathcal{M}}_2$ ($i\in \mathbb{N}$) converges to $X\in\overline{\mathcal{M}}_2$ in the Gromov-Hausdorff topology, then there exist a sequence of positive real numbers $\{\epsilon_i\}_{i\in \mathbb{N}}$ with $\lim_{i\to \infty}\epsilon_i=0$, and a sequence of $\epsilon_i$-Hausdorff approximation maps $\phi_i \colon X_i\to X$. Fix such a sequence. We say a sequence $x_i\in X_i$ converges to $x\in X$ if $\lim_{i\to \infty}\phi_i(x_i)=x$ (denote it by $x_i\stackrel{GH}{\to} x$). By the volume convergence theorem [@CC1 Theorem 5.9], $(X_i,H^n)$ converges to $(X,H^n)$ in the measured Gromov-Hausdorff sense, i.e., for all $r>0$ and all sequence $x_i\in X_i$ that converges to $x\in X$, we have $\lim_{i\to \infty}H^n(B_r (x_i))=H^n(B_r(x))$, where $H^n$ denotes the $n$-dimensional Hausdorff measure.
For all $X\in \overline{\mathcal{M}}_2$, we can consider the cotangent bundle $\pi \colon T^\ast X \to X$ with a canonical inner product by [@Ch0] and [@CC3] (see also [@Ho1 Section 2] for a short review). We have $H^n(X\setminus \pi(T^\ast X))=0$ and $T^\ast_x X:=\pi^{-1}(x)$ is an $n$-dimensional vector space for all $x\in \pi(T^\ast X)$. For all Lipschitz function $f$ on $X$, we can define $d f(x)\in T_x^\ast X$ for almost all $x\in X$, and we have $d f\in L^\infty(T^\ast X)$.
Let us recall definitions of functional spaces on limit spaces. Note that we can define such functional spaces on more general spaces than our assumption. Some of the following functional spaces are first introduced by Gigli [@Gig].
Let $X\in \overline{\mathcal{M}}_2$.
- Let ${\mathop{\mathrm{LIP}}\nolimits}(X)$ be the set of the Lipschitz functions on $X$. For all $f\in {\mathop{\mathrm{LIP}}\nolimits}(X)$, we define $\|f\|_{H^{1,2}}^2=\|f\|_2^2+\|d f\|_2^2$. Let $H^{1,2}(X)$ be the completion of ${\mathop{\mathrm{LIP}}\nolimits}(X)$ with respect to this norm.
- Define $$\begin{split}
\mathcal{D}^2(\Delta,X):=\Big\{f\in H^{1,2}(X)& : \text{there exists $F\in L^2(X)$ such that}\\
&\int_X \langle df, dh \rangle\,d H^n=\int_X F h\,d H^n \text{ for all $h\in H^{1,2}(X)$} \Big\}.
\end{split}$$ For any $f\in\mathcal{D}^2(\Delta,X)$, the function $F\in L^2(X)$ is uniquely determined. Thus, we define $\Delta f:=F$.
- Define $$\begin{split}
{\mathop{\mathrm{Test}}\nolimits}F(X):=&\left\{f\in\mathcal{D}^2(\Delta,X)\cap {\mathop{\mathrm{LIP}}\nolimits}(X):\Delta f\in H^{1,2}(X)\right\},\\
{\mathop{\mathrm{TestForm}}\nolimits}_p(X):=&\left\{\sum_{i=1}^N f_{0,i} d f_{1,i}\wedge\ldots\wedge d f_{p,i}: N\in \mathbb{N},\, f_{j,i}\in {\mathop{\mathrm{Test}}\nolimits}F(X)\right\}
\end{split}$$ for all $p\in\{1,\ldots,n\}$.
- The operator $$\nabla\colon {\mathop{\mathrm{TestForm}}\nolimits}_p(X)\to L^2(T^\ast X \otimes \bigwedge^p T^\ast X)$$ is defined by $$\begin{aligned}
&\nabla\sum_{i=1}^N f_{0,i} d f_{1,i}\wedge\ldots\wedge d f_{p,i}\\
:=&\sum_{i=1}^N \left(d f_{0,i}\otimes d f_{1,i}\wedge\ldots\wedge d f_{p,i}+ \sum_{j=1}^p f_{0,i} d f_{1,i}\wedge\ldots\wedge\nabla^2 f_{j,i}\wedge\ldots\wedge d f_{p,i}\right),\end{aligned}$$ where $\nabla^2$ denotes the Hessian ${\mathop{\mathrm{Hess}}\nolimits}$ defined in [@Gig Definition 3.3.1] or [@Ho0]. Gigli defined the functional space $W^{2,2}(X)\subset H^{1,2}(X)$, on which we can define the Hessian as an $L^2$-tensor, and showed that $\mathcal{D}(\Delta,X)\subset W^{2,2}(X)$. Honda showed that for any $f\in \mathcal{D}^2(\Delta,X)$, $f$ is weakly twice differentiable [@Ho3 Theorem 1.9] in the sense of [@Ho0], that we can define the Hessian ${\mathop{\mathrm{Hess}}\nolimits}f$ using the Levi-Civita connection defined in [@Ho0], and that ${\mathop{\mathrm{Hess}}\nolimits}f\in L^2(T^\ast X\otimes T^\ast X)$ [@Ho3 Theorem 4.11]. Moreover, Honda showed that his definition of the Hessian coincides with Gigli’s one [@Ho3 Theorem 1.9].
- For any $\omega\in {\mathop{\mathrm{TestForm}}\nolimits}_p(X)$, we define $\|\omega\|_{H_C^{1,2}}^2:=\|\omega\|_2^2+\|\nabla \omega\|_2^2$. Let $H^{1,2}_C(\bigwedge^p T^\ast X)$ be the completion of ${\mathop{\mathrm{TestForm}}\nolimits}_p (X)$ with respect to this norm.
- Define $$\begin{split}
\mathcal{D}^2(\Delta_{C,p},X):=\Big\{\omega \in& H^{1,2}_C(\bigwedge^p T^\ast X) : \text{there exists $\hat{\omega}\in L^2(\bigwedge^p T^\ast X)$ such that}\\
&\int_X \langle \nabla \omega, \nabla \eta \rangle\,d H^n=\int_X \langle\hat{\omega}, \eta\rangle \,d H^n \text{ for all $\eta\in H_C^{1,2}(\bigwedge^p T^\ast X)$} \Big\}.
\end{split}$$ For any $\omega\in\mathcal{D}^2(\Delta_{C,p},X)$, the form $\hat{\omega}\in L^2(\bigwedge^p T^\ast X)$ is uniquely determined. Thus, we put $\Delta_{C,p} \omega:=\hat{\omega}$.
- For all $k\in \mathbb{Z}_{>0}$, we define $$\begin{split}
\lambda_k(\Delta_{C,p},X):=\inf\left\{\sup_{\omega\in \mathcal{E}_k\setminus \{0\}}\frac{\|\nabla \omega\|^2_2}{\|\omega\|^2_2}: \mathcal{E}_k\subset H^{1,2}_C(\bigwedge^p T^\ast X)\text{ is a $k$-dimensional subspace}\right\}.
\end{split}$$
Similarly to the smooth case, there exists a complete orthonormal system of eigenforms of the connection Laplacian $\Delta_{C,p}$ in $L^2(\bigwedge^p T^\ast M)$, and each eigenform is an element of $\mathcal{D}^2(\Delta_{C,p},X)$ (see [@Ho2 Theorem 4.17]).
Honda [@Ho2] showed the following theorem:
\[papee\] Let $\{X_i\}_{i\in \mathbb{N}}$ be a sequence in $\overline{\mathcal{M}}_2$ and let $X\in\overline{\mathcal{M}}_2$ be its Gromov-Hausdorff limit. Then, we have $$\lim_{i\to \infty}\lambda_k(\Delta_{C,p},X_i)=\lambda_k(\Delta_{C,p},X)$$ for all $p\in\{0,\ldots,n\}$ and $k\in\mathbb{Z}_{>0}$.
Let $X\in \overline{\mathcal{M}}_2$. We say that $X$ is orientable if there exists $\omega\in L^\infty(\bigwedge^n T^\ast X)$ such that $|\omega|(z)=1$ for almost all $z\in X$ and that $$\langle\omega,\eta\rangle\in H^{1,2}(X)$$ for any $\eta \in {\mathop{\mathrm{TestForm}}\nolimits}_n(X)$. Then, we call $\omega$ an orientation of $X$.
\[papef\] Let $X\in \overline{\mathcal{M}}_2$. Then, $X$ is orientable if and only if $\lambda_1(\Delta_{C,n},X)=0$.
We first suppose that $X$ is orientable and show $\lambda_1(\Delta_{C,n},X)=0$. Let $\omega\in L^\infty(\bigwedge^n T^\ast X)$ be the orientation of $X$. By [@Hoor Proposition 6.5], for almost all $z\in X$, $\omega$ is differentiable at $z$ and $\nabla^{g_X}\omega(z)=0$, where $\nabla^{g_X}$ denotes the Levi-Civita connection defined in [@Ho0]. By Proposition 4.5 and Remark 4.7 in [@Ho2], we have $\omega\in H^{1,2}_C(\bigwedge^p T^\ast X)$. By [@Ho3 Corollary 7.10], we have $\nabla \omega(z)=\nabla^{g_X}\omega(z)=0$ for almost all $z\in X$. Thus, we get $$\lambda_1(\Delta_{C,n},X)=0$$ by the definition of $\lambda_1(\Delta_{C,n},X)$.
We next suppose $\lambda_1(\Delta_{C,n},X)=0$ and show that $X$ is orientable. Let $\{(M_i,g_i)\}_{i\in \mathbb{N}}$ be a sequence in $\mathcal{M}_2$ that converges to $X$ in the Gromov-Hausdorff topology. Then, we have $\lim_{i\to \infty}\lambda_1(\Delta_{C,n},g_i)=0$ by Theorem \[papee\]. Thus, by Corollary \[papeb\], we get that $M_i$ is orientable for sufficiently large $i$, and so $X$ is orientable by the stability of orientability [@Hoor Theorem 1.3].
Let $\{(M_i,g_i)\}_{i\in \mathbb{N}}$ be a sequence in $\mathcal{M}_2$ and let $X$ be its Gromov-Hausdorff limit. Suppose that each $M_i$ is not orientable. Then, we have $$\lambda_1(\Delta_{C,n},g_i)\geq C_1(n,K_1,2D)$$ by Lemma \[papea\]. By Theorem \[papee\], we get $$\lambda_1(\Delta_{C,n},X)\geq C_1(n,K_1,2D)>0.$$ Thus, by Lemma \[papef\], we get the theorem.
\[papeg\] Let $X\in \overline{\mathcal{M}}_2$. If $X$ is not orientable, then we have $$\lambda_1(\Delta_{C,n},X)\geq C_1(n,K_1,2D).$$
Let $\{(M_i,g_i)\}_{i\in \mathbb{N}}$ be a sequence in $\mathcal{M}_2$ that converges to $X$ in the Gromov-Hausdorff topology. By Lemma \[papef\], we have $\lambda_1(\Delta_{C,n},X)>0$, and so we get $$\lambda_1(\Delta_{C,n},g_i)>0$$ for sufficiently large $i$ by Theorem \[papee\]. Thus, $M_i$ is not orientable and $$\lambda_1(\Delta_{C,n},g_i)\geq C_1(n,K_1,2D)$$ for sufficiently large $i$ by Lemma \[papea\]. By Theorem \[papee\], we get the theorem.
We immediately get the following corollaries:
\[papeh\] Let $\{X_i\}_{i\in \mathbb{N}}$ be a sequence in $\overline{\mathcal{M}}_2$ and let $X\in\overline{\mathcal{M}}_2$ be its Gromov-Hausdorff limit. If $X_i$ is not orientable for each $i$, then $X$ is not orientable.
\[papei\] Let $\{X_i\}_{i\in \mathbb{N}}$ be a sequence in $\overline{\mathcal{M}}_2$ and let $X\in\overline{\mathcal{M}}_2$ be its Gromov-Hausdorff limit. Then, the following two conditions are mutually equivalent.
- $X_i$ is orientable for sufficiently large $i$.
- $X$ is orientable.
By Corollary \[papei\], we have that if $X_1\in\overline{\mathcal{M}}_2$ is orientable and $X_2\in\overline{\mathcal{M}}_2$ is unorientable, then $X_1$ and $X_2$ belong to different connected components in $\overline{\mathcal{M}}_2$ with respect to the Gromov-Hausdorff topology.
Now, we generalize our main theorems to Ricci limit spaces. We get the following theorem by Main Theorem 1 and Theorem \[papee\].
\[papej\] For given integers $n\geq 4$ and $2\leq p \leq n/2$ and positive real numbers $K>n-p-1$ and $v>0$, there exists a constant $C(n,p)>0$ such that $$\lambda_1(X)\geq n-p-C(n,p)\lambda_1(\Delta_{C,p},X)^{1/2}$$ holds for all $X\in \overline{\mathcal{M}}_1(n,n-p-1,K,v)$.
We get the following theorem by Main Theorem 2, Main Theorem 4, Theorem \[papee\] and Corollary \[papei\].
\[papek\] For given integers $n\geq 5$ and $2\leq p < n/2$ and positive real numbers $\epsilon>0$, $K>n-p-1$ and $v>0$, there exists a constant $\delta=\delta(n,p,\epsilon)>0$ such that if $X\in \overline{\mathcal{M}}_1(n,n-p-1,K,v)$ satisfies one of
- $\lambda_{n-p+1}(X)\leq n-p+\delta$ and $\lambda_1(\Delta_{C,p},X)\leq \delta$,
- $X$ is orientable, $\lambda_{n-p}(X)\leq n-p+\delta$ and $\lambda_1(\Delta_{C,p},X)\leq \delta$,
- $\lambda_{n-p}(X)\leq n-p+\delta$ and $\lambda_1(\Delta_{C,n-p},X)\leq \delta$,
then there exists a geodesic space $Y$ such that $d_{GH}(X,S^{n-p}\times Y)\leq \epsilon$ holds. Moreover, if $(i)$ holds, then $X$ is orientable.
Eigenvalue Estimate for $L^2$ Almost Kähler Manifolds
=====================================================
In this section, we consider $L^2$ almost Kähler manifolds, i.e., we assume that there exists a $2$-form $\omega$ which satisfies that $\|\nabla \omega\|_2$ and $\|J_\omega^2+{\mathop{\mathrm{Id}}\nolimits}\|_1$ are small, where $J_\omega\in\Gamma(T^\ast M\otimes T M)$ is defined so that $\omega=g(J_\omega\cdot,\cdot)$. The main goal is to give the almost version of (\[kae\]).
Let $(M,g)$ be a Riemannian manifold. For each $2$-form $\omega\in \Gamma(\bigwedge^2 T^\ast M)$, let $J_\omega\in\Gamma(T^\ast M\otimes T M)$ denotes the anti-symmetric tensor that satisfies $\omega=g(J_\omega\cdot,\cdot)$.
We first show the following easy lemmas.
\[pB2\] Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold. If there exists a $2$-form $\omega$ such that $\|J_\omega^2+{\mathop{\mathrm{Id}}\nolimits}\|_1<1$ holds, then $n$ is an even integer.
There exists a point $x\in M$ such that $|J_\omega^2(x)+{\mathop{\mathrm{Id}}\nolimits}_{T_x M}|<1$. For any $v\in T_x M$ with $|v|=1$, we have $|J_\omega^2(x)(v)+v|<1$, and so $|J_\omega^2(x)(v)|>0$. Thus, $J_\omega(x)$ is non-degenerate. Therefore, $(T_x M,\omega_x)$ is a symplectic vector space. This implies the lemma.
\[pB3\] Given integers $n\geq 2$, $1\leq p\leq n-1$, and positive real numbers $K>0$, $D>0$, there exists $\delta_0(n,p,K,D)>0$ such that if $(M,g)$ is an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq-K g$ and ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$, then we have $$\lambda_{\alpha(n,p)+1}(\Delta_{C,p})\geq \delta_0(n,p,K,D),$$ where we defined $$\alpha(n,p):=\binom{n}{p}=\frac{n!}{p!(n-p)!}.$$
Put $\delta:=\lambda_{\alpha(n,p)+1}(\Delta_{C,p})$. If $\delta\geq 1$, we get the lemma. Thus, we assume that $\delta<1$. Let $\omega_i\in\Gamma(\bigwedge^p T^\ast M)$ denotes the $i$-th eigenform of the connection Laplacian $\Delta_{C,p}$ acting on $p$-forms with $\|\omega_i\|_2=1$.
We have $$\label{l2es}
\|\langle \omega_i,\omega_j\rangle\|_2^2\leq \frac{1}{\lambda_1(g)}\|\nabla\langle \omega_i,\omega_j\rangle\|_2^2\leq C(n,p,K,D)\delta$$ for each $i,j=1,\ldots, \alpha(n,p)+1$ with $i\neq j$ by the Li-Yau estimate [@SY p.116] and Lemma \[Linfes\]. By Lemma \[p4c\] and (\[l2es\]), we have $$\begin{aligned}
\|\langle \omega_i,\omega_j\rangle\|_1&\leq C(n,p,K,D)\delta^{1/2} \quad (i,j=1,\ldots, \alpha(n,p)+1 \text{ with } i\neq j),\\
\||\omega_i|^2-1\|_1&\leq C(n,p,K,D)\delta^{1/2} \quad (i=1,\ldots, \alpha(n,p)+1).\end{aligned}$$ Put $$\begin{aligned}
G:=\Big\{x\in M: & ||\omega_i|^2-1|(x)\leq\delta^{1/4}\text{ for all $i=1,\ldots, \alpha(n,p)+1$, and}\\
&\left|\langle \omega_i,\omega_j\rangle\right|(x)\leq \delta^{1/4}\text{ for all $i,j=1,\ldots, \alpha(n,p)+1$ with $i\neq j$}
\Big\}.\end{aligned}$$ Then, we have ${\mathop{\mathrm{Vol}}\nolimits}(M\backslash G)\leq C_1(n,p,K,D)\delta^{1/4}{\mathop{\mathrm{Vol}}\nolimits}(M)$ for some positive constant $C_1(n,p,K,D)$ depending only on $n,p,K$ and $D$ similarly to Lemma \[pfub\].
Let us show $\delta\geq \min\left\{1/C_1(n,p,K,D)^4,1/(\alpha(n,p)+1)^4\right\}$ by contradiction. Suppose that that $\delta< \min\left\{1/C_1(n,p,K,D)^4,1/(\alpha(n,p)+1)^4\right\}$. Then, we have $G\neq \emptyset$, and so we can take a point $x_0\in G$. We show that $\omega_{1}(x_0),\ldots, \omega_{\alpha(n,p)+1}(x_0)\in \bigwedge^p T_{x_0}^\ast M$ are linearly independent. Take arbitrary $a_1,\ldots, a_k\in \mathbb{R}$ with $a_1 \omega_1(x_0)+\cdots+a_k \omega_{\alpha(n,p)+1}(x_0)=0$. Take $i$ with $|a_i|=\max\{|a_1|,\ldots,|a_k|\}$. Since we have $\langle a_1 \omega_1(x_0)+\cdots+a_k \omega_{\alpha(n,p)+1}(x_0),\omega_i(x_0)\rangle=0$, we get $$\begin{split}
0\geq |a_i||\omega_i(x_0)|^2-\sum_{i\neq j}\left|a_j\langle \omega_i(x_0), \omega_j(x_0)\rangle\right|
\geq& |a_i|(1-\delta^{1/4})-\sum_{i\neq j}|a_i|\delta^{1/4}\\
\geq& |a_i|\left(1-(\alpha(n,p)+1)\delta^{1/4}\right).
\end{split}$$ Thus, $|a_i|=0$, and so $a_1=\cdots=a_k=0$. This implies the linearly independence of $\omega_{1}(x_0),\ldots, \omega_{\alpha(n,p)+1}(x_0)$. This contradicts to $\dim\left(\bigwedge^p T^\ast_{x_0} M\right)=\alpha(n,p)$. Thus, we get $\lambda_{\alpha(n,p)+1}(\Delta_{C,p})=\delta\geq \min\left\{1/C_1(n,p,K,D)^4,1/(\alpha(n,p)+1)^4\right\}$.
\[pB4\] Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold. Suppose that a $2$-form $\omega$ satisfies
- $\|\nabla \omega\|_2^2\leq \delta\|\omega\|_2^2$,
- $\|J_\omega^2+{\mathop{\mathrm{Id}}\nolimits}\|_1\leq \delta^{1/4}\|\omega\|_2^2$
for some $0<\delta \leq 1/4$. Let $\omega_\alpha$ be its image of the orthogonal projection $$P_{\delta}:L^2\left(\bigwedge^2 T^\ast M \right)\to \bigoplus_{\lambda_i(\Delta_{C,2})\leq \delta^{1/2}} \mathbb{R}\omega_i,$$ where $\omega_i$ denotes the $i$-th eigenform of the connection Laplacian $\Delta_{C,2}$ with $\|\omega_i\|_2=1$ $(\omega_\alpha :=P_{\delta} (\omega))$. Then, we have
- $\|\nabla \omega_\alpha\|_2^2\leq 2\delta\|\omega_\alpha\|_2^2$,
- $\|J_{\omega_\alpha}^2+{\mathop{\mathrm{Id}}\nolimits}\|_1\leq 10\delta^{1/4}\|\omega_\alpha\|_2^2$.
Put $\omega_\beta:=\omega-\omega_\alpha$. Then, we have $\|\omega\|^2_2=\|\omega_\alpha\|^2_2+\|\omega_\beta\|^2_2$. By the assumption (i), we have $$\begin{aligned}
\delta \|\omega\|_2^2
\geq \|\nabla\omega\|_2^2
=\|\nabla \omega_\alpha\|_2^2+\|\nabla \omega_\beta\|_2^2
\geq \|\nabla \omega_\alpha\|_2^2+\delta^{1/2}\|\omega_\beta\|_2^2.\end{aligned}$$ Thus, we get $$\begin{aligned}
\label{AB1}\|\nabla \omega_\alpha\|_2^2&\leq \delta \|\omega\|_2^2,\\
\label{AB2}\|\omega_\beta\|_2^2&\leq \delta^{1/2}\|\omega\|_2^2,\end{aligned}$$ and so $$\label{AB3}
\|\omega_\alpha\|_2^2=\|\omega\|^2_2 -\|\omega_\beta\|^2_2\geq (1-\delta^{1/2})\|\omega\|_2^2\geq \frac{1}{2}\|\omega\|_2^2.$$ By the definitions of the norms, we have $|J_\omega|^2=2|\omega|^2$ and $|J_{\omega_\alpha}|^2=2|\omega_\alpha|^2$. Since we have $$J_\omega^2-J_{\omega_\alpha}^2
=J_\omega(J_\omega-J_{\omega_\alpha})+(J_\omega-J_{\omega_\alpha})J_{\omega_\alpha},$$ we get $$|J_\omega^2-J_{\omega_\alpha}^2|
\leq 2(|\omega|+|\omega_\alpha|)|\omega_\beta|.$$ Therefore, we have $$\|J_\omega^2-J_{\omega_\alpha}^2\|_1
\leq 4\|\omega\|_2\|\omega_\beta\|_2\leq 4\delta^{1/4}\|\omega\|_2^2$$ by (\[AB2\]), and so $$\label{AB4}
\|J_{\omega_\alpha}^2+{\mathop{\mathrm{Id}}\nolimits}\|_1
\leq \|J_{\omega}^2+{\mathop{\mathrm{Id}}\nolimits}\|_1+\|J_\omega^2-J_{\omega_\alpha}^2\|_1\leq 5 \delta^{1/4}\|\omega\|_2^2\leq 10 \delta^{1/4}\|\omega_\alpha\|_2^2$$ by (\[AB3\]). By (\[AB1\]) and (\[AB4\]), we get the lemma.
Let us show the orientability for $L^2$ almost Kähler manifolds.
For any integer $n\geq 2$ and positive real numbers $K>0$, $D>0$, there exists a constant $\delta_1(n,K,D)>0$ such that the following property holds. Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq-K g$ and ${\mathop{\mathrm{diam}}\nolimits}(M)\leq D$. If there exists a $2$-form $\omega$ such that
- $\|\nabla \omega\|_2^2\leq \delta_1\|\omega\|_2^2$,
- $\|J_\omega^2+{\mathop{\mathrm{Id}}\nolimits}\|_1\leq \delta_1^{1/4}\|\omega\|_2^2$,
then $M$ is orientable.
By Lemma \[pB2\], we have that $n=2m$ is an even integer. We first assume that $\delta_1< \min\{1/4m^2,\delta_0(n,2,K,D)^2\}$. Since $J_{\omega}$ is anti-symmetric, we have $|J_{\omega}|^2\leq \sqrt{2m}|J_{\omega}^2|$. Thus, we get $$\sqrt{\frac{2}{m}}\|\omega\|_2^2=
\frac{1}{\sqrt{2m}}\|J_{\omega}\|_2^2\leq \sqrt{2m}+\delta_1^{1/4} \|\omega\|_2^2$$ by $|{\mathop{\mathrm{Id}}\nolimits}|=\sqrt{2m}$. This and $\delta_1^{1/4}\leq \frac{1}{2}\sqrt{\frac{2}{m}}$ imply that $\|\omega\|_2\leq \sqrt{2m}$. Put $\omega_\alpha:=P_{\delta_1}(\omega)$. Note that we have that $\|\omega_\alpha\|_2\leq\|\omega\|_2\leq \sqrt{2m}$ and that $\|\omega_\alpha\|_\infty\leq C(n,K,D)$ by Lemma \[Linfes\] and Lemma \[pB3\].
We first fix $x\in M$, and consider the $\mathbb{C}$-linear map $$J_{\omega_\alpha}(x)\colon T_x M\otimes_\mathbb{R} \mathbb{C}\to T_x M\otimes_\mathbb{R} \mathbb{C}.$$ Let us extend the Riemannian metric $\langle\cdot,\cdot\rangle
$ to $T_x M\otimes_\mathbb{R} \mathbb{C}$ so that $$\langle u_1 + i v_1, u_2+iv_2 \rangle=(\langle u_1,u_2\rangle+\langle v_1,v_2\rangle)+i(\langle v_1, u_2\rangle-\langle u_1, v_2\rangle)$$ for all $u_1,u_2,v_1,v_2\in T_x M$. Since $J_{\omega_\alpha}(x)$ is anti-symmetric, there exist eigenvalues $\{\lambda_1,\overline{\lambda_1},\ldots,\lambda_m,\overline{\lambda_m}\}$ of $J_{\omega_\alpha}(x)$ and an orthogonal basis $\{E_1,\overline{E_1},\ldots,E_m,\overline{E_m}\}$ of $T_{x} M\otimes_\mathbb{R} \mathbb{C}$ such that $J_{\omega_\alpha}(x) E_i=\lambda_i E_i$, where the overline denotes the complex conjugate. Note that each $\lambda_i$ is a pure imaginary number. Let $\{E^1,\overline{E^1},\ldots,E^m,\overline{E^m}\}\subset T_x^\ast M\otimes_{\mathbb{R}} \mathbb{C}\cong(T_x M\otimes_{\mathbb{R}} \mathbb{C} )^\ast$ be the dual basis of $\{E_1,\overline{E_1},\ldots,E_m,\overline{E_m}\}$. If we extend $\omega_\alpha(x)$ to a complex bilinear form, then we have $$\omega_\alpha(x)=\sum_{i=1}^m \lambda_i E^i\wedge \overline{E^i}.$$ Thus, we get $$\omega_{\alpha}^m(x)=m! \lambda_1\cdots \lambda_m E^1\wedge\overline{E^1}\wedge E^m\wedge\overline{E^m},$$ and so $$|\omega_{\alpha}^m(x)|=m!|\lambda_1|\cdots |\lambda_m|.$$ Since we have $$|\lambda_i|^2=|J_{\omega_\alpha}^2 E_i|=|(J_{\omega_\alpha}^2+{\mathop{\mathrm{Id}}\nolimits})E_i-E_i|,$$ we get $$\left|1-|\lambda_i|^2\right|\leq|J_{\omega_\alpha}^2+{\mathop{\mathrm{Id}}\nolimits}|(x)$$ and $$|\lambda_i|\leq |J_{\omega_\alpha}|^2=2|\omega_\alpha|^2\leq C(n,K,D).$$ Therefore, we get $$\left||\omega_\alpha^m|^2-(m!)^2\right|\leq C|J_{\omega_\alpha}^2+{\mathop{\mathrm{Id}}\nolimits}|,$$ and so $$\left|\|\omega_\alpha^m\|_2^2-(m!)^2\right|\leq C\delta_1^{1/4}$$ by Lemma \[pB4\]. Since we have $$\|\nabla (\omega_\alpha^m)\|_2^2\leq \delta_1$$ by Lemma \[pB4\], we get the proposition taking $\delta_1$ sufficiently small by Corollary \[papeb\] (ii).
The following theorem is the goal of this section.
For any integer $n\geq 2$, there exists a constant $C(n)>0$ such that the following property holds. Let $(M,g)$ be an $n$-dimensional closed Riemannian manifold with ${\mathop{\mathrm{Ric}}\nolimits}\geq (n-1) g$. If there exists a $2$-form $\omega$ such that
- $\|\nabla \omega\|_2^2\leq \delta\|\omega\|_2^2$,
- $\|J_\omega^2+{\mathop{\mathrm{Id}}\nolimits}\|_1\leq \delta^{1/4}\|\omega\|_2^2$,
for some $\delta>0$, then we have $$\lambda_1(g)\geq 2(n-1)- C(n)\delta^{1/2}.$$
It is enough to prove the theorem when $\delta$ is small. Thus, we can assume that $n=2m$ is an even integer by Lemma \[pB2\]. If $n=2$, then $\lambda_1(g)\geq 2(n-1)$ is the original Lichnerowicz estimate. If $n=4$, the conclusion of the theorem can also be deduced from Main Theorem 1.
We first assume that $\delta< \min\{1/4m^2,\delta_0(n,2,K,D)^2\}$. Put $\omega_\alpha:=P_{\delta}(\omega)=\sum_{i=1}^k a_i \omega_i$. Here, $\omega_i$ is the $i$-th eigenform of the connection Laplacian $\Delta_{C,2}$ with $\|\omega_i\|_2=1$ corresponding to the eigenvalue $\lambda_i(\Delta_{C,2})\leq \delta^{1/2}$ for each $i=1,\ldots, k$. We have $k\leq \alpha(n,2)$ by Lemma \[pB3\], and $\|\omega_\alpha\|_\infty\leq C$ by Lemma \[Linfes\].
Let $f\in C^\infty(M)$ be the first eigenfunction of the Laplacian with $\|f\|_2=1$. If $\lambda_1(g)\geq 2(n-1)+1$, we get the theorem. Thus, we assume that $\lambda_1(g)\leq 2(n-1)+1$. Then, we have $\|f\|_\infty\leq C$ and $\|\nabla f\|_\infty \leq C$ by Lemma \[Linfes\]. By Lemma \[p4d\] (i) and (iii), we have $$\label{AB5}
\begin{split}
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\left|\int_M \langle\Delta (\iota(\nabla f)\omega_\alpha),\iota(\nabla f)\omega_\alpha\rangle-\lambda_1(g)\langle\iota(\nabla f)\omega_\alpha,\iota(\nabla f)\omega_\alpha\rangle\,d\mu_g
\right|\\
\leq &C\delta^{1/2}\|\omega_\alpha\|_2^2
\end{split}$$ and $$\label{AB6}
\|d^\ast (\iota(\nabla f)\omega_\alpha)\|_2^2\leq C\delta\|\omega_\alpha\|_2^2.$$ By (\[2b\]), (\[AB5\]), (\[AB6\]) and the Bochner formula, we get $$\label{AB7}
\begin{split}
&\frac{n-1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M
\langle J_{\omega_\alpha}\nabla f,J_{\omega_\alpha}\nabla f\rangle\,d\mu_g\\
\leq&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M{\mathop{\mathrm{Ric}}\nolimits}(J_{\omega_\alpha}\nabla f,J_{\omega_\alpha}\nabla f) \,d\mu_g\\
=&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle\Delta (\iota(\nabla f)\omega_\alpha),\iota(\nabla f)\omega_\alpha\rangle\,d\mu_g-\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M |\nabla (\iota(\nabla f)\omega_\alpha)|^2\,d\mu_g\\
\leq& \frac{\lambda_1(g)}{2{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle\iota(\nabla f)\omega_\alpha,\iota(\nabla f)\omega_\alpha\rangle\,d\mu_g+C\delta^{1/2}\\
=&\frac{\lambda_1(g)}{2{\mathop{\mathrm{Vol}}\nolimits}(M)}\int_M \langle J_{\omega_\alpha}\nabla f,J_{\omega_\alpha}\nabla f \rangle\,d\mu_g+C\delta^{1/2}.
\end{split}$$
Since $J_{\omega_\alpha}$ is anti-symmetric, we have $$\begin{split}
&\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\left|\int_M \langle J_{\omega_\alpha}\nabla f,J_{\omega_\alpha}\nabla f\rangle\,d\mu_g- \int_M \langle \nabla f,\nabla f\rangle\,d\mu_g\right|\\
\leq&
\frac{1}{{\mathop{\mathrm{Vol}}\nolimits}(M)}\left|\int_M \langle (J_{\omega_\alpha}^2+{\mathop{\mathrm{Id}}\nolimits})\nabla f,\nabla f\rangle\,d\mu_g\right|\\
\leq &C\|J_{\omega_\alpha}^2+{\mathop{\mathrm{Id}}\nolimits}\|_1\leq C\delta^{1/4}
\end{split}$$ by Lemma \[pB4\]. Thus, taking $\delta$ sufficiently small, we get $$\label{AB8}
\|J_{\omega_\alpha}\nabla f\|_2^2
\geq \|\nabla f\|_2^2-C\delta^{1/4}= \lambda_1(g) -C\delta^{1/4}\geq n-C\delta^{1/4}\geq \frac{n}{2}$$ by the Lichnerowicz estimate. By (\[AB7\]) and (\[AB8\]), we get the theorem.
[99]{} M. Aino, [Riemannian invariants that characterize rotational symmetries of the standard sphere,]{} Manuscripta Math. 156 (2018), 241–272. M. Aino, [Sphere theorems and eigenvalue pinching without positive Ricci curvature assumption,]{} Calc. Var. Partial Differential Equations, to appear. D. V. Alekseevsky, S. Marchiafava, [Transformations of a quaternionic Kähler manifold,]{} C. R. Acad. Sci. Paris Sér. I Math. 320 (1995), no. 6, 703–708. E. Aubry, [Pincement sur le spectre et le volume en courbure de Ricci positive,]{} Ann. Sci. École Norm. Sup. (4) 38 (2005), 387–405. F. Belgun, A. Moroianu, U Semmelmann, [Killing forms on symmetric spaces,]{} Differential Geom. Appl. 24 (2006), no. 3, 215–222. A. Besse, [Einstein manifolds,]{} Ergebnisse der Mathematik und ihrer Grenzgebiete, Springer-Verlag, Berlin (1987), xii+510 pp. J. Cheeger, [Differentiability of Lipschitz functions on metric measure spaces,]{} Geom. Funct. Anal. 9 (1999), 428–517. J. Cheeger, [Degeneration of Riemannian metrics under Ricci curvature bounds,]{} Lezioni Fermiane, Scuola Normale Superiore, Pisa, (2001). J. Cheeger, T. Colding, [Lower bounds on Ricci curvature and the almost rigidity of warped products,]{} Ann. of Math. (2) 144 (1996), 189–237. J. Cheeger, T. Colding, [On the structure of spaces with Ricci curvature bounded below. I,]{} J. Differential Geom. 46 (1997), 406–480. J. Cheeger, R. Colding, [On the structure of spaces with Ricci curvature bounded below. III,]{} J. Differential Geom. 54 (2000), no. 1, 37–74. T. Colding, [Shape of manifolds with positive Ricci curvature,]{} Invent. Math. 124 (1996), 175–191. N. Gigli, [Nonsmooth differential geometry–an approach tailored for spaces with Ricci curvature bounded from below,]{} Mem. Amer. Math. Soc. 251 (2018), no. 1196, v+161 pp. J. F. Grosjean, [A new Lichnerowicz-Obata estimate in the presence of a parallel p-form,]{} Manuscripta Math. 107 (2002), no. 4, 503–520. S. Honda, [Ricci curvature and almost spherical multi-suspension,]{} Tohoku Math. J. (2) 61 (2009), 499–522. S. Honda, [A weakly second-order differential structure on rectifiable metric measure spaces,]{} Geom. Topol. 18 (2014), 633–668. S. Honda, [Ricci curvature and $L^p$-convergence,]{} J. Reine Angew. Math. 705 (2015), 85–154. S. Honda, [Ricci curvature and orientability,]{} Calc. Var. Partial Differential Equations 56 (2017), no. 6, Art. 174, 47 pp. S. Honda, [Spectral convergence under bounded Ricci curvature,]{} J. Funct. Anal. 273 (2017), 1577–1662. S. Honda, [Elliptic PDEs on compact Ricci limit spaces and applications,]{} Mem. Amer. Math. Soc. 253 (2018), v+92 pp. P. Petersen, [On eigenvalue pinching in positive Ricci curvature,]{} Invent. Math. 138 (1999), 1–21. P. Petersen, [Riemannian geometry. Third edition,]{} Springer, Cham, 2016. xviii+499 pp. T. Sakai, [Riemannian geometry,]{} Translated from the 1992 Japanese original by the author. Translations of Mathematical Monographs, 149. American Mathematical Society, Providence, RI, (1996), xiv+358 pp. R. Schoen, S.T. Yau, [Lectures on differential geometry,]{} Conference Proceedings and Lecture Notes in Geometry and Topology, I. International Press, Cambridge, MA, (1994) v+235 pp. U. Semmelmann, [Conformal Killing forms on Riemannian manifolds,]{} Math. Z. 245 (2003), no. 3, 503–-527. Y. Tashiro, [Complete Riemannian manifolds and some vector fields,]{} Trans. Amer. Math. Soc. 117, (1965), 251–275.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We generalize the maximum likelihood method to non-Gaussian distribution functions by means of the multivariate Edgeworth expansion. We stress the potential interest of this technique in all those cosmological problems in which the determination of a non-Gaussian signature is relevant, e.g. in the analysis of large scale structure and cosmic microwave background. A first important result is that the asymptotic confidence limits on the parameters are systematically widened when the higher order correlation functions are positive, with respect to the Gaussian case.'
---
8.8in 6.1in -.45in
\#1[$^{\ref{#1}}$]{}
========
**Non-Gaussian Likelihood Function**
**Luca Amendola**
Osservatorio Astronomico di Roma [^1]
Viale del Parco Mellini, 84, Rome 00136 - Italy
: cosmology: large-scale structure of the Universe; galaxies: clustering; methods: statistics
Introduction
============
\[sec:intro\] Modern large scale astronomy is, to a large extent, the science of non-Gaussian random fields. One of the keys to understanding the formation and evolution of structure in the Universe resides infact in the statistical properties of the matter field. Rival theories of structure formation predicts different statistical features, both in the present Universe and in the primordial fluctuations encoded in the microwave background. To the scope of quantifying the statistical feature of the matter clustering, several techniques have been proposed. One of these, which is increasingly popular in astrophysics, is the estimation of parameters via the maximum likelihood method. For instance, the maximum likelihood method is currently widely employed in the analysis of cosmic microwave background (CMB) experiments, large scale surveys and cosmic velocity fields. Once a suitable likelihood function (LF) has been constructed, one estimates the best parameters simply by finding the maximum of the LF with respect to those parameters. The parameter estimates, say $\hat\alpha_i(x_i)$ (the hat is to distinguish between the estimates and the theoretically expected parameters), are then functions of the data $x_i$, i.e. of the random variables, and are therefore random variables themselves. If one is able to determine the distribution function $P(\hat\alpha_i)$ of the estimators, the confidence region (CR) of the parameters can be found as the value of the parameters for which the integral of $P(\hat\alpha_i)$ falls below a predetermined level.
There are however two problems with this approach. One is that usually we don’t know how the data are distributed, and the usual Gaussian approximation may be very poor. This is the case, for instance, in large scale structure, where we already know that the density fluctuations are not Gaussian, even on fairly large scales. The second problem is that, even if we know perfectly well the data distribution, is often not trivial to find an analytical expression for the distribution $P(\hat\alpha_i)$ of the parameter estimators $\hat\alpha_i$. Aside the simplest case in which one only needs the raw sample variance or the sample mean of normal variates (or closely related quantities), one has invariably to resort to very time-consuming MonteCarlo methods. While this second problem can be always overcome by numerical methods, the first difficulty remains, unless one takes into consideration specifically designed non-Gaussian models, and for each of these determines the confidence regions for the relevant parameters. Other than being too model-dependent, in the current astrophysical applications this procedure is in many cases prohibitively slow. It is then of interest to examine alternatives able to retain the useful features of the likelihood method while allowing more freedom in exploring different non-Gaussian (non-G, for shortness) distributions.
In this work we propose a perturbative method to estimate theoretical parameters when the higher-order multivariate moments (or $n$-points correlation functions) are non-vanishing, via an expansion around a Gaussian LF, the multivariate Edgeworth expansion (MEE). As long as the perturbative approach does not break down, i.e. as long as the departure from Gaussianity is mild, the MEE gives an answer to the first problem, the distribution function of the data, because it allows arbitrary values of the higher-order correlation functions. Then we still are left with the second problem: how do we determine in the general case the distribution function for our parameter estimators, necessary to produce the confidence regions? A first simple possibility is to approximate the $P(\hat\alpha_i)$ around its peak, previously determined by maximization of the LF, by a Gaussian distribution, multivariate in the parameter space. This allows to determine an approximate covariance matrix, whose eigenvalues give the principal axis of the parameter CR (see e.g. Kendall, Stuart & Ord 1987). Notice that this is equivalent to assume the [*parameter estimators*]{}, which are functions of all the dataset, as Gaussian distributed, but makes no assumptions on the [*data*]{} themselves. When the number of data is large, this procedure can be justified by the central limit theorem (which however cannot guarantee the asymptotic Gaussianity in the general case). This first possibility, along with its limitations, is discussed in Sect. 3. To overcome the limits relative to the Gaussian approximation of the estimator distribution, and exploiting the analytic properties of the MEE, we adopt in Sect. 4 and Sect. 5 a second, exact, way to determine the CR for some of the relevant parameters. This is a non-Gaussian generalization of the $\chi^2$ technique: instead of finding the CR by integration over the unknown distribution function of the sample estimators, we determine the CR by integrating the LF over the possible outcomes of an experiment. As in the usual $\chi^2$ method, the acceptable values of a parameter will be all those for which the data lie not too distant from the predicted values, the “distance” being measured by the quantity $\chi^2_0=x_i\lam^{ij} x_j$, where $x_i$ are the actual data and $\lam^{ij}$ is the inverse of the correlation matrix. The CR will in general depend on all the higher-order correlation functions included in the MEE, as it will be shown in Sect. 4. The formalism is then best suited to answer the question: how our results (i.e., best estimates and CR) change when the higher-order moments of the data distribution are not set to zero? If we have any reason to believe in some particular values for the non-G moments, then we can plot the CR for our parameters given those higher-order moments, and clearly the regions will be different for any set of higher-order moments. The relevance of examining how the confidence regions vary with respect to the non-G parameters is clear. Suppose in fact that two experiments, assuming Gaussianity, produce two non-overlapping CR for the same parameter, say the overall normalization of the correlation function. In general, the CR will be different in the non-Gaussian case, and it may happen that the two experiments are infact compatible when some level of non-Gaussianity is assumed. As we will show, in most cases the CR widens for positive higher-order moments, so that two non-overlapping results can be brought to agreement, [*provided some amount of non-Gaussianity is allowed*]{}. Other positive features of our formalism are that it exploits the full set of data, that it can be extended to higher and higher order moments, and that it is fully analytical.
To the order to which we limit ourselves here, we will be able to estimate the first non-G correlation function, i.e. the third-order moments. This estimate will share the good and less good properties of likelihood estimators: they are consistent estimators, but only asymptotically (i.e., for large samples) unbiassed, as we will show in Sect. 3. For the fourth-order cumulant there is not an estimator at all, since the LF is linear in it. We decided to keep track of it anyway, because it is still interesting to use the fourth-order cumulant as an external parameter, and see how our results change with different assumptions on it.
In principle one can include in the analysis all the set of higher-order moments considered relevant to the problem, but here we will limit ourselves to the first two higher-order terms, the 3- and 4-points correlation functions. For most purposes, this is the best we can do for comparing different models with observations, since current data do not permit accurate analysis of correlation functions of order higher than the fourth one.
Let us remark that we call here likelihood function the probability distribution function $f(x_i,\alpha_i)$ of the data $x_i$ (our random variables) defined in a sample space $S$, given some theoretical parameter $\alpha_i$, which can be thought to lie in the parameter space $P$. Essentially, for any point in the parameter space, i.e. for any distribution function, we will integrate the LF over the sample space $S$, i.e. over all the possible outcomes of the experiment, to determine how likely or unlikely is the possibility that the actual data set has arisen from such a parametric choice. For a discussion of the advantage and disadvantage of this approach with respect to the alternative Bayesan one, in which the integration occurs over the space of the [*theoretical*]{} parameters (as opposed to the [*sample estimators*]{} of the theoretical parameters considered in the frequentist approach), we refer to standard textbooks like Kendall, Stuart & Ord (1987).
Beside presenting the basic formalism of the non-Gaussian LF, we discuss briefly in Sect. 6 its application to large scale structure and to the CMB. In the first case the non-Gaussian nature of the galaxy distribution is a well-established fact, so that the use of a non-G LF is certainly required. In the case of CMB, the current set of data is still not accurate enough to assess the issue. The estimate of a confidence region in non-G models is however crucial in view of the discrimination among different theories of structure formation.
Formalism
=========
Let $d^i$ be a set of experimental data, $i=1,..N$, and let us form the variables $x^i=d^i-t^i$, where $t^i$ are the theoretical expected values for the measured quantities. To fix the ideas, one can think of $d^i$ as the temperature fluctuation in the $i$-th pixel in a CMB experiment, or as the number of galaxies in a given volume of the Universe. Let $c^{ij}$ be the correlation matrix \[cm\] c\^[ij]{}=<x\^ix\^j>, and let us introduce the higher-order cumulant matrices (or $n$-point correlation functions) k\^[ijk]{}&=&<x\^i x\^j x\^k>,\
k\^[ijkl]{}&=&<x\^i x\^j x\^k x\^l>-c\^[ij]{}c\^[kl]{}-c\^[ik]{}c\^[jl]{}-c\^[il]{}c\^[jk]{} (we will sometimes use the words “skewness” and “kurtosis” to refer to the 3- and 4-point correlation functions, respectively, or to their overall amplitude; in the statistical literature, the definition of skewness is actually, in our notation, $\gam_1=k^{iii}/(c^{ii})^{3/2}$, and for the kurtosis $\gam_2=k^{iiii}/(c^{ii})^2$). The correlation matrices depend in general both on a number of theoretical parameters $\alpha_j$, $j=1,..P$ (that we leave for the moment unspecified) and on the experimental errors. In most cases, we can assume the experimental errors to be Gaussian distributed (or even uncorrelated) so that they can be completely characterized by the correlation matrix $e^{ij}$, which is simply to be added in quadrature to the 2-point correlation function. It is useful to define then the matrix \_[ij]{}=(c\^[ij]{}+e\^[ij]{})\^[-1]{}. The problem of estimating the parameters $\alpha_j$ is solved by maximizing, with respect to the parameters, the likelihood function L=f(), where $f(\vx)$ is the multivariate probability distribution function (PDF) for the random variables $x_i$. Clearly, knowing the LF one can, at least in principle, determine also the parameter CR, as will be discussed in the next sections. The main difficulty to this approach, however, is that we do not know, in general, the exact form for the PDF $f(\vx)$. The usual simplifying assumption is then that $f(\vx)$ is a multivariate Gaussian distribution L\_g=f()=G(,)(2)\^[-N/2]{} ||\^[1/2]{} (-x\^i \_[ij]{} x\^j). where $|\lam|={\rm det}(\lam_{ij})$. This is usually assumed, for instance, in analysing the CMB fluctuation maps and the cosmic velocity fields. A straightforward way to generalize the LF so as to include the higher-order correlation functions, which embody the non-Gaussian properties of the data, is provided by the multivariate Edgeworth expansion (MEE). An unknown PDF $f(\vx)$ can indeed be expanded around a multivariate Gaussian $G(x,\lam)$ according to the formula (Chambers 1967; McCullagh 1984; Kendall, Stuart & Ord 1987) \[mee\] f()=G(,)\[1+[16]{}k\^[ijk]{}h\_[ijk]{}(,) +[124 ]{}k\^[ijkl]{}h\_[ijkl]{}(,) +[172 ]{}k\^[ijk]{}k\^[lmn]{}h\_[i..n]{}(,)+...\], where $h_{ij..}$ are Hermite tensors, the multivariate generalizations of the Hermite polynomial. If there are $r$ subscripts, the Hermite tensor $h_{ij..}$ is said to be of order $r$, and is given by \[defpol\] h\_[ij...]{}=(-1)\^r G\^[-1]{}(,) \_[ij...]{} G(,), where $\partial_{ij...}=(\partial/\partial x_i)(\partial/\partial x_j)...$. The Hermite polynomials are located on the main diagonal of the Hermite tensors, when $\lam_{ij}=\delta_{ij}$. Notice that the function $f(\vx)$ is normalized to unity, since the integrals of all the higher order terms from minus to plus infinity vanish. It can be shown that the MEE gives a good approximation to any distribution function provided that all the moments are defined and that the higher order correlation functions do not dominate over the Gaussian term. In other words, the MEE can be applied only in the limit of mild non-Gaussianity. More accurately, the approximation is good, in the sense that the error one makes in the truncation is smaller than the terms included, if the cumulants obey the same order-of-magnitude scaling of a standardized mean (Chambers 1967). This condition is satisfied, for instance, by the cumulants of the galaxy clustering in the scaling regime, which explains why the (univariate) Edgeworth expansion well approximates the probability distribution of the large scale density field (Juszkiewicz 1994, Kofman & Bernardeau 1994). The same expansion has been also applied to the statistics of pencil-beam surveys, in which the one-dimensional power spectrum coefficients can be written as a genuine standardized mean (Amendola 1994). Finally, it has also been used to go beyond the Gaussian approximation in calculating the topological genus of weakly non-Gaussian fields (Matsubara 1994). Let us also note that the MEE lends itself to a further generalization: if the experimental errors are [*not*]{} Gaussian distributed, then the expansion for the data given the error correlation functions $e^{ij..}$ is the same as in Eq. (\[mee\]), but with the new cumulants $K^{ij..}=k^{ij..}+e^{ij..}$. In fact, let $x_t^i$ be the theoretical values whose measure is given by the data $d^i$, and let $\xi^i=d^i-x_t^i$ be the experimental error. The theoretical values are random variables in the sense that the theory usually predicts only their distribution, not their definite value. For instance, once the monopole is subtracted, the standard cosmological models predict CMB fluctuations Gaussian distributed with zero mean, $t^i=0$. Then we are concerned with the distribution of $x^i=
x_t^i+(d^i-x_t^i)=x_t^i+\xi^i$, the sum of the theoretical values $x_t^i$ and of the experimental errors $\xi^i$, both of which are random variables. If the two are independent, the general theorems on the random variables ensure that [*the cumulants cumulate*]{}, i.e. that the cumulants of $x^i$ are the sum of the ones of $x_t^i$ and of $\xi^i$.
Two properties are of great help in dealing with the MEE. The first is that $k^{ijk...}$ and $h_{ijk...}$ are contra- and co-variant tensors, respectively, with respect to linear transformations of the variables $x_i$. It follows then that $f(\vx)d\vx$ is totally invariant with respect to the linear transformations which leave invariant the quadratic form $\chi^2=x^i\lam_{ij} x^j$. This property is very useful, because we can always diagonalize the quadratic form by choosing a linear combination $y^j=A^j_i x^i$ such that $\chi^2=x^i\lam_{ij} x^j= y^i \delta_{ij} y^j$. The MEE in the new variables $y^i$ remains formally the same as in Eq. (\[mee\]), with $x\to y$ and $\lam\to \delta$, but now $G(\vy,\delta)$ factorizes, and all the calculations are simplified. Notice that even if the new variables are uncorrelated, they are not statistically independent, since they are not (in general) Gaussian variates. The higher-order matrices are then not diagonalized. In the following we will often assume that the variable transformation has been already performed, so that we will write $y$ and $\delta$ instead of $x$ and $\lam$, leaving all the other symbols unchanged. The second useful property is that the MEE is [*analytically integrable*]{} if the integration region is bounded by $\chi^2=const$. This property will be exploited in Sect. 4.
Best estimates and asymptotic confidence regions
================================================
The likelihood estimates for the parameters are to be obtained by maximizing Eq. (\[mee\]) with respect to the parameters. To illustrate some interesting points, let us put ourselves in the simplest case, in which all data are independent, and we only need to estimate the parameters $\sigma$ and $k_3$ entering the 2- and 3-point correlation function as overall amplitudes: \[simple\] c\_[ij]{}=\^2 \_[ij]{},k\_[ijk]{}=k\_3 \_[ij]{}\_[jk]{}. Because we are in such a simplified case, we will recover several well-known formulae of sampling statistics, like the variance of the standard deviation $\sigma$ and of $k_3$. It is important to bear in mind, however, that the MEE is much more general than we are assuming in this section, since it can allow for full correlations among data, for experimental errors, and for non-linear parametric dependence.
For simplicity, we also assume that the sample kurtosis is negligible. Because of this, we can put the fourth order sample cumulant of the dataset to zero (see, e.g., Kendall, Stuart & Ord 1987): k\_4=[N(N-1)(N-2)(N-3)]{}\[(N+1)\_i x\_i\^4- 3(N-1)(\_i x\_i\^2)\^2\]=0, so that we have, for large $N$, \[need\] \_i x\_i\^4= 3(\_i x\_i\^2)\^2. We show here that the maximum likelihood estimators for the variance and for the skewness in the case of independent data and for $N\to \infty$ reduce to the usual sample quantities \[mlvar\] \^2&=&\_i x\_i\^2/(N-1),\
\[sample\] k\_3&=&[N(N-1)(N-2)]{}\_i x\_i\^3. We will assume also that the average has been subtracted from the data, i.e. that $\sum_i x_i=0$. This actually reduces the degrees of freedom, but in the limit of large $N$ we can safely ignore this problem. If the distribution function of $x_i$ is approximated in the limit of small $k_3$ by the univariate Edgeworth expansion \[uniee\] f\_i=G(x\_i,)\[1+k\_3 h\_[3i]{}/6+k\_3\^2 h\_[6i]{}/72\], where $G(x_i,\sigma)$ is a Gaussian function, then the multivariate distribution function for the dataset is L()=\_i f\_i. By the definition in (\[defpol\]) we have \[h3h6\] h\_[6i]{}&=&\^[-12]{}\[x\_i\^6-15 \^2 x\_i\^4+45 \^4 x\_i\^2 -15\^6\],\
h\_[3i]{}&=&\^[-6]{}\[x\_i\^3-3\^2 x\_i\]. Let us pause to evaluate the order-of-magnitude of the non-G corrections in the univariate Edgeworth expansion (\[uniee\]). Assuming $x_i\sim \sigma$, the first correction term is of the order of $ \gam_1\equiv k_3/\sigma^3$, which is the dimensionless definition of skewness. The general rough requirement for the truncated Edgeworth expansion is then that $\gam_1\ll 1$. This condition will be encountered several times throughout this work. The maximum likelihood estimators for $\sigma$ and $k_3$ are then the values $\hat\sigma,\hat k_3$ which maximize $L$, or, equivalently, its logarithm $\log L$. We have then the equations \[bestsig\] [dLd]{}=\_i [df\_id]{}= \_i{ [1\^[3]{}]{} \[x\_i\^2-\^[2]{}\]+[k\_3\^[7]{}]{} \[2 \^2 x\_i-x\_i\^3\] +[k\_3\^24 \^[11]{}]{}\[5\^4-16\^2 x\_i\^2+5x\_i\^4\] }=0, and \[bests3\] [dLdk\_3]{}=\_i [df\_idk\_3]{}= \_i{ /}=0. To first order in $k_3$, the latter equation gives \[der\] \_i =0, so that our estimator is \[k3estim\] k\_3=-[6\_i h\_[3i]{}\_i \[ h\_[6i]{}- h\_[3i]{}\^2\]]{}. Suppose now that the solution for $\sigma$ of Eqs. (\[bestsig\]) and (\[bests3\]) is the usual variance estimator (\[mlvar\]), with $N\approx N-1$. Then we can observe that, from (\[h3h6\]), \[mlh6\] h\_[6i]{}=h\_[3i]{}\^2- 3\^[-10]{}(3x\_i\^4-12\^2x\_i\^2 +5\^4)=h\_[3i]{}\^2-6N\^[-6]{}, where in the last step we used Eq. (\[need\]) and Eq. (\[mlvar\]). Inserting (\[mlh6\]) in (\[k3estim\]) we obtain finally (assuming that $\sum_i x_i=0$) \[k3est\] k\_3=[\_i h\_[3i]{}N\^[-6]{}]{}= [\_i x\_i\^3N]{}, which coincides with (\[sample\]) for large $N$. Finally, going back to Eq. (\[bestsig\]), and inserting $k_3=\hat k_3$ we recover the sample variance $\hat\sigma^2=\sum_i x_i^2/N$, so that our proof is complete. If needed, the small bias introduced by a finite $N$ can be easily removed just multiplying $\hat k_3,\hat \sigma$ derived from the likelihood method by suitable functions of $N$.
The same calculation can be carried out in the more general case of dipendent variables, but the search for the maximum is more simply performed numerically when the situation is more complicated (e.g., because of the presence of experimental errors, or of more parameters, or more complicate parameter dependence). We just quote the result when only an overall skewness parameter is required, as when the 3-point correlation function is given by $k_{ijk}=k_3 s_{ijk}$, and the tensor $s_{ijk}$ is known (see Section 5). The best estimate for $k_3$ is then k\_3=-6 [s\^[ijk]{}h\_[ijk]{}s\^[ijk]{}s\^[lmn]{}h\_[ijklmn]{}]{}, which reduces to the expression above when $s_{ijk}=\delta_{ij}
\delta_{jk}$, using the relation \[sums\] \_[i,j]{} h\_[iiijjj]{}=h\_[6i]{}+(\_i h\_[3i]{} )\^2- \_i h\_[3i]{}\^2, and observing that $(\sum h_{3i})^2$ is of order $\hat k_3^2$, and thus negligible.
Once we have the best estimators $\hat \alpha_i(\vx)$ of our parameters, we need to estimate the confidence regions for that paramaters, i.e. the range of values in which we expect to find our estimators to a certain probability, given that the data distribution is approximated by the MEE. The problem consists in determining the behavior of the unknown distribution $P[\hat\alpha_i(\vx)]$, when we know the distribution for the random variables $x_i$. This problem is generally unsoluble analitycally, and the common approach is to resort to MonteCarlo simulations of the data. However, we can always approximate $P(\hat\alpha_i)$ [*around its peak*]{} by a Gaussian distribution multivariate [*in the parameter space*]{}; if the number of data $N\to\infty$, this procedure can be justified by the central limit theorem. For instance, if $\hat k_3=\sum x_i^3/N$, then its distribution will tend to a Gaussian whatever the distribution of the data $x_i$ is, in the limit of large $N$. In more general cases (e.g. correlated data) the central limit theorem does not guarantee the asymptotic Gaussianity; we can expect however it to be a first reasonable approximation far from the tails. If this approximation is adopted, then it can be shown (see e.g. Kendall, Stuart & Ord 1987) that the covariance matrix of the parameters can be written as \_[ab]{}\^[-1]{}=-[L(,\_a)\_a\_b]{}|\_[\_a=\_a]{}, where $a,b$ run over the dimensionality $P$ of the parameter space. The 1$\sigma$ confidence region is then enclosed inside the $P$-dimensional ellipses with principal axis equal to $\lam_a^{1/2}$, where $\lam_a$ are the eigenvalues of $\Sigma_{ab}$. Let us illustrate this in the same simplified case as above: $N$ independent data characterized by variance $\alpha_1=
\sigma$ and skewness $\alpha_2=k_3$. To further simplify, we assume that the mixed components $\Sigma_{12}=\Sigma_{21}$ can be neglected (see below). The component $\Sigma_{22}$ is then easily calculated as \_[22]{}\^[-1]{}=-\_[i,j]{}h\_[iiijjj]{}/36, Thus, using Eq. (\[sums\]), the variance of $\hat k_3$ turns out to be (dropping the hats here and below) \_[22]{}= 6\^6/N, which, not unexpectedly, is the sample skewness variance, i.e. the scatter in the skewness of Gaussian samples (for the dimensionless skewness defined as $\gam_1=k_3/\sigma^3$ the variance is $6/N$). In other words, to this order of approximation, the variance in the sample skewness in non-G data equals the variance in the sample skewness of Gaussian distributed data. The generalization to dependent data is \[vark3\] \_[22]{}\^[-1]{}=-s\^[ijk]{}s\^[lmn]{}h\_[ijklmn]{}/36 which gives then the variance of the estimator $\hat k_3$ in the general case.
More interesting is the error in the variance parameter $\sigma$ when not only a non-zero skewness $k_3$ is present, but also a non-zero kurtosis parameter $k_4$, defined in a way similar to $k_3$ as $k_{ijkl}=k_4 s_{ijkl}$. Then the result turns out to be, in the same approximations as above, \[varvar\] \_[11]{}=[\^22N]{}, where $\gam_2=k_4/ \sigma^4$ is the dimensionless kurtosis. The mixed components amount to $\Sigma_{12}^{-1}=\Sigma_{21}^{-1}=
-Nk_3/\sigma^7$. Then we see that in the determinant of $\Sigma_{ab}$ we have $[\Sigma_{12}^{-1}]^{2}\ll
[\Sigma_{11}^{-1}\Sigma_{22}^{-1}]$ for $k_3/\sigma^3\ll 1$, which is again the mild non-Gaussianity condition we are assuming throughout this work.
Eq. (\[varvar\]) is again an expected results: it is infact the variance of the standard deviation $\sigma$ for $N$ independent data when a non-zero fourth-order moment is included. The first term in (\[varvar\]) is the usual variance of the sample variance for Gaussian, independent data. The second term is due to the kurtosis correction: it will broaden the CR for $\sigma$ when $k_4$ is positive, and will shrink it when it is negative. Depending on the relative amplitude of the higher-order corrections, the CR for the variance can extend or reduce. It is important however to remark that this estimate of the confidence regions is approximated, and that it can be trusted only around the peak of the likelihood function. This means that we cannot use the CR estimated by the method exposed here when we are interested in large deviations from the best estimates. The true CR will in general be more and more different from this simple estimate as its probability content grows: to a confidence level of, say, 99.7%, we cannot reliably associate a CR of 3$\Sigma_{11}^{1/2}$, as we would do were the $P(\hat\alpha_i)$ a perfect Gaussian distribution.
This limitation is the main motivation for the rest of this work. Adopting a $\chi^2$ technique, we will be able to give an analytical expression for the CR of the variance (or of other parameters entering $\chi^2$) when non-G corrections are present. This is useful whenever we actually measured non-vanishing higher-order cumulants and wish to quote a CR for the variance allowing for the non-Gaussianity, or when, more generally, we have reason to suspect that our data are non-Gaussian and we wish to investigate how the CR vary with different non-Gaussian assumptions. As already remarked, non-Gaussianity can also be invoked to put in agreement two experimental results reporting non-overlapping CR. The results of the next sections will confirm the approximate trend of Eq. (\[varvar\]), as long as the CR does not extend into the tails of the paramater distribution.
Non-Gaussian $\chi^2$ method
============================
If our data are distributed following the MEE, then we can measure the likelihood to have found our actual dataset integrating the LF over all the possible outcomes of our experiment. According to the $\chi^2$ method, the actual dataset is more likely to have occurred, given our assumptions, the higher is the probability to obtain values of $\chi^2$ larger than the measured $\chi^2_0$. Then the relevant integral we have to deal with is \[rel\] M(\_0)=\_[\^2\^2\_0]{} L(x,) \_i dx\_i, where the region of integration extends over all the possible data values which lie inside the region delimited by the actual value $\chi^2_0$. The function $M(\chi_0)$ gives then the probability of occurrence of a value of $\chi^2$ smaller than the one actually measured. We can then use $M(\chi_0)$ for evaluating a CR for the parameters which enter $\chi_0^2$, like the quadrupole and the primordial slope in the case of CMB. The CR will depend parametrically on the higher-order moments; however, this will not provide a CR for the higher-order moments themselves. The method of the previous section can always be employed to yield a first approximation for such moments. Both too high and too small values of $\chi^2_0$ have to be rejected; fixing a confidence level of $1-\eps$, we will consider acceptables the values of the parameters for which $M(\chi_0)$ is larger than $\eps/2$ and smaller than $1-\eps/2$. Notice that the theoretical parameters enter $M(\chi_0)$ both through the integrand $L(x,\lam)$ and the integration region $\chi^2_0$. This section is devoted to the evaluation of (\[rel\]).
Let us split the LF into a Gaussian part and a non-G correction, L=L\_g+L\_[ng]{}. The integral of the Gaussian part is the standard one, and it is easily done: G(x,) \_i dx\_i=G(y,) \_i dy\_i =\_0\^[\_0\^2]{} P\_N(\^2)d \^2=F\_N(), where $P_N(\chi^2)$ is the $\chi^2$ PDF with $N$ degrees of freedom, and $F_N$ its cumulative integral. Notice that the integral has been performed over a compact region in $x$-space whose boundary is $\chi^2=\chi_0^2$. It is then convenient to change variables in the integral, from $y^j$ to the hyperradius $\chi^2$ and $N-1$ angles \_[i]{}dx\_i(\^2)\^[p-1]{}d(\^2) d=[A\_[2p]{}2]{}(\^2)\^[p-1]{}d(\^2), where $p=N/2$ and $A_{2p}=2\pi^p/\Gamma(p)$ (the surface area of a unitary $(N-1)$-sphere). The same procedure will be applied several times in the following.
For the non-G sector we have to consider separately the three last terms in the MEE (\[mee\]). However, since it will be shown shortly that the term linear in the skewness does not contribute to the final result, we will focus only on the two last terms in the MEE. Let us consider the term in $k^{ijkl}$. Its integral can be written as K&=&[ k\^[ijkl]{}(x)24 ]{}G(x,) h\_[ijkl]{}(x,) \_i dx\_i\
&=& [ k\^[ijkl]{}(y)24 ]{}G(y,) h\_[ijkl]{}(y,) \_i dy\_i= [ k\^[ijkl]{}(y)24 ]{}\_i\_j\_k\_l \_i G\_i dy\_i, where $G_i=(2\pi)^{-1/2} \exp(-y_i^2/2)$ and where the kurtosis tensor $k^{ijkl}$ in the first line is to be calculated with respect to $x^i$, while in the last line with respect to $y^i$. Since $y^i=A^i_a x^a$, then \[tensor\] k\^[ijkl]{}(y)=A\^i\_aA\^j\_bA\^k\_cA\^l\_d k\^[abcd]{}(x), and likewise for $k^{ijk}$. Suppose now one of the subscripts, say $i$, appears an odd number of times, like in $k^{iiij}$. Let us call then $i$ an odd index. The integral $K$ will then be odd in $y_i$. Since the integration region is symmetric around the origin, $K$ would vanish. This shows that any term in $K$ containing odd indexes of $k^{ijkl}$ must vanish. This explains also why the skewness term $h_{ijk}$, which always contains some odd index, gives no contribution to the likelihood integral. The only non-zero terms in $K$ are then of the type $k^{jjkk}$ and $k^{jjjj}$ (the index order is irrelevant). We then need only two kinds of integrals for as concerns $K$. Let us evaluate the first kind: I\_1=\_[i= j,k]{} G\_i dy\_i\[dy\_j dy\_k \^2\_j\^2\_k G\_j G\_k\]. The inner integral must be evaluated inside the circle bounded by $\rho^2_{jk}=\chi_0^2-\chi_{jk}^2$, where $\chi^2_{jk}=
\sum_{(i\not= j,k)} y_i^2$. Transforming the variables $(y_j,y_k)$ to the radius $\rho_{jk}$ and the angle $\theta$, and integrating over the new variables, we obtain I\_1=(\_[i= j,k]{} G\_i dy\_i) G\_2(\_[jk]{}) f\_1(\_[jk]{}) =G\_N(\_0) f\_1(\_[jk]{}) \_[i= j,k]{} dy\_i, where we define $G_N(\cz)=(2\pi)^{-N/2} \exp(-\cz^2/2)$, and where \[f1\] f\_1(\_[jk]{})=(4\_[jk]{}\^2-\_[jk]{}\^4)/4. Changing again variables under the integral to the hyperradius $\chi^2_{jk}$ and $N-3$ angles, we obtain I\_1=G\_N(\_0) A\_[2p\_2]{}\_0\^[\_0\^2]{} f\_1(\_[jk]{}) (\^2\_[jk]{}) \^[p\_2-1]{}d \^2\_[jk]{}=q\_1(\_0) G\_N(\_0), where $p_2=(N-2)/2$, and where \[q1\] q\_1()= \^[N/2]{}\^N\[N+2-\^2\]/(2+N/2). All the other integrals we need can be obtained in similar ways. For instance, the second kind of non-vanishing integral in $K$ is I\_2=\_[i= j]{} G\_i dy\_i\[dy\_j \^4\_j G\_j \]=G\_N(\_0) A\_[2p\_1]{}\_0\^[\_0\^2]{} f\_2(\_[j]{}) (\_j\^2)\^[p\_1-1]{} d \^2\_[j]{}=q\_2()G\_N(), where $p_1=(N-1)/2$ and \[q2\] q\_2()=[32]{}\^[N/2]{}\^N/(2+N/2). For as concerns the last term in (\[mee\]), we need only to evaluate three new integrals, from the terms $h_{jjkkll}$, $h_{jjjjkk}$ and $h_{j...j}$. Let us denote these integrals by $I_3, I_4$ and $I_5$. In complete analogy to the two integrals $I_1,I_2$, we find $I_i=G_N(\cz)q_i(\cz)$ where q\_3()&=& [14]{}\^[N/2]{} \^N h(),\
q\_4()&=& [34]{}\^[N/2]{}\^N h(),\
q\_5()&=&[154]{} \^[N/2]{}\^N h() , and where h()= .
These expressions are all what we need for the complete evaluation of the likelihood integral. The general result is then \[finres\] &&M()=L dx\_i=F\_N()\
&&+[G\_N(\_0)\^[N/2]{}\^N 2(2+N/2)]{} , where $C_a=c_1+3 c_2$, and $C_b=c_3+3 c_4+15 c_5$, and the coefficients $c_i$ are formed by summing over all the even diagonals of the correlation tensors $k^{ij..}$ and multiplying for the Edgeworth coefficients $(1/24)$ for $c_1,c_2$ and $(1/72)$ for $c_3,c_4$ and $c_5$. Let us denote with ${\rm Tr}_{ab..}(k^{ij...})$ the sum over all the disjoint partitions of $a$-plets, $b$-plets, etc. of equal indexes: e.g., ${\rm Tr}_{22}(k^{ijkl})$ means summing over all terms like $k^{iikk},k^{ikik},k^{ikki}$ but without including $k^{iiii}$; or, ${\rm Tr}_{24}(k^{ijk}k^{lmn})$ means summing over terms like $k^{iii}k^{ijj}$ or $k^{iij}k^{jii}$. Then, the coefficients $c_i$ in Eq. (\[finres\]) can be written as \[fincoeff\] c\_1&=&(1/24)[Tr]{}\_[22]{}(k\^[ijkl]{}),c\_2=(1/24)[Tr]{}\_[4]{}(k\^[ijkl]{}),\
c\_3&=&(1/72)[Tr]{}\_[222]{}(k\^[ijk]{}k\^[lmn]{}),c\_4=(1/72)[Tr]{}\_[24]{}(k\^[ijk]{}k\^[lmn]{}),\
c\_5&=&(1/72)[Tr]{}\_[6]{}(k\^[ijk]{}k\^[lmn]{}). Let us make some comments on the result so far obtained. First, notice that $M(\cz)$ is a cumulative function and as such it has to be a monotonically increasing function of its argument bounded by zero and unity. This provides a simple way to check the consistency of our assumptions: when the higher-order moments are too large, the MEE breaks down, $M(\cz)$ is no longer monotonic, and can decrease below zero or above unity. Second, let us suppose that the higher-order correlation functions are positive, which is the case for the galaxy clustering (see Section 6). Then the non-G corrections in Eq. (\[finres\]) are negative for $\cz^2\gg N$. The value of $\cz(\eps)$, corresponding to a probability content $M(\cz)=1-\eps$, is a measure of how large is the confidence region associated with the threshold $\eps$, if $\cz$ is single-valued on the parameter space, a common occurrence in practice. The fact that the corrections are negative for $\cz^2\gg N$ implies that the value of $\cz=\cz(\eps)$ is larger than in the purely Gaussian case, in the limit of $\eps\to 0$. Consequently, if the higher-order correlation functions are positive, [*the confidence regions are systematically widened when the non-Gaussian corrections are taken into account*]{}. For $\eps$ not very close to zero is not possible to make such a definite statement; the regions of confidence will widen or narrow depending on the value of the moments, as will be graphically shown in the next section. Let us remember that for two-tail tests the CR is enclosed by the upper limit $\cz^{(1)}(\eps/2)$ and the lower limit $\cz^{(2)}(1-\eps/2)$, and the limits behave differently depending on $\eps$ and on the higher-order moments. Finally, it is easy to write down the result in the particular case in which all the cumulant matrices are diagonal, i.e. for statistically independent variables. In this case the variables $y^i$ are simply equal to $x^i/\sigma_i$, if $\sigma_i=(\lam^i_i)^{-1/2}$, and we can put $k^{iii}(y)=k^{iii}(x)/\sigma^3\equiv\gam_{1,i}$, and likewise $k^{iiii}(y)\equiv \gam_{2,i}$ (skewness and kurtosis coefficients). Then, we have $c_1=c_3=c_4=0$, and Eq. (\[finres\]) can be simplified to \[resind\] M()=F\_N()+G\_N(\_0) q(), where \[indp\] q()=[6\^[N/2]{}\_0\^N (N+2)(N/2)]{} { [\_224]{}+[572]{}\_1\^2 }, and where we introduced the average squared skewness, $\gam_1^2=
\sum \gam_{1,i}^2/ N,$ and the average kurtosis, $\gam_2=\sum \gam_{2,i}/N$.
Graphical examples
==================
This section is devoted to illustrate graphically some properties of the function $M(\cz)$ in its simplified version (\[resind\]) above, first putting $\gam_1=0$, then $\gam_2=0$, and assuming $N=10$ and $N=100$. In all this section we can think of $\cz$ as depending monotonically on one single parameter, for instance the overall normalization $A>0$ of the correlation function: $\cz^2(A)=x^i x^j (A c_{ij}+e_{ij})^{-1}$. Then it appears that $\cz^2$ decreases from a finite, positive value to zero as $A$ goes from zero to infinity. The range in which $A$ is bounded increases or decreases with the bounding range of $\cz$; we can then speak of a CR on $\cz$ meaning in fact the corresponding CR on the parameter $A$. In the general case, the relation between $\cz$ and its parameters can be quite more complicated. In Fig. 1[*a*]{} (for $\gamma_1=0$ and $N=10$), we show how the function $M(\cz)$ varies with respect to the non-Gaussian parameter $\gamma_2$. Schematically, for $\cz^2/N> 1$, the function $M(\cz)$ decreases when $\gamma_2>0$ and increases in the opposite case. As anticipated, for too large a $\gam_2$, $M(\cz)$ develops a non-monotonic behavior. The consequence of the behavior of $M(\cz)$ on the confidence region of $\cz$ is represented in Fig. 1[*b*]{}, where the contour plots of the surface $M(\cz,\gamma_2)$ are shown. Consider for instance the two outer contours, corresponding to $M=.01$, the leftmost, and $M=.99$, the rightmost. The important point is that the range of $\cz$ inside such confidence levels increases for increasing $\gamma_2$; with respect to the Gaussian case, $\gamma_2=0$, the acceptable region for $\cz$ widens substantially even for a small non-Gaussianity. As a consequence, a value as high as, say, $\cz^2/N=2.5$, is inside the 99% confidence level if $\gam_2>.2$. The same is true for the other contour levels, although with a less remarkable trend. This behavior confirms the approximate result of Eq. (\[varvar\]). As anticipated, this means that the non-G confidence regions will be larger and larger (if the higher moments are positive) than the corresponding Gaussian regions for higher and higher probability thresholds. Notice that in this case the parameter $\gam_2$ itself cannot be given a CR, since as we already noticed the LF has no maximum when varied with respect to it. We can use $\gam_2$ only as an external parameter, either provided by theory, or estimated from the data in some other way. Fig. 2[*a,b*]{} reports the same features for $\gamma_1=0, N=100$. Now the confidence regions are much narrower, because of the increased number of experimental data.
The situation is qualitatively different considering $\gamma_2=0$ and varying $\gam_1$, the average skewness. Fig. 3 and Fig. 4 are the plots for this case ($N=10$ and $N=100$, respectively). The contour levels are obviously symmetric for $\pm \gam_1$. Now for any given $\cz$ there is a CR for $\gam_1$ and viceversa, so that a bound can be given on each parameter given the other one, although the joint CR for [*both*]{} parameters is infinite. Now one can see two different features in the contour plots. For the outer contours, delimiting levels of 1% on both tails, the CR of $\cz$ [*increases*]{} for larger $|\gam_1|$, with a minimum for the Gaussian case. For the internal contours, however, the CR actually shrinks for larger $|\gam_1|$, being maximal at the Gaussian point. It is clear that in the general case, $\gam_1,\gam_2\not=0$, the topography of the LF can be quite complicated.
Comments on practical application
=================================
The results of the previous sections can be employed to estimate theoretical parameters and confidence regions in several interesting cases. We consider here two of these, the large scale structure (LSS) of galaxies and the CMB.
In the case of LSS surveys, the data usually consist of the fluctuations $x^i=\delta n^i/\hat n$ in the number counts of galaxies in the $i$-th cell in which the survey is partitioned. (We assume here for simplicity that the average density $\hat n$ is fixed [*a priori*]{}. Otherwise, we can include it in the set of parameters to be estimated.) The main problem in applying the formalism developed so far to real situations is to choose a “good” set of theoretical parameters $\alpha_j$. In principle we can parametrize the statistical properties of the LSS in an infinite number of ways. However, the particular set of parameters we are going to adopt has been singled out in the current literature, both theoretical and observational, with very few exceptions. Assuming for the correlation function the power-law form $\xi(r)=(r_0/r)^{\gam}$, the cell-averaged $c^{ij}$ is given by the following expression \[cfint\] c\^[ij]{}=(r\_[12]{})W\_[R\_i]{}(\_1)W\_[R\_j]{}(\_2) d\_1 d\_2, where $r_{12}=|\vr_1-\vr_2|$, and $W_{R_i}$ ($W_{R_j}$) is the normalized window function of characteristic size $R_i$ ($R_j$) relative to the $i$-th ($j$-th) cell. If the cells $i,j$ are fully characterized by a size $R$ and a separation $s_{ij}$, the integral (\[cfint\]) can be written as c\^[ij]{}=J(,R/s\_[ij]{})(R/r\_0)\^[-]{}, where $J(\gam,R/s_{ij})$ is a dimensionless function of $\gam$ and $R/s_{ij}$. Following standard work (e.g. Peebles 1980) we will then write for the higher-order correlation functions the following expressions \[lss-cf\] k\^[ijk]{}&=&Q(c\^[ij]{}c\^[jk]{}+c\^[ik]{}c\^[jk]{}+ c\^[ik]{}c\^[ij]{}),\
k\^[ijkl]{}&=&R\_a\_2 c\^[ij]{}c\^[jk]{}c\^[kl]{}+R\_b\_3c\^[ij]{}c\^[ik]{}c\^[il]{}, where $\sum_2$ ($\sum_3$) means summing over all the 12 (4) tree graphs with at most two (three) connecting lines per vertex (i.e. summing over topologically equivalent graph configurations). Note that we define $Q,R_a$ and $R_b$ in terms of the [*cell-averaged*]{} correlation functions, rather than in terms of $\xi(r)$, as currently done. Our definition has the advantage that from $Q,R_a$ and $R_b$ one can obtain directly the often quoted scaling coefficients $S_3=3Q$ and $S_4=12R_a+4R_b$, without complicated integrals over the window functions. The drawback is that our $Q,R_a,R_b$ cannot be compared directly to the values reported in literature, albeit the difference should be very small.
Several analysis of large scale surveys show that $Q,R_a,R_b$ are fairly constant over several scales, and of order unity. On scales larger that $\approx 10\hm$, however, the power-law form of $c^{ij}$ is not longer acceptable. For such scales is preferable to parametrize instead the power spectrum $P(k)$ and to use the identity \[corrpk\] (r\_[ij]{})=[12\^2r\_[ij]{}]{}\_0\^ k P(k) (kr\_[ij]{}) dk, from which, using Eq. (\[cfint\]), c\^[ij]{}=[12\^2s\_[ij]{}]{}\_0\^ k P(k) W\_k\^2(ks\_[ij]{}) dk, where $W_k$ is the Fourier transforms of the window function. Various forms of $P(k)$ have been proposed so far. For instance, one can assume the simple functional form proposed by Peacock (1991), with its two scale parameters $k_0,k_1$, or the CDM-like form of Efstathiou, Bond & White (1992), involving an overall normalization and a dimensionless parameter $\Gamma$. To give an idea of how big the non-G corrections are, let us assume to have $N$ independent data (i.e., data on cells at separations much larger than the correlation length) and let us estimate the parameters $\gamma_1,\gamma_2$ of Eq. (\[indp\]). Since $\gam_{1,i},\gam_{2,i}$ and $\sigma_i$ are the same for all the $N$ data, one has (dropping the subscript $i$) $\gam_2=k^{iiii}(x)/ \sigma^4=S_4\sigma^2$ and $\gamma_1^2=S_3^2\sigma^2$. For $S_3\approx 3$ and $S_4\approx 20$, as large scale surveys suggest, one gets $\gamma_2\approx 20\sigma^2$, and $\gamma_1^2\approx 10
\sigma^2$. For scales around 10 $\hm$ or so, where $\sigma^2\approx 1$, $\gamma_1,\gamma_2$ are then very large, but they decrease rapidly for larger scales. On scales larger than 30 $\hm$ or so, $\gamma_1,
\gamma_2$ are small enough to use the MEE also near the tails.
The non-G LF allows a determination of the parametric set in such a way that the best estimate of one parameter depends on all the other ones, unlike the common procedure of estimating one parameter fixing the others (in particular, fixing the non-Gaussian parameters to zero). For instance, the usual way of estimating $r_0,\gam$ is to find the best $\chi^2$ power-law fit to the observed correlation function, which amounts to assume a Gaussian distribution around the mean values. Both the estimate and the confidence region would then be corrected by the higher order terms. However, as already mentioned, we can use the MEE for estimating the higher-order moments themselves only if enough terms have been included in the expansion. The reason is clear by looking at the Eq. (\[mee\]): at this order of truncation, the expansion is linear in the fourth order moment, and as a consequence it has no maximum when derivated with respect to, e.g., $R_a$ or $R_b$. The best estimate does not exist at all. We can give however an estimate for $Q$, and we can expect it to be a good estimate as long as it is in the regime in which the MEE holds. A simple way to check this is to see whether for that value of $Q$ the function $M(\cz)$ is well-behaved, i.e. is a monotonic increasing function bounded by zero and unity. In principle, one can proceed further, including more and more terms in the LF, so that one can reach not only a higher degree of approximation, but also estimate the error introduced by the truncation itself. Needless to say, these goods come at the price of a factorial increase in algebraic complication.
Once we have chosen our parameter set, the only remaining difficulty is to evaluate the coefficients $c_1,..c_5$. Let us remark that the Eqs. (\[lss-cf\]) are valid with respect to the original data $x^i$, while we need the correlation functions for $y^i=A^i_j x^j$ to evaluate $c_1,..c_5$. The relation between the two sets of correlation functions is provided by Eq. (\[tensor\]). The evaluation of $c_1,..c_5$ is straightforward. One needs simply to scan all the possible combinations of indexes $i,j,k,..$ and sum only those tensor components with all equal indexes (for $c_2$ and $c_5$), or those with all paired indexes (for $c_1$ and $c_3$), or finally those with a $(2,4)$ index structure (for $c_4$). A more explicit expression for the coefficients can also be found in specific cases (e.g. exploiting the symmetry under index permutation of the tensors in (\[lss-cf\])), but the general calculation can be coded so easily on computers that we prefer to leave it in the form (\[fincoeff\]). Let us then summarize the steps needed to analyze a given set of data. First, one selects a value for the chosen parameter set inside a plausible range. Second, one diagonalizes, for that particular parameter set, the quadratic form $x^i\lam_{ij}x^j$ so to determine the matrix $A^i_j$ such that $y^i=A^i_j x^j$. Third, one evaluates the five coefficients $c_1,..c_5$ summing over all the required tensor components. Fourth, one evaluates $L$ and $M(\cz)$ for the selected parameter set. Fifth, one repeates the four previous steps spanning a reasonable range in the parameter space. Finally, the values for which $L$ has a maximum inside the range, if any, are the best estimate of the set of parameters, while the region for which $
\eps/2< M(\cz)<1-\eps/2$ defines their joint confidence region.
For the CMB, the procedure is very similar. The major difference is the set of parameters we are interested in. For simplicity, let us consider an experiment like COBE, in which the large angular beam size is mainly designed to study the Sachs-Wolfe effect of primordial fluctuations. The two-point angular correlation function can be conveniently written as \[cfa\] c\^[ij]{}= \_[l=2]{}\^ C\_l W\_l()P\_l(\_[ij]{}), where $\alpha_{ij}$ is the angular separation between the $i$-th and $j$-th pixel on the sky, $W_l(\beta)$ is the observational window function relative to a beam angular size $\beta$, $P_l$ is the Legendre polynomial of order $l$, and $C_l$ is defined in terms of the multipole coefficients $a^m_l$ as C\_l=\_[m=-l]{}\^[l]{} |a\^m\_l|\^2. For the Sachs-Wolfe effect of fluctuations with power spectrum $P=Ak^n$ we can derive the expected variance of the amplitudes $a^m_l$ as (e.g. Kolb & Turner 1989) \[sw\] \_l\^2= [()\^2 5]{}, where $\qps$ is the expected quadrupole signal derived from the correlation function. The theoretical value for $C_l$ is then $C_l=(2l+1)\sigma_l^2$, and it depends uniquely on $\qps$ and $n$. Finally, we rewrite Eq. (\[cfa\]) as \[cfa2\] c\^[ij]{}= \_[l=2]{}\^(2l+1)\_l\^2 W\_l() P\_l(\_[ij]{}). The correlation function for the Sachs-Wolfe temperature fluctuations is then parametrized by $\qps$ and $n$. The situation for the higher-order correlation functions is much less well established. Non-Gaussianity in the CMB is predicted by several models, like topological defect theories, or non-standard inflation, or can be induced by some kind of foreground contamination. There is not, however, a single, widely accepted way to parametrize non-Gaussianity in this context (see e.g. Luo & Schramm 1993 for some possible alternatives). A very simple possibility is to assume for the CMB $n$-point correlation functions the same kind of scaling observed in LSS. Preliminary constraints on the 3-point parameter from the COBE data have already been published (Hinshaw 1994). Some model of inflation predicts indeed this sort of scaling, although the expected amplitude of the non-Gaussian signal in standard models is far below observability (Falk, Rangarajan & Srednicki 1993; Gangui 1994). For small scale experiments, the $c^{ij}$ parametrization is different, and often a Gaussian shape $c^{ij}=c_0\exp(-\alpha_{ij}^2/2\alpha_c^2)$, is assumed. The formalism here presented can be clearly applied to any desired form of the correlation function.
Conclusions
===========
Let us summarize the results reported here. This work is aimed at presenting a new analytic formalism for parametric estimation with the maximum likelihood method for non-Gaussian random fields. The method can be applied to a large class of astrophysical problems. The non-Gaussian likelihood function allows the determination of a full set of parameters and their [*joint*]{} confidence region, without arbitrarily fixing some of them, as long as enough non-linear terms are included in the expansion. The CR for all the relevant parameters can be estimated by approximating the distribution function for the parameter estimators around its peak by a Gaussian, as in Sect. 3. To overcome this level of approximation, in Sect. 4 we generalized the $\chi^2$ method to include non-Gaussian corrections. The most interesting result is then that the CR for the parameters which enter $\cz^2$ is systematically widened by the inclusion of the non-Gaussian terms, in the limit of $\eps\to 0$. Two experiments producing incompatible results can then be brought to agreement when third and fourth-order cumulants are introduced. In the more general case, the CR may extend or reduce.
While we leave the analysis of real data to subsequent work, we displayed some preliminary comments on the application to two important cases, large scale structure and cosmic microwave background.
There are two main limitations to the method. One is that we obviously have to truncate the MEE to some order, and consequently the data analysis implicitly assumes that all the higher moments vanish. The second limitation is that the method is not applicable to strongly non-Gaussian field, where the MEE breaks down. This can be seen directly from Eq. (\[finres\]): for arbitrarily large constants $c_1-c_5$ the likelihood integral is not positive-definite, although always converge to unity. Assuming the scaling relation of Eq. (\[lss-cf\]), for instance, the condition $c^{ij}<1$ will ensure that the higher order terms are not dominating over the lower terms, as long as the scaling constants are of order unity. Basing upon the current understanding of the matter clustering, we expect the condition of weak non-Gaussianity to hold for scales ranging from $\sim
30\hm$ to the horizon scale.
**Acknowledgments**
It is very likely that the completion of this work would have been much more difficult, if possible at all, without the crucial suggestions of Albert Stebbins, and the useful discussions with Stéphane Colombi and Scott Dodelson. I thank them all. Further, I thank the generous hospitality at the NASA/Fermilab Astrophysics Center, where most of this work has been done.
**References**
Amendola L. 1994 , 430, L9\
Chambers J. 1967, Biometrika 54, 367\
Efstathiou G., Bond J.R. & White S.D.M. 1992 , 258, 1[P]{}\
Falk T., Rangarajan R., & Srednicki M. 1993 , 403, L1\
Gangui A., Lucchin F., Matarrese S. & Mollerach S. 1994, in press\
Hinshaw G. , 431, 1\
Juszkiewicz R., Weinberg D. H., Amsterdamski P., Chodorowski M. & Bouchet F. 1993, preprint (IANSS-AST 93/50)\
Kendall M., Stuart, A., & Ord J. K., 1987, Kendall’s Advanced Theory of Statistics, (Oxford University Press, New York)\
Kofman L. & Bernardeau, F. 1994, preprint IFA-94-19\
Kolb E. W. & Turner M. 1989, The Early Universe (Addison Wesley, Reading)\
Luo X. & Schramm D.N. 1993 , 71, 1124\
Matsubara T. 1994, preprint UTAP-183/94\
McCullagh P. 1984, Biometrika, 71, 461\
Peacock J.A., 1991, , 253, 1[P]{}\
Peebles P.J.E., 1981, The Large Scale Structure of the Universe (Princeton University Press, Princeton)\
**Figure caption**
Plot of $M(\cz)$ as a function of $\cz^2/N$ and of the dimensionless kurtosis $\gam_2$, for $\gam_1=0, N=10$. For $\gam_2=0$ we return to the usual $\chi^2$ cumulative function. Notice how for large kurtosis $\gam_2$ the cumulative function $M(\cz)$ develops minima and maxima, indicating that the MEE is breaking down. [*b)*]{} Contour levels of $M(\cz)$ corresponding to $M=.01,.1,.2,.3,.7,.8,.9,.01$, from left to right. Notice how the limits for $\cz$ broaden for increasing $\gam_2$.
[**Fig. 2**]{} [*a)*]{} Same as in Fig. 1[*a*]{}, now with more data, $N=100$. [*b)*]{} Contour levels of $M(\cz)$ for the same values as in Fig. 1[*b*]{}. The CR is now much smaller than previously.
[**Fig. 3**]{} [*a)*]{} Same as in Fig. 1[*a*]{}, now with $\gam_2=0$, $N=10$, and varying $\gam_1$. [*b)*]{} Contour levels of $M(\cz)$ for the same values as in Fig. 1[*b*]{}.
Same as in Fig. 1[*a*]{}, now with $\gam_2=0, ~N=100$, and varying $\gam_1$. [*b)*]{} Contour levels of $M(\cz)$ for the same values as in Fig. 1[*b*]{}.
[^1]: e-mail [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Ning Xu [^1]\
Department of Mathematics and Science, Huai Hai Institute of Technology,\
Jiangsu, Lianyungang 222005, P. R. China
title: ' Products of redial derivative and integral-type operators from Zygmund spaces to Bloch spaces'
---
-1cm
[**Abstract.**]{} Let $H(\mathbb{B})$ denote the space of all holomorphic functions on the unit ball $\mathbb{B}\in \mathbb{C}^n$. In this paper we investigate the boundedness and compactness of the products of radial derivative operator and the following integral-type operator $$I_\varphi^g f(z)=\int_0^1 \Re f(\varphi(tz))g(tz)\frac{dt}{t},\ z\in\mathbb{B}$$ where $g\in H(\mathbb{B}), g(0)=0$, $\varphi$ is a holomorphic self-map of $\mathbb{B}$, between Zygmund spaces and Bloch spaces.
[**Keywords:**]{} radial derivative operator; integral-type operator; Zygmund space; Bloch space
[**1. Introduction**]{}
Let $H(\mathbb{B})$ denote the space of all holomorphic functions on the unit ball $\mathbb{B}\subset\mathbb{C}^n$. Let $z=(z_1,\ldots,z_n)$ and $w=(w_1,\ldots,w_n)$ be points in the complex vector space $\mathbb{C}^n$ and $<z,w>=z_1\overline{w_1}+\ldots+z_n\overline{w_n}$. Let $$\Re f(z)=\sum z_j\frac{\partial f}{\partial z_j}(z)$$ stand for the radial derivative of $f\in H(\mathbb{B})$\[1\]. It is easy to see that, if $f\in H(\mathbb{B}),
f(z)=\sum_\alpha a_\alpha z^\alpha,$ where $\alpha$ is a multi-index, then $\Re f(z)=\sum_\alpha|\alpha|a_\alpha z^\alpha$. We write $\Re^mf =\Re(\Re^{m-1}f)$.
The Bloch space $\mathcal{B}(\mathbb{B})$ is the space of all $f\in H(\mathbb{B})$ such that $$\|f\|_\mathcal{B}:=|f(0)|+\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\nabla f(z)|<\infty,$$
The little Bloch space $\mathcal{B}_0(\mathbb{B})$ is the space of all $f\in H(\mathbb{B})$ such that $$\lim\limits_{|z|\rightarrow 1}(1-|z|^2)|\nabla f(z)|=0.$$
It is well known that $f\in \mathcal{B}(\mathbb{B})$ if and only if $$b(f):=\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re f(z)|<\infty,$$ and that $f\in \mathcal{B}_0$ if and only if $\lim\limits_{|z|\rightarrow 1}(1-|z|^2)|\Re f(z)|=0$. Moreover, the following asymptotic relation holds\[2\]: $$\|f\|_\mathcal{B}\asymp |f(0)|+b(f).$$
Let $\mathcal{Z}$ denote the class of all $f\in H(\mathbb{B})$, such that $$\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re^2 f(z)|<\infty. \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (1)$$ Therefore, $\mathcal{Z}$ is called the Zygmund class. Under the natural norm $$\|f\|_\mathcal{Z}:=|f(0)|+|f'(0)|+\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re^2 f(z)|<\infty. \ \ \ \ (2)$$ $\mathcal{Z}$ becomes a Banach space. Zygmund class with this norm will be called the Zygmund space.
The little Zygmund space $\mathcal{Z}_0$ denote the closure in $\mathcal{Z}$ of the set of all polynomials. From Theorem 7.2 of\[3\], we see that $$f\in \mathcal{Z}_0 \Leftrightarrow \lim\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re^2 f(z)|=0.\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (3)$$
Suppose that $g\in H(\mathbb{B}), g(0)=0$ and $\varphi$ is a holomorphic self-map of $\mathbb{B}$, then an integral-type operator, denote by $I_\varphi^g$ on $H(\mathbb{B})$, is defined as follows: $$I_\varphi^g f(z)=\int_0^1 \Re f(\varphi(tz))g(tz)\frac{dt}{t},\ \ g\in H(\mathbb{B}),\ z\in\mathbb{B} \ \ \ \ \ \ \ \ \ (4)$$ Operator (4) is related to the following operators $$T_g(f)(z)=\int_0^1 f(tz)\Re g(tz)\frac{dt}{t},\ \ I_g(f)=\int_0^1 \Re f(tz)g(tz)\frac{dt}{t}. \ \ \ \ \ \ \ (5)$$ acting on $H(\mathbb{B})$, introduced in \[4\] and studied in \[5-10\], as well as the operator $T_g$ introduced in \[11\]acting on holomorphic functions on the unit polydisc (see,also\[12\],\[13\],as well as \[14\] for a particular case of the operator). One of motivations for introducing operator $I_\varphi^g$ stems from the operator introduced in \[15\]. Some characterizations of the boundedness and compactness of these and some other integral-type operators mostly in $\mathbb{C}^n$, can be found, for example, in\[4,6,7-9,15-31\].
In this paper we study the boundedness and compactness of products of $\Re$ and $I_\varphi^g$ between Zygmund space and Bloch spaces on the unit Ball.
Throughout this paper, constants are denoted by $C$,they are positive and may differ from one occurence to the other.\
[**2. Auxiliary results**]{}
[**Lemma 1$^{[19]}$.**]{} Let $\Re$ be the radial derivative operator. The product of $\Re$ and $I_\varphi^g$ $$\Re[ I_\varphi^g(f)](z)=\Re f(\varphi(z))g(z)\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (6)$$
[**Lemma 2$^{[17]}$.**]{} Suppose $f\in\mathcal{Z}$.The following statements are true.\
(a). There is a positive constant $C$ independent of $f$ such that $$|\Re f(z)|\leq C\|f\|_\mathcal{Z}\ln\frac{e}{1-|z|^2}.\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (7)$$ (b). There is a positive constant $C$ independent of $f$ such that $$\|f\|_\infty =\sup_{|z|<1}|f(z)|\leq C\|f\|_\mathcal{Z}.\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (8)$$
For studying the compactness of the operator $\Re I_\varphi^g:\mathcal{Z}\rightarrow\mathcal{B},$ we need the following Lemma. The proof of the lemma is standard, hence we omit the details.
[**Lemma 3.**]{} Assume that $g\in H(\mathbb{B})$, $\varphi$ be a holomorphic self-map of $\mathbb{B}$. Let $T=\Re I_\varphi^g$ ,then $T:\mathcal{Z}(or\mathcal{Z}_0 )\rightarrow\mathcal{B} $ is compact if and only if $T$ is bounded and for any bounded sequence $(f_k)_{k\in\mathbb{N}}$ in $\mathcal{Z}(or\mathcal{Z}_0 )$ which converges to 0 uniformly on compact subsets of $\mathbb{B}$, $Tf_k\rightarrow 0$ as $k\rightarrow\infty$.
[**Lemma 4$^{[17]}$.**]{} A closed set K in $\mathcal{B}_0$ is compact if and only if it is bounded and satisfies $$\lim\limits_{|z|\rightarrow1}\sup\limits_{f\in K}(1-|z|^2)|\Re f(z)|=0.\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (9)$$ [**3.The boundedness and compactness of $\Re I_\varphi^g:\mathcal{Z}(\mathcal{Z}_0)
\rightarrow\mathcal{B}(\mathcal{B}_0)$**]{}\
[**Theorem 1.**]{} Let $\varphi$ be a holomorphic self-map of $\mathbb{B}$. Then the following statements are equivalent.
\(a) $\Re I_\varphi^g:\mathcal{Z}\rightarrow\mathcal{B}$ is bounded;
\(b) $\Re I_\varphi^g:\mathcal{Z}_0 \rightarrow$ $\mathcal{B}$ is bounded;
\(c) $$\sup\limits_{z\in\mathbb{B}}\frac{(1-|z|^2)|\Re\varphi(z)||g(z)|}{1-|\varphi(z)|^2}<\infty,\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (10)$$ and $$\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re g(z)|\ln\frac{e}{1-|\varphi(z)|^2}<\infty.\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (11)$$ [**Proof.**]{}(a)$\Rightarrow$(b) This implication is obvious.
(b)$\Rightarrow$(c) Assume that $\Re I_\varphi^g$:$\mathcal{Z}_0$ $\rightarrow$ $\mathcal{B}$ is boundedness, i.e.,there exists a constant $C$ such that $$\|\Re I_\varphi^g(f)\|_\mathcal{B}\leq C\|f\|_\mathcal{Z},$$ for all $f\in\mathcal{Z}_0$. Taking the functions $f_j(z)=z_j\in\mathcal{Z}_0$ and $f_j(z)=z_j-z_j^2\in\mathcal{Z}_0,j=1,2,\cdots,n$,we get $$\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\varphi_j(z)||\Re \varphi(z) g(z)+\Re g(z)|<\infty,\ \ \ \ (12)$$ and $$\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|(\varphi_j(z)-4\varphi^2_j(z))
(\Re\varphi(z) g(z)+\Re g(z))+2\varphi^2_j(z)\Re g(z)|<\infty.\ (13)$$ Using (12) and the boundedness of functions $\varphi_j$, we have that $$\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re \varphi(z) g(z)+\Re g(z)|<\infty.\ \ \ \ (14)$$ Then with (13),(14) and the boundedness of functions $\varphi_j$, we have that
$$\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re g(z)|<\infty ,
\ \ \sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re\varphi(z) g(z)|<\infty \ \ \ \ (15)$$
Set $$h(\zeta)=(\zeta-1)[(1+\ln\frac{1}{1-\zeta})^2+1], \zeta\in \mathbb{C},$$ and $$h_a(z)=\frac{h(<z,a>)}{|a|^2}(\ln\frac{1}{1-|a|^2})^{-1},$$ for $a\in\mathbb{B}\backslash \{0\}$. It is known that $h_a(z)\in\mathcal{Z}_0$(see\[17\]).Since $$\Re h_a(z)=\frac{<z,a>}{|a|^2}(\ln\frac{1}{1-<z,a>})^2(\ln\frac{1}{1-|a|^2})^{-1},$$ and $$\Re^2h_a(z)=\Re h_a(z)+\frac{2<z,a>^2}{|a|^2(1-<z,a>)}(\ln\frac{1}{1-<z,a>})(\ln\frac{1}{1-|a|^2})^{-1},$$ for$|\varphi(z)|>\sqrt{1-1/e}$ we have $$\begin{aligned}
C\|\Re I_\varphi^g\|_\geq||\Re I_\varphi^g(h_{\varphi(z)})\|_\mathcal{B}
&\geq& (1-|z|^2)\ln\frac{1}{1-|\varphi(z)|^2}\Re g(z)|\\
&-&\frac{2(1-|z|^2)}{1-|\varphi(z)|^2}|\varphi(z)|^2||\Re\varphi(z)||g(z)|\\
&-&(1-|z|^2)\ln\frac{1}{1-|\varphi(z)|^2}|\Re\varphi(z)||g(z)|.\end{aligned}$$ Hence $$\begin{aligned}
(1-|z|^2)\ln\frac{1}{1-|\varphi(z)|^2}|\Re g(z)|&\leq & C+\frac{2(1-|z|^2)}{1-|\varphi(z)|^2}|\varphi(z)|^2||\Re\varphi(z)||g(z)|\\
&+&(1-|z|^2)\ln\frac{1}{1-|\varphi(z)|^2}|\Re\varphi(z)||g(z)|\\
&\leq &C+(2+e)\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re\varphi(z)||g(z)|,\ \ \ \ \ (16)\end{aligned}$$ which using the fact of $(1-|\varphi(z)|^2)\ln\frac{1}{1-|\varphi(z)|^2}\leq e$.\
For $|a|>\sqrt{1-1/e}$, set $$f_a(z)=\frac{h(<z,a>)}{|a|^2}(\ln\frac{1}{1-|a|^2})^{-1}-\int_0^1\ln\frac{1}{1-<tz,a>}\frac{dt}{t}.$$ Then $f_a\in \mathcal{Z}_0$.It is easy to see that $$\Re f_a(z)=\frac{<z,a>}{|z|^2}(\ln\frac{1}{1-<z,a>})^2(\ln\frac{1}{1-|a|^2})^{-1}-\ln\frac{1}{1-<z,a>},$$ $$\begin{aligned}
\Re^2f_a(z)&=&\Re f_a(z)+\frac{2<z,a>^2}{|a|^2(1-<z,a>)}(\ln\frac{1}{1-<z,a>})(\ln\frac{1}{1-|a|^2})^{-1}\\
&-&\frac{<z,a>}{1-<z,a>}+\ln\frac{1}{1-<z,a>}.\end{aligned}$$ Therefore $$\begin{aligned}
C\|\Re I_\varphi^g\|\geq\|\Re I_\varphi^g(f_{\varphi(z)})\|_\mathcal{B}
&=&\sup\limits_{z\in \mathbb{B}}(1-|z|^2)|\Re^2(I_\varphi^gf_{\varphi(z)})(z)|\\
&=&(1-|z|^2)(\frac{|\varphi(z)|^2}{1-|\varphi(z)|^2}+\ln\frac{1}{1-|\varphi(z)|^2})|\Re \varphi(z)||g(z)|\\
&\geq&(1-|z|^2)(\frac{|\varphi(z)|^2}{1-|\varphi(z)|^2}+1)|\Re \varphi(z)||g(z)|\\
&=&\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re \varphi(z)||g(z)|.\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (17)\end{aligned}$$ On the other hand, from (15) we have that $$\sup\limits_{|\varphi(z)|\leq\sqrt{1-1/e}}(1-|z|^2)|\Re g(z)|\ln\frac{1}{1-|\varphi(z)|^2}\leq
\sup\limits_{|\varphi(z)|\leq\sqrt{1-1/e}}(1-|z|^2)|\Re g(z)|<\infty.\ \ \ (18)$$ Hence from (15),(16),(17)and (18),we obtain (11).Further, from (17), we have $$\sup\limits_{|\varphi(z)|>\sqrt{1-1/e}}\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re \varphi(z)||g(z)|\leq C.\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (19)$$ On the other hand, from (15) we have that $$\sup\limits_{|\varphi(z)|\leq\sqrt{1-1/e}}\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re \varphi(z)||g(z)|\leq
e, \ \sup\limits_{|\varphi(z)|\leq\sqrt{1-1/e}}(1-|z|^2)|\Re \varphi(z)||g(z)|<\infty.\ \ \ \ \ \ \ \ \ \ \ (20)$$ Combining (19)and (20),(10) follows.
[**Theorem 2.**]{} Let $\varphi$ be a holomorphic self-map of $\mathbb{B}$. Then the following statements are equivalent.
(a)$\Re I_\varphi^g:\mathcal{Z}\rightarrow\mathcal{B}$ is compact;
(b)$\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is compact;
(c)$\Re I_\varphi^g:\mathcal{Z}\rightarrow\mathcal{B}$ is bounded, $$\lim_{|\varphi(z)|\rightarrow1}\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re \varphi(z)||g(z)|=0,\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (21)$$ and $$\lim_{|\varphi(z)|\rightarrow1}(1-|z|^2)|\Re g(z)|\ln\frac{e}{1-|\varphi(z)|^2}=0.\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (22)$$
[**Proof.**]{} (a)$\Rightarrow$ (b) This is obvious.
(b)$\Rightarrow$ (c) Assume that $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is compact, then it is clear that $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is bounded. By theorem 1, we know that $\Re I_\varphi^g:\mathcal{Z}\rightarrow\mathcal{B}$ is bounded. Let $(z^k)_{k \in\mathbb{N}}$ be a sequence in $\mathbb{B}$ such that $|\varphi(z^k)|\rightarrow1$ as $k\rightarrow\infty$ and $\varphi(z^k)\neq 0, k\in\mathbb{N}$. Set $$h_k(z)=\frac{h(<z,\varphi(z^k)>)}{|\varphi(z^k)|^2}(\ln\frac{1}{1-|\varphi(z^k)|^2})^{-1}, k\in\mathbb{N}.$$ Then from the proof of theorem 1,we see that $h_k\in\mathcal{Z}_0$,for each $k\in\mathbb{N}$. Moreover $h_k\rightarrow 0$ uniformly on compact subsects of $\mathbb{B}$, as $k\rightarrow\infty$.
Since $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is compact,by Lemma 3 $$\lim_{k\rightarrow\infty}\|\Re [I_\varphi^g(h_k)]\|_\mathcal{B}=0.$$ On the other hand, similar to the proof of Theorem 1, we have $$\begin{aligned}
0 \leftarrow\|\Re I_\varphi^g(h_k)\|_\mathcal{B}
&\geq& (1-|z^k|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}|\Re g(z^k)|\\
&-&\frac{2(1-|z^k|^2)}{1-|\varphi(z^k)|^2}|\varphi(z^k)|^2\Re\varphi(z^k)||g(z^k)|\\
&-&(1-|z^k|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}|\Re\varphi(z^k)||g(z^k)|\\
&=&(1-|z^k|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}|\Re g(z^k)|\\
&-&M_1\frac{(1-|z^k|^2)}{1-|\varphi(z^k)|^2}|\Re\varphi(z^k)||g(z^k)|,\end{aligned}$$ where $M_1:=2|\varphi(z^k)|^2-(1-|\varphi(z^k)|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}$.
From this we obtain $$\lim\limits_{k\rightarrow\infty}(1-|z^k|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}|\Re g(z^k)|=
\lim\limits_{k\rightarrow\infty}\frac{(1-|z^k|^2)}{1-|\varphi(z^k)|^2}|\Re\varphi(z^k)||g(z^k)|=0,\ \ \ \ \ (23)$$ if one of these two limits exists,which use the case of $$\lim\limits_{k\rightarrow\infty}[2|\varphi(z^k))^2|+(1-|\varphi(z^k)|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}]=2.$$
Next, set $$\begin{aligned}
f_k(z)&=&\frac{h(<z,\varphi(z^k)>)}{|\varphi(z^k)|^2}(\ln\frac{1}{1-|\varphi(z^k)|^2})^{-1}\\
&-&\int_0^1\ln^3\frac{1}{1-<tz,\varphi(z^k)>}\frac{dt}{t}(\ln\frac{1}{1-|\varphi|(z^k)|^2})^{-2}.\end{aligned}$$ Since $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is compact,we have $\|\Re I_\varphi^g(f_k)\|_{\mathcal{B}}\rightarrow0$ as $k\rightarrow\infty.$ Thus,similar to the proof of Theorem 1,when$\sqrt{1-\frac{1}{e}}<|\varphi(z^k)|<1$ $$\begin{aligned}
0\leftarrow\|\Re I_\varphi^g(f_k)\|_\mathcal{B}
&\geq&(1-|z^k|^2)|\ln\frac{1}{1-|\varphi(z^k)|^2}-\frac{|\varphi(z^k)|^2}{1-|\varphi(z^k)|^2}||\Re \varphi(z^k)||g(z^k)|\\
&\geq&(1-|z^k|^2)|\frac{|\varphi(z^k)|^2}{1-|\varphi(z^k)|^2}||\Re \varphi(z^k)||g(z^k)|\\
&-&(1-|z^k|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}|\Re \varphi(z^k)||g(z^k)|\\
&\geq&(1-\frac{1}{e})\frac{1-|z^k|^2}{1-|\varphi(z^k)|^2}|\Re \varphi(z^k)||g(z^k)|\\
&-&\frac{1-|z^k|^2}{1-|\varphi(z^k)|^2}(1-|\varphi(z^k)|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}|\Re \varphi(z^k)||g(z^k)|\\
&=&M_2\frac{1-|z^k|^2}{1-|\varphi(z^k)|^2}|\Re \varphi(z^k)||g(z^k)|,\end{aligned}$$ where$ M_2:=1-\frac{1}{e}-(1-|\varphi(z^k)|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}$.
Hence $$\lim\limits_{k\rightarrow\infty}\frac{1-|z^k|^2}{1-|\varphi(z^k)|^2}|\Re \varphi(z^k)||g(z^k)|=
\lim\limits_{k\rightarrow\infty}(1-|z^k|^2)\ln\frac{1}{1-|\varphi(z^k)|^2}|\Re g(z^k)|=0.\ (24)$$
From (24) easily following that $\lim\limits_{k\rightarrow\infty}(1-|z^k|^2)|\Re g(z^k)|=0$,which altogether imply (21)and (22).
(c)$\Rightarrow$ (a) $$C_1=(1-|z|^2)|\Re \varphi(z)||g(z)|<\infty,\ \ \ C_2=(1-|z|^2)|\Re g(z)|<\infty.\ \ \ \ \ \ \ \ \ \ \ (25)$$
For every $\varepsilon>0$,there is a $\delta\in (0,1)$, such that $$\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re \varphi(z)||g(z)|<\varepsilon,\ \ \
(1-|z|^2)|\Re g(z)|\ln\frac{e}{1-|z|^2}<\varepsilon.\ \ \ \ \ \ \ \ \ \ \ \ \ (26)$$
Assume that $(f_k)_{k\in\mathbb{N}}$ is a sequence in $\mathcal{Z}$ such that $\sup\limits_{k\in\mathbb{N}}\|f_k\|_{\mathcal{Z}}\leq L$ and $f_k$ converges to 0 uniformly on compact subsets of $\mathbb{B}$ as $k\rightarrow\infty$. Let $K=\{z\in\mathbb{B}:|\varphi(z)|\leq\delta\}$. Then by (25) and (26) ,we have that $$\begin{aligned}
\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|(\Re^2(I_{\varphi}^g(f_k))(z)|
&=&\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re^2f_k(\varphi(z))\Re\varphi(z)g(z)+\Re f_k(\varphi(z))\Re g(z)|\\
&\leq&\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re^2f_k(\varphi(z))\Re\varphi(z)g(z)|\\
&+&\sup\limits_{z\in\mathbb{B}}(1-|z|^2)|\Re f_k(\varphi(z))\Re g(z)|\\
&\leq&\sup\limits_{z\in K}(1-|z|^2)|\Re^2f_k(\varphi(z))\Re\varphi(z)g(z)|\\
&+&\sup\limits_{z\in K}(1-|z|^2)|\Re f_k(\varphi(z))\Re g(z)|\\
&+&\sup\limits_{z\in\mathbb{B}\backslash K}(1-|z|^2)|\Re^2f_k(\varphi(z))\Re\varphi(z)g(z)|\\
&+&\sup\limits_{z\in\mathbb{B}\backslash K}(1-|z|^2)|\Re f_k(\varphi(z))\Re g(z)|\\
&\leq&\sup\limits_{z\in K}(1-|z|^2)|\Re^2f_k(\varphi(z))\Re\varphi(z)g(z)|\\
&+&\sup\limits_{z\in K}(1-|z|^2)|\Re f_k(\varphi(z))\Re g(z)|\\
&+&\sup\limits_{z\in\mathbb{B}\backslash K}\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re\varphi(z)g(z)|\|f_k\|_{\mathcal{Z}}\\
&+&C \sup\limits_{z\in\mathbb{B}\backslash K}\ln\frac{e}{1-|\varphi(z)|^2}|\Re g(z)|\|f_k\|_{\mathcal{Z}}\\
&\leq&C_1\sup\limits_{z\in K}|\Re^2f_k(\varphi(z))|+C_2\sup\limits_{z\in K}|\Re f_k(\varphi(z))|+(C+1)\varepsilon\|f_k\|_{\mathcal{Z}}\\\end{aligned}$$
Hence $$\begin{aligned}
\|\Re I_{\varphi}^g(f_k)\|_\mathcal{B}&\leq& C_1\sup\limits_{z\in K}|\Re^2f_k(\varphi(z))|+C_2\sup\limits_{z\in K}|\Re f_k(\varphi(z))|\\
&+&(C+1)\varepsilon\|f_k\|_{\mathcal{Z}} +|f'_k(\varphi(0))||\varphi'(0)|\end{aligned}$$
Since $(f_k)_{\mathbb{N}}$ converges to 0 uniformly on compact subsets of $\mathbb{B}$ as $k\rightarrow\infty$, Cauchy’s estimate gives that $\Re f_k\rightarrow 0$ and $\Re^2 f_k\rightarrow 0$ as $k\rightarrow\infty$ on compact subsets of $\mathbb{B}$. Hence, letting $k\rightarrow\infty$, we obtain $$\lim\limits_{k\rightarrow\infty}\|\Re I_{\varphi}^g(f_k)\|_\mathcal{B}=0.$$
[**Theorem 4.**]{} Let $\varphi$ be a holomorphic self-map of $\mathbb{B}$. Then $\Re I_\varphi^g:
\mathcal{Z}_0\rightarrow\mathcal{B}_0$ is bounded if and only if $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is bounded $$\lim_{|z|\rightarrow1}(1-|z|^2)|\Re g(z)|=0,\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (27)$$ $$\lim_{|z|\rightarrow1}(1-|z|^2)|\Re \varphi(z)||g(z)|=0.\ \ \ \ \ \ \ \ \ \ \ \ \ (28)$$ [**Proof:**]{} Assume that $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}_0$ is bounded. Then, it is clear that $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is bounded.Taking the function $f_j(z)=z_j$ and $f_j(z)=z_j-z_j^2,j=1,2,\cdots,n,$ we obtain (27),(28).
Assume that $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is bounded and (27),(28) holds. Then for each polynomial $p$, we have that $$\begin{aligned}
(1-|z|^2)|\Re^2( I_\varphi^gp)(z))|&\leq &(1-|z|^2)|\Re^2p(\varphi(z))||\Re \varphi(z)||g(z)|\\
&+ &(1-|z|^2)|\Re p(\varphi(z))||\Re g(z)|,\ \ \ \ \ \ \ \ \ \ \ \ \ (29)\end{aligned}$$ from (27),(28) it follows that $\Re I_\varphi^g p \in\mathcal{B}_0$. Since the set of all polynomials is dense in $\mathcal{Z}_0$, we have that for every $f\in\mathcal{Z}_0$, there is a sequence of polynomials $(p_n)_{n\in\mathbb{N}}$ such that $\|f-p_n\|_\mathcal{Z}\rightarrow0$ as $n\rightarrow\infty$. Hence $$\|\Re I_{\varphi}^g(f)-\Re I_{\varphi}^g(p_n)\|_\mathcal{B}
\leq \|\Re I_{\varphi}^g\|_{\mathcal{Z}_0\rightarrow\mathcal{B}}\|f-p_n\|_\mathcal{Z}\rightarrow 0$$ as $n\rightarrow\infty$. Since the operator $\Re I_{\varphi}^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is bounded, hence $\Re I_{\varphi}^g(\mathcal{Z}_0)\subseteq \mathcal{B}_0$.
[**Theorem 5.**]{} Let $\varphi$ be a holomorphic self-map of $\mathbb{B}$.Then the following statements are equivalent.
(a)$\Re I_\varphi^g:\mathcal{Z}\rightarrow\mathcal{B}$ is compact;
(b)$\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}_0$ is compact;
\(c) $$\lim_{|z|\rightarrow1}\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re \varphi(z)||g(z)|=0,\ \ \ \ \ \ \ \ \ \ \ \ \ (30)$$ and $$\lim_{|z|\rightarrow1}(1-|z|^2)|\Re g(z)|\ln\frac{e}{1-|\varphi(z)|^2}=0.\ \ \ \ \ \ \ \ \ \ \ \ \ (31)$$
[**Proof:**]{} (a)$\Rightarrow$(b). It is clear.
(b)$\Rightarrow$(c).Assume that $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}_0$ is compact,then $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}_0$ is bounded.From the proof of Theorem 4,we known that $$\lim_{|z|\rightarrow1}(1-|z|^2)|\Re g(z)|=0,$$ $$\lim_{|z|\rightarrow1}(1-|z|^2)|\Re \varphi(z)||g(z)|=0,$$
Hence, if $\|\varphi\|<1$, $$\lim_{|z|\rightarrow1}\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re \varphi(z)||g(z)|\leq
\frac{1}{1-\|\varphi\|_\infty}\lim_{|z|\rightarrow1}(1-|z|^2)|\Re g(z)|=0,$$ $$\lim_{|z|\rightarrow1}(1-|z|^2)|\Re g(z)|\ln\frac{e}{1-|\varphi(z)|^2}\leq
\ln\frac{e}{1-\|\varphi\|^2_\infty}\lim_{|z|\rightarrow1}(1-|z|^2)|\Re g(z)|=0.$$ from which the result follows in this case.
Assume $\|\varphi\|=1$. Let $(\varphi(z^k))_{k\in\mathbb{N}}$ be a sequence such that $|\varphi(z^k)|\rightarrow 1$ as $k\rightarrow\infty$. Since $\Re I_\varphi^g:\mathcal{Z}_0\rightarrow\mathcal{B}$ is compact, by Theorem 2, $$\lim_{|\varphi(z)|\rightarrow1}\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re \varphi(z)||g(z)|=0,\ \ \ \ \ \ \ \ \ \ \ \ \ (32)$$ and $$\lim_{|\varphi(z)|\rightarrow1}(1-|z|^2)|\Re g(z)|\ln\frac{e}{1-|\varphi(z)|^2}=0.\ \ \ \ \ \ \ \ \ \ \ \ \ (33)$$ It is not difficult to see that (28),(32) implies(30). Similar, (27) and (33) imply (31).
(c)$\Rightarrow$(a). Let $f\in \mathcal{Z}$, we have $$(1-|z|^2)|\Re^2(I_{\varphi}^g(f))(z)|\leq(\frac{1-|z|^2}{1-|\varphi(z)|^2}|\Re \varphi(z)||g(z)|+
(1-|z|^2)\ln\frac{e}{1-|\varphi(z)|^2}|\Re g(z)|\|f\|_{\mathcal{Z}}.$$ Taking the supremum in this inequality over all $f\in \mathcal{Z}$ such that $\|f\|_{\mathcal{Z}}\leq 1$. Letting $|z|\rightarrow1$ and using (30),(31) $$\lim\limits_{\|z\|\rightarrow1}\sup\limits_{\|f\|_{\mathcal{Z}}\leq 1}(1-|z|^2)|\Re^2(I_{\varphi}^g(f))(z)|=0.$$ Using Lemma 3,we obtain that the operator $\Re I_\varphi^g:\mathcal{Z}\rightarrow\mathcal{B}_0$ is compact.
[REFERENCES]{}
1. W.Rudin, Function Theory in the Unit Ball of $\mathbb{C}^n$, Spring-Verlay, New York, 1980.
2. D.Clahane, S.Stevi$\acute{c}$, Norm equivalence and composition operators between Bloch/Lipschitz spaces of the unit ball, J.Inequal. Appl. 2006(2006).Article ID 61068, 11 pp.
3. K.Zhu, Spaces of Holomorphic Functions in the Unit Ball, Springer, New York, 2005.
4. Z.Hu, Extended Ces$\grave{a}$ro operators on the Bloch spaces in the unit ball of $\mathbb{C}^n$, Acta. Math. Sci. Ser. B Engl. Ed.23(4)(2003)561-566.
5. S.Stevi$\acute{c}$, On an integral operator on the unit ball in $\mathbb{C}^n$,J.Inequeal.Appl.\
2005(1)81-88.
6. N.Xu, Extended Ces$\grave{a}$ro operators on $\mu$-Bloch spaces in $\mathbb{C}^n$, J. of Math. Research and Exposition, 29(5)(2009)913-922.
7. Z.Hu, Extended Ces$\grave{a}$ro operators on mixed norm spaces, Proc. Amer. Math. Soc. 131(7)(2003)2171-2179.
8. Z.Hu, Extended Ces$\grave{a}$ro operators on Bergman spaces, J. Math. Anal. Appl. 296(2004)435-454.
9. S. Li, S.Stevi$\acute{c}$, Ces$\grave{a}$ro-type operators on some spaces of analytic functions on the unit ball,Appl. Math. and Comput. 208(2009)378-388.
10. S.Stevi$\acute{c}$, Ces$\grave{a}$ro averaging operators,Math.Nachr.248-249(2003)185-189.
11. S.Stevi$\acute{c}$, Boundedness and compactness of an integral operator on a weighted space on the polydisc,Indian J. Pure Appl. Math. 37(6)(2006) 343-355.
12. S.Stevi$\acute{c}$, Boundedness and compactness of an integral operator on mixed norm spaces on polydisc, Sibirsk. Mat. Zh. 48(3)(2007)694-706.
13. D.C. Chang, S.Stevi$\acute{c}$, The generalized Ces$\grave{a}$ro operator on the unit polydisk, Taiwan. J. Math. 7(2)(2003)293-308.
14. D.C. Chang, S.Stevi$\acute{c}$, Estimates of an integral operator on function spaces, Taiwan. J. Math. 7(3)(2003)423-432.
15. S. Li, S.Stevi$\acute{c}$, Generalized composition operators on Zygmund spaces and Bloch type spaces, J. Math. Anal. Appl. 338(2008)1282-1295.
16. D.C. Chang, S.Stevi$\acute{c}$, On some integral operators on the unit polydisk and the unit ball, Taiwan. J. Math. 11(5)(2007)1251-1286.
17. S. Li, S.Stevi$\acute{c}$, Riemann-Stieltjes operators between mixed norm spaces, Indian J. Math. 50(1)(2008)177-188.
18. S. Li, S.Stevi$\acute{c}$, Products of composition and differentiation operators from Zygmund spaces to Bloch spaces and Bers spaces,Appl. Math. and Comput. 217(2010)3144-3154.
19. S.Stevi$\acute{c}$, On an integral operator between Bloch-type spaces on the unit ball, Bull. Sci. Math. 134(2010)329-339.
20. S.Stevi$\acute{c}$, S.I.Ueki, Integral-type operators acting between weighted-type spaces on unit ball,Appl.Math.Comput. 215(2009)2464-2471.
21. K.Avetisyan, S.Stevi$\acute{c}$, Extended Ces$\grave{a}$ro operator between different Hardy spaces, Appl.Math.Comput.(2009)346-350.
22. W.Yang, On an integral-type operator between Bloch-type spaces,\
Apll.Math.Comput.215(3)(2009)954-960.
23. X.Zhang, Weighted composition operators between $\mu$-Bloch spaces on the unit ball,Sci.China(Ser.A)48(2005)1349-1368.
24. X.Zhu, Integral-type operators from iterated logarithmic Bloch spaces to Zygmund-type spaces,Appl.Math.Comput.215(2009)1170-1175.
25. S.Stevi$\acute{c}$, Generalized composition operators from logarithmic Bloch spaces to mixed-norm spaces,Util.Math.77(2008)167-172.
26. S.Stevi$\acute{c}$, On an integral operator from the Zygmund space to the Bloch-type space on the unit ball, Glasg.J.Math.51(2009)275-287.
27. S.Stevi$\acute{c}$, On an integral operator from the logarithmic Bloch-type and mixed-norm spaces to Bloch-type spaces, Nonlinear Anal. TMA 71(2009)6323-6342.
28. S.Stevi$\acute{c}$, Weighted composition operators from the logarithmic weighted-type space to the weighted Bergman space in $\mathbb{C}^n$, Appl.Math.Comput.\
216(2010)924-928.
29. S.Stevi$\acute{c}$, On a new operator from H$^\infty$ to the Bloch-type space on the unit ball,Util.Math.77(2008)257-263.
30. S.Stevi$\acute{c}$, On a new operator from the logarithmic Bloch space to the Bloch-type space on the unit ball,Appl.Math.Comput.206(2008)313-320.
31. S.Stevi$\acute{c}$, Products of integral-type operators and composition operators from the mixed-norm space to Bloch-type spaces,Siberian J.Math.50(4)(2009)726-736.
[^1]: Corresponding author, e-mail: [email protected].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a gravitational hierarchical $N$-body code that is designed to run efficiently on Graphics Processing Units (GPUs). All parts of the algorithm are exectued on the GPU which eliminates the need for data transfer between the Central Processing Unit (CPU) and the GPU. Our tests indicate that the gravitational tree-code outperforms tuned CPU code for all parts of the algorithm and show an overall performance improvement of more than a factor 20, resulting in a processing rate of more than $2.8$ million particles per second.'
author:
- 'Jeroen Bédorf$^1$, Evghenii Gaburov$^{1,2}$, and Simon Portegies Zwart$^1$'
bibliography:
- 'Bedorf\_J.bib'
title: 'Bonsai: A GPU Tree-Code'
---
Introduction {#sect:introduction}
============
A populair method for simulating gravitational $N$-body systems is the hierarchical tree-code algorithm orginaly introduced by [@1986Natur.324..446B]. This method reduces the computational complexity of the simulation from ${\cal O}(N^2)$ to ${\cal O}(N\log N)$ per crossing time. The former, though computationally expensive, can easily be implemented in parallel for many particles. The later requires more communication and book keeping when developing a parallel method. Still for large number of particles ($N \gtrsim 10^5$) hierarchical[^1] methods are more efficient than brute force methods. Currently parallel octree implementations are found in a wide range of problems, such as: self gravitating systems, smoothed particle hydrodynamics, molecular dynamics, clump finding, ray tracing and voxel rendering. All of these problems require a high amount of computation time. For high resolution simulations ($N \gtrsim 10^5$) 1 Central Processing Unit (CPU) is not sufficient, one has to use computer clusters or even supercomputers, both of which are expensive and scarce. A GPU provides an atractive alternative to such systems.
The GPU is a massively parallel processor which is specialised in performing independend parallel computations. For an overview and more details see [@2007NewA...12..641P; @2007astro.ph..3100H; @2008NewA...13..103B; @Gaburov2009630]. In this work we have implemented all the seperate parts of the Barnes-Hut tree-code algorithm on the GPU. This includes the tree-construction and computation of the tree-properties (multipole moments). By doing so we remove the need to communicate large amounts of data between the CPU and GPU. Since there is no time lost by CPU-GPU communication we can make optimal use of the GPU in a block time-step algorithm. In previous work [@OctGravICCS10; @Hamada:2010:TAN:1884643.1884644] parts of the algorithm were executed on the CPU which only allows for efficient (parallel) execution if shared time-steps are used.
Full implementation details, performance and accuracy characteristics can be found in [@2011arXiv1106.1900B]
Results {#Sect:Results}
=======
The algorithms are implemented as part of the gravitational $N$-body code [Bonsai[^2]]{}. We compare the performance of the GPU algorithms with (optimized) CPU implementations of comparable algorithms. The performance of the tree-traverse depends critically on the multipole acceptance critera (MAC) ($\theta$) which sets the trade-off between speed and accuracy. In our implementation we use a combination of the method introduced by Barnes (1994) and the method used for tree-traversion on vector machines see [@Barnes1994; @1990JCoPh..87..161B]. To compare the CPU and GPU implementations we measure the wall-clock time for the most time critical parts of the algorithms. For the tree-construction we distinguish three parts; Sorting of the particles along a Space Filling Curve ([@Morton]) (sorting in the figure) , reordering of particle properties based on the sort (moving) and construction of the tree-datastructure (tree-construction). Furthermore, timings are presented for the computation of the multipole moments and tree-traverse. As for hardware we used a Xeon E5620 CPU with 4 physical cores and a NVIDIA GTX480 GPU. The resuls are presented in Fig. \[Image:GenScaling\].
![Wall-clock time spent by the CPU and the GPU on various primitive algorithms. The bars show the time spent on the five selected sections of code. The results indicate that our GPU code outperforms the CPU code on all fronts and is between 2 and 30 times faster. Note that the y-axis is in logscale. (Timings using a $2^{20}$ million body Plummer sphere with $\theta=0.75$)[]{data-label="Image:GenScaling"}](Bedorf_J_Fig1){width="0.75\columnwidth"}
To measure the scaling of the implemented algorithms we execute simulations using Plummer spheres ([@1915MNRAS..76..107P]) with $N=2^{15}$ ( 32k) up to $N=2^{22}$ ( 4M) particles. We measure the performance of the same algorithms as in the previous paragraph. The results can be found in Fig. \[Image:Scaling\]. The wall-clock time spent in the sorting, moving, tree-construction and multipole computation algorithms scales linearly with $N$ for $N \gtrsim 10^6$. For smaller $N$, however, the scaling is sub-linear, because the algorithms require more than $10^5$ particles to saturate the GPU. In theory the tree-traverse scales as ${\cal O}(N\log N)$, whereas empirically the wall-clock time scales almost linearly with $N$. This is explained in the inset of Fig. \[Image:Scaling\], which shows the average number of interactions per particle during the simulation. The average number of particle-cell interactions doubles between $N \gtrsim $ 32k and $N \lesssim $ 1M and keeps gradually increasing for $N \gtrsim $ 1M. This break is not clearly visible in the timing results since for small particle numbers ($N \lesssim $ 1M ) not all GPU resources are satured. Finally, more than 90% of the wall-clock time is spent on tree-traversion with $\theta=0.75$. This allows for block time-step execution where the tree-traverse time is reduced by a factor $N/N_{\rm active}$, where $N_{\rm active}$ is the number of particles that have to be updated.
![The wall-clock time spent by various parts of the program versus the number of particles $N$. We used Plummer models as initial conditions and varied the number of particles over two orders of magnitude. The scaling of the tree-walk is between ${\cal O}(N)$ (shown with the black solid line) and the theoretical ${\cal O}(N\log N)$ and is due to the average number of interactions staying roughly constant (see inset). The asymptotic complexity of the tree-construction approaches ${\cal O}(N)$, as expected, since all the constituent primitives share the same complexity. The timings are from the GTX480 GPU with $\theta$ = 0.75.[]{data-label="Image:Scaling"}](Bedorf_J_Fig2){width="0.75\columnwidth"}
Discussion and Conclusions {#sect:discussion}
==========================
We have presented an efficient gravitational $N$-body tree-code. In contrast to other existing GPU tree-codes, this implementation is executed completely on the GPU. On a GTX480 the number of particles processed per unit time is $2.8$ million particles per second with $\theta=0.75$. This allows us to routinely carry out simulations on the GPU. Since the current version can only use 1 GPU, the limitation is the amount of memory. For 5 million particles $\sim 1$ gigabyte of GPU memory is required.
Even though the sorting, moving and tree-construction parts of the code take up roughly 10% of the execution time in the presented timings, these methods do not have to be executed during each time-step when using the block time-step method. It is sufficient to only recompute the multipole moments of tree-cells that have updated child particles, and only when the tree-traverse shows a considerable decline in performance does the complete tree-structure has to be rebuild. This decline is the result of inefficient memory reads and an increase of the average number of particle-cell and particle-particle interactions. This quantity increases because the tree-cell size increases, which causes more cells to be opened by the MAC.
Although the implemented algorithms are designed for a shared-memory architecture, they can be used to construct and traverse tree-structures on parallel GPU clusters using the methods described in [@169640; @1996NewA....1..133D]. Furthermore, in case of a parallel GPU tree-code, the CPU can exchange particles with the other nodes, while the GPU is traversing the tree-structure of the local data, making it possible to hide most of the communication time.
The implemented algorithms are not limited to the evaluation of gravitational forces, but can be applied to a variety of problems, such as neighbour search, clump finding algorithms, fast multipole method and ray tracing. In particular, it is straightforward to implement Smoothed Particle Hydrodynamics in such a code, therefore having a self-gravitating particle based hydrodynamics code implemented on the GPU.
This work is supported by NOVA and NWO grants (\#639.073.803, \#643.000.802, and \#614.061.608, VICI \#643.200.503, VIDI \#639.042.607). The authors would like to thank Massimiliano Fatica and Mark Harris of NVIDIA for the help with getting the code to run on the Fermi architecture
[^1]: Tree data-structures are commonly referred to as hierarchical data-structures. In this work we use an octree data-structure.
[^2]: The code is publicly available at:
| {
"pile_set_name": "ArXiv"
} |
---
author:
-
bibliography:
- 'ConstantRecognition.bib'
title: Criteria for the numerical constant recognition
---
[Introduction]{}
Experimental mathematics become very important nowadays [@BorweinBailey]. Using modern computers we are able to perform sequential brute-force searches of various mathematical objects and models at stunning scale. Multi-core processors [@epyc], GPUs, vectorization, FPGA and future quantum computers allow us to reach depths incomprehensible by humans. In essence, this is some form of artificial intelligence [@Eurega], because obtained results are indistinguishable from those provided by reasoning or more sophisticated algorithms [@SR]. Particularly interesting problem is a numerical constant recognition [@plouffe]. We are given decimal expansion, and ask if there is equivalent formula? There is strong demand from users of mathematical software to provide such a feature (identify/Maple [@Maple], nsimplify/SymPy [@symPy], RIES [@RIES], FindFormula [@Mathematica] ). In typical situation we encounter some decimal number, resulting from numerical simulation or experiment, e.g: 1.82263, and ask ourselves if it is equivalent to some symbolic expression, like e.g: $\sqrt{2}^{\sqrt{3}}$. Unfortunately, such a problem without additional constraints is ill-posed, and provided ”answers” nonsense. Often ridiculously complicated [@ISCwayback]. This is not a surprise, as Cardinality of real/complex transcendental numbers is uncountably infinite (continuum $\mathfrak{c}$), while number of formulas and symbols is countable ($\aleph_0$). Therefore probability for randomly chosen real number to be equivalent to some formula is zero. We must therefore restrict to numbers with finite decimal expansion. This is also practical approach, due to widespread of floating-point hardware, with double precision [@double] being *de facto* standard. However, in this case we encounter infinity from the other side. Assuming floating-point constant is a truncated real number[^1], we can produce arbitrary number of formulas which numerically differ only at more distant decimal places: $1.8226300001, 1.8226300000009 ,1.822630000100017$ etc. Fortunately, one still can ask which one of these formulas has *lowest* Kolmogorov complexity [@complexity]. But to do this, we must specify ”programing language” first. This lead to the following formulation of the numerical constant recognition problem. Imagine a person with hand-held scientific calculator, who secretly push a sequence of buttons, and print out the decimal/numerical result. We ask: could this process be reversed? In other words, given numerical result, are we able to recover sequence of some particular calculator buttons? Since every meaningful combination of buttons is equivalent to some explicit, closed-form [@closed], mathematical formula, above thought experiment becomes practical formulation of the constant recognition problem. In real world implementation, human and calculator are replaced by a computer program[^2], increasing speed billion-fold. Mathematically, results which can be explicitly computed using hand-held scientific calculator are members of so-called $\exp-\log$ (EL) numbers [@chow]. Above formulation has many advantages over set-theoretic or decimal string matching approach. It is well-posed, tractable, and can be either answered:\
a) positively,\
b) in terms of probability, or\
c) falsified.\
This depend on both maximum length of button sequence (code length, Kolmogorov complexity), as well as precision of the numerical result. One could anticipate, that for high-precision numbers (arbitrary precision in particular) and short code length identification must be unambiguous. For intermediate case (double precision, unspecified code length) we expect to provide some probability measure for identification. For low-accuracy numbers (e.g. of experimental or Monte Carlo origin) failure seems inevitable. Article is devoted to quantify above considerations. Certain statistical criteria are presented for practical number identification. Hopefully, this will convince researchers, that constant recognition problem can be solved practically. Instead of software used mainly for recreational mathematics and random guessing, it could become reliable tool for numerical analysis and computer algebra systems.
Article is organized as follows. In Sect. \[ELnumbers\] we discuss simplified computer language (in replacement for physical scientific calculator) used to define Kolmogorov complexity of the formulas. In Sect. \[enum\] we discuss some practical issues regarding numerical implementation of the sequential formula generators. Then, we combine experience/knowledge of Sect. \[ELnumbers\] and \[enum\] to find properties and typical behavior of the constructed EL numbers. Sections \[Criterion1\]-\[Criterion3\] propose three criteria for number recognition: (1) instant error drop-off to machine epsilon instead statistically expected $e$-folding; exponential growth of number of formulas indicates failure of search (2) maximum likelihood formula in the view of statistical process (3) compression ratio of the decimal constant in terms of RPN calculator code.
[Elementary complex $\exp-\log$ numbers \[ELnumbers\]]{}
Our first task is to precisely define set of formulas/numbers we want to identify using decimal expansion. We restrict ourselves to explicit formulas [@chow] i.e. those computable using hand-held scientific calculator. We assume root-finding procedure is absent. Therefore, as notable exception, most of algebraic numbers are not in above class, because polynomials of the order 5 and larger are not explicitly solvable in general. To be specific, we consider any complex number created from:
1. integers: $0,1,-1,2,-2, \ldots$
2. rationals: $\frac{1}{2},\frac{1}{3}, \frac{3}{2}, \ldots$
3. mathematical constants: $0, 1, 2, e,\pi, i, \phi, \ldots$
4. addition/subtraction, multiplication/division, exponentiation/logarithm (arbitrary base)
5. elementary functions of one variable:
- trigonometric: $\sin, \cos, \tan$
- inverse trigonometric (cyclometric): $\arcsin, \arccos, \arctan$
- hyperbolic: $\sinh, \cosh, \operatorname{arsinh}, \operatorname{arcosh}, \operatorname{tgh}, \operatorname{artgh}$
- $\sqrt{\quad} \;, \; \lg_2 \; , \; \ln \;, \; \exp$
- usually unnamed functions: $1/x, x+1, -x, x^2, 2x, x^x \ldots$
6. function composition
The choice of complex instead of real field is justified by mathematical simplicity. In practical numerical implementation it might be sub-optimal, due to reduced numerical performance and still missing implementation of some elementary functions (e.g: `clog2`) in standard libraries [@glibc]. Last item in the above list, function composition, is important for completeness. E.g: $\ln{(\ln{(\ln{x})})}$ is well-defined elementary expression, although rarely used in practice. No special functions are used.
Particularly troublesome are sets of symbols related to integers and rationals, as they both are potentially infinite. Without proper handling enumeration quickly get stuck on rational approximations with large integer numerator/denominator. Therefore, we must restrict to some ”small” subset of them. Standard computer languages (C, C++, Fortran) use ”small” integers still too large from our point of view, e.g. 8-bit signed and unsigned ones. Complexity of numbers 0, $2^{7}=128$, 188, …is identical using such an approach. The other extreme follows reductionist definition [@Tarski], staring from e.g: $-1$, and constructing all other integers as follows: $$1 = (-1) \times (-1), \; 0 = 1 + (-1), \; 2=1+1, \ldots$$ In calculators, ten digits are present: 0,1,2,3,4,5,6,7,8,9 and integers are entered in sequence using standard positional numeral system. RIES [@RIES] follow similar way, except for 0 and 1. Use of positional base-10 numerals is simple reflection of human anatomy/history, and difficult to include in our virtual calculator. Therefore we restrict to really small magnitude integers, $-1,0,1,2$ in particular. All other integers must be constructed from above four. Noteworthy, these integers are also among essential mathematical constants included in famous [@e] Euler formula: $$\label{euler}
e^{i \pi} + 1 = 0, \qquad e^{2 \pi i} - 1 = 0.$$ Two is also smallest possible integer base for logarithm $\log_2$. Binary system is now used in virtually all modern computer hardware.
The same reasoning applies to rationals. One might enumerate them using Cantor diagonal method, Stern-Brocot tree, or generate unambiguously by repeated composition[^3] of a function: $$\operatorname{next}(x) = \left(1 + 2 \lfloor x \rfloor -x \right)^{-1},$$ starting from zero: $$0,1,\frac{1}{2},2,\frac{1}{3},\frac{3}{2},\frac{2}{3},3,\frac{1}{4},\frac{4}{3},\frac
{3}{5},\frac{5}{2},\frac{2}{5}, \ldots$$ Usually the best is to leave language without explicit rationals, and let them appear by division of integers. Alternatively, addition of reciprocal $\operatorname{inv}(x) \equiv 1/x$ and successor $\operatorname{suc}(x) \equiv x+1$ functions allow for construction of continued fractions, see Appendix B in Supplemental Material.
Set of (named) mathematical constant certainly must include $e$, $\pi$, and $i = \sqrt{-1}$ due to , as long as we consider complex numbers. Besides this, some mathematicians consider other constants [@const] as ”important”, like first square roots ($\sqrt{2}, \sqrt{3}, \sqrt{5}$), golden ratio $\phi = (1+\sqrt{5})/2$ [@RIES], $\ln{2}$, $e^{1/e}$ and so on. Noteworthy, all of them are itself in $\exp-\log$ class, i.e., we can generate them from integers and elementary functions. They do not define any new numbers. Their inclusion only alter language definition and therefore Kolmogorov complexity. Other situation is with inclusion of mathematical constants of unknown status/memberships (unknown to be rational or transcendental), like Euler gamma $\gamma = 0.577216\ldots $ or Glaisher constant $A=1.28243\ldots$. They may, or may not, extend $\exp-\log$ class. This is itself an application for constant recognition.
Commutative operations like addition and multiplication are in general $n$-ary. To simulate calculator behavior we must treat them as repeated binary operations: $$x+y+z = (x+y) + z.$$
Elementary functions are not independent, e.g: $$\sinh{x} = \frac{e^x - e^{-x}}{2}, \tan{x} = \frac{\sin{x}}{\cos{x}}.$$ Scientific calculators sometimes distinguish less important, secondary functions which require pressing two buttons. But mathematically they are completely equivalent. You can compute $\sinh, \cosh$ using $\exp$, or *vice versa*. This is especially true in complex domain, where all elementary functions are reducible to $\exp$ and $\ln$.
Reduction is possible for binary operations as well, e.g: $$x \times y = e^{\ln{x} + \ln{y}}, \quad x \cdot y = \log_x{\left[ (x^y)^x \right] }.$$ Replacing multiplication by logarithms and addition is an achievement of medieval mathematics [@briggs] used without changes for 300 years. It was replaced with slide rule used in XX century. They both went extinct with modern computers. However, we point out, that using logarithms and addition you can not only do bottom-up translation to multiplication/exponentiation. You might go opposite way as well, from high-rank (grade) hyper-operation down do additions. Expressing addition/multiplication by exponentiation/logarithms only (up-bottom) is therefore possible, but tricky, see Appendix C in Supplemental Material.
Taking above considerations into account, important question arises: how many constants, functions and binary operations are required for our virtual calculator to be still fully operational? How far reduction process can go, without impairing our computational abilities? This is also known as ,,broken calculator problem” [@RIES]. Another related question is, if such a reduced calculator is optimal for the task of constant recognition. Obviously, maximally reduced button set makes theoretical and statistical analysis convenient, therefore it is very useful as mathematical model. In practice, as we will show later, Kolmogorov complexity of formulas in maximally reduced language, which are considered as simple by humans, might become surprisingly large. On the contrary, formulas *simple* in reduced operation set, like nested power-towers, look inhuman, and are out of scope traditional mathematical aesthetics, despite small Kolmogorov complexity.
The simplest possible language we found has a length three ($n=3$). It is still able to perform all operations of the scientific calculator. It includes:
\[base-3\] $$\begin{aligned}
\label{e}
\text{either} & \qquad e \quad \text{or} \quad \pi. \end{aligned}$$ $$\begin{aligned}
\text{binary exponentiation} & \qquad x^y,\end{aligned}$$ $$\begin{aligned}
\text{arbitrary base (two-argument) logarithm} &\qquad \log_x{y}.\end{aligned}$$
Another very simple language of length four is:
\[base-4\] $$\begin{aligned}
\text{any constant}& \qquad x.\end{aligned}$$ $$\begin{aligned}
\label{e1}
\text{natural exponential function}& \qquad \exp{x} \equiv e^x,\end{aligned}$$ $$\begin{aligned}
\label{e2}
\text{natural logarithm}& \qquad \ln{x} \equiv \log_e{x},\end{aligned}$$ $$\begin{aligned}
\text{subtraction}& \qquad x-y.\end{aligned}$$
Detailed proofs by exhaustion are presented in Appendices C,D in Supplemental Material. Above two examples clearly justify name $\exp-\log$ for class of explicit elementary numbers discussed in this section. We were unable to find any shorter languages. At least one (noncommutative?) binary operation seem required to start abstract syntax (AST) tree growth (see however Appendix B), and at least one constant to terminate leafs. One of the essential constants $e$ or $\pi$ probably also is required, either explicitly, like in , or hidden in function/operation definition (\[e1\],\[e2\]). But we cannot prove that further reduction to 2 buttons is impossible. We also cannot prove, that above language of length three is unique. Therefore, one cannot estimate Kolmogorov complexity of EL number unambigously. Existence of both smallest and unique language, generating all $\exp-\log$ numbers, would provide natural enumeration of mathematical formulas, similar to Peano arithmetic for natural numbers. Anyway, or are the sets of irreducible operations, for now. Calculator/language can be extended, but under no circumstances any of the buttons defined by or can be removed. Failure to abide by above requirement might result in catastrophic failure of the number recognition software even in the simplest test cases. This is plague of existing implementations, letting unaware end-users to choose base building blocks on their own. This leaves impression of random failures without any obvious reason. Users can use their own set of constants, functions and binary operations, but they must be merged with irreducible set or .
It is illustrative to compare and with *Mathematica* core language [@Mathematica], composed of:
\[mma\] $$\begin{aligned}
\text{addition (Plus)} & \qquad x+y,\end{aligned}$$ $$\begin{aligned}
\text{multiplication (Times)} & \qquad x \times y,\end{aligned}$$ $$\begin{aligned}
\text{exponentiation (Power)} & \qquad x^y,\end{aligned}$$ $$\begin{aligned}
\text{natural logarithm (Log)} &\qquad \ln{x},\end{aligned}$$ $$\begin{aligned}
\text{mathematical constants (E, Pi, I)} \qquad e, \pi, i, \ldots\end{aligned}$$
augmented with arbitrarily large integers, including their complex combinations (Complex) and rationals (Rational). Visual comparison of , and is presented in Fig. \[BaseSet\].
![\[BaseSet\] Illustration of base constants, functions and binary operations used to define $\exp-\log$ number class. Irreducible set is labeled ”up-botton”, because it is based on highest rank complex (hyper-)operation currently implemented (exponentiation), while (bottom-up) starts with low-rank arithmetic (subtraction). *Mathematica* core language is included for comparison. All three sets allow generation of any explicit elementary transcendental constant in the EL class. Question mark symbolize unknown ”holy-grail” complex function, which could, if exists, by repeated composition, generate all $\exp-\log$ numbers in unique order. ](figs/BaseSet){width="50.00000%"}
Accounting for the above considerations, for further analysis we have selected four base sets (buttons) to attack the inverse calculator problem:
\[CALC\]
1. `E`, `POW`, `LOG` \[$n=3$ buttons, eq. \] $$\label{CALC1}$$
2. `X`, `EXP`, `LN`, `-` \[$n=4$ buttons, eq. \] $$\label{CALC2}$$
3. `Pi`, `E`, `I`, `-1`, `2`, `1/2`, `Log`, `Plus`, `Times`, `Power` \[$n=10$ buttons, eq. \] $$\label{CALC3}$$
4. full scale scientific calc \[$n=36$ buttons, see below\] $$\label{CALC4}$$
Calculator 1 (up-bottom), based on is the simplest found. It use one constant ($e\simeq 2.71828\ldots$, PUSH on stack operation), no functions, and only two non-commutative binary operations: arbitrary base logarithm $\log_x{y}$ and exponentiation $y^x$. Noteworthy, AST is a binary tree in case of .
Calculator 2 (bottom-up), based on is the second shortest, and have remarkable property: it can use any numerical constant/symbol $x$ to generate all other numbers. For example, $0 = x - x$, $1 = \exp{(0)}$, $e=\exp{(1)}$ and so on. We can use any of $x=0$, $x=-1$, $x=e$, $x=i$, $x=\phi$, …This property can be exploited in two applications: (1) use our calculator to operate on vectors, (2) use $x$ with large Kolmogorov complexity to ”shift” formula generator. The former can be used e.g. to take advantage of modern CPU AVX extensions, the latter to use $x$ as some form of random seed for enumeration procedure. This can be done extending any of calculators with $x$ as well, but only for it is visible *explicite* in definition. In presentation of results we used $x=2$.
Calculator 3 is designed to mimic *Mathematica* behavior. Constants beyond $\pi,e$ and $i$ were chosen as follows. To enable rapid integer generation via addition, multiplication and exponentiation we use -1 and 2. Incidentally, these numbers are also are initial values for Lucas numbers. Rational numbers are easily generated via continued fractions, thanks to $-1$ constant, allowing for construction of reciprocal $1/x \equiv x^{-1}$. Rational constant $1/2$ together with $i$ form trigonometric functions $e^{ix} = \cos{x} + i \sin{x}$ and square roots $\sqrt{x} = x^{1/2}$. Number of instructions equal to ten is selected intentionally. It will allow for instant estimate of the compression factor in Sect. \[Criterion3\]. Both target numerical constant and RPN calculator code can be expressed as string of base-10 digits `0123456789`, see next Sect. \[enum\].
Fourth calculator is the largest one. Maximum number of 36 buttons is limited by a current implementation of the fast `itoa` function used to convert string variables into base-36 numbers, i.e. alphanumeric lowercase digits and letters. It is the most close to what people usually expect from scientific hand-held device. Full list of buttons:
- constants: $1,2,3,4,5,6,7,8,9,e,\pi,i,\phi$
- functions:$\ln, \exp, \operatorname{inv}, \operatorname{minus}, \mathrm{sqrt}, \operatorname{sqr},$ $\sin, \arcsin, \cos, \arccos, \tan, \arctan,$ $\sinh, \operatorname{arsinh}, \cosh, \operatorname{arcosh}, \operatorname{tgh}, \operatorname{artgh},$
- binary operations: $+, -, \times, / , y^x$.
”Full” calculator defined above use 13 constants, 18 functions of one variable, and 5 binary operations.
[Efficient formula enumeration \[enum\]]{}
Once calculator buttons, equivalent to some truncated computer language specification, are fixed, we face the next task: enumeration of all possible formulas. We exclude recursive implementation. While short and elegant, it quickly consumes available memory, and is hard to parallelize [@RIES]. For sequential generation, we notice that formulas are equivalent to set of all abstract syntax trees (AST) or valid reverse polish [@RPNorig] notation (RPN, [@RPN]) calculator codes. The former description has an advantage in case where efficient tree enumeration algorithm exists [@Knuth], e.g: binary trees [@tree]. Unfortunately, this is applicable only to the simplest calculator . Other mix binary and unary trees. Therefore, to handle variety of possible calculators, including future extension to genuine $n$-ary special functions (e.g: hypergeometric, Meijer G, Painleve transcendents) our method of choice is enumeration of RPN codes. This has two major advantages. First, enumeration is trivially provided by `itoa` function with base-$n$ numbers (including leading zeros), where $n$ is number of buttons. Second, while majority of enumerated codes is invalid, checking RPN syntax is very fast, almost negligible compared to `itoa` itself, let alone to computation of complex exponential and logarithmic functions. Procedure has two loops: outer for code length $K$, inner enumerating $n^K$ codes of length $K$. Inner loop is trivial to parallelize using OpenMP directives without effort, scaling linearly with number of physical cores. Moreover, hyper-threading is utilized as well, although scaling is only at quarter of core count. Therefore, prospects for high-utilization of modern high-end multi-core CPU’s [@epyc] are looking good. Digits are associated with RPN calculator buttons. For simplest case they are ternary base digits `012`, which were assigned to three RPN calculator buttons as: 0 $\to$ E ($e$), 1 $\to$ LOG ($\log_x{y}$), 2 $\to$ POW ($x^y$). Detailed example of the algorithm for is presented in Appendix A of Supplemental Material.
Combinatorial growth of the enumeration is characterized, in addition to Kolmogorov complexity (RPN code length) $K$, by a total number of possible codes tested so far $k_1$. It grows with $K$ as: $$k_1 = \sum_{K'=1}^{K} n^{K'} = \frac{n^K-1}{n-1} n.$$ Total number of syntactically correct codes $k_2$ and total number of *unique* numbers $k_3$ are additonal useful characterizations of the search depth. Obviously, $k_3 \leq k_2 \leq k_1$. For perfectly efficient enumeration $k_1=k_2=k_3$. Unfortunately, this is currently possible only for rational numbers, see Appendix B. Case $k_1=k_2$ is equivalent to knowledge of unique tree enumeration algorithm. To achieve $k_2=k_3$ ($p=1$ in Fig. \[k2k3\], dashed diagonal line) algorithm must magically somehow know in advance all possible mathematical simplifications. Not only those trivial, like $e^{ \ln{x} } = \ln{(e^x)}=x$ but anything mathematically imaginable, e.g, $2^{1-\ln{\ln{\pi}}} = 2 (\ln{\pi})^{-\ln{2}}$. We doubt this is possible at all. Some ideas, based on solved rational numbers enumeration, are presented in Appendix B. In practice, we achieved $k_3/k_2 = 0.59, 0.06, 0.08, 0.49$ for calculators 1-4 , respectively. Large $k_3/k_2$ ratio for is a result of very simple tree structure, in which only every nine-th odd-$K$ RPN codes are valid. Without taking this into account, $k_3/k_2=0.59/2/9 = 0.03$, i.e., the worst of four. Full calculator perform surprisingly well.
![\[k2k3\] Number of unique formulas $k_3$ *versus* number of valid codes $k_2$. Solid lines show $k_3 = {k_2}^p$ fits, with $p=0.96491, 0.826455, 0.848559, 0.948259$ for calculators 1-4, respectively. Dashed diagonal line shows perfect case $k_3=k_2$. ](figs/k2k3){width="50.00000%"}
[Quasi-convergence of enumerated formulas and constant recognition \[Criterion1\] ]{}
Enumerated formulas provide sequence of elementary transcendental $\exp-\log$ (EL) numbers. For they are, in order of increasing Kolmogorov complexity $K$: $$\begin{aligned}
\label{ELseq}
K=1&\qquad e, \\
K=2&\qquad \text{none}, \nonumber \\
K=3&\qquad 1, e^e, \nonumber \\
K=4&\qquad \text{none}, \nonumber \\
K=5&\qquad 0, 1/e, e^{e^e}, e^{e^2}, \nonumber \\
\ldots \nonumber\end{aligned}$$ where repeated numbers were omitted. Position within the same level of $K$ is unspecified, but follows from presumed enumeration loop and base-$n$ digits association with buttons. From sequence we can extract subsequence of progressively better approximations for target number $z$, e.g., using example from the introduction, $z=1.8226346549662422143937682155941\ldots$. Analysis of sequences obtained this way is main goal of current section. We are interested in convergence properties and criteria for termination of sequence.
Let us assume for the moment, that we are able to provide on demand as many decimal digits for target number $z$ as required. There are two mutually exclusive possibilities: (i) number $z$ is in $\exp-\log$ class, defined e.g. by or , or (ii) $z$ is true non-elementary transcendental constant outside EL class. In the former case, we expect that error eventually, at some finite $K$, will drop to an infinitesimally small value, limited only by currently used precision. Search algorithm could then terminate and switch to some high-precision or symbolic algebra verification method. Unfortunately, equivalence problem is undecidable in general [@richardson]. One cannot exclude, that both numbers deviate at some far more distant decimal place. In the case of (ii) i.e. truly transcendental number, however, sequence will converge indefinitely, by generation of approximations, with progressively more complicated formulas. Realistic convergence examples, computed using extended precision (`long double` in C) to prevent round-off errors, are presented in Figure \[SmokingGun\]. The example target $z=\sqrt{2}^{\sqrt{3}}$ from introduction was selected. Number is obviously in $\exp-\log$ class. But is not *explicite* listed in any of , so it must be entered using appropriate sequence of buttons. For RPN calculator it is obvioulsy 2, SQRT, 3, SQRT, POW with $K=5$. In Fig. \[SmokingGun\] is is visible at $K=5.28$, where decimal part show inner loop progress.
![\[SmokingGun\] Typical behavior of absolute error $\epsilon$ for subsequence of progressively better approximations in terms of $\exp-\log$ formulas generated with ”calculators” , as a function of Kolmogorov complexity $K$. For we have used $x=2$. For (blue) and (black) sudden error drop-off to machine epsilon (zero) is observed for small $K$, marking possible ”exact” formula match. Calculations were done using extended precision (`long double` in C) while horizontal line mark machine epsilon for IEEE 754-2008 (binary64). ](figs/Convergence1){width="50.00000%"}
Two types of behavior can be observed in Fig. \[SmokingGun\]. Vertical absolute error $\epsilon$ scale is adjusted to range $\{ 2^{-63}, 1 \}$, i.e., `long double` machine epsilon for this and related Figs. \[Convergence2\], \[Convergence3\]. In typical situation absolute error $\epsilon$ for best approximation decreases exponentially with code length. However, if ”true” $\exp-\log$ formula is encountered, error instantly drops off to limiting value. In example from Fig. \[SmokingGun\] it is small multiple of machine epsilon for `long double` precision ($2^{-63}$) or binary zero.
Criterion directly based on Kolmogorov complexity is inconvenient, if one deals with languages of various size, like our calculators 1-4, eq. (\[CALC\]). Convergence rate also depends on language size $n$, cf. Fig. \[SmokingGun\]. One could use Kolmogorov complexity corrected for language size, or compile formulas generated in extended calculators or down to one of primitive forms given by or . The former approach usually underestimate, and the latter heavily overestimates true complexity. Therefore we propose another criterion, independent of the language used to generate formulas. Instead of plotting error as a function complexity, we plot $N$-th best approximation (Fig. \[Convergence2\]). Now all curves nearly overlap, and observed lower error limit is $\epsilon > e^{-N}$.
Above considerations provide the first criterion for constant identification:
Criterion 1:\
Identification candidate: *if absolute error $\epsilon$ in subsequence of progressively better approximations in terms of $\exp-\log$ formulas deviates ”significantly” from estimated upper limit $e^{-N}$ (drops to numerical ”zero”/machine epsilon in particular) we can stop search and return formula code, as possible identification candidate.*
Failure of search: *if absolute error in subsequence of progressively better approximations in terms of $\exp-\log$ formulas follow $e^{-N}$ and reach numerical precision limit, or computational resources are exhausted, search failed.*
Candidate code must be then verified using symbolic methods, high-precision numerical confirmation test, and ultimately proved using standard mathematical techniques. Above criterion do not provide any numerical estimate for probability of successful identification. However, we point out, that even in case of possible misidentification, unexpected drop of error to machine epsilon marks stop of the search anyway. This is because finding better approximation would require formula with complexity already above threshold, given by intersection of dotted lines in Fig. \[Convergence2\]. Beyond that, number of formulas with identical decimal expansion grows exponentially. This behavior mark search STOP criterion for cases, where ”smoking gun” feature from Fig. \[SmokingGun\] was not encountered. In practice this still require a lot of computational resources, beyond capabilities of mid-range PC/laptop. That is why curves in Figs. \[SmokingGun\]-\[Convergence3\] are still far from double epsilon, marked with `DBL_EPSILON` dotted line.
![\[Convergence2\] Similar to Fig. \[SmokingGun\], but $K$ has been replaced by number of subsequent best approximations $N$ found so far. This allows for direct comparison of languages of different length. Possible identification is marked by error for next approximation significantly below $e^{-N}$.](figs/Convergence2){width="50.00000%"}
Expected decrease of approximation error for next best one is $1/e \simeq 0.37$ of previous. Therefore, we use $e$-folding name for Criterion 1.
![\[Convergence3\] Similar to Fig. \[Convergence2\], but $N$ has been replaced by number $k_2$ of valid RPN codes tested so far. Power-law is observed, with $\epsilon \propto k_2^{-1}$. Upper axis show estimated by power-law fit $k_3 \simeq k_2^{0.83}$ number of *unique* formulas.](figs/Convergence3){width="50.00000%"}
Remarkable observation is provided by yet another Figure \[Convergence3\]. Instead of $K$ or $N$ we used total number $k_2$ of *valid* RPN codes tested before encountering next approximation. Power law behavior is found, and band of data points (Fig. \[Convergence3\]) can be roughly approximated as proportional to $1/k_2$. Therefore, obtaining definite negative answer for constant known with machine precision of $\sigma$ in terms of Criterion 1 require testing of $\sigma^{-1}$ codes. For double precision machine epsilon it is above $10^{15}$. You need either hundreds of CPU cores, or a lot of patience (days of search). Positive identification can be much faster, of course, like for calculators 3, 4 in Figures \[SmokingGun\], \[Convergence2\].
[Statistical properties of the $\exp-\log$ numbers \[Criterion2\] ]{}
Observation from Fig. \[Convergence3\] and results of previous section lead to another view of constant identification problem. It can be described as a random process. Consider following numerical *Monte Carlo* experiment. We generate pseudo-random numbers $\xi>0$ from exponential distribution with some scale $\lambda$ and Probability Distribution Function (PDF): $$\label{Pexp}
P(\xi) = \frac{e^{-\xi/\lambda}}{\lambda}.$$ Next, we compare $\xi$ with our target number $z>0$, and generate sequence of progressively better approximations. Surprisingly, observed behavior is similar to Fig. \[Convergence2\], but with larger fluctuations. Let’s further assume $z$ is known with numerical precision $\sigma < z$. In other words, true number is in the interval $[z-\sigma,z+\sigma]$. Assuming we can easily obtain probability of random hit into vicinity $z$: $$P(z,\sigma, \lambda) = \int_{z-\sigma}^{z+\sigma} P(\xi) \; d \xi$$ what gives average number of tries: $$k_2 = \frac{e^{z/ \lambda}}{2 \sinh{\sigma/\lambda}} \sim \frac{const}{\sigma}.$$ Above agrees qualitatively with result presented in Fig. \[Convergence3\] if $\sigma \ll z$.
However, statistical distribution of $\exp-\log$ numbers is unknown, PDF was chosen intuitively. Statistical properties of EL numbers should not depend on on language used to generate them, at least in the limiting case of large complexity. Numerical evidence, obtained by collecting real numbers generated according to procedure presented in Sect. \[enum\], is presented in Fig. \[StatReals\]. Numbers with $\operatorname{Im}(z) \neq 0$ were discarded[^4] to simplify analysis and presentation.
![\[StatReals\] Histogram of first 20k real numbers generated with use of calculators 1-4. Dashed line show exponentially transformed Cauchy distribution . ](figs/StatReal){width="50.00000%"}
None of the widespread distributions match numbers presented in Fig. \[StatReals\]. The most close are: Pareto, Levy and Cauchy distributions. Noteworthy, all of them have undefined mean and infinite dispersion. If we look at statistics of base 10 logarithm, it mimic t-Student, Cauchy and Laplace (double $\exp$) distributions. Probably this is indeed a brand new statistics, following some empirical distribution. True distribution of real EL numbers (Fig. \[StatReals\]) is, however, quite well described by transformed Cauchy distribution with PDF: $$\label{cauchy}
P(x) = \frac{1}{\pi x} \frac{1}{1+(\ln{x})^2}.$$ Therefore, later, statistical distribution of (real) EL numbers will be approximated by Cauchy-like distribution .
Let us assume statistical distribution of real EL constants has PDF $P(x)$ empirically derived from histogram in Fig. \[StatReals\] or is given by . Let the target number $z$ is know with precision $\sigma$, and statistical distribution of error for $z \pm \sigma$ is $Q(x, z, \sigma)$. $Q$ might be normal or uniform distribution, for example. Then conditional probability $p_k$, that $k$-th tested formula **is not random** match for target value $z \pm \sigma$ is: $$\begin{aligned}
p_k = \prod_{i=1}^k \left( 1- \frac{\int_{-\infty}^{\infty} P(\xi) Q(\xi) \; d \xi}{Q(z)} \right) = \\ \nonumber
=
\left( 1- \frac{\int_{-\infty}^{\infty} P(\xi) Q(\xi) \; d \xi}{Q(z)} \right)^k.\end{aligned}$$
Then, replacing $P(\xi) \simeq P(z)$ and integrating, likelihood $\mathcal{L}$ for proper identification of target constant $z$ with i-th unique value $x_i$ is: $$\label{Id_likelihood}
\mathcal{L} = \left[ 1- \frac{P(z)}{Q(z)} \right]^{k_3} Q(x_i, z, \sigma),$$ where $i=k_3$, i.e. number of unique values tested so far. To understand , it might be approximated by $$\label{Id_likelihood2}
\left( 1- \kappa \, \sigma \; P(z) \; k_3 \right) Q(x_i, z, \sigma),$$ assuming $\sigma \ll z$ and using Maclaurin expansion $(1+\alpha)^k \simeq 1 + \alpha k + \ldots$. Value of $\kappa$ depends on error distribution $Q$, e.g., $\kappa=2$ for uniform and $\kappa=\sqrt{2 \pi} \simeq 2.5$ for normal distribution.
Likelihood becomes zero for large $k_3$, where search for $k_3>1/\sigma$ is pointless anyway due to Criterion 1 (Sect. \[Criterion1\]). The most intriguing application of is when $\sigma$ is still large compared with machine epsilon, and has uncertainty of statistical nature. This allow for selection of maximum likelihood formula(s) in marginal sense, i.e. for $k_3 \sim \sigma^{-1}$. Because calculation of $k_3$ requires storage of all previously computed values, it is reasonable to replace it with $k_2$ using power-law fit from Fig. \[k2k3\]:
$$\label{Id_likelihood3}
\left[ 1- \kappa \; {k_2}^p \; \sigma\; P(z) \right] \times Q(x_i, z, \sigma).$$
Noteworthy, as a function of $k_2$, i.e. number of tested syntactically correct RPN codes, must have a maximum. Formula for $x_i$ with maximum probability is well defined. For limited precision maximum is very pronounced (Fig. \[likelihood\]), but for very small $\sigma$ we might not be able to reach it at all, due to limited computational resources (e.g: Fig. \[likelihood\], solid blue). Equations (\[Id\_likelihood\]-\[Id\_likelihood3\]) can be understood as follows. If we find remarkably simple and elegant formula at the very beginning of the search, which is however quite far from target $z$, e.g., several $\sigma$’s in Gaussian distribution, we reject it on the basis of improbable error in measurement/calculation. Similarly, if we find formula well within error range, e.g. $4 \sigma$, but after search covering billions of formulas, we reject it expecting random coincidence. Somewhere in the middle lays optimal formula, with maximum likelihood, estimated with use of . In practice, likelihood is very small, and $\log$-likelihood is more convenient: $$\log \mathcal{L} = -k_3 \ln \left(1 - \frac{P(z)}{Q(z,z,\sigma)} \right) + \ln Q(x_i,z,\sigma).$$ Expanding $\ln{(1-\epsilon)} \simeq -\epsilon$ and assuming $Q$ is Gaussian, we have following useful approximation for $\log$-likelihood: $$\label{logL}
\log \mathcal{L} = k_3 \frac{P(z)}{Q(z,z,\sigma)} - \ln{(\sqrt{2 \pi} \sigma)} - \frac{(x_i-z)^2}{2 \sigma^2}.$$
We note, that probability of the identification related to could be, in principle, estimated directly, especially for single precision floats, as there are only $2^{32}$ of them. Using floats as bins for sequentially generated EL numbers, we can fill them with numbers associated with their complexity.
![ \[likelihood\] Log-likelihood of subsequent best $\exp-\log$ approximations, as estimated from . Red, green, blue and black lines are for calculators 1,2,3 and 4, respectively. Target values are approximations to our introduction example $z=\sqrt{2}^{\sqrt{3}}$ in form $z \pm \sigma$ with: $z=1.82 \pm 0.005$ (dot-dashed, x), $z=1.823 \pm 0.0005$ (dotted, o), $z=1.8226 \pm 0.00005$ (dashed, +), $z=1.82263 \pm 0.000005$ (solid, \#). ](figs/Likelihood1 "fig:"){width="50.00000%"}\
![ \[likelihood\] Log-likelihood of subsequent best $\exp-\log$ approximations, as estimated from . Red, green, blue and black lines are for calculators 1,2,3 and 4, respectively. Target values are approximations to our introduction example $z=\sqrt{2}^{\sqrt{3}}$ in form $z \pm \sigma$ with: $z=1.82 \pm 0.005$ (dot-dashed, x), $z=1.823 \pm 0.0005$ (dotted, o), $z=1.8226 \pm 0.00005$ (dashed, +), $z=1.82263 \pm 0.000005$ (solid, \#). ](figs/Likelihood2 "fig:"){width="50.00000%"}\
![ \[likelihood\] Log-likelihood of subsequent best $\exp-\log$ approximations, as estimated from . Red, green, blue and black lines are for calculators 1,2,3 and 4, respectively. Target values are approximations to our introduction example $z=\sqrt{2}^{\sqrt{3}}$ in form $z \pm \sigma$ with: $z=1.82 \pm 0.005$ (dot-dashed, x), $z=1.823 \pm 0.0005$ (dotted, o), $z=1.8226 \pm 0.00005$ (dashed, +), $z=1.82263 \pm 0.000005$ (solid, \#). ](figs/Likelihood3 "fig:"){width="50.00000%"}\
![ \[likelihood\] Log-likelihood of subsequent best $\exp-\log$ approximations, as estimated from . Red, green, blue and black lines are for calculators 1,2,3 and 4, respectively. Target values are approximations to our introduction example $z=\sqrt{2}^{\sqrt{3}}$ in form $z \pm \sigma$ with: $z=1.82 \pm 0.005$ (dot-dashed, x), $z=1.823 \pm 0.0005$ (dotted, o), $z=1.8226 \pm 0.00005$ (dashed, +), $z=1.82263 \pm 0.000005$ (solid, \#). ](figs/Likelihood4 "fig:"){width="50.00000%"}
Discussion above provides second criterion for constant identification:
Criterion 2:\
Identification candidate: *if likelihood given by has reached maximum, formula has the highest probability, and should be returned.*
One might return a few highest likelihood formulas near maximum as well, or one with largest value so far, if computational resources are limited. Noteworthy, likelihood value provide also relative quantitative estimate of identification probability. We may estimate likelihood , using calculator , for $z=1.82263\pm0.00005$ (Fig. \[likelihood\], solid blue line) to be identified as $\sqrt{2}^{\sqrt{3}}$ $$\log \mathcal{L} = 8.33 \qquad (\mathcal{L} = 4 \times 10^{3})$$ compared to $$\log \mathcal{L} = -62.3 \qquad (\mathcal{L} = 9 \times 10^{-28})$$ for $(\ln{4})^{\ln{2\pi}}$. From our Criterion 3, the former is more than $10^{30}$ times more probable than the latter. Without this sort of ”ranking”, discussion, if e.g. $\sinh ^{-1}\left(\phi \sinh ^{-1}(\pi )\right)$ is ”simpler” or ”more elegant” compared to, e.g., $\sin( \cosh{4} ) + \sqrt{\tanh{5}}$ might continue indefinitely, leading to nothing. Likelihood or provide numerical form of the Occam’s razor. It can be applied automatically, within software (CAS) environment.
[Constant recognition as data compression \[Criterion3\] ]{}
Sequence of progressively better approximations in form of RPN calculator codes can be viewed as a form of lossy compression. If exact formula is found, then compression becomes lossless. In the intermediate case, compression ratio provides measure, how good is some formula to recover decimal expansion. This process is illustrated in Table \[compress\]. Calculator 3 defined in eq. has been used, because both decimal expansion of the target number $z$ and RPN code are strings of the same base-10 digits, i.e. `0123456789`. Therefore, compression ratio is simply a number of correct digits divided by code length $K$.
--------------------------------- ----------- ------------- --------------------------------
Numerical Code Compression Formula
value ratio
\[2mm\]
\[2mm\] 3.141592653589793238512 0 0.00 $\pi$
2.718281828459045235428 1 0.00 $e$
0.000000000000000000000 2 0.00 $0$
2.000000000000000000000 7 0.00 $2$
**1**.718281828459045235428 164 0.33 $e-1$
**1**.772453850905516027310 809 0.33 $\sqrt{\pi}$
**1**.648721270700128146893 819 0.33 $\sqrt{e}$
**1.8**37877066409345483606 0043 0.50 $\ln{(\pi+\pi)}$
**1.8**37877066409345483606 7053 0.50 $\ln{(2 \pi)}$
**1.82**0796326794896619256 08485 0.60 $\frac{1}{2}(\pi+\frac{1}{2})$
**1.82**0796326794896619256 80485 0.60
**1.82**0796326794896619256 80845 0.60
**1.82**0796326794896619256 88045 0.60
**1.82**1126701185962651818 0338975 0.43 $2^{1-\ln{\ln{\pi}}}$
**1.82**1126701185962651818 7033895 0.43
**1.82**1126701185962651818 8303975 0.43 $2 (\ln{\pi})^{-\ln{2}}$
**1.82**4360635350064073446 8819745 0.43
**1.82**4360635350064073446 8781945 0.43
**1.82**4360635350064073446 8197485 0.43
**1.82**4360635350064073446 7819485 0.43
**1.82**1126701185962651818 7830395 0.43
**1.82**1662858741926632288 6091579 0.43
**1.822**361069544464599575 2298979 0.57
**1.822**413909696397869321 77408934 0.50
**1.822**722133555469366033 80790539 0.50
**1.8226**90334737686312645 004377539 0.56
**1.822634654966242214**488 888854979 2.11
--------------------------------- ----------- ------------- --------------------------------
: \[compress\]
In Table \[compress\], compression ratio of the target number $z=\sqrt{2}^{\sqrt{3}} \simeq 1.8226346549662422143937682155941\ldots$ in form of $n=10$ RPN calculator codes is shown.
Buttons/operations were assigned as follows:\
0 $\to$ ,\
1 $\to$ ,\
2 $\to$ ,\
3 $\to$ ,\
4 $\to$ ,\
5 $\to$ ,\
6 $\to$ ,\
7 $\to$ ,\
8 $\to$ ,\
9 $\to$ .
In fact, last code in Table \[compress\] `888854979`, with pronounced compression ratio of $19/9\simeq2.11 > 1$, is equivalent to exact formula, with RPN sequence , , , , $\times$, $+$, , 2, , i.e., $2^{\sqrt{3}/2}$.
In general case, compresion ratio $r$ is given by: $$\label{r}
r = \frac{-\log_{10} \max{(\epsilon, \sigma)}}{K \; \log_{10} n},$$ where $\epsilon$ is absolute precision of the approximation, $K$ - RPN code length, $n$ - number of calculator buttons.
Now we can formulate the last criterion for constant recognition:
Criterion 3:\
Identification candidate: *if compression ratio given by is $r \gg 1$, or $r$ reaches maximum in the course of search, then formula should be returned.*\
Failure: *if $r \ll 1$ formula/code is unlikely match.*
Criterion 3 has an advantage of being very simple. It do not require recorded history of search, like Criterion 1 (Sect. \[Criterion1\]), of searching for maximum and knowledge of statistics, like Criterion 2 (Sect. \[Criterion2\]). If indeed $r$ has a maximum, then it strenghten identification, but this is not required. You might ask for compresion ratio of any combination of decimal expansion and formula. The only required action is to compile formula to RPN code or similar one. Therefore, it will work with any searching method, e.g: *Monte Carlo*, genetic or shortest path tree algorithms. It is weak compared to $e$-folding or statistical criteria, but could easily exclude most of formulae produced in variety of software, especially very complicated ones, and those including large ($i \gg 2$) integers.
[Conclusions]{}
Three criteria proposed in the article provide robust tool for decimal constant identification.
![\[3C\] Recognition indicators as a function of $N$. Solid red line shows absolute $e$-folding $e_1$, while dotted red line relative $e$-folding $e_2$ (upper panel). Green line show likelihood (middle panel), and blue compression ratio (bottom panel).](figs/ThreeCriteria){width="50.00000%"}
Process of recognition, applying all three criteria, is illustrated in Fig. \[3C\], using blind-test target value of $z=201.06192983$. Assuming all digits are correct, we adopted (gaussian) error $\sigma=0.000000005$. Using calculator , we plot three indicators of matching quality. For $e$-folding measure we use $$e_1 = \frac{z}{\epsilon_N} e^{-N},$$ and additionally: $$e_2 = \frac{\epsilon_{N-1}}{\epsilon_N} \frac{1}{e},$$ where $N$-th best approximation error is $\epsilon_N$. For likelihood, we used , and for compression ratio . Clearly, three subsequent data points, for $N=16,17,18$, stand out (Fig. \[3C\]).
From first criterion perspective (Sect. \[Criterion1\]), error dropped nearly $10^5$ times compared to expected $e^{-N}$ (Fig. \[3C\], upper panel, solid red). Is also was $10^8$ times smaller compared to previously found $x_{15}=e^e (1/2+\pi)^2$, while statistically anticipated decrease was $1/e$ (Fig. \[3C\], upper panel, dotted red). Using criterion 2 (Sect. \[Criterion2\]) we found likelihood many many orders of magnitude larger compared to any other formulas. Three data points with nearly equal $\log L\sim 18.193$ (Fig. \[3C\], middle panel, green crosses), 16,17,18-th best approximations, are in fact the same formula typed using different RPN sequences. Differences are due to round-off errors. Shortest one is preferred in terms of criterion 3 (Sect. \[Criterion3\]), see Fig. \[3C\], lower panel. Constant $z$ is then unambiguously recognized by all three criteria as $64 \pi$, with no other candidates in sight.
Above example show possible way to precise formulation of decimal constant recognition problem, as a task reverse to pushing button sequence, and its solution. Using three proposed criteria we are able to judge which formulas, matching given floating-point constant, are the most likely. In original formulation they require enumeration of all possible codes with growing complexity, but statistical and compression criteria are in fact independent of the method used to obtain expression. They require only formula to be compiled into RPN code. Then, Criterion 3 can be applied directly, and Criterion 2 indirectly, using code length $K$ to find upper limit on $k_1,k_2,k_3$ (Fig. \[k2k3\]).
Calculator used to enumerate codes can be arbitrary, as long as it includes ”irreducible” buttons equivalent to or . However, search results (required depth in particular) depend on calculator definition. This is especially visible in search of human-provided test cases, with strong bias towards decimal numerals, and repulsive reaction to power-towers and nested (exponential) function compositions. Unfortunately, the latter are strong at approximation, with multi-layer neutral networks being notable example using sigmoidal functions. Therefore, for numbers provided by humans, full calculator is usually the best, for mathematical results and for randomly generated expressions or .
Actually, while searching for constants, we are implicitly dealing with enumeration of all EL functions of one complex variable. This is especially visible in definition of calculator . Constant functions are just subset of them. Therefore, mathematical and statistical tools developed to solve constant identification problem, could be extended to identification of functions of one variable. This is a task for further research, with many more potential applications.
Supplemental Materials
Enumeration example \[enum\_example\]
=====================================
Detailed example of formula enumeration algorithm. Top-down base system (see main text) has been used for simplicity. Ternary base digits were assigned to three RPN calculator buttons as: 0 $\to$ E ($e$), 1 $\to$ LOG ($\log_x{y}$), 2 $\to$ POW ($x^y$). After code length 3 invalid codes were omitted to save space.
------ ------- --------- ----------------------- -----------------------
Enum CODE Syntax RPN sequence formula
0 0 VALID E $e$
1 1 INVALID
2 2 INVALID
3 00 INVALID
4 10 INVALID
5 20 INVALID
6 01 INVALID
7 11 INVALID
8 21 INVALID
9 02 INVALID
10 12 INVALID
11 22 INVALID
12 000 INVALID
13 100 INVALID
14 200 INVALID
15 010 INVALID
16 110 INVALID
17 210 INVALID
18 020 INVALID
19 120 INVALID
20 220 INVALID
21 001 VALID E, E, LOG $\log_e{e}=1$
22 101 INVALID
23 201 INVALID
24 011 INVALID
25 111 INVALID
26 211 INVALID
27 021 INVALID
28 121 INVALID
29 221 INVALID
30 002 VALID E, E, POWER $e^e$
31 102 INVALID
32 202 INVALID
33 012 INVALID
34 112 INVALID
35 212 INVALID
36 022 INVALID
37 122 INVALID
38 222 INVALID
…
210 00101 VALID E, E, LOG, E, LOG $\log_e{1} = 0$
219 00201 VALID E, E, POWER, E, LOG $\log_e{e^e} = e$
228 00011 VALID E, E, E, LOG, LOG $nan$
255 00021 VALID E, E, E, POWER, LOG $\log_{e^e}{e} = 1/e$
291 00102 VALID E, E, LOG, E, POWER $e^1 = e$
300 00202 VALID E, E, POWER, E, POWER $e^{(e^e)}$
309 00012 VALID E, E, E, LOG, POWER $1^e = 1$
336 00022 VALID E, E, E, POWER, POWER $(e^e)^e = e^{e^2}$
…
------ ------- --------- ----------------------- -----------------------
Enumeration of integer and rational numbers \[rat\]
===================================================
If we restrict ourselves to simplest case of integer and rational numbers, enumeration procedure without repetition, i.e., one-to-one mapping of non-negative integers $i$ into integers $j$ is known: $$j = \frac{1}{4} (-1)^i \left( -2 i+(-1)^i-1 \right).$$
Positive rationals $r$ can be enumerated by repeated composition of the function: $$\operatorname{next}{(r)} = \frac{1}{1 + 2 \lfloor r \rfloor - r },$$ starting with zero.
Surprisingly, function $\operatorname{next}$ is composition of two *self-inverse* functions: $$\operatorname{inv}(x) = \frac{1}{x}, \quad \operatorname{ladder}{(x)} = 1 + 2 \lfloor x \rfloor - x.$$
![\[SelfInv\] Self-inverse function, able to generate all integers and rationals without repetitions. ](figs/SelfInv){width="66.60000%"}
Illustrative example is provided as follows. Let us define two additional self-inverse functions: $$\operatorname{minus}(x) = -x,$$ and $$\operatorname{pre}(x) = -x + 1.$$
Any integer and rational (including negative) can be obtained be repeated composition of functions $\operatorname{inv}, \operatorname{minus}$ and $\operatorname{pre}$, cf. Fig. \[SelfInv\], starting with number (or a function) zero. Moreover, the appear in non-repetitive order: $$0,1, \infty, -1,2,\frac{1}{2},-2,3,-\frac{1}{2},\frac{3}{2},\frac{1}{3},
-3,\frac{2}{3},4,-\frac{1}{3},-\frac{3}{2},\frac{4}{3},\frac{5}{2},\frac{1}{4},-\frac{2}{3},
-4,\frac{3}{4},\frac{5}{3},5,\frac{2}{5},-\frac{1}{4},-\frac{4}{3},-\frac{5}{2}, \ldots$$
Above properties are remarkable, and suggest possible way to non-repetitive generation of $\exp-\log$ numbers by function composition, in unique order. However, it is unclear what kind of function(s) is could be. For example, self-inverse function related to exponentiation is: $$e^{-1/\ln{x}},$$ where -1 can be replaced by any other constant. In fact, it could be generalized to: $$p^{q/\log_p{x}},$$ with arbitrary $p,q$. So far, our attempts to find $p,q$ failed. We only guess they are somehow related to $e, \pi$ and possibly $i$.
Proof of completeness of up-bottom base set \[base-3-proof\]
=============================================================
Goal of his section is to show, that all explicit elementary complex numbers can be reduced to three elements.
We start with symbols: $$e, x^y, \log_x{y}.$$
One can compute: $$\ln{x} = \log_e{x}, \quad 1=\ln{e}, \quad 0=\ln{1}, \quad 1/e = \log_{e^e}{e}, -1 = \ln{\log_{e^e}{e}}, 2 = \ln{\ln{(e^e)^e}}.$$
Reversing role of logarithm base and argument, we also get: $$1/2 = \log_{e^{e^2}}{e}.$$
This way one can compute all natural numbers and their reciprocals (egyptian fractions).
Multiplication can be computed by: $$x \cdot y = \log_{x}{ \left[ (x^y)^x \right]},$$ while division is: $$\frac{x}{y} = \log_{x^y}{x^x}.$$
Doing addition is tricky, but possible: $$x+y = \log_x \; \log_{x^x} \left( \left( (x^x)^{x^y} \right)^{x^x} \right).$$
Reciprocal is: $$1/x = \log_{x^x}{x},$$ and sign change: $$-x = \log_x { \log_{x^{x^x}}{x} }.$$
This complete basic 6 binary operations. We need only square root: $$\sqrt{x} = x^{\log _{\log _x\left(\left(x^x\right)^x\right)}(x)}$$ of -1: $$i = \log_x\left(\log _{x^x}(x)\right)^{\log _{\log _x\left(\left(x^x\right)^x\right)}(x)},$$ to compute remaining trigonometric functions. In particular (see Fig. \[pi\] for more readable form): $$\pi = \log \left(e^{\log \left(\log \left(e^e,e\right),e\right)^{\log \left(\log
\left(\left(e^e\right)^e\right),e\right)}},\log \left(\log
\left(e^e,e\right),e\right)\right).$$
This completes the proof.
![\[pi\] Tree form of $\pi$ computed using primitive calculator CALC1. ](figs/PiELogPowera.png){width="50.00000%"}
Proof of completeness of bottom-up base set \[base-4-proof\]
=============================================================
Goal of his section is to show, that all explicit elementary complex numbers can be reduced to calulator with four buttons.
We start with symbols: $$\exp, \ln, x, -.$$
One can calculate: $$0 = x-x, \quad 1 = \exp{0} \equiv \exp{(x-x)}, \quad e=\exp{(1)} \equiv \exp{(\exp{(x-x)})}.$$
Now, we know how to compute: $$\exp, \ln, x, -, 0, 1, e.$$
Addition is: $$x+y = x-(0-y) = x-((x-x)-y).$$
One can change sign and compute reciprocal with: $$-x = 0-x = (x-x)-x, \quad 1/x = \exp{(-\ln{x})}.$$ So far we have: $$\exp, \ln, x, -, 0, 1, e, +, 1/x,$$ succesor is: $$x+1 = x-(0-1) = x - \left( (x-x) - \exp{(x-x)} \right).$$
Using succesor and reciprocal on can compute all integers and rationals. Multiplication and division using logarithms are well-known: $$x \cdot y = \exp{(\ln{x} + \ln{y})}, \quad \frac{x}{y} = \exp{(\ln{x} - \ln{y})}.$$
Binary exponentiation and logarithm are: $$x^y = \exp{(y \cdot \ln{x})}, \quad \log_x{y} = \frac{\ln{y}}{\ln{x}}.$$
Let’s proceed to square root and $i$: $$\sqrt{x} = x^{1/2}, \quad i = \sqrt{-1}.$$
Number $\pi$ is: $$\pi = -i \ln{(-1)}.$$
Now, calculating trigonometric functions is straightforward: $$\sin{x} = \frac{\exp{(i x)} - \exp{(-i x) }}{2 i}, \ldots$$
Above shows, that all constants, functions and binary operations from Sect. 2 can be computed using CALC2.
Distribution of EL numbers on complex plane \[complex\]
=======================================================
Distrubution of the EL numbers generated by sequence on complex plane is presented in Figs. \[ComplexPlane\] and \[ComplexPlane\_FractionalPart\]. Visually, it is far from random. However, it is not fractal, a because EL numbers include rationals, which are everywhere dense.
![\[ComplexPlane\] Distribution of the EL numbers on complex plane for calculators 1-4. Calculator 1 (top-down) is shown in the upper-left panel. Only formulas with $\operatorname{Im}(z) \neq 0$ are shown, up to Kolmogorov complexity $K \leq 23$, what gives 1500k unique complex numbers. Upper-right panel is for CALC2 (down-top), up to $K \leq 16$; lower-left shows CALC3 (Mathematica) for $K \leq 9$; lower-right present full calculator CALC4 up to $K \leq 5$. ](figs/CALC1_complex_2Pi.png "fig:"){width="50.00000%"} ![\[ComplexPlane\] Distribution of the EL numbers on complex plane for calculators 1-4. Calculator 1 (top-down) is shown in the upper-left panel. Only formulas with $\operatorname{Im}(z) \neq 0$ are shown, up to Kolmogorov complexity $K \leq 23$, what gives 1500k unique complex numbers. Upper-right panel is for CALC2 (down-top), up to $K \leq 16$; lower-left shows CALC3 (Mathematica) for $K \leq 9$; lower-right present full calculator CALC4 up to $K \leq 5$. ](figs/CALC2_complex_2Pi.png "fig:"){width="50.00000%"}\
![\[ComplexPlane\] Distribution of the EL numbers on complex plane for calculators 1-4. Calculator 1 (top-down) is shown in the upper-left panel. Only formulas with $\operatorname{Im}(z) \neq 0$ are shown, up to Kolmogorov complexity $K \leq 23$, what gives 1500k unique complex numbers. Upper-right panel is for CALC2 (down-top), up to $K \leq 16$; lower-left shows CALC3 (Mathematica) for $K \leq 9$; lower-right present full calculator CALC4 up to $K \leq 5$. ](figs/CALC3_complex_2Pi.png "fig:"){width="50.00000%"} ![\[ComplexPlane\] Distribution of the EL numbers on complex plane for calculators 1-4. Calculator 1 (top-down) is shown in the upper-left panel. Only formulas with $\operatorname{Im}(z) \neq 0$ are shown, up to Kolmogorov complexity $K \leq 23$, what gives 1500k unique complex numbers. Upper-right panel is for CALC2 (down-top), up to $K \leq 16$; lower-left shows CALC3 (Mathematica) for $K \leq 9$; lower-right present full calculator CALC4 up to $K \leq 5$. ](figs/CALC4_complex_2Pi.png "fig:"){width="50.00000%"}
![\[ComplexPlane\_FractionalPart\] Same as in Fig. \[ComplexPlane\], but his time distribution of the fractional part of the EL numbers on complex plane is shown. ](figs/CALC1_complex_frac.png "fig:"){width="50.00000%"} ![\[ComplexPlane\_FractionalPart\] Same as in Fig. \[ComplexPlane\], but his time distribution of the fractional part of the EL numbers on complex plane is shown. ](figs/CALC2_complex_frac.png "fig:"){width="50.00000%"}\
![\[ComplexPlane\_FractionalPart\] Same as in Fig. \[ComplexPlane\], but his time distribution of the fractional part of the EL numbers on complex plane is shown. ](figs/CALC3_complex_frac.png "fig:"){width="50.00000%"} ![\[ComplexPlane\_FractionalPart\] Same as in Fig. \[ComplexPlane\], but his time distribution of the fractional part of the EL numbers on complex plane is shown. ](figs/CALC4_complex_frac.png "fig:"){width="50.00000%"}
[^1]: If it is not, then it is a rational number, e.g 1.82263 = 182263/100000.
[^2]: Alternatively, database of the pre-computed results [@ISCwayback].
[^3]: See also Appendix B in Supplemental Material.
[^4]: See Appendix E in Supplemental Material for distribution of EL numbers on the complex plane.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Herschel “DUst around NEarby Stars (DUNES)” survey has found a number of debris disk candidates that are apparently very cold, with temperatures near 22K. It has proven difficult to fit their spectral energy distributions with conventional models for debris disks. Given this issue we carefully examine the alternative explanation, that the detections arise from confusion with IR cirrus and/or background galaxies that are not physically associated with the foreground stars. We find that such an explanation is consistent with all of these detections.'
author:
- András Gáspár
- 'George H. Rieke'
title: |
The Herschel Cold Debris Disks:\
Confusion with the Extragalactic Background at $160~\micron$
---
Introduction
============
Debris disks play a vital role in our understanding of the exterior parts of planetary systems. While inner orbit ($< 5~{\rm AU}$) planets are now readily observed with various techniques (i.e., radial velocity and planetary transit surveys), wider orbit planets are significantly more difficult to detect, with only a handful of them discovered by direct imaging [@marois08; @kalas08; @lagrange10; @rameau13; @currie14; @kraus14; @bailey14]. However, with their large surface areas, even low mass and low density debris disks are relatively easy to detect in the mid- to far-infrared wavelengths at larger stellocentric distances, providing ways to study the outer parts of the systems.
Debris disks have a number of characteristic temperatures, of which the most prominent are $190~{\rm K}$ [@morales11] and 45 – $80~{\rm K}$, with a weak dependence on the spectral type of the star [e.g., @ballering13]. Our solar system is an example, with the Asteroid belt at $~2.3 - 3.3~{\rm AU}$ and Kuiper belt at $30 - 50~{\rm AU}$ [@backman95; @vitense12].
An intriguing new result from [*Herschel*]{} was the discovery of a new class of cold debris disks [@eiroa11; @eiroa13], with characteristic temperatures of $22~{\rm K}$. Scaling from the models for Kuiper Belt dust by [@yamamoto98], such a disk would be located at about 120 AU, in an environment dramatically different from those normally assumed for debris disks. The properties of such disks have been studied by [@krivov13], who concluded that they would need to be made up of particles that are larger than a few millimeters and smaller than $10~{\rm km}$, that are dynamically quiescent, and have orbital eccentricities and inclinations $ \le 0.01$. Systems with such specific parameters are not just difficult to form, but also challenging to maintain, when one considers all the destructive external effects disks at large stellocentric regions (especially ones outside the “stello-pause”) may experience (e.g., erosion by the interstellar medium, stellar fly-bys, etc.).
In this paper, because of the issues detailed above, we re-evaluate the possibility that these excesses are not intrinsic to the stars but result from confusion with unrelated sources. In section \[sec:false\], we show that both likely forms of confusion noise, IR cirrus and distant background galaxies, would match the apparent temperature of the cold excesses. In section \[sec:confusion\], we show that standard treatments of confusion noise suggest that such sources may significantly affect the apparent detection of cold debris disks. In section \[sec:MC\], we follow up this possibility with a Monte Carlo analysis, which we then use in section \[sec:stat\] to evaluate the null hypothesis that the apparent cold disks are instead drawn from the populations of confusing sources. In section \[sec:params\], we investigate the dependence of the results on the interval of the background galaxy fluxes considered in the statistical analysis, while in section \[sec:comps\], we compare our results to previous statistical analyses. Finally, in section \[sec:sum\], we summarize our results.
Possible Sources of False Cold Disk Signatures {#sec:false}
==============================================
The cold debris disks have a characteristic temperature of $22~{\rm K}$ [@eiroa11]. At this temperature an excess by a factor of two at $160~\micron$ yields an excess by only a factor of 1.2 at $100~\micron$. That is, an excess below typical detection limits at $100~\micron$ and shorter wavelengths can be substantially above the stellar output at $160~\micron$. We now consider whether the SEDs of the possible confusing sources are consistent with this value.
First, we consider confusion by infrared cirrus. There are a number of relevant measurements: 1.) [@roy10] use BLAST data to find temperatures of $19.9\pm1.3~{\rm K}$ and $16.9\pm0.7~{\rm K}$ for cold interstellar dust in two regions; 2.) [@martin10] fit early Herschel data with a temperature of $23.6\pm1.0~{\rm K}$; 3.) [@bracco11] find $T = 19.0\pm2.4~{\rm K}$, using a different set of early Herschel data; 4.) [@veneziani13] use Bayesian methods with a broad set of data to find temperatures in the ISM cold dust of $\sim 20~{\rm K}$ with a range of about $4~{\rm K}$ around this value. Therefore, IR cirrus is a viable candidate to contribute to emission at the appropriate temperature for the apparent cold disks.
We now turn to confusion by background galaxies. There are a number of systematic changes in the SEDs of luminous galaxies with increasing redshift (and increasing luminosity at the detection threshold [e.g., @rujopakarn13; @berta13; @symeonidis13 and references therein]). We have quantified these trends as in [@rujopakarn13]. We have fitted a blackbody to the appropriate galaxy SED for direct comparison with the assumed disk SED in [@eiroa11]. We take luminosities between the lower envelope of the distribution of detection limits with redshift in @magnelli13 ([-@magnelli13], Figure 8) and twice this value, to obtain luminosities as a function of redshift, characteristic of the faintest sources detected with PACS. We then redshift blackbody fits to the SEDs for the appropriate luminosities by the appropriate values to obtain apparent temperatures of faint 160 $\micron$ detections vs. redshift. We find that the values range from about 25 K at $z = 0.4$ to about 29 K near $z = 0.8$, from where they decline to about 20 K at $z = 2$. [@magnelli13] give redshifts of $z = 1.22^{+0.68}_{-0.41}$ and $z = 0.94^{+0.52}_{-0.38}$ (interquartile ranges) respectively for 160 $\micron$ sources fainter and brighter than . Thus, the faint detections should fall within the 20 - 29 K apparent temperature range. The temperatures estimated from the 100 and 160 $\micron$ measurements of the six sources identified as having cold excesses in [@eiroa13] range from 22.5 to 31 K for the three with probable weak 100 $\micron$ excesses (HIP 73100, 92043, and 109378) and from 2 $\sigma$ upper limits of 21 to 26.5 K for the three with no indicated 100 $\micron$ excesses (HIP 171, 29271, and 49908). We conclude that the expected spectral behavior of faint background galaxies is consistent with the temperatures assigned to the cold debris disks.
Estimates of the Effects of Confusion Noise {#sec:confusion}
===========================================
Confusion with distant background galaxies becomes an increasing issue with increasing wavelength for the Herschel instruments. A conventional definition of the confusion limit is the “source density criterion (SDC)”, when 10% of the sources of a given flux are so tightly crowded that they cannot be measured. [@dole03] find that this limit corresponds to 16.7 beams per source, where the definition of the beam area is based on that by [@condon74]. [@berta11] estimate that the SDC is reached for Herschel at source flux densities of 0.4, 1.5-2, and $8~{\rm mJy}$ respectively at 70, 100, and $160~\micron$. Given the fall of a stellar photospheric output inversely as the square of the wavelength, the SDC is a significant issue for the [@eiroa13] sample only at the longest wavelength band of these three. However, the flux densities attributed to the cold disks are very similar to the limit there. Of the $\sim 100$ sources without $100~\micron$ excesses, the confusion statistic would imply that roughly 6 would be confused with background galaxies at $8~{\rm mJy}$ or brighter, compared with the six sources identified by [@eiroa13] as cold disk sources, some of which have excess fluxes less than $8~{\rm mJy}$. This agreement calls for a more detailed investigation.
[@hogg98] show that sources near the confusion limit and with low ratios of signal to noise tend to be biased too high in apparent brightness. They derive a correction dependent on the slope of the source counts and the signal to noise ratio of the source, to remove this bias and assign the maximum likelihood flux to a source. In the case of the $160~\micron$ galaxy measurements, a slope of $q=0.9$ can be derived for the source counts from the [@magnelli13] data between 1 and 10 mJy, while the signal to noise ratios can be obtained from [@eiroa13], Table 14. Table 1 shows the six cold disk candidate stars with the resulting estimates of the fluxes from the disks alone at $160~\micron$. We have left the error estimates as in [@eiroa13], although [@hogg98] argue that the errors should be expected to increase in these nearly-confusion-limited cases [see their Figure 2 and also Figure 3 in @hogg01]. Four of the six candidates have dropped below the usual $\chi_{160} > 3$ detection criterion[^1], suggesting that a more detailed treatment of the confusion effects may be critical in evaluating the reality of the cold disks.
Monte Carlo Analyses {#sec:MC}
====================
The preceding sections indicate that confusion noise may play a significant role in mimicking the signature of a hypothetical extremely cold debris disk. We therefore perform various Monte Carlo analyses, allowing a relatively easy exploration of the confusion noise in more detail within the full parameter space. Our analyses considers two main sources of obtained flux: cirrus noise and background galaxies. We detail these in the following subsections, while in Table \[tab:var\] we summarize the parameters of the analyses with the default values given. The numerical variables of the model (i.e. the size of the artificial field, log bin size in the galaxy distribution, Airy pattern bin size) were determined with convergence tests to ensure fast computational speeds with reliable results.
[llr]{} D & Size of artificial field & $0.5~{\rm sq.\ deg.}$\
$G_{\rm min}$ & Minimum galaxy flux considered & $1~{\rm mJy}$\
$G_{\rm max}$ & Maximum galaxy flux considered & $225.42~{\rm mJy}$\
$b_{\rm gal}$ & log bin size in the galaxy distribution & $0.02$\
$\sigma_{\rm cirrus}$ & Std. dev. of cirrus noise & $0.505~{\rm mJy}$\
$\mu_{\rm limit}$ & Location par. of noise log-norm distr. & 0.67\
$\sigma_{\rm limit}$ & Scale par. of noise log-norm distr. & 0.33\
$N_{\ast}$ & Number of positions tested & $10^6$\
$r_t$ & Target radius & $6\arcsec$\
$r_p$ & Photometry radius & $8\arcsec$\
$S_{\rm in}$ & Sky aperture inner radius & $18\arcsec$\
$S_{\rm out}$ & Sky aperture outer radius & $28\arcsec$\
$\Delta B_{\ast}$ & Bin size in photometry distribution & $0.1~{\rm mJy}$
Cirrus noise
------------
Determining the value of the cirrus noise is difficult. Because of this, our goal was to assign a highly conservative value to it, without neglecting it. This was also appropriate, as the DUNES survey was designed to observe sources in low cirrus background regions. We used Equation 22 of [@miville07] to calculate the confusion noise, which is based on the power-spectrum of the far-infrared dust emission and calibrated to low levels. The spectral index in the equation is given by Equation 4 in their paper. Using HSpot, we estimated the average ISM flux background for the DUNES sources to be $<I_{160}> = 7.02~{\rm MJy}~{\rm sr}^{-1}$, and a $<I_{160}>/<I_{100}>$ ratio of 1.845. Assuming the standard [@condon74] definition of beam size, we derived a cirrus noise of $\sigma_{\rm cirrus} = 0.505~{\rm mJy}$. This estimate is only about half as large as is indicated in the scaling relations in HSpot. We dealt with the small number of sources with much stronger than average cirrus by eliminating them from our test sample, rather than trying to estimate the cirrus noise more accurately. In the Monte Carlo simulations the cirrus noise value at each test location was determined by choosing a value following a Gaussian probability function centered at zero with a standard deviation of $\sigma_{\rm cirrus}$.
[l|c|cc|cc]{} $6^{\prime\prime}$ Target Radius & 6 & 6.7 & 6.4 & 6.1 & 5.6\
$7^{\prime\prime}$ Target Radius & 6 & 8.6 & 7.4 & 7.9 & 6.5\
$8^{\prime\prime}$ Target Radius & 7 & 10.6 & 8.3 & 9.9 & 7.3
Background galaxy contribution
------------------------------
As introduced in section \[sec:confusion\], background galaxies can dominate the confusion noise at far-IR wavelengths. For our Monte Carlo analyses, we randomly distributed galaxies on a 0.5 sq. degree area, with a fiducial galaxy flux interval of 1 to $225~{\rm mJy}$, although for certain tests we extended the lower limit to $0.012~{\rm mJy}$. The galaxy number counts were adopted from three separate studies. Between 0.012 and $1.25~{\rm mJy}$ we used the modeling results found in Table B.2 of [@franceschini10], between 1.42 and $28.38~{\rm mJy}$ we adopted the observed galaxy counts of the GOODS-S ultradeep Herschel survey from [@magnelli13], while the number counts of the brightest galaxies were adopted from Table 5 of [@berta11] (all fields combined). Although these are three independent studies, their differential number count curves connect smoothly. Total number counts were calculated in logarithmically evenly spaced flux bins, with the number counts appropriately interpolated (in log space) at the bin boundaries and integrated (also in log space) with a simple second order trapezoid method. Between 1 and $225~{\rm mJy}$ the artificial field (0.5 sq. degree) has altogether 19146 galaxies, and between 6 and $13~{\rm mJy}$ it has $\sim 2776$ galaxies (or $\sim 5552$ galaxies per sq. degree), which agrees with the estimated 5500 galaxies per sq. degree in this interval cited by [@krivov13].
When considering confusion with background galaxies, [@eiroa13] only used the differential count value determined at $6~{\rm mJy}$, resulting in a smaller total number of estimated background sources ($\sim 2000$ per sq. degree). As detailed in section \[sec:params\], one of the key differences between our analyses and the previous ones is that we use a larger interval of background galaxy fluxes (and integrate the differential distribution). As we will show, galaxies fainter than the detection threshold ($\le 6~{\rm mJy}$) contribute to the confusion noise, as their spatial distribution is not isotropic enough for their contribution to the total flux to be canceled out by sky subtraction, even when considering a completely random field as we do here. Natural clustering of galaxies will likely even enhance their contributions [@fernandez08].
The results of our model will depend predominantly only on a single parameter, the beam solid angle ($\Omega$) (i.e., the confusion beam size). The value of the beam solid angle is a matter of definition. The classic [@condon74] definition of the effective beam solid angle is $$\Omega_e = (\frac{1}{4}\pi\Theta_1\Theta_2)\frac{1}{\left(\gamma-1\right)\ln 2}\;,$$ where $\Theta_1$ and $\Theta_2$ are the half-power axes of the elliptical Gaussian beam and $\gamma$ is the slope of the differential distribution of sources. According to Table 3.1 of the PACS Observer’s Manual, $\Theta_1=10.65^{\prime\prime}$ and $\Theta_2=12.13^{\prime\prime}$ at a scan speed of $20^{\prime\prime}~{\rm s}^{-1}$, while the value of $\gamma$ is around 1.9 at low fluxes ($1$ - $10~{\rm mJy}$), according to the [@magnelli13] data. These yield an effective beam solid angle of $162^{\prime\prime 2}$, or a confusion beam radius of $7.19^{\prime\prime}$.
The DUNES team uses the images to identify potentially confusing sources (of similar brightness to the target) at $\ge 6^{\prime\prime}$. Hereafter, we refer to this distance as the [*target radius*]{}. After excluding targets with confusing sources, they perform photometry in a photometry radius of $8^{\prime\prime}$. The sky background was subtracted based on a value measured in an annulus outside the photometry radius. Our models were constructed to reproduce this measurement strategy. We assumed aperture photometry carried out within a radius of $8^{\prime\prime}$. We tested a variety of target radii inward of which we assumed it was no longer possible to distinguish a background source from the target, besides the $6^{\prime\prime}$ assumed by DUNES. In all models, we rejected targets with sources lying in the annulus between the target and photometry radii that also were more than $2.5 \sigma$ above a value chosen with a log-norm probability function with $\mu_{\rm limit}$ and $\sigma_{\rm limit}$ parameters (see Table \[tab:var\]) that describes the distribution of photometry errors for the DUNES sample.
We integrate the flux of the sources within the photometry aperture and the corresponding sky annulus using two methods: treating the galaxies as point sources in one of them, and convolving their emissions with the Herschel Airy pattern at 160 $\micron$ in the other. The first is the traditionally used method when considering confusion, however, we have found that smoothing the emissions with the point spread functions (PSF) will affect the results of the confusion estimates. We introduce the results of both calculations for completeness and also to allow comparisons with previous work.
After generating the artificial background galaxy map, our code determined random positions and performed the previously described “aperture photometry”. For the smoothed model all partial fluxes contained within the apertures were included (i.e., fluxes from sources both within and outside the apertures). For the point source method the total fluxes of sources within the apertures were added to determine the total flux, but only for sources that were located within the apertures. To censor bright galaxies within the sky annuli, as was done by [@eiroa13], we removed bright galaxies from the sky annulus above a simulated upper limit. As with the photometry aperture, this limit was set at $2.5\times$ a $\sigma$ value that was randomly chosen from the photometry error distribution described above. Finally, the flux within the sky background was normalized by the ratio of the aperture area to the sky annulus area. The flux at the test location was then determined by subtracting the “sky background” from the flux determined within the confusion beam aperture and adding the cirrus noise.
Statistics {#sec:stat}
==========
In this section, we compare the results of our model to the DUNES observations. We first define the DUNES sample we compare our models to, and then compare the model results to the observations with various statistical methods, while varying the target radius. In Table \[tab:stats\], we summarize the detection statistics of the observations and the models.
The DUNES sample {#sec:sample}
----------------
There are 133 sources in the DUNES sample [@eiroa13] of which 131 have data at $160~\micron$. Of these, 100 sources do not have detectable excesses at either PACS wavelengths. From these 100, we removed 6 sources whose limits on their $160~\micron$ excess were higher than the typical value within the sample (HIP 71681, HIP 71683, HIP 88601, HIP 104214, HIP 104217, and HIP 108870). Of the remaining 31 excess sources, [@eiroa13] list 6 as harboring cold debris disks. Although it is listed as a cold disk candidate, the excess for HIP 92043 is detected at $70~\micron$ (both MIPS and PACS) and at $100~\micron$ and $160~\micron$ [@eiroa13], so its identification as a cold disk candidate depends on the relatively weak $160~\micron$ result in Table 1. We computed a weighted average of the MIPS and PACS $70~\micron$ data, obtaining an excess of $11.9 \pm 3.3~{\rm mJy}$, took the $100~\micron$ result from [@eiroa13] and the maximum likelihood value at $160~\micron$ from Table \[tab:george\] and then fitted the excess spectral energy distribution at all three wavelengths with a modified blackbody with $\beta = 0.65$ [@gaspar12b]. We found that a disk temperature of $62~{\rm K}$ fitted within the errors ($\chi^2_{\rm reduced} = 1.35$), so there is no need to hypothesize a cold disk for this star and we remove it from the cold disk sample. Of the original 6 cold debris disk candidates [@eiroa13], we only consider HIP 171, HIP 29271, and HIP 49908 most likely to have alternative explanations for their apparent far infrared excesses. HIP 73100 and HIP 109378 show evidence for excess emission at $100~\micron$, but the rapid increase in their SEDs to $160~\micron$ probably arises from confusion.
[cccc]{} 171& 6.2& 2.5& 2.5\
29271& 6.0& 2.2& 2.7\
49908& 4.8& 2& 2.4\
73100& 8.3& 2.5& 3.3\
92043& 9.3& 4& 2.3\
109378& 9.8& 2& 4.9
Apart from the five cold disk candidates, additional spurious sources were listed in Table D.1 of [@eiroa13]. Two of the spurious sources have heavy cirrus contamination \[HIP 29568 (“structured background”) and HIP 71908 (“emission strip”; HSpot indicates a high interstellar background level of $59.7~{\rm MJy}~{\rm sr}^{-1}$)\] and one is probably a spurious detection (HIP 38784). Four of the remaining sources (HIP 40843, 85295, 105312, and 113576) are potentially contaminated by background galaxies. Of these four sources, two have the peaks of the emission of their $160~\micron$ component within the $8^{\prime\prime}$ photometry aperture of the survey (HIP 85295 at $4.8^{\prime\prime}$ and HIP 105312 at $7.16^{\prime\prime}$). For our models to stay consistent with the observational sample, we include these two sources in the cold disk sample (one or two of them, depending on the size of the target radius). This means that there are a total of six/seven sources with apparent cold excesses (HIP 171, HIP 29271, HIP 49908, HIP 73100, HIP 85295, HIP 105312, and HIP 109378).
This leaves us a total sample of 93/94 sources (6/7 with excess and 87 without), depending on the considered target radius, with the boundary at $7.16^{\prime\prime}$. Of the non-excess sources, 33 have measured fluxes, while the remaining 54 only have upper limits. The remaining 25 sources with detected debris disk excesses were not included in the statistical analysis, as estimating a possible level of contamination for them is not possible. Of these sources, three (HIP 4148, HIP 27887, and HIP 51502) have equilibrium temperatures around 30 K, close to the levels of the cold disk candidates. The final results of the paper would indicate an additional 1.98 cold sources remaining in the sample of 25, possibly also explaining the far-IR excesses observed at these three sources.
Method 1: Via Realization of Data {#sec:real}
---------------------------------
The first method we apply realizes artificial datasets and counts the number of detections within the dataset. First, 93/94 source locations are randomly selected within our artificial field and the total flux at these locations calculated according to the procedure described in section \[sec:MC\]. Then a detection threshold (determined at $3\sigma_F$) is randomly paired to each artificial location from the sample of 93/94 DUNES sources. If the total flux is larger than the detection threshold then the number of detections in the realized dataset is incremented by one. We realized $10^5$ datasets of 93/94 sources for each of the tested target radii. In Figure \[fig:sampling\], we show the results of these tests for both the “point sources” and “PSF smoothed sources” models and in Table \[tab:stats\], we summarize the detection statistics of the model. The probability of finding more sources increases with larger target radii, as expected.
The “point-sources model” yields probability curves that are strongly dependent on the target radii. With over 40% of the photometry area located between 6 and 8$^{\prime\prime}$, sensoring confusing sources in the outer aperture is critical. The probability curves are wide, for example at a target radius of $7^{\prime\prime}$, the model predicts $8.55\pm2.79$ sources, meaning that detecting 5.7 sources is just as likely as detecting 11.3.
A closer representation of the measurements is performed by the “PSF smoothed-sources model”. The distributions are narrower and the peaks are closer and at lower values than for the simpler “point-sources model”. The peaks shifting to lower values is due to the generally higher sky background values, which is a result of contributions to the sky flux from sources outside the reference sky annulus, which are now smoothed into the sky area. As the sky annulus is larger in area than the aperture photometry area, and it also receives contributions from sources inside of it as well as from outside of it, this is a significant effect. Moreover, the distributions are also narrower due the PSF smoothing, as background levels become more homogenous. As an example, at a target radius of $7^{\prime\prime}$, the “PSF smoothed-sources” model predicts $7.38\pm2.61$ sources. This agrees well with the 6/7 sources expected at $7.16^{\prime\prime}$ according to the observations.
The significant number of potentially confusing sources among the stars without excesses or with cold ones raises a question of the contamination among those with debris disks. A rough estimate can be obtained by noting that among the stars with normal disks, there are five with apparent detections at 160 $\micron$ ($\chi_{160} > 3$) and with flux densities less than , within the range where confusion is a risk. From the statistics above, these numbers suggest that no more than one of the normal disk stars may have a 160 $\micron$ flux density dominated by a background galaxy.
Method 2: Via Distribution Functions
------------------------------------
With the second method, we generated distributions of the artificial fluxes by testing $N_{\ast}$ number of random positions. The flux values of our sample of $N_{\ast}$ test points were then binned with a bin size of $\Delta B_{\ast} = 0.1~{\rm mJy}$. These distributions were then compared to the observed distribution of fluxes with various methods.
### Percentages with Cumulative Distributions
The simplest test that can be performed is determining the percent of sources above given thresholds using the cumulative distribution function (CDF) of the model [as in @krivov13; @eiroa13]. This test is not rigorous (e.g., it does not take account of upper limits above the sample detection threshold). However, for illustration and to allow comparison with previous statistical analysis, we begin with the results of this test. In Figure \[fig:DF\], we show binned distribution functions of background fluxes of our nominal model, while varying the target radius, for both the “point sources” and “PSF smoothed sources” models. Increasing the target radius, as expected, will widen the distribution and yield more high flux sources. The faintest cold disk candidate has an excess of $6.39~{\rm mJy}$. The number of predicted sources above this limit at various target radii is also summarized in Table \[tab:stats\]. The results of the CDF analysis compare fairly well to the observed number of sources with cold disk signatures, especially when considering the classic [@condon74] definition of confusion beam size and the more realistic “PSF smoothed sources” model.
The peak of the distribution at negative values in Figure \[fig:DF\] results because more brighter galaxies will be located within the larger area sky annulus than within the search area. Unless the area of the sky annulus is equal to the photometry aperture area, this will always result in a negative bias. We have tested this by using a sky annulus with the same area as the photometry aperture, resulting in a peak at zero. The effect is less prominent in the “PSF smoothed sources” model compared with the “point sources model”.
### Kaplan-Meier estimates {#sec:KM}
The Kaplan-Meier (KM) estimates [@km58] of both the modeled and observed distributions provide a systematic method to compare these distributions while taking account of the upper limits (or censoring) in the observations. The KM method has been adopted for astronomical data analysis [e.g., @feigelson85], where it is useful for randomly picked datasets, such as the background distribution in the DUNES survey. We used the ASURV package [@feigelson85] to calculate the KM estimates and compare the KM curves of the observations and models at various target radii in Figure \[fig:SF\]. The DUNES data we compare our models to depends on the target radius with the addition of the extra seventh source when comparing to the $8^{\prime\prime}$ model. The bottom panels in the Figure show these calculations, while the top panels show the comparisons at $6$ and $7^{\prime\prime}$ with the KM curve of the observations using six excess sources. For the observations, we have set all sources apart from the cold disk candidates (the remaining 87 sources) as upper limits. The upper limits were set to ${\rm UL} = F-P+3\sigma_F$ for sources where the photospheres were detected and kept at their original published upper limit value minus the estimated photosphere where they were not. Here, $F$ is the measured flux density, $P$ is the expected value from the stellar photosphere, and $\sigma_F$ is the quoted uncertainty. The models generally appear to agree closely with the distribution of the observations.
### Kolmogorov-Smirnov test on the incompleteness-corrected sample {#sec:KSKM}
There is no standard method to determine the probability of agreement between censored data and a numerical model. We have therefore proceeded as follows. The Kaplan-Meier estimator introduced in the previous subsection can be thought of as an incompleteness-corrected CDF, as it carries on the probabilities of previous events occurring with the knowledge of the censoring. To obtain a rigorous test making use of the upper limits, we perform a Kolmogorov-Smirnov (KS) test on the incompleteness-corrected sample, by increasing the weight of the surviving sample members exactly the way an incompleteness correction would. The KS statistic was only calculated for sources above the detection threshold of $6.39~{\rm mJy}$ (as below we do not have any data) and the probabilities were calculated by scaling with the complete distribution. In Figure \[fig:KSKM\], we show the probabilities obtained as a function of the target radius with these methods. The probability curve indicates that the data are consistent with being drawn from the confusion-limited model at $> 80\%$ confidence for all target radii between 2 and $8^{\prime\prime}$ (the drop in probability for large radii is because the model predicts too many detections, so this case is not of interest). This range of target radii includes all plausible definitions for the PACS beam. The figure shows that even when considering a smaller target radius, as long as the photometry is performed up to $8^{\prime\prime}$, the model results will be consistent with the observational statistics.
![Probability of agreement between the incompleteness corrected data and the model as a function of the confusion beam radius.[]{data-label="fig:KSKM"}](f4.eps)
We also performed the Anderson-Darling K-sample test [@scholz87] on the incompleteness corrected sample, as it is more sensitive at the edges of the distributions than the KS test, using the statistical analysis software package [R]{}. The observed data has many upper limits above the detection threshold of $6.39~{\rm mJy}$, and hence has significant corrections for incompleteness. These corrections introduce pseudo-ties in the data, to which the Anderson-Darling test is sensitive. Therefore, we used the method that assumes ties within the data, described in section 5 of their paper. The analysis showed that the two distributions are indistinguishable within target radii of $4.2$ and $6.3^{\prime\prime}$ for the “point sources” model and for all target radii larger than $5.3^{\prime\prime}$ for the “PSF smoothed sources” model at a 95% confidence level.
Parameter dependence {#sec:params}
====================
Although the results mainly depend on the choice of target radius, here we investigate how the results depend on the range of galaxy fluxes considered. The main motivations for this study are the previous analyses [@eiroa13; @krivov13] that rejected the hypothesis that all of these systems could be explained by confusion, but only used a limited range of galaxy fluxes, between 6 and $13~{\rm mJy}$.
We simulated 900 models, with both the minimum and maximum galaxy fluxes ranging between 0.012 and $225.49~{\rm mJy}$ and a target radius of $7.19^{\prime\prime}$ using the PSF smoothed approach, and calculated the completeness corrected KS test (as in section \[sec:KSKM\]) for each of them. In Figure \[fig:SminSmax\], we show the results of these KS tests as a 2D plot, contouring the 1, 2, and 3 $\sigma$ probabilities and also plotting the ranges considered by the previous studies and ours. Compared with the full-range estimate, the limited flux interval produces less sources through the omission of noise due to the cumulative effects of faint sources. This result demonstrates that the difference between our work and the previous conclusions about the cold disks can largely be explained by the inappropriate limitation in confusing source fluxes assumed by [@eiroa13] and [@krivov13].
![The probability of agreement between the incompleteness corrected data and the model as a function of the minimum and maximum galaxy fluxes considered in the model. The blue star shows the interval considered by [@eiroa13] and [@krivov13] and the red circle shows the interval considered by our nominal model.[]{data-label="fig:SminSmax"}](f5.eps)
Comparison to previous work {#sec:comps}
===========================
The DUNES and DEBRIS Herschel Open Time Key Program surveys were the first surveys ever conducted with the specific goals of detecting debris disks at wavelengths between $100$ and $800~\micron$. At these wavelengths, as shown in Section \[sec:false\], confusion with the extragalactic background and/or infrared cirrus can be an important effect. While simple galactic number count statistics suffices for confusion studies at shorter wavelengths, due to the larger confusion beam sizes and the high number of confusing sources at the detection threshold, a more sophisticated analysis is necessary at these longer wavelengths.
In the discovery paper, [@eiroa13] analyze the likelihood of these sources originating from confusion with the extragalactic background in their Section 7.2.1. After excluding spurious sources with obvious high background/cirrus contamination, they conclude with a list of six sources requiring an alternate explanation. Based on the [@berta11] galaxy counts at $160~\micron$, they perform count statistics. Based on their artificial data tests, they assume a confusion beam radius of $5^{\prime\prime}$, where they were able to separate two equal sources with fluxes near the detection threshold value. For multiple sources that are fainter than the detection threshold, this may be an inadequate confusion beam radius value, especially when considering the classic [@condon74] definition of confusion beam size. They also use the differential number density of galaxies at the detection threshold as a total source count, yielding a low number of possible contaminating sources. In Section \[sec:params\], we show the importance of using the full range of background galaxy fluxes when calculating the effects of confusion. Finally, they considered their complete observational catalog for the statistics (133 sources), including systems that were shown to harbor debris disks. Although systems with debris disks may also have background confusion at $160~\micron$, the contribution from the background will be difficult to distinguish from the debris disk component, requiring these systems to be removed from the analysis sample. These approximations resulted in a prediction of only 1.2% of the sources having background confusion.
The theoretical analysis in [@krivov13] focused on explaining the physical likelihood of cold debris disks existing and deem it unlikely that all of these sources could originate from confusion with cirrus, which we agree with. They performed searches for strong X-ray and/or optical galactic counterparts, but the results from these tests were largely inconclusive within the confusion beam. As in [@eiroa13], they also performed statistical tests, mostly with the same arguments. They show that the offsets between the assumed position of the sources and the $160~\micron$ fluxes are all within $5^{\prime\prime}$, however, as per the definition of confusion beam ($7.19^{\prime\prime}$), all positions within it are not separable. This is also noted in [@krivov13], which is why they search for sources of background confusion within a radius of $6^{\prime\prime}$ in their statistical analysis. However, they only look at extragalactic sources within the flux range of the cold sources (6 to 13 mJy), not accounting for possible confusion originating from multiple fainter sources. They calculate a confusion probability of 4.8%, and scaling to the complete DUNES sample (133 sources) predict 6.4 false detections. Assuming that all of the seven spurious sources in Table D.1 of [@eiroa13] are a result of extragalactic background contamination, they determine that there is a 69% probability that the remaining six cold disks are true debris detections. This argument, however, does not take into account that two of the seven sources are obviously contaminated by high cirrus noise (HIP 29568 and HIP 71908), while three of the remaining five (HIP 40843, HIP 105312, and HIP 113576) have the peaks of their $160~\micron$ emission outside of the $6^{\prime\prime}$ confusion beam radius used in their analysis. Of the remaining two sources, HIP 38784 is a spurious detection with double $160~\micron$ peaks (of which one is also outside of the $6^{\prime\prime}$ radius. There is only a single source from their Table D.1, HIP 85295, that needs to be counted as a source in the statistical analysis, as in our paper.
The most detailed work on confusion estimates for debris disk studies at longer wavelengths were performed by [@sibthorpe13]. As a first step, they convolve the [@berta11] number counts with Gaussians with various error estimates as a way of accounting for the Eddington bias. They present two calculations, one that calculates the probability of a single source producing the confusion and one that calculates the probability of one or more sources producing it. They also introduce a Monte Carlo style algorithm to calculate the probability of confusion. However, their algorithm considers the survey limiting flux density not just as a detection threshold, but also as the minimum galaxy flux in the model. For the $7^{\prime\prime}$ model, at $S_{\rm lim}=6.39~{\rm mJy}$, they predict a probability of 7.8%, which is close to the value given by our “point-sources” model. However, the methods of sky subtraction are not introduced in the paper, therefore we are unable to access the final results presented in it.
The cold disk candidate HIP 92043 is analyzed in [@marshall13]. They also include a statistical argument whether the source can plausibly be a cold disk source in their section 4.1. Although they add a cold component to their model to fit at $160~\micron$, they describe their excess detection at this wavelength as marginal. Their statistical analysis is along the lines of that of [@eiroa13] and predict 1% of the sources having a contamination at the 12.9 mJy level. As a comparison, our CDF model predicts 2.1% of the sources having a contamination above the 12.9 mJy level for the “point-sources”, and 1.8% of them for the “PSF smoothed-sources” model at their assumed $11.3^{\prime\prime}$ target aperture. Our higher values are due to the same effects as previously. They also cite the work of [@sibthorpe13], however, only consider their model where confusion with a single bright source is calculated. The MC models of [@sibthorpe13] show higher probabilities of confusion than their single source confusion analytic estimates.
Summary {#sec:sum}
=======
In this paper, we evaluate the hypothesis of a newly discovered class of [*Herschel*]{} cold debris disks [@eiroa11; @eiroa13; @krivov13]. We test whether the apparent temperature and flux distributions are instead consistent with confusion noise. Although this scenario has been considered by previous work, there are a few differences between our analyses:
we simulate confusion noise using the full relevant range of background galaxy fluxes and allow for confusion from multiple sources,
we account for the smoothing of the emissions by the PSF of the telescope,
we develop an analysis method that accounts for the censorship of the data due to the limitations in signal to noise ratio in the DUNES $160~\micron$ data.
We test the hypothesis that the distribution of cold debris disks is entirely due to confusion with background galaxies (after rejecting cases with elevated noise from IR cirrus). We evaluate the hypothesis as a function of the target radius used to measure sources at $160~\micron$. We find that there is a greater-than-80% probability that the two distributions (confusion noise and the proposed cold debris disks) are indistinguishable, so long as the confision beam is between 2 and $8^{\prime\prime}$ in radius. This range of beam size includes all plausible values for the DUNES measurements. We conclude that the background confusion hypothesis is a viable alternative to the cold debris disk explanation for the $160~\micron$ detections of these sources.
We thank Benjamin Weiner, Brandon Kelly, Eric Feigelson, and Ewan Cameron for inputs on the statistical analysis as well as the Astrostatistics and Astroinformatics Portal (ASAIP) hosted by Pennsylvania State University. Support for this work was provided by NASA through Contract Number 1255094 issued by JPL/Caltech.
[27]{} natexlab\#1[\#1]{}
Backman, D. E., Dasgupta, A., & Stencel, R. E. 1995, , 450, L35
Bailey, V., Meshkat, T., Reiter, M., et al. 2014, , 780, L4
, N. P., [Rieke]{}, G. H., [Su]{}, K. Y. L., & [Montiel]{}, E. 2013, , 775, 55
, S., [et al.]{} 2011, , 532, A49
, S., [et al.]{} 2013, , 551, A100
, A., [et al.]{} 2011, , 412, 1151
, J. J. 1974, , 188, 279
Currie, T., Daemgen, S., Debes, J., et al. 2014, , 780, L30
, H., [Lagache]{}, G., & [Puget]{}, J.-L. 2003, , 585, 617
, C., [et al.]{} 2011, , 536, L4
, C., [et al.]{} 2013, , 555, A11
, E. D., & [Nelson]{}, P. I. 1985, , 293, 192
, N., [Lagache]{}, G., [Puget]{}, J.-L., & [Dole]{}, H. 2008, , 481, 885
, A., [Rodighiero]{}, G., [Vaccari]{}, M., [Berta]{}, S., [Marchetti]{}, L., & [Mainetti]{}, G. 2010, , 517, A74
, A., [Psaltis]{}, D., [Rieke]{}, G. H., & [[Ö]{}zel]{}, F. 2012, , 754, 74
, C., [et al.]{} 2013, , 432, 23
, D. W. 2001, , 121, 1207
, D. W., & [Turner]{}, E. L. 1998, , 110, 727
, P., [et al.]{} 2008, Science, 322, 1345
, E. L., & [Meier]{}, P. 1958, Journal of the American Statistical Association, 53, 457
Kraus, A. L., Ireland, M. J., Cieza, L. A., et al. 2014, , 781, 20
Krivov, A. V., Eiroa, C., L[ö]{}hne, T., et al. 2013, , 772, 32
, A.-M., [et al.]{} 2010, Science, 329, 57
, B., [et al.]{} 2013, , 553, A132
, C., [Macintosh]{}, B., [Barman]{}, T., [Zuckerman]{}, B., [Song]{}, I., [Patience]{}, J., [Lafreni[è]{}re]{}, D., & [Doyon]{}, R. 2008, Science, 322, 1348
Marshall, J. P., Krivov, A. V., del Burgo, C., et al. 2013, , 557, A58
, P. G., [et al.]{} 2010, , 518, L105
, M.-A., [Lagache]{}, G., [Boulanger]{}, F., & [Puget]{}, J.-L. 2007, , 469, 595
, F. Y., [Rieke]{}, G. H., [Werner]{}, M. W., [Bryden]{}, G., [Stapelfeldt]{}, K. R., & [Su]{}, K. Y. L. 2011, , 730, L29
Rameau, J., Chauvin, G., Lagrange, A.-M., et al. 2013, , 772, L15
, A., [et al.]{} 2010, , 708, 1611
, W., [Rieke]{}, G. H., [Weiner]{}, B. J., [P[é]{}rez-Gonz[á]{}lez]{}, P., [Rex]{}, M., [Walth]{}, G. L., & [Kartaltepe]{}, J. S. 2013, , 767, 73
, F. W. & [Stephens]{}, M. A. 1987, Journal of the American Statistical Association, 82, 399
, B., [Ivison]{}, R. J., [Massey]{}, R. J., [Roseboom]{}, I. G., [van der Werf]{}, P. P., [Matthews]{}, B. C., & [Greaves]{}, J. S. 2013, , 428, L6
, M., [et al.]{} 2013, , 431, 2317
, M., [Piacentini]{}, F., [Noriega-Crespo]{}, A., [Carey]{}, S., [Paladini]{}, R., & [Paradis]{}, D. 2013, , 772, 56
Vitense, C., Krivov, A. V., Kobayashi, H., L[ö]{}hne, T. 2012, , 540, A30
Yamamoto, S., & Mukai, T. 1998, Earth, Planets, and Space, 50, 531
[^1]: $\chi_{160} = \left(F_{160} - P_{160}\right)/\sigma_{160}$, where $F$ is the measured flux, $P$ is the estimated photosphere, and $\sigma$ is the error of photometry.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We analyze the linear stability of the base state of the problem of coupled flow and deformation in a long and shallow rectangular soft hydraulic conduit with a thick top wall. Specifically, the steady base state is computed at low but finite Reynolds number. Then, we show that with the upstream flux fixed and the outlet pressure set to gauge, the flow is linearly stable to infinitesimal flow-wise perturbations. Multiple oscillatory but stable eigenmodes are computed in a range of the reduced Reynolds number, $\hat{Re}$, and the so-called fluid–structure interaction (FSI) parameter, $\lambda$, indicating the stiffness of this FSI system. These results provide a framework to address, in future work, the individual effects of various aspects of two-way FSI coupling on instability and flow transition in soft hydraulic conduits.'
author:
- Xiaojia Wang
- 'Ivan C. Christov'
bibliography:
- 'mendeley\_refs.bib'
title: |
Soft Hydraulics in Channels with Thick Walls:\
The Finite-Reynolds-Number Base State and Its Stability
---
\[sec:intro\]Introduction
=========================
The fluid–structure interactions (FSIs) between external or internal flows (either viscous or inviscid) and elastic structures, as well as the linear stability of such coupled mechanics problems, is a research subject with a a time-honored history [@P16]. While FSI topics such as aeroelasticity [@BAH96] and blood flow in large arteries [@P80] are now quite classical, the mechanical interaction between *slow* viscous flows and compliant conduits [@CPFY12] has opened new avenues of FSI research [@DS16; @KCC18], both at the microscale for, e.g., for lab-on-a-chip applications [@FZPN19], and at the macroscale for, e.g., soft robotics applications [@MEG17].
In the present work, motivated by recent “ultrafast mixing” experimental studies in compliant microchannels [@VK13; @KB16], we wish to determine the linear stability of finite-Reynolds-number perturbations to the steady flow and deformation solution for FSI in a rectangular soft hydraulic conduit with a thick top wall. We derived the vanishing-Reynolds-number steady FSI solution in our previous work [@WC19]. Unlike the prior study [@VK13], herein we do not use experimental, computational, or other empirical information to derive our linear stability model (beyond the standard assumptions on separation of length scales, and the smallness of relevant parameters in the system). In doing so, we address the linear stability consequences of different FSI effects in soft-walled microchannels, such as the non-constant axial pressure gradient and the non-flat (deformed) base state of the flow conduit, by extending the results from our recent rigorous mathematical theory [@WC19].
Furthermore, we investigate the relative importance and effect of the flow inertia (quantified by the reduced Reynolds number, $\hat{Re}$), and the compliance of the top wall (quantified by the FSI parameter, $\lambda$), on the linear stability problem. In particular, the base state is found to be stable in the range of $\hat{Re}$ and $\lambda$ considered herein, which is a typical range for microfluidic systems. We conclude with a discussion of possible extensions to the present theory.
![Diagram of one-half of an $x$-symmetric thick-walled microchannel, labelled with the dimensional variables (lower case) of the problem. The origin of the coordinate system (labeled with a red a dot) is set at the centerline ($x=0$) of the rigid bottom wall of the channel. Here, $h_0$, $w$, and $\ell$ represent the undeformed channel height, width and length, respectively, while $t$ is the top wall’s thickness. The deformed fluid–solid interface is defined as $y=h_0+u_y^0(x,z)$, where the compliant top wall’s $y$-displacement evaluated at $y=h_0$ is denoted by $u_y^0$. The Newtonian fluid flow, with a given volumetric flow rate $q$, is in the positive $z$-direction, as indicated by arrows, from the inlet at $z=0$ to the outlet at $z=\ell$. The reduced Reynolds number introduced in Eq. can be defined using the dimensional variables in the figure as $\hat{Re}=\epsilon Re={qh_0}/(\nu w\ell)$, where $\nu$ is the kinematic viscosity of the fluid, and $\epsilon=h_0/\ell$ is the axial aspect ratio. Reproduced and adapted with permission from Ref. [@WC19] 2019 The Author(s) (X.W. and I.C.C.) Published by the Royal Society.[]{data-label="fig:schematic"}](Schematic-confs.eps){width="60.00000%"}
\[sec:gov\_eq\]Governing Equations
==================================
To consider finite-Reynolds-number perturbations to the steady $Re=0$ base flow, we allow a finite *reduced* Reynolds number: $\hat{Re} = \epsilon Re = \mathcal{O}(1)$ as $\epsilon\to0$, where $\epsilon\ll1$ is the undeformed-height-to-length ratio of the long and shallow microchannel (see Fig. \[fig:schematic\] for notation and schematic of the physical setup). Then, the leading-order (in $\epsilon$) governing incompressible Navier–Stokes flow equations are as follows (see Ref. [@WC19] for the derivation and discussion):
\[NS-O1\] $$\begin{aligned}
\frac{\partial V_X}{\partial X}+\frac{\partial V_Y}{\partial Y}+\frac{\partial V_Z}{\partial Z}&=0,\label{COM-O1}\displaybreak[3]\\
-\frac{\partial P}{\partial X}&=0,\label{COLM-X-O1}\\
-\frac{\partial P}{\partial Y}&=0,\label{COLM-Y-O1}\\
\hat{Re}\left(\frac{\partial V_Z}{\partial T}+V_X\frac{\partial V_Z}{\partial X}+V_Y\frac{\partial V_Z}{\partial Y}+V_Z\frac{\partial V_Z}{\partial Z}\right)&=-\frac{\partial P}{\partial Z}+\frac{\partial^2 V_Z}{\partial Y^2}\label{COLM-Z-O1}.\end{aligned}$$
These equations, and all capital letters used herein are dimensionless. The non-dimensionalization is standard and discussed in Ref. [@WC19]. For the present purposes, since we will not use the dimensional variables at all in the discussion below, we do not go over the non-dimensionalization. Equation is the continuity (conservation of mass) equation, which is balanced at the leading order. Equations , , and are the conservation of linear momentum equations in the $X$, $Y$, and $Z$ directions respectively. Owing to the long and shallow nature of the microchannel, the $X$ and $Y$ equations simply state there is no pressure gradients in those directions at the leading order in $\epsilon$, and the flow is primarily unidirectional in the $Z$ direction.
We are interested in the flow regime in which the characteristic time scale set by the compliant wall’s inertia is much smaller than the characteristic flow time scale. In other words, we assume that the inertia of the elastic solid is negligible, and the unsteadiness in this FSI system is fully determined by the fluid flow. This assumption is often invoked when studying the relaxation time [@PYDHD09] or the start-up time [@MCSPS19] of compliant microchannels. Note, however, it is also possible that, in some regimes, the inertia of the compliant wall may play a role in the unsteady inflation or relaxation of the soft wall, due to the interplay between the deformation and flow [@IWC20; @MCSPS19].
Here, having restricted to a prototypical microsystem in which we can neglect the inertia of the elastic wall, the displacement field developed in Ref. [@WC19] can be transferred smoothly into the unsteady problem. Specifically, for a thick top wall, as considered herein, with $(t/w)^2\gg1$, the (dimensionless) deformation profile at the fluid–solid interface (again, see Ref. [@WC19] for the derivation and discussion) is $$\label{uyinf_dimless}
U_Y^0(X,Z,T)=P(Z,T)\underbrace{\sum_{m=1}^{\infty}\frac{2A_m}{m\pi}\sin\left[m\pi\left(X+\frac{1}{2}\right)\right]}_{=:\mathfrak{G}(X)},$$ where $A_m=\frac{2}{m\pi}[1-(-1)^m]$. Thus, the deformed channel height is $$\label{H}
H(X,Z,T) = 1 + \lambda U_Y^0(X,Z,T) \stackrel{\text{by Eq.~\eqref{uyinf_dimless}}}{=} 1 + \lambda P(Z,T)\mathfrak{G}(X).$$ Here, $\lambda = u_c/h_0$, which is the ratio of the characteristic deformation scale $u_c$ of the elastic solid to the undeformed channel height $h_0$, is termed the *FSI parameter*; for $\lambda=0$, there is no deformation, while for $\lambda=\mathcal{O}(1)$ significant FSI-induced deformation of the flow conduit occurs.
Unlike the case in Ref. [@WC19], here we retain the $\hat{Re}$ terms as $\epsilon\to0$, which yields a nonlinear governing equation for $V_Z$. To make progress, it is standard to integrate Eqs. across a deformed axial cross-section (fixed $Z$) and to introduce the flow rate $$Q(Z,T) \equiv \int_{-1/2}^{+1/2}\int_{0}^{H(X,Z,T)} V_Z(X,Y,Z,T) \,dY\,dX$$ into the formulation (see, e.g., [@SWJ09] and the references therein). However, after this integration, we still need a relation between $V_Z$ and $Q$ to deal with the integral in $Y$. Here, motivated by prior studies on inertial fluid effects in microchannels [@SWJ09; @IWC20], we apply the von Kármán–Polhausen approximation [@panton] for the velocity profile: $$\label{Vz-Q}
V_Z(X,Y,Z,T) = \frac{6Q\big[H(X,Z,T)-Y\big]Y}{\int_{-1/2}^{+1/2}H(X,Z,T)^3 \, dX}.$$ Essentially, this assumption enforces a parabolic (Poiseuille) profile in each axial cross-section, while simultaneously accounting for the flow-wise variation of the height $H$. Also, note that the assumed closure relation is consistent with the previous result [@WC19] in the limit $\hat{Re}\to 0$. Furthermore, the kinematic boundary condition is imposed at the moving fluid–solid interface: $$\label{kinematic}
\frac{\partial H}{\partial T}=\left.V_Y\right|_{Y=H(X,Z,T)}.$$
Then, performing the cross-sectional integration of the governing equations , substituting the ansatz , using the condition , and simplifying, we obtain
\[inteq\]$$\begin{aligned}
\frac{\partial Q}{\partial Z}+\lambda\mathfrak{I}_1\frac{\partial P}{\partial T}&=0,\label{inteq1}\\
\hat{Re}\left[\frac{\partial Q}{\partial T}+\frac{6}{5}\frac{\partial}{\partial Z}\left(\frac{\mathfrak{C}}{\mathfrak{B}^2}Q^2\right)\right]&=-\frac{\partial P}{\partial Z}(1+\lambda\mathfrak{I}_1 P)- \frac{12\mathfrak{A}}{\mathfrak{B}}Q\label{inteq2},\end{aligned}$$
where
$$\begin{aligned}
\mathfrak{I}_i=&\int_{-1/2}^{+1/2}\mathfrak{G}^i(X)\,dX, \quad i = 1,2,\hdots,5,\label{Ii}\\
\mathfrak{A}[P(Z)]=&1+\lambda \mathfrak{I}_1 P(Z),\label{A}\\
\mathfrak{B}[P(Z)]=&1+3\lambda\mathfrak{I}_1P(Z)+3\lambda^2\mathfrak{I}_2P^2(Z)+\lambda^3\mathfrak{I}_3 P^3(Z)\label{B}\\
\mathfrak{C}[P(Z)]=&1+5\lambda\mathfrak{I}_1P(Z)+10\lambda^2\mathfrak{I}_2P^2(Z)+10\lambda^3\mathfrak{I}_3 P^3(Z)+5\lambda^4\mathfrak{I}_4P^4(Z)+\lambda^5\mathfrak{I}_5P^5(Z).\label{C}\end{aligned}$$
Equations and describe the coupling between the fluid flow and the elastic wall’s deformation. Note that $H(X,Z,T)$ is completely determined by the pressure profile, $P(Z,T)$, because $\mathfrak{G}(X)$ is a known function defined by the Fourier series in Eq. . Likewise, the constants $\{\mathfrak{I}_i\}_{i=1}^5$ are known; their values are pre-computed and listed in Table \[tab:table-I\]. Meanwhile, $\mathfrak{A}$, $\mathfrak{B}$ and $\mathfrak{C}$ are functionals of the pressure $P$ and, thus, implicitly functions of $Z$.
$\mathfrak{I}_1$ $\mathfrak{I}_2$ $\mathfrak{I}_3$ $\mathfrak{I}_4$ $\mathfrak{I}_5$
------------------ ------------------ ------------------ ------------------ ------------------
0.542710 0.333333 0.215834 0.143959 0.097864
: \[tab:table-I\] The values of the constants $\{\mathfrak{I}_i\}_{i=1}^5$ defined by Eq. .
Fixing the flow rate upstream, and keeping the outlet of the channel open to atmospheric conditions, we can impose the following boundary conditions: $$\label{bc}
Q|_{Z=0}=1, \qquad P|_{Z=1}=0.$$ Note that no restrictions are imposed on the wall’s deformation at the inlet and outlet. Those would require a matched asymptotic calculation taking into account axial bending (see, e.g., Ref. [@AC18b] for a discussion of this issue in the context of a slender microtube), which is beyond the scope of the present work.
The Base State at Finite $\hat{Re}$
-----------------------------------
At steady state, the boundary conditions on the flow rate indicates that $Q(Z)\equiv1$, while $P_0(Z)$ and $H_0(X,Z)$ should satisfy
\[steady\] $$\begin{aligned}
\frac{d}{d Z}\left[ \hat{Re}\frac{6}{5}\frac{\mathfrak{C}}{\mathfrak{B}^2} + \left(1+\frac{1}{2}\lambda\mathfrak{I}_1 P_0\right)P_0 \right]=&-\frac{12\mathfrak{A}}{\mathfrak{B}}\label{steady1},\\
H_0(X,Z)=&1+\lambda P_0(Z)\mathfrak{G}(X).\label{steady2}\end{aligned}$$
The unknown in Eqs. is $P_0(Z)$, subject to the outlet boundary condition $$\label{steadybc}
P_0(Z=1)=0.$$ If $\hat{Re}\to 0$, Eq. can be rewritten as $-(\mathfrak{B}/12){d P_0}/{d Z} =1$, where $\mathfrak{B}[P(Z)]$ is given by Eq. . This ordinary differential equation can be easily shown to match the previous result in Ref. [@WC19].
Equation subject to Eq. are solved together numerically as a “final value problem” using the classical fourth-order Runge–Kutta (RK4) method implemented using the python package SciPy [@SciPy]. In particularly, within each step of the RK4 method, a nonlinear algebraic problem must be solved because the functionals $\mathfrak{A}$, $\mathfrak{B}$ and $\mathfrak{C}$ depend on the solution $P_0(Z)$. This nonlinear solution step is accomplished using `optimize.fsolve` from SciPy. The scheme is validated for $\hat{Re}=0$ against the analytical result from Ref. [@WC19].
As shown in Fig. \[fig:P0H0\](a), we observe that the inclusion of flow inertia ($\hat{Re}=\mathcal{O}(1)$) results in a larger total pressure drop, $\Delta P \equiv P(1)-P(0)$, and a steeper pressure gradient $dP/dZ$ at the outlet ($Z=1$). After obtaining the pressure distribution $P_0(Z)$, the shape of the deformed channel $H_0(X,Z)$ is just a linear function of $P_0(Z)$ found from Eq. . Thus, as shown in Fig. \[fig:P0H0\](b), the interface deformation at the channel mid-plane, $X=0$, has an identical shape to the pressure distribution.
Perturbation and Linear Stability Problem
=========================================
Let us introduce the following perturbations to the steady finite-$\hat{Re}$ base state $\{Q=1,P=P_0(Z)\}$ derived in the previous section:
\[perturb\] $$\begin{aligned}
Q(Z,T)=&1+\alpha \widetilde{Q}(Z,T),\\
P(Z,T)= &P_0(Z)+\alpha \widetilde{P}(Z,T),\end{aligned}$$
where $\alpha\ll 1$ is an arbitrary small parameter quantifying the magnitude of the axial perturbations $\widetilde{Q}$ and $\widetilde{P}$. Then, it follows that $$H(X,Z,T) = H_0(X,Z) + \alpha \lambda \widetilde{P}(Z,T)\mathfrak{G}(X).$$ Since the actual boundary conditions were imposed on the base state, the perturbations should satisfy homogeneous boundary conditions: $$\label{perturb_bc1}
\widetilde{Q}|_{Z=0}=0, \qquad \widetilde{P}|_{Z=1}=0.$$
Next substituting Eqs. into the governing equations , and only keeping terms up to $\mathcal{O}(\alpha)$, we obtain the linearized equations governing the evolution of perturbations:
\[stability\] $$\begin{aligned}
\frac{\partial \widetilde{Q}}{\partial Z}+\lambda\mathfrak{I}_1\frac{\partial \widetilde{P}}{\partial T}&=0,\label{stability1}\\
\hat{Re}\frac{\partial\widetilde{Q}}{\partial T} +\frac{6}{5}\hat{Re} \frac{\partial}{\partial Z}\left[\frac{2\mathfrak{C}_0}{\mathfrak{B}_0^2}\widetilde{Q}+\left(\frac{\mathfrak{C}_0^{\prime}}{\mathfrak{B}_0^2}
-\frac{2\mathfrak{B}_0^{\prime}\mathfrak{C}_0}{\mathfrak{B}_0^3}\right)\widetilde{P}\right]\label{stability2}
&= -12\frac{\mathfrak{A}_0}{\mathfrak{B}_0}\widetilde{Q}\\
&\phantom{=}+\left(-\lambda\mathfrak{I}_1\frac{dP_0}{dZ}-12\frac{\lambda\mathfrak{I}_1}{\mathfrak{B}_0}+12\frac{\mathfrak{A}_0\mathfrak{B}_0^{\prime}}{\mathfrak{B}_0^2}\right)\widetilde{P}-\mathfrak{A}_0\frac{\partial\widetilde{P}}{\partial Z},\nonumber\end{aligned}$$
where $\mathfrak{A}_0 \equiv \mathfrak{A}[P_0(Z)]$, $\mathfrak{B}_0\equiv \mathfrak{B}[P_0(Z)]$ and $\mathfrak{C}_0\equiv \mathfrak{C}[P_0(Z)]$ are evaluated via Eqs. –, and
$$\begin{aligned}
\mathfrak{B}_0^{\prime}=&3\lambda\mathfrak{I}_1+6\lambda^2\mathfrak{I}_2P_0(Z)+3\lambda^3\mathfrak{I}_3P_0^2(Z),\label{B0prime}\\
\mathfrak{C}_0^{\prime}=&5\lambda\mathfrak{I}_1+20\lambda^2\mathfrak{I}_2P_0(Z)+30\lambda^3\mathfrak{I}_3P_0^2(Z)+20\lambda^4\mathfrak{I}_4P_0^3(Z)+5\lambda^5\mathfrak{I}_5P_0^4(Z).\label{C0prime}\end{aligned}$$
Note that the variables with the subscripts “0" are obtained from the base state solution discussed in the previous section. Thus, they are known for the purposes of the upcoming linear stability calculation.
We restrict our analysis to asymptotic stability of modal perturbations (excluding any effects of transient growth arising from fact that the base state is non-constant and the linear operator is non-normal [@Sch07]). To this end, let $$\label{perturb2}
\widetilde{Q}(Z,T)=Q_1(Z)\mathrm{e}^{-\mathrm{i}\omega T}, \qquad \widetilde{P}(Z,T)=P_1(Z) \mathrm{e}^{-\mathrm{i}\omega T}.$$ Further applying $dQ_1/dZ=\mathrm{i}\omega\lambda\mathfrak{I}_1P_1$, Eqs. can be rewritten in the matrix form: $$\label{stab_matrix}
\begin{pmatrix}
\frac{d}{dZ} & 0\\[1mm] \mathcal{L}_Q & \mathcal{L}_P
\end{pmatrix}\begin{pmatrix} Q_1 \\[1mm] P_1\end{pmatrix}
=\omega\begin{pmatrix}0 & \mathrm{i}\lambda\mathfrak{I}_1\\[1mm] \mathrm{i}\hat{Re} & -\mathrm{i}\frac{6}{5}\hat{Re}\frac{2\mathfrak{C}_0}{\mathfrak{B}_0^2}\lambda\mathfrak{I}_1
\end{pmatrix}\begin{pmatrix} Q_1 \\[1mm] P_1\end{pmatrix},$$ where we have defined the following operators for convenience:
$$\begin{aligned}
\mathcal{L}_Q =& \frac{6}{5}\hat{Re} \frac{d}{dZ}\left(\frac{2\mathfrak{C}_0}{\mathfrak{B}_0^2}\right)+12\frac{\mathfrak{A}_0}{\mathfrak{B}_0},\\[1mm]
\mathcal{L}_P =& \frac{6}{5}\hat{Re}\left[\frac{d}{dZ}\left(\frac{\mathfrak{C}_0^{\prime}}{\mathfrak{B}_0^2}-\frac{2\mathfrak{C}_0\mathfrak{B}_0^{\prime}}{\mathfrak{B}_0^3}\right)+\left(\frac{\mathfrak{C}_0^{\prime}}{\mathfrak{B}_0^2}-\frac{2\mathfrak{C}_0\mathfrak{B}_0^{\prime}}{\mathfrak{B}_0^3}\right)\frac{d}{dZ}\right]+\lambda\mathfrak{I}_1\frac{d P_0}{dZ}+12\frac{\lambda\mathfrak{I}_1}{\mathfrak{B}_0}-12\frac{\mathfrak{A}_0\mathfrak{B}_0^{\prime}}{\mathfrak{B}_0^2}+\mathfrak{A}_0\frac{d}{dZ}.\end{aligned}$$
The corresponding boundary conditions, obtained from Eq. , are $$\label{stab_bc1}
Q_1(0)=0, \qquad P_1(1)=0.$$ Substituting the latter into Eqs. , we obtain two further boundary conditions: $$\label{stab_bc2}
\left.\frac{d Q_1}{dZ}\right|_{Z=1}=0, \qquad \left.\left[\mathcal{L}_P P_1+\frac{6}{5}\hat{Re}\frac{2\mathfrak{C}_0}{\mathfrak{B}_0^2}\frac{dQ_1}{dZ}\right]\right|_{Z=0}=0.$$
Equation and the boundary conditions in Eqs. and constitute a *generalized eigenvalue problem*, in which $\omega\in\mathbb{C}$ is the eigenvalue and $[Q_1,P_1]^\top$ is the eigenfunction. The system is said to be linearly unstable if there exist eigenvalues with $\operatorname{Im}(\omega)>0$ for a combination of the parameters. To solve this eigenvalue problem, we shall employ the Chebyshev pseudospectral numerical method. In this way, we can resolve the eigenspectra to determine if the system exhibits linear stability (or instability).
Results and Discussion
======================
The Chebyshev pseudospectral method [@SH01; @Boyd00] for the linear stability problem is implemented as described in [@IWC20], using the python package SciPy [@SciPy]. Simply speaking, the eigenfunctions $Q_1$ and $P_1$ are approximated with an $N$-th degree polynomial each, then the generalized eigenvalue problem (Eqs. , and ) is discretized by enforcing the satisfaction of the equations at $N+1$ Gauss–Lobatto points. Specifically, Eq. is required to be satisfied at $N-1$ interior Gauss–Lobatto points while the boundary conditions and are imposed at the two end points. Furthermore, since the boundary conditions are homogeneous, the right-hand-side matrix in Eq. is singular.
The eigenspectra for our genearalized eigenvalue problem are discrete. Since the left-hand-side matrix is real while the right-hand-matrix is purely imaginary, the resulting eigenspectra in $\mathbb{C}$ are symmetric about the imaginary axis. Multiple eigenvalue pairs, which are complex conjugates and thus have the same magnitude, are observed in our calculations (see Figs. \[fig:eig1\] and \[fig:eig2\]). The eigenvalues are ordered with ascending magnitude and thus, the eigenvalue pairs share the same position in the $\mathbb{C}$ plane.
In the following discussion, different modes are referred to as the eigenfunctions corresponding to eigenvalues with different magnitudes $|\omega|$. For example, the first mode corresponds to the eigenvalue with the smallest magnitude, and the second mode has the eigenvalue with the second smallest magnitude, and so on. Furthermore, it is worth pointing out that, for our generalized eigenvalue problem , in principle, the eigenspectra should consist of an infinite number of discrete points, as the differential operators are infinite dimensional objects. However, since we numerically solve the problem by pseudospectral discretization, the resolution of the eigenspectra is determined by the number of Guass–Lobatto points. Therefore, considering the limits numerical linear algebra algorithms, the eigenspectra shown are the part for which the magnitudes of the eigenvalues are relatively small, whose computation is tractable using a finite number of grid points. The following results are calculated with $N=1000$ Gauss–Lobatto points for both eigenfunctions, $Q_1$ and $P_1$, with only the first 500 eigenvalues shown in Figs. \[fig:eig1\] and \[fig:eig2\]. The accuracy of the calculations is assured by comparing the latter results to those with $N=800$ Gauss–Lobatto points for verification.
\
First, we investigate the eigenspectra by varying $\hat{Re}$ and fixing $\lambda=0.5$, as shown in Fig. \[fig:eig1\]. With this value of $\lambda$, appreciable deformation is observed in the base state (see Fig. \[fig:P0H0\](b)). With the increase of $\hat{Re}$, ranging from $0.01$ to $3$, the imaginary parts of the majority of eigenvalues increase. However, no instabilities are observed as $\operatorname{Im}(\omega)<0$ for all cases considered. Several modes with purely imaginary eigenvalues are found. Specifically, for $\hat{Re}=0.01$, there are 6 purely decaying modes, while only 2 such modes are observed for the other three cases. Among these modes, the one closest to the real axis is of interest because it represents the slowest decaying mode of the system. Table \[tab:table-II\] lists the largest imaginary part of all modes for the four values of $\hat{Re}$ considered. Interestingly, we do not observe a monotonic trend with the increase of $\hat{Re}$. Indeed, even without FSI, it is expected that a duct flow becomes more unstable as $\hat{Re}$ increases [@SH01].
Let us now take a look at the real part of the eigenvalues. For each case, the difference in the magnitudes of the real parts of two different modes is much larger than that of their imaginary parts, which is why the eigenspectra have a “seagull" shape with a pair of relatively flat wings. The multiple eigenvalues with large-magnitude real parts evidence the existence of the highly-oscillatory eigenmodes, indicating the inherent stiffness of this FSI system. Comparing the cases of different $\hat{Re}$ in Fig. \[fig:eig1\], the real parts of the eigenvalues display a decreasing trend with the increase of $\hat{Re}$.
Next, we keep $\hat{Re}=1$ fixed while varying the FSI parameter, $\lambda$. Note that our system is governed by two dimensionless groups, unlike classical hydrodynamics stability problems [@SH01], which is the result of the coupled physics involved in two-way FSI. Still, as shown in Fig. \[fig:eig2\], no instabilities are observed by varying $\lambda$, but there are some interesting differences with respect to varying $\hat{Re}$. For instance, in a less compliant system with $\lambda=0.1$, there are no purely decaying modes; all modes have non-zero real parts, meaning they are intrinsically oscillatory. It is also observed that $\operatorname{Re}(\omega)$ decreases as $\lambda$ increases.
As for the eigenfunctions, in Fig. \[fig:eigfunc\] we show the first four modes for the case of $\hat{Re}=1$ and $\lambda=0.5$ as an example. The first two modes (labelled “mode1" and “mode2") correspond to two eigenvalues with $\operatorname{Re}(\omega)=0$ and $\operatorname{Im}(\omega)<0$ from Fig. \[fig:eig1\](c), for which the eigenfunctions are found to be real. In particular, $Q_1$ is monotonically increasing from the inlet to the outlet, while $P_1$ is relatively flat for most of the channel, displaying a sharp decrease near the outlet. For the other two modes (labelled “mode3" and “mode4"), the corresponding eigenfunctions exhibit spatially-varying crests or troughs. The eigenfunctions of the fourth mode are “wavier” than the third mode. This observation is typical, and more humps would be observed in the higher modes, if we were to plot them.
$\hat{Re}$ 0.01 0.1 1 3
----------------------------- ----------- ----------- ----------- -----------
$\operatorname{Im}(\omega)$ $-4.1459$ $-4.1582$ $-4.4004$ $-1.8194$
: \[tab:table-II\] The largest imaginary part of the eigenvalues corresponding to different $\hat{Re}$ with $\lambda=0.5$ fixed.
Conclusion
==========
In this preliminary assessment of linear stability of the novel coupled flow and deformation solution in a thick-walled rectangular microchannel from Ref. [@WC19], we found that, within a range of the reduced Reynolds number, $\hat{Re}$, and the FSI parameter, $\lambda$, the inflated base state is linearly stable to infinitesimal flow-wise perturbations. With the Chebyshev pseudospectral method, we were able to resolve multiple highly oscillatory but stable eigenmodes, which indicates the stiffness of the FSI system. Unlike problems of classical hydrodynamic stability of duct flows [@SH01], this FSI problem is governed by two dimensionless groups ($\hat{Re}$ and $\lambda$), and they both have a non-trivial influence on the eigenspectrum.
Importantly, unlike previous work [@VK13], wherein the linear stability analysis was conducted locally on an approximately flat base state and the nonlinear pressure gradient was imported from computational fluid dynamics (CFD) simulations in a static but deformed geometry, the base state that we perturbed herein is non-flat, computed consistently from two-way coupled FSI theory [@WC19]. This base state, featuring a nonlinear pressure gradient, was derived from the closed-form 1D model consisting of Eqs. and . Indeed, in Ref. [@VK13], the nonlinear pressure gradient was conjectured to be the most important factor in triggering instability. The velocity profile, which was chosen in Ref. [@VK13] to be a quartic because it was closer to the output of CFD simulations than the parabolic profile, was thought to be slightly less significant. In this respect, even though the velocity profile in our analysis is parabolic (to be consistent with the $\hat{Re}\to0$ solution [@WC19]), other profiles shapes (as function of $Y$) are allowed within the von Kármán–Polhausen approximation in Eq. , as long as the boundary conditions at the top and bottom walls are satisfied.
Admittedly, our different formulation of the linear stability problem led us to a different conclusion from Ref. [@VK13], and we did not reproduce the instabilities observed therein. Nevertheless, the experiments [@VK13] are reproducible [@KB16] and the ultra-fast mixing phenomenon at low Reynolds number is striking. Therefore, the phenomenon of low-Reynolds-number FSI-induced instabilities remains relevant to understand from scratch (without “infusing” the linear stability calculation with CFD or experimental results) due to its potential relevance as new modality of mixing in microfluidics [@OW04]. One of the possible reasons that our reduced model did not predict an instability is that we fixed the upstream flow rate and set the outlet pressure to gauge. These boundary conditions might not perfectly match the experimental conditions in Ref. [@VK13]. Another possibility may be that, the inertia of the elastic solid, which we have neglected, plays a role in triggering the instability.
In future work, we would like to address the effect of different boundary conditions on the linear (in)stability problem formulated herein. For example, we might consider fixing the total pressure drop $\Delta P$ across the length of the channel, leaving the inlet flow rate to be “free.” Another extension of the present theory can be accomplished by properly introducing the compliant wall’s inertia (and unsteadiness) into the formulation. This extension requires updating the current solid mechanics model by properly justifying an independent time scale over which the elastic deformation varies.
This research was supported by the US National Science Foundation under grant No. CBET-1705637. I.C.C. is grateful to Prof. Michail Todorov for his invitation to present this research at the Twelfth Conference of the Euro-American Consortium for Promoting the Application of Mathematics in Technical and Natural Sciences (AMiTaNS’20).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $G$ be a graph obtained by taking $r\geq 2$ paths and identifying all first vertices and identifying all the last vertices. We compute the Castelnuovo–Mumford regularity of the quotient $S/I(X)$, where $S$ is the polynomial ring on the edges of $G$ and $I(X)$ is the vanishing ideal of the projective toric subset parameterized by $G$. The case we consider is the first case where the regularity was unknown, following earlier computations (by several authors) of the regularity when $G$ is a tree, cycle, complete graph or complete bipartite graph, but specially in light of the reduction of the computation of the regularity in the bipartite case to the computation of the regularity of the blocks of $G$. We also prove new inequalities relating the Castelnuovo–Mumford regularity of $S/I(X)$ with the combinatorial structure of $G$, for a general graph.'
address:
- 'CMUC, Department of Mathematics, University of Coimbra, Apartado 3008, EC Santa Cruz 3001–501 Coimbra, Portugal. '
- |
Departamento de Matemática\
Instituto Superior Técnico\
Universidade de Lisboa\
Avenida Rovisco Pais, 1\
1049-001 Lisboa, Portugal
- |
Departamento de Matemáticas\
Centro de Investigación y de Estudios Avanzados del IPN\
Apartado Postal 14–740\
07000 Mexico City, D.F.
author:
- Antonio Macchia
- Jorge Neves
- |
\
Maria Vaz Pinto
- 'Rafael H. Villarreal'
title: |
Regularity of the vanishing ideal over a\
parallel composition of paths
---
[^1]
Introduction {#sec: intro}
============
Let $K$ be a field and denote by $S$ the polynomial ring $K[t_1,\dots,t_s]$ with the standard grading. If $M$ is a finitely generated graded $S$-module and $$\label{eq: minimal graded free resolution}
0\to F_c\stackrel{\phi_c}{{\longrightarrow}} \cdots {{\longrightarrow}} F_2 \stackrel{\phi_2}{{\longrightarrow}} F_1 \stackrel{\phi_1}{{\longrightarrow}} F_0 {\longrightarrow}M$$ is a minimal graded free resolution, the Castelnuovo–Mumford regularity of $M$ is the integer: $${\operatorname{reg}}M = \max_{i,j} {\left \{ j-i \mid b_{ij} \neq 0 \right \}},$$ where $b_{ij}$ are the graded Betti numbers of $M$, defined by $F_i \cong \oplus_{j\in \mathbb{Z}} S(-j)^{b_{ij}}$. The regularity of $M$ reflects the size of the degrees of the entries of the matrices in (\[eq: minimal graded free resolution\]), and therefore, in a certain sense, the complexity of $M$ as a graded module. In the case when $M=S/I$, with $I$ a Cohen–Macaulay homogeneous ideal, we know that (cf. [@monalg Proposition 4.2.3]): $$\label{eq: relation of reg with Hilbert Series}
{\operatorname{reg}}S/I= \max_j {\left \{ j-c\mid b_{cj}\not = 0 \right \}} = \deg F_{S/I}(t) + \dim S/I,$$ where $F_{S/I}(t)$ is the Hilbert Series of the module $S/I$ in rational function form.
Recently, many authors have studied the Castelnuovo-Mumford regularity of ideals associated to some combinatorial structure. For square free monomial ideals generated in degree $2$, so-called *edge ideals* as their generators correspond to the edges of a graph (cf. [@monalg Chapter 6]), the regularity can be bounded using the induced matching number of the associated graph (cf. [@HaVT08], [@Ka06 Lemma 2.2] and [@Wo14]). For chordal graphs, it has been shown that the regularity actually coincides with this graph invariant (see [@HaVT08 Corollary 6.9]). Several families of binomial ideals associated with a combinatorial structure have also been studied. The class of *toric ideals*, i.e., the ideal of relations of the edge subring of a graph, whose generators correspond to *even closed walks* on the graph (cf. [@monalg Chapter 8]), is one such example. For a complete graph $\mathcal{K}_n$, the regularity of its edge subring is equal to $\lfloor n/2 \rfloor$, while for a complete bipartite graph $\mathcal{K}_{a,b}$, this invariant coincides with $\min \{a,b\}-1$ (cf. [@monalg]). Lower and upper bounds for the regularity of toric ideals, in terms of the structure of the underlying graph, have recently been established (cf. [@BiO'KVT]). Another class of binomial ideals which has been extensively studied in recent times is the class of *binomial edge ideals*. These ideals are generated by the maximal minors of a $2 \times s$ generic matrix, whose column indices correspond to the edges of a graph. The regularity of these ideals can also be expressed and bounded in terms of graph-theoretic invariants (cf. [@EnZa15; @KiSM16; @MaMu13]).
For the purposes of this work, $K$ will be a finite field of cardinality $q$. In the rest of the paper all *graphs* will be undirected and without loops; multiple edges are allowed. The vertex set of a graph $G$ will be denoted by $V_G$ and its edge set by $E_G$. We denote the number of edges by $s$ and we fix an ordering of the set of edges given by an identification of $E_G$ with the set of variables of $K[t_1,\dots,t_s]$. If $H$ is a subgraph of $G$ we denote by $K[E_H]$ the polynomial subring on the variables of $E_H$, under the above identification. To $G$ we associate a set $X$ defined by $$\label{eq: definition of X}
X = {\left \{ ({\mathbf{x}}^{t_1},{\mathbf{x}}^{t_2},\dots,{\mathbf{x}}^{t_s})\in {\mathbb{P}}^{s-1} \mid {\mathbf{x}}\in (K^*)^{V_G} \right \}},$$ where, if ${\mathbf{x}}=\sum_{v\in V_G} x_v v$, with $x_v\in K^*$, for all $v\in V_G$, and $t_i$ is the edge ${\left \{ v,w \right \}}$ (with $v\not = w$), we set ${\mathbf{x}}^{t_i}=x_{v}x_{w}$. As ${\mathbf{x}}^{t_i}\not = 0$, for all $i$, $X$ is a subset of the projective torus ${\mathbb{T}}^{s-1}\subset {\mathbb{P}}^{s-1}$. We refer to $X$ as *the projective toric subset parameterized by $G$*. Denote by $I(X)$ the vanishing ideal of $X$. Observe that $$I(\mathbb{T}^{s-1})=(t_1^{q-1}-t_s^{q-1},\dots,t_{s-1}^{q-1}-t_s^{q-1}) \subset I(X).$$
The notion of parameterized projective toric subsets and the study of their vanishing ideals was introduced in [@ReSiVi11]. Unlike in the case of the edge ideal of $G$, we know that $I(X)$ is always a Cohen–Macaulay homogeneous binomial ideal of height $s-1$ (Cf. [@ReSiVi11 Theorem 2.1]).
In the original definition of a parameterized projective toric subset, $G$ is assumed to be a simple graph. However, on the one hand, we note that multiple edges play no part in the invariants of $K[E_G]/I(X)$. More precisely, if $G'$ is the simple graph obtained from $G$ by removing all extra edges through any two given vertices and $X'$ is the projective toric subset parameterized by $G'$, then $$K[E_{G'}]/I(X') \cong K[E_{G}]/I(X),$$ simply because $t_{j}-t_i\in I(X)$, for every extra edge $t_j$ between the endpoints of $t_i$. On the other hand, allowing extra edges eases notation and simplifies statements and proofs.
As $X$ is a finite set, the value of the Hilbert function of $K[E_G]/I(X)$ is eventually equal to $|X|$, the cardinality of $X$; therefore, $\deg K[E_G]/I(X)=|X|$. A formula for the degree was first given in [@ReSiVi11] for connected graphs and then generalized to any graph in [@NeVPVi15 Theorem 3.2]: $$\label{eq: degree}
\deg K[E_G]/I(X) =
\begin{cases}
\left (\frac{1}{2}\right)^{\gamma -1 } (q-1)^{n-m+\gamma-1}, \text{ if } \gamma\geq 1\text{ and } q\text{ is odd,}\\
(q-1)^{n-m+\gamma-1},\text{ if } \gamma\geq 1\text{ and } q\text{ is even,}\\
(q-1)^{n-m-1}, \text{ if }\gamma=0,
\end{cases}$$ where ($q$ is the cardinality of $K$), $n$ is the cardinality of $V_G$, $m$ is the number of connected components of $G$ and $\gamma$ the number of those that are non-bipartite.
Using the identity (\[eq: relation of reg with Hilbert Series\]) and the fact that $\dim K[E_G]/I(X)=1$, we deduce that the regularity of $K[E_G]/I(X)$ coincides with its *regularity index*, i.e., the minimum degree $d$ for which the value of the Hilbert function at $k$ is equal to the value of the Hilbert polynomial at $k$, for every (Cf. [@monalg Corollary 4.1.12].) Since the Hilbert function of $K[E_G]/I(X)$ is strictly increasing for and the Hilbert polynomial is equal to $|X|=\deg K[E_G]/I(X)$ we conclude that ${\operatorname{reg}}K[E_G]/I(X)$ is the minimum $d$ for which the value of the Hilbert function at $d$ is equal to $|X|=\deg K[E_G]/I(X)$.
In Table \[table: values of reg\] we list cases for which this invariant is known.
${\operatorname{reg}}K[E_G]/I(X)$
------------------------------------------- ----------------------------------------------------------------------------------------
$X={\mathbb{T}}^{s-1}$ $(s-1)(q-2)$
$G=\mathcal{K}_n$ $\lceil (n-1)(q-2)/2 \rceil$
$G=\mathcal{K}_{a,b}$ $(\max{\left \{ a,b \right \}}-1)(q-2)$
$G=C_{2k}$ $(k-1)(q-2)$
$G=\mathcal{K}_{\alpha_1,\dots,\alpha_r}$ $\max{\left \{ \alpha_1(q-2),\dots,\alpha_r(q-2),\lceil (n-1)(q-2)/2\rceil \right \}}$
: Known values of ${\operatorname{reg}}K[E_G]/I(X)$
\[table: values of reg\]
When $X$ coincides with the projective torus ${\mathbb{T}}^{s-1}$ (which, from (\[eq: degree\]), is the case, for example, if $G$ is a tree or an odd cycle), $$I(X) = (t_1^{q-1}-t_s^{q-1},\dots,t_{s-1}^{q-1}-t_s^{q-1}).$$ Thus the regularity can be computed from (\[eq: relation of reg with Hilbert Series\]), (see also [@SaVPVi11]). The regularity in the case is given in [@GoReSa13 Remark 3]. The case $G= \mathcal{K}_{a,b}$ is given in [@GoRe08 Corollary 5.4] and the case of an even cycle, $G=C_{2k}$, in [@NeVPVi15 Theorem 6.2]. In the case of a complete multipartite graph, this invariant was computed in [@NeVP14 Theorem 4.3]. (Here $r \geq 3$ and the $n$ in the formula is )
A graph $G$ is said to be $2$-connected if $|V_G|>2$ and, for every vertex $v\in V_G$, the graph $G-v$ is connected. Any graph decomposes into *blocks*, which consist of either maximal $2$-connected subgraphs, single edges or isolated vertices. When $G$ is bipartite, we know that ${\operatorname{reg}}(E_G)/I(X)$ can be computed from its block decomposition. More precisely, if $G$ is a simple bipartite graph with no isolated vertices and $H_1,\dots,H_r$ are the blocks of $G$, then $$\label{eq: reg additive on blocks}
{\operatorname{reg}}K[E_G]/I(X) = \sum_{k=1}^r {\operatorname{reg}}K[E_{H_k}]/I(X_k) + (r-1)(q-2),$$ where $X_k$ is the projective toric subset parameterized by the graph $H_k$, for each $k=1,\dots,r$ (cf. [@NeVPVi14 Theorem 7.4]). This reduces the problem of computing ${\operatorname{reg}}K[E_G]/I(X)$ for a bipartite graph to the case of $2$-connected graphs. Notice that (\[eq: reg additive on blocks\]), together with the formula for the regularity in the case of even cycles, gives the regularity for any bipartite *cactus graph* (a simple graph the blocks of which are edges or even cycles).
A $2$-connected graph can be reconstructed from one of its cycles by adding a path by its endpoints (also known as an *ear*) to the cycle and successively repeating this operation (a finite number of times) to the graphs obtained (cf. [@Di10 Proposition 3.1.1]). The simplest $2$-connected graph is a cycle. The second simplest $2$-connected graph is a cycle with an attached ear. This graph can also be obtained by identifying the endpoints of $3$ paths, which, in turn, is also known as the *parallel composition* of $3$ paths. Therefore the parallel composition of $3$ paths is the first case of a $2$-connected graph for which the regularity of $K[E_G]/I(X)$ was not known.
The aim of this work is to compute the Castelnuovo–Mumford regularity of $K[E_G]/I(X)$, when $X$ belongs to the family of projective toric subsets parameterized by a graph given as the parallel composition of $r\geq 2$ paths, as illustrated in Figure \[fig: G v2\]. (Notice that this graph may well have multiple edges if more than one $P_i$ has length equal to $1$.)
(0,-.6) circle (1pt); (3.45,-.6) circle (1pt);
(0,-.6)..controls (0.15,0)..(.75,0); in [0.75,1.5]{} [(,0) circle (1pt) – (+.75,0) ]{}; (2.5,0) node[$\cdots$]{}; (2.7,0)..controls (3.25,0)..(3.45,-.6); (3,.15) node [$P_1$]{};
(0,-.6)..controls (0.15,-.45)..(.75,-.45); in [0.75,1.5]{} [(,-.45) circle (1pt) – (+.75,-.45) ]{}; (2.5,-.45) node[$\cdots$]{}; (2.7,-.45)..controls (3.25,-.45)..(3.45,-.6); (3,-.3) node [$P_2$]{};
(0,-.6)..controls (0.15,-1.2)..(.75,-1.2); in [0.75,1.5]{} [(,-1.2) circle (1pt) – (+.75,-1.2) ]{}; (2.5,-1.2) node[$\cdots$]{}; (2.7,-1.2)..controls (3.25,-1.2)..(3.45,-.6); (3,-.67) node[$\vdots$]{}; (3,-1.05) node [$P_r$]{};
Our first main result concerns the bipartite case.
\[th: main theorem\] Let $X$ be the projective toric subset parameterized by the parallel composition of $r\geq 2$ paths, the lengths of which, $k_1,\dots,k_r$, have the same parity. Then $${\operatorname{reg}}K[E_G]/I(X) =
\begin{cases}
(\lfloor k_1/2\rfloor +\cdots +\lfloor k_r/2 \rfloor)(q-2),\; \text{if $k_i$ are odd,}\\
( k_1/2 +\cdots +k_r/2 -1)(q-2),\; \text{if $k_i$ are even.}
\end{cases}$$
We prove this result in Section \[sec: main thm\], by proving the two inequalities involved. The lower bound is a straightforward consequence of the fact that $G$ is bipartite (cf. (\[eq: lower bound on reg for bipartite G\]) and Lemma \[lemma: easy inequality\], below). For the upper bound we divide the proof into two cases. The case of $k_i$ even is worked out by induction on $r$ and arguing using suitable coverings of $G$ (cf. Proposition \[prop: opp inequality in the case k\_i even\]). The case of $k_i$ odd is harder and relies on a characterization of the homogeneous binomials in $I(X)$ (cf. Theorem \[thm: opp inequality in the case k\_i odd\]).
With Theorem \[th: main theorem\] we are able to study the non-bipartite case.
\[th: main theorem 2\] Let $X$ be the projective toric subset parameterized by a graph $G$ that is the parallel composition of $r\geq 2$ paths, the lengths of which have mixed parities. Then $${\operatorname{reg}}K[E_G]/I(X) = {\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2),$$ where $H_1$ is the parallel composition of the paths of odd lengths, $H_2$ is the parallel composition of the paths of even lengths, and $X_1, X_2$, respectively, are the projective toric subsets they parameterize.
We point out that the formula of Theorem \[th: main theorem 2\] includes the case when only one path has length of different parity. In this situation, the corresponding summand of the formula does not follow from Theorem \[th: main theorem\], rather, it can be retrieved from the first formula of Table \[table: values of reg\]; more precisely, if $H_i$ consists of a path of length $k$ then ${\operatorname{reg}}K[E_{H_i}]/I(X_i) = (k-1)(q-2)$.
The proof of Theorem \[th: main theorem 2\] occupies the second half of Section \[sec: main thm\]. As with our other main result we prove the two inequalities separately (cf. Lemma \[lemma: easy inequality in the non-bipartite case\] and Theorem \[thm: harder inequality in the non-bipartite case\]). This time, the easier inequality is the one giving the upper bound. For the lower bound inequality we need to use different techniques to those used in the proof of Theorem \[th: main theorem\].
Section \[sec: prelim\] provides the background theory and the results that are used in our proofs. We single out the new contributions of Proposition \[prop: general lower bound in contraction\], Proposition \[prop: regularity inequality over subgraphs\] and Proposition \[prop: lower bound using independent set\], as we believe these results will prove useful in the study of the regularity for a general graph.
Preliminaries {#sec: prelim}
=============
Let $K$ be a finite field of cardinality $q$. As in Section \[sec: intro\], $G$ will denote a graph with edge set $E_G$ of cardinality $s$ (we always assume that $G$ has no isolated vertices). We fix an identification of the variables of $K[t_1,\dots,t_s]$ with $E_G$ and denote the former by $K[E_G]$. Let $X$ be the projective toric subset parameterized by $G$, as defined in (\[eq: definition of X\]). If $a=(a_1,\dots,a_s)\in {\mathbb{N}}^s$, $t^a$ denotes the monomial $t_1^{a_1}\cdots t_s^{a_s}\in K[E_G]$.
We start by recalling a criterion for membership in $I(X)$ of a homogeneous binomial that only involves the combinatorics of $G$. It involves checking a linear congruence at every vertex of the graph.
(0,0) – (1.5,2.5); (1.5,2.5) – (2.25,3); (1.5,2.5) – (2.25,2);
(0,0) – (-2.5,1.5); (-2.5,1.5) – (-2,2.25); (-2.5,1.5) – (-3,2.25);
(0,0) – (-3,-1); (-3,-1) – (-3.5,-1.75); (-3,-1) – (-3.8,-0.75);
(0,0) – (1.5,-2.5); (1.5,-2.5) – (2.25,-2.75); (1.5,-2.5) – (1.75,-3.25);
(0,0) – (3,-.5); (3,-.5) – (3.75,0);
(0,0) circle (14pt); (0,0) node [$v$]{};
(1.5,2.5) circle (14pt); (-2.5,1.5) circle (14pt); (-3,-1) circle (14pt); (1.5,-2.5) circle (14pt); (3,-.5) circle (14pt);
(2,.1) node [$t_{i_1}$]{}; (.5,1.85) node [$t_{i_2}$]{}; (-1.65,.6) node [$t_{i_3}$]{}; (-1.5,-1.1) node [$t_{i_4}$]{};
(-.45,-1.77) node [$.$]{}; (-.2,-1.85) node [$.$]{}; (0.07,-1.84) node [$.$]{};
(1.4,-1.45) node [$t_{i_r}$]{};
Let $v\in V_G$ and let $t_{i_1},\dots,t_{i_r}$ be the edges incident to $v$ (cf. Figure \[fig: congruence\]). Then by [@NeVP14 Lemma 2.3], a homogeneous binomial $t^a-t^b\in K[E_G]$ belongs to $I(X)$ if and only if, for every vertex $v\in V_G$, if $i_1,\dots,i_r$ are the indices of the edges incident to it, the congruence $$\label{eq: congruence}
a_{i_1}+\cdots +a_{i_r} \equiv b_{i_1}+\cdots + b_{i_r} \pmod{q-1}$$ is satisfied. It follows easily from this criterion, that if $H$ is a subgraph of $G$ and $Y$ is the projective toric subset parameterized by $H$, then $I(Y)= I(X)\cap K[E_H]$.
The following lemma will be used in the proof of Theorem \[thm: opp inequality in the case k\_i odd\], below. Recall that an ear of $G$ is a path which is maximal with respect to the condition that all of its interior vertices have degree $2$ in $G$.
\[lemma: swaping exponents\] Let $t^a-t^b\in K[E_G]$ be a homogeneous binomial. Let $t_i$ and $t_j$ be edges along an ear of $G$ in a same parity position along this path. Let $\sigma \colon K[E_G] \to K[E_G]$ be the automorphism defined by swapping the two edges $t_i$ and $t_j$. Then $$t^a-t^b \in I(X) \iff \sigma(t^a) - \sigma(t^b) \in I(X).$$
It is clear we can reduce to the case illustrated in Figure \[fig: edges along ear\].
(0,0) – (9,0);
(0,0) – (-.9,-.5); (0,0) – (-.8,.5); (-1,.2) node [$\vdots$]{};
(9,0) – (9.9,-.5); (9,0) – (9.8,.5); (10,.2) node [$\vdots$]{};
in [0,3,6,9]{} [(,0) circle (14pt) ]{};
(0,0) node [$v_1$]{}; (3,0) node [$v_2$]{}; (6,0) node [$v_3$]{}; (9,0) node [$v_4$]{};
(1.5,.5) node [$t_{i}$]{}; (4.5,.5) node [$t_{k}$]{}; (7.5,.5) node [$t_{j}$]{};
Since $\sigma(t^{{a}})-\sigma(t^{{b}})$ is homogeneous if and only if $t^a-t^b$ is, it suffices to check the equivalence of the system of $4$ linear congruences given by the $4$ vertices $v_1,v_2,v_3$ and $v_4$. Let ${E(v_i)}$ denote the set of edges incident to $v_i$ and denote by $E_1$ the set $E(v_1)\setminus {\left \{ t_i \right \}}$ and, likewise, $E_4=E(v_4)\setminus {\left \{ t_j \right \}}$. Let $$A_1 =\sum_{t_\ell\in E_1} a_\ell,\quad
A_4 =\sum_{t_\ell\in E_4} a_\ell,\quad
B_1 =\sum_{t_\ell\in E_1} b_\ell,\quad \text{and}\quad
B_4 =\sum_{t_\ell\in E_4} b_\ell.$$ Then, we need to show that the two systems of congruences modulo $q-1$ $$\begin{cases}
A_1+a_i \equiv B_1 + b_i \\
a_i+a_k \equiv b_i + b_k \\
a_k+a_j \equiv b_k + b_j \\
a_j+A_4 \equiv b_j + B_4 \\
\end{cases} \quad \text{and}\quad\:\:
\begin{cases}
A_1+a_j \equiv B_1 + b_j \\
a_j+a_k \equiv b_j + b_k \\
a_k+a_i \equiv b_k + b_i \\
a_i+A_4 \equiv b_i + B_4 \\
\end{cases}$$ are equivalent, which is clearly true.
Our approach to computing ${\operatorname{reg}}K[E_G]/I(X)$ is to consider an Artinian quotient $K[E_G]/I(X,g)$, where $g\in K[E_G]$ is a suitable monomial.
\[prop: membership of a zero dimensional reduction\] Let $g\in K[E_G]$ be a monomial.
1. There exists a monomial order and a binomial Gröbner basis $\mathcal B$ of $I(X)$ such that $\mathcal B \cup {\left \{ g \right \}}$ is a Gröbner basis for the ideal $(I(X),g)\subset K[E_G]$.
2. A monomial $t^a\in K[E_G]$ belongs to $(I(X),g)$ if and only if there exists a monomial such that $t^a-gt^b$ is homogeneous and belongs to $I(X)$.
Since $I(X)$ is generated by homogeneous binomials, the Gröbner basis obtained from such a set, after fixing any monomial order, consists of homogeneous binomials, by Buchberger’s Algorithm. Let $t_{i_1},\dots,t_{i_r}$ be the variables dividing $g$. Fix the graded reverse lexicographical order after reordering the variables in way such that are the last variables of the ring. Let $\mathcal{B}$ be a binomial Gröbner basis of $I(X)$ with respect to such order. To prove (i) it suffices to show that $S(f,g)$ reduces to $0$ modulo $\mathcal{B}\cup {\left \{ g \right \}}$, for every $f\in \mathcal{B}$. Let $f=t^a-t^b\in \mathcal{B}$. Assume, without loss of generality, that $\operatorname{lt}(f)=t^a$. If $t_{i_r}$ divides $t^a$, then $t_{i_r}$ does not divide $t^b$ (we may assume the generating set we start with consists of irreducible binomials). This implies that $t^b\succ t^a$, hence $t_{i_r}$ does not divide $t^a$. Arguing in the same way, by induction, we conclude that none of $t_{i_1},\dots,t_{i_r}$ divides $t^a$ and thus $\gcd(g,t^a)=1$. Accordingly, $$S(f,g) = g(t^a-t^b) - t^ag = -gt^b$$ which reduces to zero modulo $\mathcal{B}\cup {\left \{ g \right \}}$. This completes the proof of (i).
Let $t^a$ be a monomial. One direction of the equivalence in (ii) is clear. Assume that Then, considering the Gröbner basis $\mathcal{B}\cup {\left \{ g \right \}}$ obtained in (i), $t^a$ has zero remainder after division with $\mathcal{B}\cup {\left \{ g \right \}}$. Since the division of a monomial by a binomial is still a monomial, the division algorithm stops the first time $g$ is used. Thus, the partial quotients of division are monomials $t^a = h_0$, $h_1,\dots,h_k$ such that $h_{i}-h_{i-1}\in I(X)$, for all $i=1,\dots,k$ and such that $g$ divides $h_k$. Writing $h_k=gt^b$, we get a homogeneous binomial $t^a-gt^b$ which belongs to $I(X)$, as required.
\[prop: computing reg by reducing to Artinian quotient\] Let $g\in K[E_G]$ be a monomial. Then $K[E_G]/(I(X),g)$ is zero in degree $d$ if and only if $d\geq {\operatorname{reg}}K[E_G]/I(X) + \deg(g)$.
We denote $K[E_G]/I(X)$ by $R$ and, by abuse of notation, $K[E_G]/(I(X),g)$ by $R/g$. Since $g$ is an $R$-regular element and $R$ is Cohen–Macaulay, $$\dim R/g = \dim R-1 = 0.$$ Moreover, since $R/g$ is a quotient of a polynomial ring with the standard grading by a homogeneous ideal, its regularity index is the minimum degree $d$ for which $(R/g)_d=0$. (It is easy to see that $(R/g)_d=0$, for some $d$, implies $(R/g)_{d+k}=0$, for all $k\geq 0$.) Hence we need to show that the regularity index of $R/g$ is equal to ${\operatorname{reg}}K[E_G]/I(X) + \deg(g)$. Consider the following exact sequence of graded $K[E_G]$-modules: $$\nonumber
0\to R[-\deg(g)]\stackrel{\cdot g}{\longrightarrow} R \to R/g\to 0.$$ Comparing the degree of the Hilbert series of the three terms and using the identity (\[eq: relation of reg with Hilbert Series\]), we get $\deg F_{R/g} +1 = {\operatorname{reg}}R +\deg(g)$, where $F_{R/g}$ is the Hilbert Series of the $K[E_G]$-module $R/g$ in rational function form. As $\deg F_{R/g} +1$ is the regularity index (cf. [@monalg Corollary 4.1.12]), we have proved the claim.
We note that the following proposition can be easily derived from [@NeVPVi14 Theorem 7.4] in the bipartite case, and from [@SaVPVi11 Corollary 3.10] and [@GoReHe03 Lemma 1] in the non-bipartite case, when G is a unicyclic connected graph and the only cycle of G is odd. Here, we do not assume G is bipartite nor a unicyclic connected graph with an odd cycle.
\[prop: aditivity on leaves\] Let $v\in V_G$ be a vertex of degree $1$. Assume that $|E_G|>1$. Consider the graph and denote by $X'$ the projective toric subset parameterized by it. Then $${\operatorname{reg}}K[E_G]/I(X) = {\operatorname{reg}}K[E_{G'}]/I(X') + (q-2).$$
Let $t_i\in E_G$ be incident to $v$ and let $t_j\in E_G\setminus t_i$. According to Proposition \[prop: computing reg by reducing to Artinian quotient\], to show that $${\operatorname{reg}}K[E_G]/I(X) \leq {\operatorname{reg}}K[E_{G'}]/I(X') + (q-2)$$ it suffices to show that for any monomial $t^a\in K[E_G]$ of degree ${\operatorname{reg}}K[E_{G'}]/I(X') + (q-2)+1$ we have $t^a \in (I(X),t_j)$. Let $t^a$ be such a monomial. If $a_i\geq q-1$ then writing $t^a = t^{a'}t_i^{q-1}$ for some $a'\in {\mathbb{N}}^s$, we get: $$t^a = t^{a'}(t_i^{q-1}-t_j^{q-1}) + t^{a'}t_j^{q-1} \in (I(X),t_j).$$ Assume now that $a_i<q-1$. Consider $a'\in {\mathbb{N}}^s$, with $a'_i=0$, such that $t^a = t^{a'}t_i^{a_i}$. Then $\deg t^{a'}=\deg t^a - a_i \geq {\operatorname{reg}}K[E_{G'}]/I(X')+1$, by our assumptions. As $t^{a'}$ belongs to $K[E_{G'}]$, using Proposition \[prop: computing reg by reducing to Artinian quotient\] we get $t^{a'} \in (I(X'),t_j)\subset K[E_{G'}]$. As $G'$ is a subgraph of $G$ we have $I(X')\subset I(X)$ and therefore $t^{a'} \in (I(X),t_j)$.
Using the same idea, let us now show that $${\operatorname{reg}}K[E_{G'}]/I(X') \leq {\operatorname{reg}}K[E_{G}]/I(X) - (q-2).$$ Let $t^a\in K[E_{G'}]$ be a monomial of degree ${\operatorname{reg}}K[E_{G}]/I(X) - (q-2)+1$. Then $t^at_i^{q-2}$ belongs to $K[E_G]$ and has degree ${\operatorname{reg}}K[E_{G}]/I(X)+1$. We deduce that $t^a t_i^{q-2} \in (I(X),t_j)$. By Proposition \[prop: membership of a zero dimensional reduction\], there exists a monomial $t^b\in K[E_G]$ such that $t^at_i^{q-2} - t_jt^b \in I(X)$. However the congruence at vertex $v$ gives $b_i= q-2 + k(q-1)$, for some $k\geq 0$. Let $b'\in {\mathbb{N}}^s$ be such that $b'_i=0$ and $t^b = t^{b'}t_i^{b_i}$. Then: $$t^at_i^{q-2} - t_jt^b \in I(X) \implies t^a-t_jt_i^{k(q-1)}t^{b'}\in I(X)\implies t^a - t_j^{1+k(q-1)}t^{b'} \in I(X).$$ Since $t^a - t_j^{1+k(q-1)}t^{b'} \in K[E_{G'}]$ and $I(X')=I(X)\cap K[E_{G'}]$, we deduce that $t^a \in (I(X'),t_j)$.
Let $G$ be a connected graph and a spanning subgraph of a bipartite graph $H$. Let $Y$ be the projective toric subset parameterized by $H$. Then, by [@VPVi13 Lemma 2.13], if $|X|=|Y|$, it follows that $${\operatorname{reg}}K[E_G]/I(X)\geq {\operatorname{reg}}K[E_H]/I(Y).$$ Hence if $G$ is a connected bipartite spanning subgraph of $\mathcal{K}_{a,b}$, by (\[eq: degree\]) the assumption on the cardinality of the associated parameterized projective toric subsets holds and we obtain: $$\label{eq: lower bound on reg for bipartite G}
{\operatorname{reg}}K[E_G]/I(X)\geq (\max{\left \{ a,b \right \}}-1)(q-2).$$
In the remainder of this section we introduce three new inequalities involving ${\operatorname{reg}}K[E_G]/I(X)$. They will play an important role in the proofs of Theorem \[th: main theorem\] and Theorem \[th: main theorem 2\].
\[prop: general lower bound in contraction\] Let $v_1$ and $v_2$ be two vertices of $G$ such that ${\left \{ v_1,v_2 \right \}}$ is a non-edge of $G$. Let $G'$ be the graph obtained by identifying $v_1$ with $v_2$ and denote by $X'$ the projective toric subset parameterized by it. Then ${\operatorname{reg}}K[E_G]/I(X)\geq {\operatorname{reg}}K[E_{G'}]/I(X')$.
The edge sets of $G$ and $G'$ have the same cardinality. Morevoer, there is an induced identification of the edges of $G'$ with the variables of the polynomial ring $K[t_1,\dots,t_s]$ under which $K[E_G]=K[E_{G'}]$.
(-0.5,-.5) node [$G$]{};
(0,.5) circle (1pt); (0,-.5) circle (1pt); (-.7,0) circle (1pt); (-.7,0) – (-.75,.15); (-.7,0) – (-.8,-.1);
(.5,.4) circle (1pt); (.5,.4) – (.55,.55); (.5,.4) – (.65,.45);
(.5,-.4) circle (1pt); (.5,-.4) – (.55,-.55); (.5,-.4) – (.65,-.45);
(-.3,.75) circle (1pt); (-.3,.75) – (-.45,.8); (-.3,.75) – (-.2,.85);
(-.15,-.7) circle (1pt); (-.15,-.7) – (-.25,-.75); (-.15,-.7) – (-.1,-.8);
(-.7,0).. controls (-.65,.15) and (-.05,.55) .. (0,.5); (-.7,0).. controls (-.65,-.15) and (-.05,-.55) .. (0,-.5); (0,.5).. controls (.1,.55) and (.45,.45) .. (.5,.4); (0,-.5).. controls (.1,-.55) and (.45,-.45) .. (.5,-.4); (.5,.4).. controls (.6,.4) and (.6,-.4) .. (.5,-.4);
(-.3,.75)–(0,.5); (-.15,-.7)–(0,-.5);
(0,.35) node [$v_1$]{}; (0,-.35) node [$v_2$]{};
(0,.25)–(0,-.25);
(.9,0).. controls (1,.07) and (1,-.07) ..(1.1,0); (1.1,0).. controls (1.2,.07) and (1.2,-.05) ..(1.3,0);
(2,-.5) node [$G'$]{};
(2.5,0) circle (1pt);
(1.8,0) circle (1pt); (1.8,0) – (1.75,.15); (1.8,0) – (1.7,-.1);
(3,.4) circle (1pt); (3,.4) – (3.05,.55); (3,.4) – (3.15,.45);
(3,-.4) circle (1pt); (3,-.4) – (3.05,-.55); (3,-.4) – (3.15,-.45);
(2.27,.3) circle (1pt); (2.27,.3) – (2.11,.36); (2.27,.3) – (2.38,.4);
(2.37,-.22) circle (1pt); (2.37,-.22) – (2.25,-.24); (2.37,-.22) – (2.44,-.33);
(1.8,0).. controls (1.85,.05) and (2.45,.05) .. (2.5,0); (1.8,0).. controls (1.85,-.05) and (2.45,-.05) .. (2.5,0);
(2.5,0).. controls (2.6,.15) and (2.95,.45) .. (3,.4); (2.5,0).. controls (2.6,-.15) and (2.95,-.45) .. (3,-.4); (3,.4).. controls (3.1,.4) and (3.1,-.4) .. (3,-.4);
(2.27,.3)–(2.5,0); (2.37,-.22)–(2.5,0);
(2.3,.75)..controls (2.75,.7) and (2.5,.3) .. (2.5,.1);
(2,.75) node [$v_1=v_2$]{};
Since the parameterization of $X'$ is obtained by adding the restriction that the coefficient of $v_1$ in the formal sum $\sum_{v\in V_G} x_v v$ be equal to the coefficient of $v_2$ we obtain $X'\subset X$ (cf. (\[eq: definition of X\])), and thus, $I(X)\subset I(X')$. Let $t_1$ be an edge. According to Proposition \[prop: computing reg by reducing to Artinian quotient\], to show that $${\operatorname{reg}}K[E_{G'}]/I(X') \leq {\operatorname{reg}}K[E_{G}]/I(X)$$ it suffices to prove that for any monomial $t^a$ of degree ${\operatorname{reg}}K[E_G]/I(X) +1$ we have $t^a \in (I(X'),t_1)$. Let $t^a$ be such a monomial. Then, using again Proposition \[prop: computing reg by reducing to Artinian quotient\], we deduce that $(t^a\in I(X),t_1)$. Since $I(X)\subset I(X')$ we get $(t^a\in I(X'),t_1)$.
\[prop: regularity inequality over subgraphs\] Let $H_1,H_2\subset G$ be subgraphs such that $E_G=E_{H_1}\cup E_{H_2}$ and $E_{H_1}\cap E_{H_2}\not = \emptyset$. Let $X_1$ and $X_2$ be the projective toric subsets parameterized by $H_1$ and $H_2$ and $I(X_1)\subset K[E_{H_1}]$, $I(X_2)\subset K[E_{H_2}]$ their corresponding vanishing ideals. Then $${\operatorname{reg}}K[E_G]/I(X)\leq {\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2).$$
Let $t_i\in E_{H_1}\cap E_{H_2}$. According to Proposition \[prop: computing reg by reducing to Artinian quotient\], it suffices to show that any monomial $t^a\in K[E_G]$, of degree ${\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2) + 1$, belongs to $(I(X),t_i)$. Let us write $t^a=t^bt^c$ for some $t^b\in K[E_{H_1}]$ and $t^c\in K[E_{H_2}]$. Since , we have or By Proposition \[prop: computing reg by reducing to Artinian quotient\] it follows that $t^b \in (I(X_1),t_i)\subset K[E_{H_1}]$ or $t^c \in (I(X_2),t_i)\subset K[E_{H_2}]$, respectively. In both cases we conclude that $t^a\in (I(X),t_i)$.
\[prop: lower bound using independent set\] Let ${\left \{ v_1,\dots,v_r \right \}}$ be an independent set of vertices of $G$. Assume that there is an edge in $G-{\left \{ v_1,\dots,v_r \right \}}$. Then ${\operatorname{reg}}K[E_G]/I(X) \geq r(q-2)$.
By Proposition \[prop: computing reg by reducing to Artinian quotient\], to show that ${\operatorname{reg}}K[E_G]/I(X) \geq r(q-2)$ it suffices to show that there exists and edge $t_i$ and a monomial $t^a\in K[E_G]$ of degree $r(q-2)$ that does not belong to $(I(X),t_i)$. Let $t_i$ be an edge of $G-{\left \{ v_1,\dots,v_r \right \}}$ and, for every $i=1,\dots,r$, let $t_{j_i}$ be an edge incident to $v_i$. Such edges exist since we assume that $G$ has no isolated vertices. Notice also that since ${\left \{ v_1,\dots,v_r \right \}}$ is an independent set the edges $t_{j_1},\dots,t_{j_r}$ are distinct. Consider the monomial: $$t^a = (t_{j_1}\cdots t_{j_r})^{q-2}$$ and let us show that $t^a\not \in (I(X),t_i)$. Suppose the contrary holds. Then, by Proposition \[prop: membership of a zero dimensional reduction\], there exists a monomial $t^b$ such that $t^a-t_it^b$ is homogeneous and belongs to $I(X)$. Since $t_i$ is not incident to any of the vertices of ${\left \{ v_1,\dots,v_r \right \}}$, evaluating the congruence at a particular vertex of this set, we conclude that the degree of $t^b$ in the edges incident to it is $\geq q-2$. Since, by assumption, these vertices possess no common incident edges we deduce that the degree of $t^b$ in edges incident to the vertices of ${\left \{ v_1,\dots,v_r \right \}}$ is $\geq r(q-2)$. In particular, $\deg(t^b)\geq r(q-2)$. But this implies that $t^a-t_it^b$ is not homogeneous, which is a contradiction.
We note that Proposition \[prop: lower bound using independent set\] implies (\[eq: lower bound on reg for bipartite G\]).
Proof of the main results {#sec: main thm}
=========================
The aim of this section is to prove Theorem \[th: main theorem\] and Theorem \[th: main theorem 2\]. In what follows $G$ is the parallel composition of $r\geq 2$ paths $P_1,\dots,P_r$ of lengths $k_1,\dots,k_r$. In a first instance, we assume that these integers have the same parity, so that $G$ is bipartite. If $r=2$ and one of $k_1,k_2$ is $>1$, then $G$ is an even cycle of length $k_1+k_2$. In this case, by [@NeVPVi15 Theorem 6.2], we know that ${\operatorname{reg}}K[E_G]/I(X) = ((k_1+k_2)/2-1)(q-2)$. If $r=2$ and $k_1=k_2=1$, then $G$ is a graph on $2$ vertices with $2$ multiple edges. Hence the value of the regularity is the same as in the case of a tree with a single edge, which is $(s-1)(q-2)=0$ (cf. Table \[table: values of reg\]). Both cases agree with the formula in Theorem \[th: main theorem\].
\[lemma: easy inequality\] $${\operatorname{reg}}K[E_G]/I(X) \geq
\begin{cases}
(\lfloor k_1/2\rfloor +\cdots +\lfloor k_r/2 \rfloor)(q-2),\; \text{if $k_i$ are odd,}\\
( k_1/2 +\cdots +k_r/2 -1)(q-2),\; \text{if $k_i$ are even.}
\end{cases}$$
If $k_i$ are odd, then $G$ is a connected spanning subgraph of $\mathcal K_{\rho,\rho}$, where $\rho$ is the integer If $k_i$ are even, then $G$ is a connected spanning subgraph of $\mathcal{K}_{(\rho-r+2),\rho}$ where $\rho$ is the integer $k_1/2+\cdots +k_r/2$. Hence the claim follows from (\[eq: lower bound on reg for bipartite G\]).
In the next two results we prove the opposite inequalities in each case. We need to fix some notation. For each $i\in {\left \{ 1,\dots,r \right \}}$, let $\sigma_i = k_1+\cdots + k_{i-1}$, so that, in particular, $\sigma_1=0$.
(0,-.6) circle (1pt); (3.45,-.6) circle (1pt);
(0,-.6)..controls (0.15,0)..(.75,0); in [0.75,1.5]{} [(,0) circle (1pt) – (+.75,0) ]{}; (2.48,-.006) node[$\cdots$]{}; (2.7,0)..controls (3.25,0)..(3.45,-.6); (.45,.12) node [$t_{\sigma_1+1}$]{}; (1.12,.12) node [$t_{\sigma_1+2}$]{}; (1.85,.12) node [$t_{\sigma_1+3}$]{}; (2.92,.12) node [$t_{\sigma_1+k_1}$]{};
(0,-.6)..controls (0.15,-.45)..(.75,-.45); in [0.75,1.5]{} [(,-.45) circle (1pt) – (+.75,-.45) ]{}; (2.48,-.456) node[$\cdots$]{}; (2.7,-.45)..controls (3.25,-.45)..(3.45,-.6); (.45,.-.33) node [$t_{\sigma_2+1}$]{}; (1.12,-.33) node [$t_{\sigma_2+2}$]{}; (1.85,-.33) node [$t_{\sigma_2+3}$]{}; (2.92,-.33) node [$t_{\sigma_2+k_2}$]{};
(0,-.6)..controls (0.15,-1.2)..(.75,-1.2); in [0.75,1.5]{} [(,-1.2) circle (1pt) – (+.75,-1.2) ]{}; (2.48,-1.206) node[$\cdots$]{}; (2.7,-1.2)..controls (3.25,-1.2)..(3.45,-.6); (3,-.7) node[$\vdots$]{}; (.45,.-1.05) node [$t_{\sigma_r+1}$]{}; (1.12,-1.05) node [$t_{\sigma_r+2}$]{}; (1.85,-1.05) node [$t_{\sigma_r+3}$]{}; (2.92,-1.05) node [$t_{\sigma_r+k_r}$]{};
Let us label the edges of $G$ as in Figure \[fig: labelling of edges\]. For each $i\in {\left \{ 1,\dots,r \right \}}$, let $f_i,g_i\in K[E_G]$ be: $$\label{eq: def of f_i and g_i}
f_i=t_{\sigma_i+1}\cdot t_{\sigma_i+3} \cdots t_{\sigma_i+2\lceil k_i/2 \rceil-1}\quad \text{and} \quad
g_i=t_{\sigma_i+2}\cdot t_{\sigma_i+4} \cdots t_{\sigma_i+2\lfloor k_i/2 \rfloor}.$$ (In other words, $f_i$ is the product of every other edge in $P_i$ starting with $t_{\sigma_i+1}$ and $g_i$ is the product of every other edge in $P_i$ starting with $t_{\sigma_i+2}$.) We notice that, for all $i\not = j$, $$\label{eq: relation between fi and gj}
f_ig_j - f_jg_i \in I(X).$$
\[prop: opp inequality in the case k\_i even\] If $k_i$ are even, then ${\operatorname{reg}}K[E_G]/I(X)\leq (k_1/2+\cdots +k_r/2 -1)(q-2)$.
According to Proposition \[prop: computing reg by reducing to Artinian quotient\], it suffices to show that any monomial $t^a\in K[E_G]$ of degree belongs to $(I(X),t_1)$. We may assume $t_1$ does not divide $t^a$. We will argue by induction on $r$. For $r=2$, as observed earlier, the result holds true. Assume now that $r\geq 3$. Let $H$ be the subgraph of $G$ given by ${\left \{ t_1 \right \}}\cup P_2\cup \cdots \cup P_r$ and $Y$ be the projective toric subset parameterized by $G$. By induction and [@NeVPVi14 Theorem 7.4], $${\operatorname{reg}}K[E_H]/I(Y) = (k_2/2+\cdots + k_r/2)(q-2).$$ Set $t^a = t^bt^c$, with $t^b\in K[E_{P_1}]$ and $t^c\in K[E_{H}]$. If , then, by Proposition \[prop: computing reg by reducing to Artinian quotient\], $t^c\in (I(Y),t_1)\subset (I(X),t_1)$, which implies that $t^a \in (I(X),t_1)$. Assume that . Then $\deg(t^b)\geq (k_1/2-1)(q-2)+1$. Consider now the subgraphs of $G$ given by $$H_1=P_1\cup P_2 \quad \text{and}\quad H_2=P_1\cup P_3\cup \cdots \cup P_r$$ and denote by $X_1$ and $X_2$, respectively, the projective toric subsets parameterized by them. Set $t^c=t^dt^e$ with $t^bt^d\in K[E_{H_1}]$ and $t^bt^e\in K[E_{H_2}]$. By the induction hypothesis, $$\renewcommand{{1.5}}{1.3}
\begin{array}{c}
{\operatorname{reg}}K[E_{H_1}]/I(X_1) = (k_1/2+k_2/2-1)(q-2)\quad \text{and} \\
{\operatorname{reg}}K[E_{H_2}]/I(X_2) = (k_1/2+k_3/2+\cdots + k_r/2-1)(q-2).
\end{array}$$ Hence, if $\deg(t^bt^d)\geq (k_1/2+k_2/2-1)(q-2) +1$, we get $t^bt^d \in (I(X_1),t_1)\subset (I(X),t_1)$ which implies that $t^a\in (I(X),t_1)$. Similarly, if $\deg(t^bt^e)\geq (k_1/2+k_3/2+\cdots + k_r/2-1)(q-2) +1$. Suppose that $$\renewcommand{{1.5}}{1.3}
\begin{array}{c}
\deg(t^bt^d)\leq (k_1/2+k_2/2-1)(q-2)\quad \text{and} \\
\deg(t^bt^e)\leq (k_1/2+k_3/2+\cdots + k_r/2-1)(q-2).
\end{array}$$ Since $\deg(t^a)=\deg(t^bt^d)+\deg(t^bt^e)-\deg(t^b)$, we deduce that $$\renewcommand{{1.5}}{1.3}
\begin{array}{c}
\deg(t^a) \leq (k_1/2+\cdots+k_r/2-1)(q-2)-1,
\end{array}$$ which is a contradiction.
\[thm: opp inequality in the case k\_i odd\] If $k_i$ are odd, then ${\operatorname{reg}}K[E_G]/I(X) \leq (\lfloor k_1/2\rfloor +\cdots+\lfloor k_r/2\rfloor)(q-2)$.
We will use induction on $k_1+\dots+k_r$. In the base case, $r=2$ and $k_1=k_2=1$. As we mentioned earlier, ${\operatorname{reg}}K[E_G]/I(X) = 0$.
Assume that $k_1+\cdots+k_r>3$ and, as induction hypothesis, that the statement of the theorem holds for any $k'_1,\dots,k'_{r'}$ and $r'\geq 2$ such that $k'_1+\cdots +k'_{r'}<k_1+\cdots+k_r$. If $r=2$, then, as observed in the beginning of this section, $G$ is an even cycle of length $k_1+k_2$ and accordingly $${\operatorname{reg}}K[E_G]/I(X) = ((k_1+k_2)/2-1)(q-2)=(\lfloor k_1/2 \rfloor + \lfloor k_2/2 \rfloor)(q-2).$$ Hence, we may assume $r\geq 3$. If, for some $i$, $k_i=1$, let $G'$ be the subgraph of $G$ given as the parallel composition of all $P_1,\dots,P_r$ but $P_i$. We note that $G'$ is a connected spanning subgraph of $G$ and hence, if $X'$ is the projective toric subset parameterized by $G'$, by the induction hypothesis, since $\lfloor k_i/2 \rfloor=0$, we get $${\operatorname{reg}}K[E_G]/I(X) \leq {\operatorname{reg}}K[E_{G'}]/I(X') \leq (\lfloor k_1/2\rfloor +\cdots+\lfloor k_r/2\rfloor)(q-2).$$ Thus, we may assume $k_i\geq 3$, for all According to Proposition \[prop: computing reg by reducing to Artinian quotient\], to show that ${\operatorname{reg}}K[E_G]/I(X)\leq (\lfloor k_1/2\rfloor +\cdots+\lfloor k_r/2\rfloor)(q-2)$, it suffices to show that any monomial $t^a\in K[E_G]$ of degree $$\label{eq: degree of t^a}
(\lfloor k_1/2\rfloor +\cdots+\lfloor k_r/2\rfloor)(q-2)+\lfloor k_1/2 \rfloor
+\cdots + \lfloor k_r/2 \rfloor$$ belongs to the ideal $(I(X),g)\subset K[E_G]$, where $g=g_1\cdots g_r$ and $g_i$ were defined in (\[eq: def of f\_i and g\_i\]). Let $t^a$ be one such monomial and write it as the product of monomials, $h_1\cdots h_r$, where $h_i\in K[E_{P_i}]$. By (\[eq: degree of t\^a\]), we have $\deg(h_i)\leq \lfloor k_i/2 \rfloor (q-1)$, for some $i\in{\left \{ 1,\dots,r \right \}}$. Without loss of generality we assume $i=1$. In particular, $$\label{eq: inequality4}
\deg(h_2\cdots h_r)\geq (\lfloor k_2/2\rfloor+\cdots +\lfloor k_r/2\rfloor)(q-2) + \lfloor k_2/2 \rfloor +\cdots + \lfloor k_r/2 \rfloor.$$ Since $g$ is invariant under the swapping of variables corresponding to edges of $P_1$ in a same parity position, using Lemma \[lemma: swaping exponents\], we may assume that $$\label{eq: inequality5}
a_1\leq a_3 \leq \cdots \leq a_{2\lceil k_1/2\rceil -1}\text{ and }
a_2\leq a_4 \leq \cdots \leq a_{2\lfloor k_1/2\rfloor}.$$ Let $H$ be the subgraph of $G$ given by $P_2\cup\cdots \cup P_r$ and denote by $Y$ the projective toric subset parameterized by it. By induction, ${\operatorname{reg}}K[E_H]/I(Y)=(\lfloor k_2/2 \rfloor +\cdots+\lfloor k_r/2 \rfloor)(q-2)$. Then, by (\[eq: inequality4\]), Proposition \[prop: computing reg by reducing to Artinian quotient\] and Proposition \[prop: membership of a zero dimensional reduction\], there exists a monomial $t^b\in K[E_H]$, for some $b\in {\mathbb{N}}^s$ supported on the edges of $H$, such that $h_2\cdots h_r-g_2\cdots g_rt^b \in I(Y)\subset I(X)$ and hence $$\label{eq: inequality6}
t^a-h_1g_2\cdots g_rt^b \in I(X).$$ If $a_2\not = 0$, then from (\[eq: inequality5\]) we deduce that $g_1$ divides $h_1$ and we are done. If $a_1\not = 0$, then there exists $c\in {\mathbb{N}}^s$ such that $h_1=f_1t^c$. Accordingly, $h_1g_2\cdots g_rt^b = f_1g_2\cdots g_r t^{b+c}$. Since $f_1g_2-f_2g_1 \in I(X)$, we deduce that $f_1g_2\cdots g_rt^{b+c} - f_2g_1g_3\cdots g_r t^{b+c} \in I(X)$, which, together with (\[eq: inequality6\]), implies that $$\label{eq: membership1}
t^a - f_2g_1g_3\cdots g_r t^{b+c} \in I(X).$$ Consider $a'\in {\mathbb{N}}^s$ such that $t^{a'}=f_2g_1g_3\cdots g_rt^{b+c}$. Since $h_1=f_1t^c$ and the monomials $f_2, g_3,\dots, g_r, t^b$ are supported away from the edges of $P_1$, we see that, if $1\leq i\leq k_1$, $a'_i = a_i-1$, when $i$ is odd, and $a'_i=a_i+1$, when $i$ is even. In particular, $a'_2\not =0$ and, in the corresponding decomposition $t^{a'}=h'_1\cdots h'_r$ with monomials $h'_i \in K[E_{P_i}]$, we get $\deg(h'_1) = \deg(h_1)-1$. Repeating the previous argument, we deduce that $t^{a'}\in (I(X),g)$, which, using (\[eq: membership1\]) implies that $t^{a}\in (I(X),g)$.
We are left with the case of $a_1=a_2=0$. We regard $t^a$ as a monomial in $K[E_{G'}]$, where $G'$ is the graph obtained as the parallel composition of $P_1\setminus {\left \{ t_1,t_2 \right \}}$, $P_2,\dots, P_r$.
(0,-.6)..controls (0.05,-.1) and (0.15,0)..(0.7,0); (0.804,0)–(1.1,0); (0,-.6).. controls (0.15,-.5)..(.75,-.5); (0.95,-.505) node[$\cdots$]{};
(1.45,0)..controls (1.75,0)..(1.75,-.6); (1.3,-.006) node[$\cdots$]{};
(0,-.6)..controls (0.15,-1.2)..(.75,-1.2); (0.95,-1.206) node[$\cdots$]{};
(0,-.6) circle (3.5pt); (0,-.6) node [$v_1$]{}; (-.045,-.28) node [$t_{1}$]{};
(0.17,-.12) circle (3.5pt); (0.17,-.12) node [$v_2$]{}; (0.4,.1) node [$t_{2}$]{};
(0.7,0) circle (3.5pt); (0.7,0) node [$v_3$]{}; (0.95,.12) node [$t_{3}$]{};
(1.15,-.5)..controls (1.7,-.5)..(1.75,-.6); (.48,.-.4) node [$t_{\sigma_2+1}$]{};
(1.15,-1.2)..controls (1.75,-1.2)..(1.75,-.6); (.48,.-1.05) node [$t_{\sigma_r+1}$]{}; (0.5,-0.7) node [$\vdots$]{};
(1.75,-.6) circle (3.5pt);
(3,-.6)..controls (3.05,-.1) and (3.15,0)..(3.7,0); (3,-.6)..controls (3.15,-.5)..(3.75,-.5); (3.95,-.505) node[$\cdots$]{};
(4.15,0)..controls (4.5,0)..(4.5,-.6); (3.95,-.006) node[$\cdots$]{};
(3,-.6)..controls (3.15,-1.2)..(3.75,-1.2); (3.95,-1.206) node[$\cdots$]{};
(3,-.6) circle (3.5pt); (3,-.6) node [$v'_1$]{}; (3.02,-.07) node [$t_{3}$]{};
(4.15,-.5)..controls (4.4,-.5)..(4.5,-.6); (3.48,.-.4) node [$t_{\sigma_2+1}$]{};
(4.15,-1.2)..controls (4.5,-1.2)..(4.5,-.6); (3.48,.-1.05) node [$t_{\sigma_r+1}$]{}; (3.5,-0.7) node [$\vdots$]{};
(4.5,-.6) circle (3.5pt);
Let $X'$ be the projective toric subset parameterized by $G'$. By the induction hypothesis $${\operatorname{reg}}K[E_{G'}]/I(X') = (\lfloor k_1/2 \rfloor+\cdots+\lfloor k_r/2 \rfloor -1)(q-2).$$ Hence, by Proposition \[prop: computing reg by reducing to Artinian quotient\] and Proposition \[prop: membership of a zero dimensional reduction\], there exists $t^d\in K[E_{G'}]$, where $d\in {\mathbb{N}}^s$ is supported on the edges of $G'$, such that $$\label{eq: binomial in G'}
t^a-g'_1g_2\cdots g_rt_3^{q-1}t^d \in I(X'),$$ where $g'_1=g_1/t_2\in K[E_{G'}]$. We claim there exists $k\in {\left \{ 1,2,\dots,q-1 \right \}}$ such that $$\label{eq: binomial in G}
t^a -t_1^{\widehat{k}}t_2^k g'_1g_2\cdots g_rt^d \in I(X)$$ with $\widehat{k}=q-1-k$. We define $k$ using the congruence at vertex $v'_1$ of $G'$ (see Figure \[fig: graphs G and G’\]) which, according to [@NeVP14 Lemma 2.3], is satisfied for the binomial in (\[eq: binomial in G’\]). This congruence is: $$\nonumber
\renewcommand{{1.5}}{1.3}
\begin{array}{c}
a_3+a_{\sigma_2+1}+\cdots + a_{\sigma_r+1}\equiv q-1 + d_3 + d_{\sigma_2+1} +\cdots + d_{\sigma_r+1} \pmod{q-1} \\
\iff a_3-d_3 \equiv d_{\sigma_2+1}+\cdots +d_{\sigma_r+1} - (a_{\sigma_2+1} +\cdots + a_{\sigma_r+1} ) \pmod{q-1}.
\end{array}$$ Let $k\in {\left \{ 1,2,\dots,q-1 \right \}}$ to be such that: $$\label{eq: def of k}
k \equiv a_3-d_3 \equiv d_{\sigma_2+1}+\cdots +d_{\sigma_r+1} - (a_{\sigma_2+1} +\cdots + a_{\sigma_r+1} ) \pmod{q-1}.$$ Let us now show that (\[eq: binomial in G\]) holds. Since $t^a -t_1^{\widehat{k}}t_2^k g'_1g_2\cdots g_rt^d$ is homogeneous, it will suffice to check the congruences at each vertex of $G$. Since for the binomial in (\[eq: binomial in G’\]), from which we obtain this binomial, the congruences are satisfied at all vertices of $G'$, it will be enough to check the congruences for the vertices $v_1,v_2$ and $v_3$. At $v_1$, we have: $$a_{\sigma_2+1}+\cdots + a_{\sigma_r+1}\equiv (q-1)-k + d_{\sigma_2+1}+\cdots + d_{\sigma_r+1} \pmod{q-1},$$ at $v_2$, $0\equiv (q-1)-k+k \pmod{q-1}$ and at $v_3$, $a_3 \equiv k + d_{3} \pmod{q-1}$, all of which hold, by virtue of (\[eq: def of k\]). This completes the proof of the theorem.
The proof of Theorem \[th: main theorem\] follows from Lemma \[lemma: easy inequality\], Proposition \[prop: opp inequality in the case k\_i even\] and Theorem \[thm: opp inequality in the case k\_i odd\].
We now turn to the proof of Theorem \[th: main theorem 2\]. In this case, $G$ is the parallel composition of paths $P_1,\dots,P_r$ the lengths of which have mixed parity. We assume, without loss of generality, that $P_1,\dots,P_l$ have odd lengths and $P_{l+1},\dots,P_r$ have even lengths, for some $1\leq l <r$. We will keep the notation for the edges of $G$ as in the beginning of this section and recall that (as in the statement of Theorem \[th: main theorem 2\]) we will be denoting by $H_1$ the parallel composition of the paths of odd lengths, by $H_2$ the parallel composition of the paths of even lengths and by $X_1,X_2$, respectively, the projective toric subsets they parameterize.
\[lemma: easy inequality in the non-bipartite case\] ${\operatorname{reg}}K[E_G]/I(X) \leq {\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2)$.
Consider the cover of $G$ given by $H_1$ and $H'_2$ where $H'_2$ is given by ${\left \{ t_1 \right \}}\cup H_2$. Then and therefore by Proposition \[prop: regularity inequality over subgraphs\], $$\label{eq: proof of easy inequality in the non-bipartite case}
{\operatorname{reg}}K[E_G]/I(X) \leq {\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H'_2}]/I(X'_2),$$ where $X'_2$ is the projective toric subset parameterized by $H'_2$. By Proposition \[prop: aditivity on leaves\], we know that ${\operatorname{reg}}K[E_{H'_2}]/I(X'_2)= {\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2)$. Combining this with (\[eq: proof of easy inequality in the non-bipartite case\]) completes the proof of the lemma.
\[thm: harder inequality in the non-bipartite case\] ${\operatorname{reg}}K[E_G]/I(X) \geq {\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2)$.
We divide the proof into cases.
*The case $l=1$ and $r=2$.* In this case $G$ is a cycle of (odd) length $k_1+k_2$. Accordingly, $X$ coincides with ${\mathbb{T}}^{k_1+k_2-1}$ and, by the formula in Table \[table: values of reg\], On the other hand $H_1$ and $H_2$ are paths of lengths $k_1$ and $k_2$ and the projective toric subsets they parameterized are the tori ${\mathbb{T}}^{k_1-1}$ and ${\mathbb{T}}^{k_2-1}$ so that, again by the same formula, ${\operatorname{reg}}K[E_{H_1}]/I(X_1)=(k_1-1)(q-2)$ and ${\operatorname{reg}}K[E_{H_2}]/I(X_2) = (k_2-1)(q-2)$. We deduce that $${\operatorname{reg}}K[E_G]/I(X) = {\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2).$$
In the other cases, we will use vertex identifications and Proposition \[prop: general lower bound in contraction\]. For this purpose, let us denote the terminal vertices of the parallel composition yielding $G$ by $v$ and $w$.
*The case $l=1$, $k_1=1$ and $r-l>1$.* Consider the vertices of $P_2,\dots,P_r$ at an odd number of edges away from $v$ (or $w$). They form an independent set of vertices of cardinality $k_2/2+\cdots +k_r/2$. Then, by Proposition \[prop: lower bound using independent set\], we get $$\label{eq: inequality when there is only one odd path}
{\operatorname{reg}}K[E_G]/I(X) \geq (k_2/2+\cdots+k_r/2)(q-2).$$ Now, by Theorem \[th: main theorem\], the right-hand of (\[eq: inequality when there is only one odd path\]) is equal to $$0+{\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2) = {\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2).$$
*The case $l=1$, $k_1>1$ and $r-l>1$.* Let $G'$ be the graph obtained by identifying all the vertices in the paths $P_2,\dots,P_r$ at an even number of edges away from $v$ (or $w$) with the vertex $v$. The resulting graph $G'$ consists of an odd cycle of length $k_1$ with a set of $k_2/2+\cdots +k_r/2$ double edges incident to one of its vertices (cf. Figure \[fig: First reduction\]).
(0,0) circle (1pt); (0,0).. controls (-.15,.3) and (-.35,.5) ..(-.5,.5); (-.5,.5) circle (1pt);
(0,0).. controls (-.15,-.3) and (-.35,-.5) ..(-.5,-.5); (-.5,-.5) circle (1pt);
(-.5,.5).. controls (-.6,.5) and (-.8,.35) ..(-.9,.15); (-.5,-.5).. controls (-.6,-.5) and (-.8,-.35) ..(-.9,-.15);
(-.9,.06) node [$\vdots$]{};
(0,0).. controls (.1,.15) and (.4,.45) ..(.5,.5); (0,0).. controls (.1,.05) and (.45,.4) ..(.5,.5); (.5,.5) circle (1pt);
(0,0).. controls (.1,.05) and (.6,.12) ..(.7,.1); (0,0).. controls (.1,0) and (.6,.04) ..(.7,.1); (.7,.1) circle (1pt);
(.65,-.10) node [$\cdot$]{}; (.62,-.18) node [$\cdot$]{}; (.59,-.26) node [$\cdot$]{};
(0,0).. controls (.1,-.15) and (.4,-.45) ..(.5,-.5); (0,0).. controls (.1,-.05) and (.45,-.4) ..(.5,-.5); (.5,-.5) circle (1pt);
(0,.5) – (0,.1); (0,.6) node [$v=w$]{};
(-1.3,.4).. controls (-.95,.4) and (-.65,.3).. (-.6,.2); (-1.7,.4) node ;
(.8,.4).. controls (1.05,.2) and (1.05,-.2).. (.8,-.4); (1.7,0) node [$k_2/2+\cdots + k_r/2$]{};
Let $X'$ the projective toric subset parameterized by $G'$. The regularity of $K[E_{G'}]/I(X')$ is the same as if in $G'$ all double edges were single edges. Hence by Proposition \[prop: aditivity on leaves\] and the formula for the odd cycle case we get: $$K[E_{G'}]/I(X') = (k_1-1)(q-2) + (k_2/2+\cdots + k_r/2)(q-2)$$ which coincides with ${\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2)$. Since, by Proposition \[prop: general lower bound in contraction\], ${\operatorname{reg}}K[E_G]/I(X)\geq {\operatorname{reg}}[E_{G'}]/I(X')$ we obtain the desired inequality.
*The case $l>1$ and $r-l=1$.* In this case we construct a graph $G'$ by identifying all vertices in $P_1,\dots,P_l$ at an even number of edges away from $v$ with the vertex $v$.
(0,0) circle (1pt); (0,0)..controls (.1,.15 ) and (.9,.15) .. (1,0); (.5,.05) node [$\vdots$]{}; (0,0)..controls (.1,-.15) and (.9,-.15) .. (1,0); (1,0) circle (1pt);
(1.1,.1) node [$w$]{}; (-.075,-.125) node [$v$]{};
(0,0)..controls (.1,-.6) and (.9,-.6) .. (1,0); (.5,-.45) circle (1pt); (.15,-.3) circle (4pt); (.1,-.25) node [$\cdot$]{}; (.16,-.3105) node [$\cdot$]{}; (.22,-.37) node [$\cdot$]{}; (.8,-.5) node [$P_r$]{};
(0,0)..controls (-.15,.04) and (-.55,.04) .. (-.7,0); (0,0)..controls (-.15,-.04) and (-.55,-.04) .. (-.7,0); (-.7,0) circle (1pt);
(0,0)..controls (-.15,.15) and (-.55,.33) .. (-.62,.35); (0,0)..controls (-.15,.04) and (-.55,.26) .. (-.62,.35); (-.62,.35) circle (1pt);
(0,0)..controls (-.04,.1) and (-.04,.6) .. (0,.7); (0,0)..controls (.04,.1) and (.04,.6) .. (0,.7); (0,.7) circle (1pt);
(-.43,.50) node [$\cdot$]{}; (-.33,.55) node [$\cdot$]{}; (-.23,.60) node [$\cdot$]{};
(.65,0).. controls (.5,.3) and (.65,.4).. (.85,.4); (1.45,.4) node ;
(.75,-.25).. controls (.8,-.4) and (1,-.4).. (1.35,-.3); (1.9,-.3) node ;
(-.8,.175).. controls (-.83,.5) and (-.5,.83).. (-0.175,.8); (-1.45,.7) node [$\lfloor k_1/2 \rfloor + \cdots + \lfloor k_{r-1}/2 \rfloor$]{};
This graph consists of an odd cycle of length $k_r+1$ (given by $P_r$ and (a choice of) an edge ${\left \{ v,w \right \}}$) that has $l$ multiple edges between $v$ and $w$ and of a set of $\lfloor k_1/2 \rfloor + \cdots \lfloor k_{r-1}/2 \rfloor$ double edges incident at $v$ (cf. Figure \[fig: Second reduction\]). Arguing as above, we get: $$\renewcommand{{1.5}}{1.3}
\begin{array}{c}
{\operatorname{reg}}K[E_G]/I(X) \geq {\operatorname{reg}}K[E_{G'}]/I(X') = (\lfloor k_1/2 \rfloor + \cdots \lfloor k_{r-1}/2 \rfloor)(q-2) + k_r(q-2)\\
= {\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2).
\end{array}$$
*The case $l>1$ and $r-l>1$.* As in the previous case, let $G'$ be the graph obtained by identifying the vertices in $P_1,\dots,P_l$ at an even number of edges away from $v$ with this vertex. We notice that the subgraph of $G'$ consisting of the paths $P_{l+1},\dots,P_r$ and (a choice of) an edge ${\left \{ v,w \right \}}$ belongs to the second case, above. Consequently, $$\renewcommand{{1.5}}{1.3}
\begin{array}{c}
{\operatorname{reg}}K[E_G]/I(X) \geq (\lfloor k_1/2 \rfloor + \cdots + \lfloor k_l/2 \rfloor)(q-2) + (k_{l+1}/2+\cdots+k_r/2)(q-2)\\
= {\operatorname{reg}}K[E_{H_1}]/I(X_1) + {\operatorname{reg}}K[E_{H_2}]/I(X_2) + (q-2). \qedhere
\end{array}$$
The proof of Theorem \[th: main theorem 2\] is obtained by combining Lemma \[lemma: easy inequality in the non-bipartite case\] and Theorem \[thm: harder inequality in the non-bipartite case\]. In Table \[table: values of reg for a parallel composition\] we give explicit formulas for the regularity of $K[E_G]/I(X)$ when $G$ is a parallel composition of $r\geq 2$ paths of lengths $k_1,\dots,k_r$, of which $k_1,\dots,k_l$ are odd and $k_{l+1},\dots,k_r$ are even.
${\operatorname{reg}}K[E_G]/I(X)$
-------------- --------------------------------------------------------------------------------------------
$l=0$ $(k_1/2+\cdots + k_r/2 - 1)(q-2)$
$l=r$ $(\lfloor k_1/2 \rfloor + \cdots + \lfloor k_r/2 \rfloor)(q-2)$
$l=1,r=2$ $(k_1+k_2-1)(q-2)$
$l=1, r>2$ $(k_1+k_2/2+\cdots +k_r/2 -1)(q-2)$
$l>1, r=l+1$ $(\lfloor k_1/2 \rfloor + \cdots + \lfloor k_{r-1}/2 \rfloor + k_r)(q-2)$
$l>1, r>l+1$ $(\lfloor k_1/2 \rfloor + \cdots + \lfloor k_l/2 \rfloor + k_{l+1}/2+\cdots + k_r/2)(q-2)$
: Values of ${\operatorname{reg}}K[E_G]/I(X)$ when $G$ is a parallel composition of paths.
\[table: values of reg for a parallel composition\]
[99]{}
J. Biermann, A. O’Keefe and A. Van Tuyl, *Bounds on the regularity of toric ideals of graphs*, Preprint arXiv:1605.06980
R. Diestel, *Graph theory*, Fourth edition. Graduate Texts in Mathematics, **173**. Springer, Heidelberg, 2010. xviii+437 pp. ISBN: 978-3-642-14278-9
V. Ene and A. Zarojanu, *On the regularity of binomial edge ideals*, Math. Nachr. **288** (2015), no. 1, 19–24.
M. González and C. Rentería *Evaluation codes associated to complete bipartite graphs*, Int. J. Algebra **2** (2008), no. 1–4, 163–170.
M. González, C. Rentería and M. Hernández de la Torre, *Minimum Distance and Second Generalized Hamming Weight of Two Particular Linear Codes*, Congr. Numer. **161** (2003), 105–116.
M. González, C. Rentería and E. Sarmiento, *Parameterized codes over some embedded sets and their applications to complete graphs*, Math. Commun. **18** (2013), no. 2, 377–391.
H. T. Há and A. Van Tuyl, *Monomial ideals, edge ideals of hypergraphs, and their graded Betti numbers*, J. Algebraic Combin. **27** (2008), no. 2, 215–24.
M. Katzman, *Characteristic-independence of Betti numbers of graph ideals*, J. Combin. Theory A **113** (2006), 435–454.
D. Kiani and S. Saeedi Madani, *The Castelnuovo–Mumford regularity of binomial edge ideals*, J. Comb. Theory A **139** (2016), 80–86.
K. Matsuda and S. Murai, *Regularity bounds for binomial edge ideals*, J. Commut. Algebra **5** (2013), no. 1, 141–149.
J. Neves and M. Vaz Pinto, *Vanishing ideals over complete multipartite graphs*, J. Pure Appl. Algebra, **218** (2014), 1084–1094.
J. Neves, M. Vaz Pinto and R. H. Villarreal, *Vanishing ideals over graphs and even cycles*, Comm. Algebra, Vol. [**43**]{}, Issue 3, (2015) 1050–1075.
J. Neves, M. Vaz Pinto and R. H. Villarreal, *Regularity and algebraic properties of certain lattice ideals*, Bull. Braz. Math. Soc., Vol. **45**, N. 4, (2014) 777–806.
C. Rentería, A. Simis and R. H. Villarreal, *Algebraic methods for parameterized codes and invariants of vanishing ideals over finite fields*, Finite Fields Appl., [**17**]{} (2011), no. 1, 81–104.
E. Sarmiento, M. Vaz Pinto and R. H. Villarreal, *The minimum distance of parameterized codes on projective tori*, Appl. Algebra Engrg. Comm. Comput. **22** (2011), no. 4, 249–264.
M. Vaz Pinto and R. H. Villarreal, *The degree and regularity of vanishing ideals of algebraic toric sets over finite fields*. Comm. Algebra, **41** (2013), no. 9, 3376–3396.
R. Woodroofe, *Matchings, coverings, and Castelnuovo–Mumford regularity*, J. Commut. Algebra **6** (2014), no. 2, 287–304.
[^1]: This work was partially supported by the Centre for Mathematics of the University of Coimbra – UID/MAT/00324/2013, funded by the Portuguese Government through FCT/MCTES and co-funded by the European Regional Development Fund through the Partnership Agreement PT2020. The third author was partially supported by the Center for Mathematical Analysis, Geometry and Dynamical Systems of Instituto Superior Técnico, Universidade de Lisboa, funded by FCT/Portugal through UID/MAT/04459/2013.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Nikolay P. Zotov\
SINP, Lomonosov Moscow State University\
E-mail:
title: Small $x$ physics and hard QCD processes at LHC
---
Itroduction
===========
The so-called small $x$ regime of QCD is the kinematic region, where the characteristic hard scale of the process $\mu^2 \,\,\sim p_T^2 \,\,\sim M_T^2 = M^2 + p_T^2$,( $M$ and $p_T^2$ are the mass and the transverse momentum of the final state) is large as compared to the $\Lambda_{QCD}$ but $\mu$ is much less than the total c.m.s. energy $\sqrt s$ of the process: $\Lambda_{QCD}\,\, \ll \,\, \mu \,\,\ll\,\,\sqrt s$.
In this sense, the HERA was the first small $x$ machine, and the LHC is more of a small $x$ collider. Typical $x$ values probed at the LHC in the central rapidity region are almost two orders of magnitude smaller than $x$ values probed at the HERA at the same scale. Hence, the small $x$ corrections start being relevant even for a final state with a characteristic electroweak scale $M \sim 100$ GeV.
It means the pQCD expansion any observable quantity in $\alpha_s$ contains large coefficients\
$(\ln^n(S/M^2)) \sim(\ln^n(1/x))$ (besides the usual renorm group ones $(\ln^n(\mu^2/\Lambda^2_{QCD})) $. The resummation of these terms $(\alpha_s(\ln(1/x))^n \,\, \sim 1$ at $x \to 0$) in the framework of the Balitsky-Fadin-Kuraev-Lipatov (BFKL) theory [@BFKL] results in the so called unintegrated gluon distribution ${\cal F}(x,{\mathbf k}_T^2)$. The unintegrated gluon distribution determines the probability to find a gluon carrying the longitudinal momentum fraction $x$ and transverse momentum $k_T$. This generalized factorization is called “$k_T-$factorization” [@GLR; @LRSS]. If the terms proportional to $\alpha_s^n ln^n (\mu^2 /\Lambda^2_{QCD})$ and $\alpha_s ln^n (\mu^2 /\Lambda^2_{QCD} ) ln^n (1/x)$ are also resummed, then the unintegrated gluon distribution function depends also on the probing scale $\mu$. This quantity depends on more degrees of freedom than the usual collinear parton density, and is therefore less constrained by the experimental data. Various approaches to model the unintegrated gluon distribution have been proposed. One such approach, valid for both small and large $x$, has been developed by Ciafaloni, Catani, Fiorani and Marchesini, and is known as the CCFM model [@CCFM]. It introduces angular ordering of emissions to correctly treat gluon coherence effects. In the limit of asymptotic energies, it is almost equivalent to BFKL [@BFKL], but also similar to the collinear (DGLAP) evolution for large $x$ and high $\mu^2$. The resulting unintegrated gluon distribution functions depend on two scales, the additional scale $\bar q$ being a variable related to the maximum angle allowed in the emission.
The BFKL evolution equation predicts rapid growth of gluon density ($\sim x^{-\Delta}$, where $1 + \Delta$ is the intercept of so-called hard BFKL Pomeron). However it is clear that this growth cannot continue for ever, because it would violate the unitarity constraint [@GLR]. Consequently, the parton evolution dynamics must change at some point, and new phenomenon must come into play. Indeed as the gluon density increases, non-linear parton interactions are expected to become more and more important, resulting eventually in the slowdown of the parton density growth (known as “saturation effect”) [@GLR; @MQ]. The underlying physics can be described by the non-linear Balitsky-Kovchegov (BK) equation [@BK].These nonlinear interactions lead to an equilibrium-like system of partons with some definite value of the average transverse momentum $k_T$ and the corresponding saturation scale $Q_s(x)$. This equilibrium-like system is the so called Color Glass Condesate (CGC) [@CGC]. Since the saturation scale increases with decreasing of $x$: $Q^2_s(x,A) \sim x^{-\lambda}A^{\delta}$ (A is an atomic number ) with $\lambda \sim 0.3, \delta \sim 1/3$ [@Ian], one may expect that the saturation effect will be more clear at LHC energies.
Unintegrated parton distributions (uPDF or TMD)
===============================================
The basic dynamical quantity in the small $x$ physics is transverse-momentum-dependent (TMD) (${\mathbf k}_T$-dependent) or unintegrated parton distribution (uPDF) ${\cal A}(x,{\mathbf k}_T^2,\mu^2)$. For example to calculate the cross sections of photoproduction process the uPDF ${\cal A}(x,{\mathbf k}_T^2,\mu^2)$ has to be convoluted with the relevant partonic cross section $\hat \sigma_{\gamma g}$: $$\begin{aligned}
\sigma = \int {dz\over z} \int d{\mathbf k}_{T}^2 \,
\hat\sigma_{\gamma g}(x/z,{\mathbf k}_{T}^2,\mu^2) {\cal A}(x,{\mathbf
k}_{T}^2,\mu^2). \end{aligned}$$ For the uPDF there is no unique definition, and as a cosequence it is for the phenomenology of these quantities very important to identify uPDF which are used in description of high energy processes. For a general introduction to small $x$ physics and the small $x$ evolution equations, as well as tools for calculation in terms of MC programs, we refer to the reviews [@Small-x].
During roughly the last decade, there has been steady progress toward a better understanding of the $k_T$-factorization (high energy factorization) and the uPDF (for example [@uPDF]. Workshop on Transvere Momentum Distributions (TMD 2010), which was held in Trento (Italy), was dedicated to the recent developments in small $x$ physics, based on the $k_T$-factorization and the uPDF [@TMD].
Recently the definition for the TMDs determined by the requirement of factorization, maximal universality and internal consistency have been done by Collins [@JC]. The results obtained in previous works are reduced to the following: $k_T$(TMD)-factorization is valid in
- Back-to-back hadron or jet production in $e^+e^-$-annihilation,
- Drell-Yan process ($ P_A + P_B \to (\gamma^*, W/Z) + X$),
- Semi-inclusive DIS ($e + P \to e + h + X$).
In hadroproduction of back-to-back jets or hadrons ( $h_1 + h_2 \to H_1 + H_2 + X$) TMD-factorization is problematic.\
For expample, partonic picture gives the following $q_T$-dependent hadronic tensor for DY cross section: $$\begin{aligned}
W^{\mu \nu} = \Sigma_f |H_f (Q; \mu_R)|^{\mu \nu}\\ \nonumber
\int {d^2{\mathbf k}_{1T} d^2{\mathbf k}_{2T} {\cal A}_{f/P_1}(x_1, k_{1T}; \mu_R; {\zeta}_1) {\overline{\cal A}}_{{\overline f}/P_2}(x_2,k_{2T};\mu_R;{\zeta}_2)\delta ({\mathbf k}_{1T} + {\mathbf k}_{2T} - {\mathbf q}_T)} + Y(Q, q_T). \end{aligned}$$
The hard part $H_f (Q; \mu_R)$ is calculable to arbitrary order in $\alpha_s$, $\mu_R$ - the renormalization scale. The term $Y (Q, q_T)$ describes the matching to large $q_T$, where the approximations of TMD-factorization break down. The scales $\zeta_1, \zeta_2$ are related to the regulation of light-cone divergences and $\zeta_1 \zeta_2=Q^4$. The soft factors connected with soft gluons are contained in the definitions of the TMDs, which cannot be predicted from the theory and must be fitted to data.
The $k_T$-factorization approach in hadroproduction
===================================================
The $k_T$-factorization approach in hadroproduction is based on the work by Catani, Ciafaloni and Hautman (CCH) (see [@LRSS]) The factorization formula for $pp$-collision in physical gauge ($nA = 0, n^{\mu}= aP_1^{\mu} +
bP_2^{\mu}$) is $$\begin{aligned}
\sigma = {1\over 4M^2} \int d^2{\mathbf k}_{1T} \int {dx_1\over x_1} \int d^2{\mathbf k}_{2T} \int {dx_2\over x_2} {\cal F}(x_1,{\mathbf k}_{1T})
\hat\sigma_{gg}(\rho /(x_1 x_2),{\mathbf k} _{1T},{\mathbf k}_{2T}) {\cal F}(x_2,{\mathbf k}_{2T}),\end{aligned}$$ where $\rho = 4M^2/s$, $M$ is the invariant mass of heavy quark, and $\cal F$ are the unintegrated gluon distributions, definded by the BFKL equation: $$\begin{aligned}
{\cal F}(x,{\mathbf k}; Q_0^2) = {1\over\pi}\delta (1 - x)\delta({\mathbf k}^2 - Q_0^2) +\\ \nonumber
+ \overline\alpha_s \int {d^2{\mathbf q}\over{\pi\mathbf q}^2} \int {dz\over z}
[{\cal F}(x/z, {\mathbf k} + {\mathbf q}; Q_0^2) -
{\Theta (k - q )}{\cal F}(x/z, {\mathbf k}; Q_0^2)],\end{aligned}$$ were $\overline\alpha_s = \alpha_s N_c/\pi$. It means that the rapidity divergencies are cut off since there are an implicit cuts in the BFKL formalism. Effectively one introduces a cuts $\zeta_1, \zeta_2$, and then sets $\zeta_1 = x_1, \zeta_2 = x_2$ in (3.1).\
The declaration in CCH [@LRSS] that $\cal F$ is defined via the BFKL equation (3.2) means that the BFKL unintegrated gluon distribution reduces to the dipole gluon distribution [@BGN]. The connections between different uPDF recently were analysed in [@EA].
The procedure for resumming inclusive hard cross-sections at the leading non-trivial order through $k_T$-factorization was used for an increasing number of processes: photoproduction ones, DIS ones, DY and vector boson production, direct photon production, gluonic Higgs production both in the point-like limit, and for finite top mass $m_t$. Please look, for example [@Marzani].
The hadroproduction of heavy quarks was considered in [@BE] and recently in [@Ball]. In last paper it was shown that when the coupling runs the dramatic enhancements seen at fixed coupling, due to infrared singularities in the partonic cross sections, are substantially reduced, to the extent that they are largely accounted for by the usual NLO and NNLO perturbative corrections. It was found that resummation modifies the $B$ production cross section. at the LHC by at most 15$\%$, but that the enhancement of gluonic $W$ production may be as large 50$\%$ at large rapidities.\
In our previous papers we have used the $k_T$-factorization approach to describe experimental data on:
- heavy quark photo- and electroproduction at HERA
- $J/\psi$ production in photo- and electroproduction at HERA with taking into account the color singlet and color octet states.
- $D^*$, $D^* + jet$, $D^* + 2jet$ photoproduction and $D^*$ production in DIS
- charm contribution to the structure function $F_2^c(x, Q^2), F_L^c, F_L$
- $B$-meson and $b\bar b$ pair production at the Tevatron
- charm, beauty, $D^*$ and $J/\psi$ production in two-photon collisions at LEP2
- Higgs production at the Tevatron and LHC
- prompt photon production at the HERA and Tevatron
- W/Z production at the Tevatron
Here I want to present the results of $b$-quark and $J/\psi$ production at the LHC [@JKLZ; @BLZ]in comparison with first experimental data obtained by ATLAS, CMS and LHCb Collaborations. The description of prompt photon production and DY lepton pairs was done by M. Malyshev [@Mal].
Ingredients of our $k_T$-factorization numerical calculations
=============================================================
To calculate the cross section of any physical process in the $k_T-$factorization approach according to the formula (3.1) the partonic cross section $\hat\sigma$ has to be taken off mass shell (${\mathbf k}_{T}$-dependent) and the polarization density matrix of initial gluons has to be taken in the so called BFKL form [^1] : $$\begin{aligned}
\sum \epsilon^{\mu} \epsilon^{*\,\nu} = {k_T^{\mu} k_T^{\nu} \over{\mathbf k}_T^2}.\end{aligned}$$
Concerning the uPDF in a proton, we used two different sets. First of them is the KMR one. The KMR approach represent an approximate treatment of the parton evolution mainly based on the DGLAP equation and incorpotating the BFKL effects at the last step of the parton ladder only, in the form of the properly defined Sudakov formfactors $T_q({\mathbf k}_T^2,\mu^2)$ and $T_g({\mathbf k}_T^2,\mu^2)$, including logarithmic loop corrections [@KMR]: $$\begin{aligned}
\displaystyle {\cal A}_q(x,{\mathbf k}_T^2,\mu^2) = T_q({\mathbf k}_T^2,\mu^2){\alpha_s({\mathbf k}_T^2)\over 2\pi}
\times\atop
{\displaystyle \times \int\limits_x^1 dz \left[P_{qq}(z) {x\over z} q\left({x\over z},{\mathbf k}_T^2\right) \Theta\left(\Delta - z\right) + P_{qg}(z) {x\over z}
g\left({x\over z},{\mathbf k}_T^2\right) \right]},
\label{KMR_q}\end{aligned}$$ $$\begin{aligned}
\displaystyle {\cal A}_g(x,{\mathbf k}_T^2,\mu^2) = T_g({\mathbf k}_T^2,\mu^2)
{\alpha_s({\mathbf k}_T^2)\over 2\pi} \times \atop {
\displaystyle \times \int\limits_x^1 dz \left[\sum_q P_{gq}(z) {x\over z}
q\left({x\over z},{\mathbf k}_T^2\right) + P_{gg}(z) {x\over z} g\left({x\over z},
{\mathbf k}_T^2\right)\Theta\left(\Delta - z\right) \right]},
\label{KMR_g}\end{aligned}$$ where $\Theta$-functions imply the angular-ordering constraint $\Delta =\mu/(\mu+k_T)$ specifically to the last evolution step (to regulate the soft gluon singularities). For other evolution steps the strong ordering in transverse momentum within DGLAP equation automatically ensures angular ordering. $T_a({\mathbf k}_T^2,\mu^2)$ - the probability of evolving from ${\mathbf k}_T^2$ to $\mu^2$ without parton emission. $T_a({\mathbf k}_T^2,\mu^2)=1$ at ${\mathbf k}_T^2 > \mu^2$. Such definition of the ${\cal A}_a(x,{\mathbf k}_T^2,\mu^2)$ is correct for ${\mathbf k}_T^2 > \mu_0^2$ only, where $\mu_0 \sim 1$ GeV is the minimum scale for which DGLAP evolution of the collinear parton densities is valid.\
We use the last version of KMRW uPDF obtained from DGLAP equations [@KMRW]. In this case ($a(x, \mu^2) = xG$ or $a(x, \mu^2) =xq$) the normalization condition $$\begin{aligned}
a(x,\mu^2) = \int\limits_0^{\mu^2} {\cal A}_a(x,{\mathbf k}_T^2,\mu^2)d{\mathbf k}_T^2\end{aligned}$$ is satisfied, if $$\begin{aligned}
{\cal A}_a(x,{\mathbf k}_T^2,\mu^2)\vert_{{\mathbf k}_T^2 < \mu_0^2} = a(x,\mu_0^2) T_a(\mu_0^2,\mu^2),\end{aligned}$$ where $T_a(\mu_0^2,\mu^2)$ are the quark and gluon Sudakov form factors. Then the uPDF ${\cal A}_a(x,{\mathbf k}_T^2,\mu^2)$ is defined in all ${\mathbf k}_T^2$ region.
Another uPDF was obtained using the CCFM evolution equation. The CCFM evolution equation has been solved numerically using a Monte-Carlo method [@HJ]\
According to the CCFM evolution equation the emission of gluons during the initial cascade is only allowed in an angular-ordered region of phase space. The maximum allowed angle $\Xi$ related to the hard quark box sets the scale $\mu$: $\mu^2 = \hat s + {\mathbf Q}_{T}^2(=\mu^2_f)$.\
The unintegrated gluon distribution are determined by a convolution of the non-perturbative starting distribution ${\cal{A}}_0(x)$ and CCFM evolution denoted by $\bar{\cal A}(x,{\mathbf k}_{T}^2,\mu^2)$: $$\begin{aligned}
x{\cal A}(x,{\mathbf k}_{T}^2,\mu^2)~=~\int dz {\cal A}_0(z){x\over z} {\bar{\cal A}}({x\over z},{\mathbf k}_{T}^2,\mu^2), \end{aligned}$$ where $$\begin{aligned}
x{\cal{A}}_0(x)~=~Nx^{p_0}(1-x)^{p_1}\exp(-{\mathbf k}_{T}^2/k^2_0).\end{aligned}$$ The parameters were determined in the fit to $F_2$ data.
Heavy quark production in $pp$-interaction
===========================================
The hard partonic subprocess $g^*g^*\rightarrow Q\bar Q$ is described by the Feynman’s diagrams presented in Fig. 1.
We used Sudakov decopmosition for the momenta of heavy quarks and the initial gluons: $p_i = \alpha_i P_1 + \beta_i P_2 + p_{i\perp}, \,\,\, k_1 = \alpha P_1 + k_{1\perp},\,\,\,k_2 =\beta P_2 + k_{2\perp}$, $p_1^2=p_2^2=M^2,\,\,\,k_1^2=k_{1T}^2,\,\,\,k_2^2=k_{2T}^2$, where in the center of mass frame of colliding particles
$P_1=(E,0,0,E),\,\,\, P_2=(E,0,0,-E),\,\,\, E=\sqrt s/2,\,\,\, P_1^2=P_2^2=0,\,\,\, (P_1P_2)=s/2$.
Sudakov variables are $$\begin{aligned}
\alpha_1 = \frac{M_{1T}}{\sqrt s}\exp(y_1^\ast),\,\,\,
\alpha_2 =\frac{M_{2T}}{\sqrt s}\exp(y_2^\ast),\,\,\,
\beta_1 = \frac{M_{1T}}{\sqrt s}\exp(-y_1^\ast),\,\,\,
\beta_2 = \frac{M_{2T}}{\sqrt s}\exp(-y_2^\ast),\nonumber\end{aligned}$$
$k_{1T}+k_{2T}=p_{1T}+p_{2T},\,\,\, \alpha =\alpha_1 +\alpha_2,\,\,\, \beta =\beta_1 +\beta_2$.
To guarantee gauge invariance, the process with off-shell incoming particles has to be embedded into the scaterring of on-shell particles. The second row of Fig. 1 includes non-factorizing diagrams which are factorized only in the sum. To make this factorization one can sum up these diagrams with the last diagram in the first row leading to one diagram with an effective Lipatov vertex by working in Feynman gauge [@Lipatov]: $$\begin{aligned}
\Gamma^{\nu}(k_1, k_2) = \frac{2P_1 P_2}{s}
\left(\frac{2t_1 + M^2_T}{\beta s} P_1^{\nu} - \frac{2t_2 +
M^2_T}{\alpha s} P_2^{\nu} - (k_{1T} - k_{2T})^{\nu}\right),\end{aligned}$$ where $t_1 = k_1^2 = - {\mathbf k}_{1T}^2,\,\,\, t_2 = k_2^2 = - {\mathbf k}_{2T}^2, \,\,\,
M_{T}^2 = \hat s + ({\mathbf k}_{1T}^2 + {\mathbf k}_{2T}^2)$.\
This vertex obeys the Ward identity: $\Gamma^{\mu}(k_1, k_2)k_{1\mu}=0$. Then the last five diagrams in Fig. 1 are replaced by one diagram (Fig. 2).
By neglecting the exchanged momentum in the coupling of gluons to incoming particles, we get an eikonal vertex which does not depend on the spin of the particle: $$\begin{aligned}
{\bar u}(\lambda_1', P_1 - k_1)\gamma^{\mu}u(\lambda_1, P_1)\,\,\,\,\,\to 2P_1^{\mu}\delta_{\lambda_1',\lambda_1}.\end{aligned}$$ Then it is possible to remove the external particle lines and attach so-called “non-sense” polarization to the incoming gluons: $$\begin{aligned}
\epsilon_{k_1}^{\mu} = {\sqrt 2}P_1^{\mu}/{\sqrt s},\,\,
\epsilon_{k_2}^{\mu} = {\sqrt 2}P_2^{\mu}/{\sqrt s}.\end{aligned}$$ Instead of Feynman gauge, one can choose an appropriate axial gauge $n\bullet A = 0$($n^{\mu} = a P_1^{\mu} + b P_2^{\mu}$).\
The contraction of the eikonal coupling with the gluon polarization in this gauge $$\begin{aligned}
d_{\mu\nu}^{(n)}(k) = - g_{\mu\nu} + \frac{n_{\mu} k_{\nu} + k_{\mu} n_{\nu}}{nk} - n^2\frac{k_{\mu}k_{\nu}}{(nk)^2}\end{aligned}$$ then reads $$\begin{aligned}
P_1^{\mu}d_{\mu\nu}^{(n)}(k_1) = k_{1T\nu}/\alpha,\,\,\
P_2^{\mu}d_{\mu\nu}^{(n)}(k_2) = k_{2T\nu}/\beta.\end{aligned}$$ In such a physical gauge the non-factortizing diagrammes vanish since the direct connection of two eikonal couplings gives $P_1^{\mu}d_{\mu\nu}^{(n)}P_2^{\nu} = 0$. It means the Lipatov vertex is to be replaced by the usual three gluon vertex. Then we can use the following matrix elements according to the diagrams in Fig. 1: $$\begin{aligned}
M_1 = \bar u(p_1)(-ig\gamma^{\mu})
\varepsilon_{\mu}(k_1)i{\hat p_1 - \hat k_1 + M\over (p_1 - k_1)^2 - M^2}(-ig\gamma^{\nu})\varepsilon_{\nu}(k_2)v(p_2),\\
M_2 = \bar u(p_1)(-ig\gamma^{\nu})
\varepsilon_{\nu}(k_2)i{\hat p_1 - \hat k_2 + M\over (p_1 - k_2)^2 -
M^2}(-ig\gamma^{\mu})\varepsilon_{\mu}(k_1)v(p_2), \\
M_3 = \bar u(p_1)C^{\mu\nu\lambda}(-k_1,-k_2,k_1+k_2)\,
{g^2\varepsilon_{\mu}(k_1)\varepsilon_{\nu}(k_2)
\over (k_1 + k_2)^2}\gamma_{\lambda}\,v(p_2),\end{aligned}$$ where $$\begin{aligned}
C^{\mu\nu\lambda}(k_1,k_2,k_3)=i((k_2 - k_1)^{\lambda}
g^{\mu\nu} + (k_3 - k_2)^{\mu}g^{\nu\lambda} + (k_1 - k_3)^{\nu}g^{\lambda\mu})\end{aligned}$$ is the standard three gluon vertex.
Numerical results
=================
Recently we have demonstrated reasonable agreement between the $k_T$-factorization predictions and the Tevatron data on the $b$-quarks, $b \bar b$ di-jets, $B^+$- and $D$-mesons [@JKLZ1]. Based on these results, here we give here analysis of the CMS [@CMS4; @CMS5; @CMS6] and LHCb [@LHCB] data in the framework of the $k_T$-factorization approach. We produced the relevant numerical calculations in two ways:
- We performed analytical parton-level calculations (which are labeled as LZ).
- The measured cross sections of heavy quark production was compared also with the predictions of full hadron level Monte Carlo event generator CASCADE.
In our numerical calculations we have used three different sets, namely the CCFM A0 (B0) and the KMR ones. The difference between A0 and B0 sets is connected with the different values of soft cut and width of the intrinsic ${\mathbf k}_{T}$ distribution. A reasonable description of the $F_2$ data can be achieved by both these sets. For the input, we have used the standard MSTW’2008 (LO) [@MSTW] (in LZ calculations) and the MRST 99 [@MRST] (in CASCADE) sets. The unintegrated gluon distributions depend on the renormalization and factorization scales $\mu_R$ and $\mu_F$. We set $\mu_R^2 = m_Q^2 + ({\mathbf p}_{1T}^2 + {\mathbf p}_{2T}^2)/2$, $\mu_F^2 = \hat s + {\mathbf Q}_T^2$, where ${\mathbf Q}_T$ is the transverse momentum of the initial off-shell gluon pair, $m_c = 1.4 \pm 0.1$ GeV, $m_b = 4.75 \pm 0.25$ GeV. We use the LO formula for the coupling $\alpha_s(\mu_R^2)$ with $n_f = 4$ active quark flavors at $\Lambda_{\rm QCD} = 200$ MeV, such that $\alpha_s(M_Z^2) = 0.1232$.\
We begin the discussion by presenting our results for the muons originating from the semileptonic decays of the $b$ quarks. The CMS collaboration has measured the transverse momentum and the pseudorapidity distributions of muons from $b$-decays. The measurements have been performed in the kinematic range $p_T^\mu > 6$ GeV and $|\eta^\mu| < 2.1$ at the total center-of-mass energy $\sqrt s = 7$ TeV.\
To produce muons from $b$-quarks, we first convert $b$-quarks into $B$-mesons using the Peterson fragmentation function with default value $\epsilon_b = 0.006$ and then simulate their semileptonic decay according to the standard electroweak theory taking into account the decays $b \to \mu $ as well as the cascade decay $b\to c\to \mu$.\
The results of our calcilations are shown in Figs. 3 – 8 in the comparison with the LHC data (see [@JKLZ] for more details). We obtain a good description of the data when using the CCFM-evolved (namely, A0) gluon distribution in LZ calculations. The shape and absolute normalization of measured $b$-flavored hadron cross sections at forward rapidities are reproduced well (see Fig. 6). The KMR and CCFM B0 predictions are somewhat below the data. In contrast with $b$ hadron and decay muon cross sections, the results for inclusive $b$-jet production based on the CCFM and KMR gluons are very similar to each other anda reasonable description of the data is obtained by all unintegrated gluon distributions under consideration.
The CASCADE predictions tend to lie slightly below the LZ ones and are rather close to the [<span style="font-variant:small-caps;">M</span>C@NLO]{} calculations (not shown). The observed difference between the LZ and CASCADE connects with the missing parton shower effects in the LZ evaluations. We have checked additionaly that the LZ and CASCADE predictions coincide at parton level..
Figs. 5 and 8 show the role of fragmentation and off-shelness effects in our calculations.
(0.0,1)\
(0.0,0.5)\
\
\
\
Quarkonium production in the $k_T-$factorization approach
=========================================================
The production of prompt $J/\psi$($\Upsilon$)-mesons in $pp$-collisions can proceed via either direct gluon-gluon fusion or the production of $P$-wave states $\chi_c (\chi_b$) and $S$-wave state $\psi'$ followed by their radiative decays $\chi_c (\chi_b)\to J/\psi(\Upsilon) + \gamma$. In the $k_T-$factorization approach the direct mechanism corresponds to the partonic subprocess $g^* + g^*\to J/\psi(\Upsilon) + g$. The production of $P$-wave mesons is given by $g^* + g^*\to\chi_c(\chi_b)$, and there is no emittion of any additional gluons. The feed-down contribution from $S$-wave state $\psi'$ is described by $g^* + g^*\to \psi' + g$.\
The cross sections charmonium states depend on the renormalization and factorization scales $\mu_R$ and $\mu_F$. We set $\mu_R^2 = m^2 + {\mathbf p}_T^2$ and $\mu_F^2
= \hat s + {\mathbf Q}_T^2$, where ${\mathbf Q}_T^2$ is the transverse momentum of initial off-shell gluon pair. Following to PDG [@PDG], we set $m_{J/\psi} = 3.097$ GeV, $m_{\chi c1} = 3.511$ GeV, $m_{\chi c2} = 3.556$ GeV, $m_{\psi '} = 3.686$ GeV and use the LO formula for the coupling constant $\alpha_s(\mu^2)$ with $n_f =4$ quark flavours at $\Lambda_{QCD} = 200$ Mev, such that $\alpha(M^2_Z) = 0.1232$.\
The charmonium wave functions are taken to be equal to $|{\cal}R_{J/\psi}(0)|^2/4{\pi} = 0.0876$ GeV$^3$, $|{\cal}{R'}_{\chi}(0)|^2 = 0.075$ GeV$^5$, $|{\cal}R_{\psi '}(0)|^2/4{\pi} = 0.0391$ GeV$^3$ and the following branching fractions are used $B(\chi_{c1} \to J/\psi +
\gamma) = 0.356, B(\chi_{c2} \to J/\psi + \gamma) = 0.202,
B(\psi ' \to J/\psi + X) = 0.561$ and $B(J/\psi \to
\mu^+\mu^-) = 0.0593$. Since the branching fraction for $\chi_{c0} \to J/\psi
+ \gamma $ decay is more than an order of magnitide smaller than for $\chi_{c1}$ and $\chi_{c2}$, we neglect its contribution to $J/\psi$ production. As $\psi' \to J/\psi + X$ decay matrix elements are unknown, these events were generated according to the phase space.
Comparison the results of our calculations with the CMS [@CMSJ], ATLAS [@ATLASJ] and LHCb [@LHCbJ] data are shown in Figs. 9 - 11 [@BLZ]. We see that the taking into account sole direct production is not sufficient to descibe the LHC data.We have obtained a good overall agreement between our predictions and the data when summing up the direct and feed-down contributions. The dependence of our numerical results on the uPDF is rather weak and the CCFM and KMR predictions are practically coincide. The difference between them can be observed at small $p_T$ or at large rapidities probed at the LHCb measerements. We have evaluated the polarizations parameters of prompt $J/\psi$ mesons in the kinematical region of CMS, ATLAS and LHCb measurements in the Collins-Soper and helicity frames (see [@BLZ]). We have took into account the contributions from the direcct and feed-dowm mechanisms. The qualitative predictions for the $J/\psi$ meson polarization are stable with respect to variations in the model parameters. Therefore future price measurements of the polarizations parameters of the $J/\psi$ mesons at the LHC will play crucial role in discriminating the different theoretical approaches.
\
\
\
Conclusions
===========
In the present time there is steady progress toward a better understanding of the $k_T$-factorization (high energy factorization) and the uPDF (TMD).\
We have described the first exp. data of $b$-quark and $J/\psi$ production at LHC in the $k_T$-factorization approach. We have obtained reasonable agreement of our calculations and the first experimental data taken by the CMS and ATLAS Collaborations.\
The dependence of our predictions on the uPDF appears at small transverse momenta and at large rapidities in $H_b$ and $J/\psi$ production covered by the LHCb experiment.\
Our study has demonstrated also that in the framework of the $k_T$-factorization approach there is no need in a color octet contributions for the charmonium production at the LHC.\
As it was shown in [@BLZ] the future experimental analyses of quarkonium polarization at LHC are very important and informative for discriminating the different theoretical models.
Acknowledgements {#acknowledgements .unnumbered}
================
I’d like to thank L.N. Lipatov for very useful discussion of different problems connected with subject of this talk. I’m very grateful to S.P. Baranov, H. Jung, M. Krämer and A.V. Lipatov for fruitful collaboration. I thank S.P. Baranov for careful reading the manuscript and very useful remarks. This research was supported by DESY Directorate in the framework of Moscow – DESY project on Monte-Carlo implementation for HERA – LHC, by the FASI of Russian Federation (grant NS-1456.2008.2), FASI state contract 02.740.11.0244, RFBR grant 11-02-01454-a and also by the RMES (grant the Scientific Research on High Energy Physics).
[99]{} E.A. Kuraev, L.N. Lipatov and V.S. Fadin, Sov. Phys. JETP [**44**]{}, 443 (1976);\
E.A. Kuraev, L.N. Lipatov and V.S. Fadin, Sov. Phys. JETP [**45**]{}, 199 (1977);\
I.I. Balitsky and L.N. Lipatov, Sov. J. Nucl. Phys. [**28**]{}, 822 (1978). L.V. Gribov, E.M. Levin, and M.G. Ryskin, Phys. Rep. [**100**]{}, 1 (1983). E.M. Levin, M.G. Ryskin, Yu.M. Shabelsky and A.G. Shuvaev, Sov. J. Nucl. Phys. [**53**]{}, 657 (1991);\
S. Catani, M. Ciafoloni and F. Hautmann, Nucl. Phys. B [**366**]{}, 135 (1991);\
J.C. Collins and R.K. Ellis, Nucl. Phys. B [**360**]{}, 3 (1991). M. Ciafaloni, Nucl. Phys. B [**296**]{}, 49 (1988);\
S. Catani, F. Fiorani and G. Marchesini, Phys. Lett. B [**234**]{}, 339 (1990);\
S. Catani, F. Fiorani and G. Marchesini, Nucl. Phys. B [**336**]{}, 18 (1990);\
G. Marchesini, Nucl. Phys. B [**445**]{}, 49 (1995). A.H. Mueller and J. Qiu, Nuscl. Phys. B [**268**]{}, 427 (1986);\
K. Golec-Biernat and M. Wusthoff, Phys. Rev D [**59**]{}, 014017 (1999); D [**60**]{}, 114023 (1999). I.I. Balitsky, Nucl. Phys. B [**463**]{}, 99 (1996);\
Y.V. Kovchegov, Phys. Rev. D [**60**]{}, 034008 (1999). M. Gyulassy and L. McLerran, Nucl. Phys. A [**750**]{}, 30 (2005);\
A.V. Leonidov, Uspekhi Fiz. Nauk [**175**]{}, 345 (2005). E. Iancu, Nucl. Phys. Proc. Suppl. [**191**]{}, 281 (2009), arXiv:0901.0986 \[hep-ph\];\
J.P. Blaizot, Nucl. Phys. A [**854**]{}, 237 (2011). B. Andersson [*et al.*]{} (Small-x Collaboration), Eur. Phys. J. C [**25**]{}, 77 (2002);\
J. Andersen [*et al.*]{} (Small-x Collaboration) Eur. Phys. J. C [**35**]{}, 67 (2004);\
J. Andersen [*et al.*]{} (Small-x Collaboration) Eur. Phys. J. C [**48**]{}, 53 (2006). F. Dominguez [*et al.*]{}, Phys. Rev. D [**83**]{}, 105005 (2011);\
S.M. Aybat, T.C. Rogers, Phys. Rev. D [**83**]{}, 114042 (2011);\
I.O. Cherednikov, arXiv:1102.0892 \[hep-ph\]; I.O. Cherednikov and N.G. Stefanis, Int. J. Mod. Phys. Conf. [**Ser.4**]{}, 135 (2011), arXiv:1108.0811 \[hep-ph\]. http://www.pv.infn.it/$~$bacchett/TMDprogram.htm. J.C. Collins, [*Foundations of Perturbati ve QCD*]{} (Cambridge University Press, Cambridge, 2011). V. Barone, M. Genovese, N.N. Nikolaev, E. Predazzi and B.G. Zakharov, Phys. Lett. B [**326**]{}, 161 (1994);\
A. Bialas, H. Navelet and R. Peschanski, Nucl. Phys. B [**593**]{}, 438 (2001). E. Avsar, arXiv:1108.1181 \[hep-ph\]. S. Marzani, Nucl. Phys. Proc. Suppl. [**205 - 206**]{}, 25 (2010), arXiv:1006.2314 \[hep-ph\]. R.D. Ball and R.K. Ellis, JHEP [**0105**]{}, 053 (2001). R.D. Ball, Nucl. Phys. B [**796**]{}, 137 (2008). H. Jung, M. Krämer, A.V. Lipatov and N.P. Zotov, DESY 11-180, arXiv:1111.1942 \[hep-ph\]. S.P. Baranov, A.V. Lipatov and N. P. Zotov, DESY 11-143, arXiv:1108.2856 \[hep-ph\]. A.V. Lipatov, M.A. Malyshev and N.P. Zotov, this Proccedings. M.A. Kimber, A.D Martin and M.G. Ryskin, Phys. Rev. D [**63**]{}, 114027 (2001). G. Watt, A.D. Martin and M.G. Ryskin, Eur. Phys. J. C [**31**]{}, 73 (2003). H. Jung, Comp. Phys. Comm. [**143**]{}, 100 (2002); H. Jung and G. Salam, Eur. Phys. J. C [**19**]{}, 359 (2001);\
H. Jung, S.P. Baranov, M. Deak [*et al.*]{}, Eur. Phys. J. C [**70**]{}, 1237 (2010). L.N. Lipatov, Sov. J. Nucl. Phys. [**23**]{}, 338 (1976). H. Jung, M. Krämer, A.V. Lipatov and N.P. Zotov, JHEP [**1101**]{}, 085 (2011). CMS Collaboration, JHEP [**1103**]{}, 090 (2011). CMS Collaboration, JHEP [**1103**]{}, 136 (2011). V. Chiochia, Nucl. Phys. A [**855**]{}, 436 (2011). LHCb Collaboration, Phys. Lett. B [**694**]{}, 209 (2010). A.D. Martin, W.J. Stirling, R.S. Thorne and G. Watt, Eur. Phys. J. C [**63**]{}, 189 (2009). A.D. Martin, R.G. Roberts, W.J. Stirling and R.S. Thorne, Eur. Phys. J. C [**14**]{}, 133 (2000). C. Amsler [*et al.*]{} (PDG Collaboration), Phys. Lett. B [**667**]{}, 1 (2008). CMS Collaboration, Eur. Phys. J. C [**71**]{}, 1575 (2011). G. Aad et al. (ATLAS Collaboration), Nucl. Phys. B [**850**]{}, 387 (2011). R. Aaij et al. (LHCb Collaboration), Eur. Phys. J. C [**71**]{}, 1645 (2011).
[^1]: The problem of choicing of proper gauge will be discussed in more details in Sec. 5.
| {
"pile_set_name": "ArXiv"
} |
---
address: Orange Labs
bibliography:
- 'DistributedLocalization.bib'
title: 'Multilateration - source localization from range difference measurements: Literature survey'
---
Introduction
============
As the technologies relying on distributed (individual) sensor arrays (like Internet Of Things) gain momentum, the questions regarding efficient exploitation of such acquired data become more and more important. A valuable information that could be provided by these arrays is the location of the source of the signal, *e.g.* the RF emitter, or a sound source. In this document, the focus on the latter use case - localizing a sound source, but the reader is reminded that the discussed methods are essentially agnostic to the signal type, as long as the RD measurements are available.
Specifically, assume a large aperture array of distributed mono microphones with potentially different gains, as opposed to distributed (compact) microphone arrays. The array geometry is assumed known in advance, and the microphones are already synchronized/syntonized. We further assume that all captured audio streams are readily available (*i.e.* centralized processing architecture). Lastly, we assume the presence of a direct path (line-of-sight) and the overdetermined setting, *i.e.* the number of speech sources $\ing{S}$ is smaller than the number of available microphones in the distributed array $\ing{M}$.
This scenario imposes several techical constraints:
1. The large aperture size implies significant spatial aliasing, which, along with the relatively small number of microphones, seriously degrades performance of beamforming-based techniques, at least in the narrowband setting [@dmochowski2008spatial]. The approaches based on *distributed* beamforming, *e.g.* [@valenzise2008resource; @yao1998blind], could still be appealing if they operate in the wideband regime: unfortunately, the literature on beamforming by distributed mono microphones is scarce.
2. The absence of compact arrays prevents the traditional Direction-of-Arrival (DOA) estimation.
3. No knowledge of the sources’ ignition times prohibits the Time-of-Flight (TOF) estimation.
Due to these constraints, the scope of the review is limited to the family of *multilateration methods* [@fresno2017survey] based on the TDOA estimation. Fortunately, it has been shown [@kaune2012accuracy] that the TOF and TDOA features perform similarly in terms of localization accuracy. Of particular interest is speaker localization within reverberant (indoor) and/or noisy environments. However, the TDOA estimation in these conditions is a challenging problem in its own right (especially in the multisource setting), and is out of the scope of this document - the interested reader may consult appropriate references, *e.g.* [@chen2006time; @blandin2012multi].
Distances between microphones are considered to be of the same order as the distances between microphones and source(s), hence we are in the near-field setting. The general formulation of the time-domain signal $y_{\ing{m}}(t)$, recorded at the $\ing{m}$^th^ microphone is given by the convolutional sum: $$\label{eqConvGeneral}
y_{\ing{m}}(t) = \sum\limits_{\ing{s}=1}^{\ing{S}} \int\limits_{0}^{\infty} \iter{a_{\ing{s}}}{m} (t, \tau) x_{\ing{s}}(t - \tau) d\tau + n_{\ing{m}}(t),$$ where $\iter{a_{\ing{s}}}{m} (t, :)$ is the time-variant Room Impulse Response (RIR) filter, relating the $\ing{m}$^th^ microphone position $\vect{r}_{\ing{m}}$ with the $\ing{s}$^th^ source position $\vect{r}_{\ing{s}}$, $x_{\ing{s}}(t)$ is the signal corresponding to the $\ing{s}$^th^ source, and $n_{\ing{m}}(t)$ is the additive noise of the considered microphone. In , the microphone gains are absorbed by RIRs. In practice, various simplifications are commonly used instead of the general expression . Commonly, a free-field, time-invariant approximation is adopted - in the single source case, it is given as follows [@gustafsson2003positioning]: $$y_{\ing{m}}(t) = \iter{a_{\ing{1}}}{m}x(t - \iter{\tau_{\ing{m}}}{1}) + n_{\ing{m}}(t),$$ where the offset $\iter{\tau_{\ing{m}}}{1}$ denotes the TOF value, which is proportional to the source-microphone distance.
The TDOA, corresponding to the difference in propagation delay between the microphones $\ing{m}$ and $\ing{m'}$, with respect to the source $\ing{s}$, is defined as ${\iter{\tau_{\ing{m},\ing{m'}}}{s} = \iter{\tau_{\ing{m'}}}{s} - \iter{\tau_{\ing{m}}}{s}}$. Naturally, the TDOA measurements could be corrupted by various types of noise, which negatively affects the performance of localization algorithms. Another cause of TDOA localization errors is the inexact knowledge of microphone positions. As shown in [@ho2007source], the Cramér-Rao lower bound (CRB) [@theodoridis2015machine] of the source location estimate increases rather quickly with the increase in the microphone position “noise” (fortunately, somewhat less fast in the near field, than in the far field setting). Finally, the localization accuracy also depends on the array geometry [@yang2006theoretical], which is assumed arbitrary in our case.
In homogeneous propagation media, the TDOA values $\iter{\tau_{\ing{m},\ing{m}'}}{s}$ , directly translate into Range Differences (RD) $\iter{d_{\ing{m},\ing{m'}}}{s}$, provided the sound speed $c$: $$\label{eqRDmulti}
\iter{d_{\ing{m},\ing{m'}}}{s} := \iter{D_{\ing{m'}}}{s} - \iter{D_{\ing{m}}}{s} = \norm{\vect{r}_{\ing{m'}} - \vect{r}_{\ing{s}}}{} - \norm{\vect{r}_{\ing{m}} - \vect{r}_{\ing{s}}}{} = c\cdot \iter{\tau_{\ing{m},\ing{m'}}}{s},$$ where $\iter{D_{\ing{m}}}{s}$ denotes the distance between the source $\ing{s}$ and the microphone $\ing{m}$. Thus, the observed RDs also suffer from measurement errors, usually modeled as an additive noise. Note that the observation model defines the two-sheet hyperboloid with respect to $\vect{r}_{\ing{s}}$, with foci in $\vect{r}_{\ing{m}}$ and $\vect{r}_{\ing{m}'}$ [@robots2011; @rodriguez2011theoretical].
Given the observations $\{ \iter{d_{\ing{m},\ing{m'}}}{s} \}$, and microphone positions $\left\{ \vect{r}_{\ing{m}} \right\}$, the goal now is to estimate source position(s) $\left\{ \vect{r}_{\ing{s}} \right\}$. In the multisource setting, multiple *sets* of RDs are assumed available, and the localization of each source is to be done independently of the rest. Such measurements could be obtained by multisource TDOA estimation algorithms, *e.g.* [@lombard2010tdoa].
Thus, without loss of generality, we will only discuss the single-source setting ($\ing{s}=1$). In the noiseless case, the number of linearly independent RD observations is equal to $\ing{M} - 1$, but considering the full set of observations (of size $\ing{M}(\ing{M}-1)/2$) may be useful for alleviating the harmful effects of measurement noise [@hahn1973optimum; @schmidt1972new]. Usually, the first microphone is chosen to be a reference point: *e.g.* $\vect{r}_{\ing{1}} = \vect{0}$, where $\vect{0}$ is the null vector. By denoting $\vect{r} := \vect{r}_{\ing{s}}$, from we have $$\label{eqRD}
d_{\ing{1},\ing{m'}} = \norm{\vect{r}_{\ing{m'}} - \vect{r}}{} - \norm{\vect{r}}{}.$$
In the following sections, we discuss different types of source location estimators and methods to calculate them.
Maximum likelihood estimation
=============================
Since the observations are non-linear, a statistically efficient estimate (*i.e.* the one that attains CRB) may not be available. The common approach is to seek the maximum likelihood (ML) estimator instead.
Let $\hat{\vect{r}}$ and $\hat{d}_{\ing{1},\ing{m'}}(\hat{\vect{r}})$ denote the estimated source position, and the corresponding RD, respectively: $$\hat{d}_{\ing{1},\ing{m'}}(\hat{\vect{r}}) = \norm{\vect{r}_{\ing{m'}} - \hat{\vect{r}}}{} - \norm{\hat{\vect{r}}}{} .$$ Under the hypothesis that the observation noise is Gaussian, the ML estimator is given as the minimizer of the negative log-likelihood [@chan1994simple; @huang2001real] $$\label{eqML}
\mathcal{L}(\vect{r}) = \transp{\left( \vect{d} - \hat{\vect{d}}(\hat{\vect{r}}) \right)} \mtrx{\Sigma}^{-1} \left( \vect{d} - \hat{\vect{d}}(\hat{\vect{r}}) \right),$$ $\vect{d} = \transp{\left[ d_{\ing{1},\ing{2}} \; d_{\ing{1},\ing{3}} \hdots d_{\ing{1},\ing{M}} \right]}$, $\hat{\vect{d}}(\hat{\vect{r}}) = \transp{ \left[ \hat{d}_{\ing{1},\ing{2}}(\hat{\vect{r}}), \; \hat{d}_{\ing{1},\ing{3}}(\hat{\vect{r}}) \hdots \hat{d}_{\ing{1},\ing{M}}(\hat{\vect{r}}) \right]}$, and $\mtrx{\Sigma}$ is the covariance matrix of the measurement noise.
Note, however that the Gaussian noise assumption for the RD measurements may not hold. For instance, the digital quantization effects can induce RD errors on the order of $2$ cm [@huang2008time]. Moreover, the ML estimators are proven to attain the CRB in the asymptotic regime, while the number of microphones (*i.e.* the number of RDs) is often small. Therefore, non-statistical estimators, such as least squares, are often used in practice instead. Anyhow, in this section we discuss two families of methods proposed for the TDOA maximum likelihood estimation: the ones that aim at solving the non-convex problem directly[^1], and the ones based on convex relaxations.
Direct methods
--------------
The problem is difficult to solve directly, due to nonlinear dependence of the RDs $\{\hat{d}_{\ing{1},\ing{m'}}(\hat{\vect{r}})\}$ on the position variable $\hat{\vect{r}}$.
Early approaches, based on iterative schemes, such as linearized gradient descent and Levenberg–Marquardt algorithm [@foy1976position; @ajdler2004acoustic], suffer from sensitivity to initialization, increased computational complexity and ill-conditioning (though the latter could be improved using regularization techniques [@mantilla2015localization]). The method proposed in [@bishop2008optimal] exploits correlation among noises within different TDOA measurements, and defines a constrained ML cost function tackled by a Newton-like algorithm. According to simulation results, it is more robust to adverse localization geometries [@bishop2010optimality; @yang2006theoretical] than [@foy1976position], or the least squares methods [@smith1987closed; @schmidt1972new; @li2004least; @gillette2008linear]. Another advantage of this method is the straightforward way to provide the initial estimate (however, as usual, global convergence cannot be guaranteed).
In the pioneering article [@chan1994simple], the authors proposed a closed-form, two-stage approach, that approximates the solution of . Firstly, the (weighted) unconstrained least-squares solution (to be explained in the next section) \[fnLS\] is computed, which is then improved by exploiting the relation between the estimates of the position vector and its magnitude. The minimal number of microphones, due to the unconstrained LS estimation is $5$ in three dimensions. It has been shown [@chan1994simple] that the method attains the CRB at high to moderate Signal-to-Noise-Ratios (SNRs). Unfortunately, it suffers from a nonlinear “threshold effect” - its performance quickly deteriorates at low SNRs. Instead, an approximate, but more stable version of this ML method has been proposed in [@chan2006exact]. In addition, the estimator [@chan1994simple] comes with a large bias [@mantilla2015localization], which cannot be reduced by increasing the amount of measurements. This bias has been theoretically evaluated and reduced in [@ho2012bias].
The method proposed in [@wang2011importance] uses Monte Carlo importance sampling techniques [@theodoridis2015machine] to approximate the solution of the problem . As an initial point, it uses the estimate computed by a convex relaxation method. According to simulation experiments, its localization performance is on pair with the convex method [@yang2009efficient], but the computational complexity is much lower.
A very recent article [@larsson2019optimal] proposes the linearization approach that casts the original into an eigenvalue problem, which can be solved optimally in closed form. Additionally, the authors propose an Iterative Reweighted Least Squares scheme that approximates the ML estimate for different noise distributions.
Convex relaxations
------------------
Another important line of work are the methods based on convex relaxations of ML estimation problems. In other words, the original problem is approximated by a convex one [@boyd2004convex], which is usually far easier to solve. Two families of approaches dominate this field: methods based on semidefinite programming (SDP), and the ones relaxing the original task into a second-order cone optimization problem (SOCP). In the former, the non-convex quadratic problem is first *lifted* such that the non-convexity appears as a rank 1 constraint, which is then substituted by a positive semidefinite one [@ma2010semidefinite]. Lifting is a problem reformulation by variable substitution $\mtrx{G} = \vect{g}\transp{\vect{g}}$, where $\vect{g}$ is the original optimization variable (the term *lifting* is used to emphasize that the problem is now defined in a high-dimensional space). On the other hand, solving the SDP optimization problems can be computationally expensive, and the SOCP framework has been proposed as a compromise between the approximation quality and computational complexity (*cf.* [@kim2001second] for technical details).
One of the first convex relaxation approaches for the TDOA localization is [@lui2008semidefinite], based on SDP. The algorithm requires the knowledge of the microphone closest to the source, in order to ensure that all RDs (with that microphone as a reference) are positive. The article [@yang2009efficient] discusses three convex relaxation methods. The first one, based on SOCP relaxation is computationally efficient, but restricts the solution to the convex hull [@biswas2004semidefinite; @boyd2004convex] of microphone positions. The other two SDP-based remove this restriction, but are somewhat more computationally demanding. In addition, one of these is the *robust* version - it minimizes the worst-case error due to imprecise microphone locations. The latter requires tuning of several hyperparameters, among which is the variance of the microphone positioning error. All three versions are based on the white Gaussian noise model for the TDOA measurements, however, whithening could be applied in order to support the correlated noise case. However, the SDP solutions are not the final output of the algorithms, but are used to initialize nonlinear iterative scheme, such as [@foy1976position].
Interestingly, a recent article [@qu2016efficient] has shown that the ideas of the direct approach [@chan1994simple] and the constrained least-squares approach could be mixed together. Moreover, the cost function can be casted to a convex problem, for which an interior-point method has been proposed. However, in practice, it is a compound algorithm which iteratively solves a sequence of convex problems in order to re-calculate a weighting matrix dependant on the estimated source position. The accuracy depends on the number of iterations, which, in turn, increases computational complexity. As for [@chan1994simple], it requires $5$ microphones for the 3D localization.
Least-squares estimation {#secLS}
========================
Largely due to computational convenience, the least-squares (LS) estimation is often a preferred parameter estimation approach. It is noteworthy that all LS approaches optimize a somewhat “artificial” estimation objective, which can induce large errors in very low SNR conditions, when the measurement noise is not white, and/or for some adverse array geometries [@bishop2008optimal; @sirola2010closed; @ho2012bias].
Three types of cost functions are discussed: hyperbolic, spherical and conic LS.
Hyperbolic LS
-------------
The goal is to minimize the sum of squared distances $\epsilon_{\mathrm{h}}$ between the true and estimated RDs: $$\begin{gathered}
\label{eqCostHyp}
\epsilon_{\mathrm{h}}( \hat{\vect{r}} ) := \sum\limits_{\ing{m'}=2}^{\ing{M}} \left( d_{\ing{1},\ing{m'}} - \norm{\vect{r}_{\ing{m'}} - \hat{\vect{r}}}{} + \norm{\hat{\vect{r}}}{} \right)^2 \\ = \transp{\left( \vect{d} - \hat{\vect{d}}(\hat{\vect{r}}) \right)} \left( \vect{d} - \hat{\vect{d}}(\hat{\vect{r}}) \right) ,\end{gathered}$$ which is analogous to the ML estimation problem for $\mtrx{\Sigma} = \mtrx{I}$, with $\mtrx{I}$ being the identity matrix. Thus, in the case of *white* Gaussian noise, the hyperbolic LS solution coincides with the ML solution. Otherwise, solving comes down to finding the point $\hat{\vect{r}}$ whose distance to all hyperboloids $d_{\ing{1},\ing{m'}}$, defined in , is minimal.
However, the hyperbolic LS problem is also non-convex, and its global solution cannot be guaranteed. Instead, local minimizers are found by iterative procedures, such as (nonlinear) gradient descent or particle filtering [@torrieri1984statistical; @gustafsson2003positioning]. Obviously, the quality of the output result of such algorithms depends on their initial estimates, the choice of which is usually not mathematical, but rather application-based.
Spherical LS
------------
By squaring the idealized RD measurement expression , followed by some simple algebraic manipulations, we have $$d_{\ing{1},\ing{m'}} \norm{\vect{r}}{} + \transp{\vect{r}_{\ing{m'}}}\vect{r} - \underbrace{\frac{1}{2} \left( \norm{\vect{r}_{\ing{m'}}}{}^2 - d_{\ing{1},\ing{m'}}^2 \right)}_{b_{\ing{m'}}} = 0.$$ The interest of this operation is in decoupling of the position vector and its magnitude, which are to be replaced by their estimates $\hat{\vect{r}}$ and $\hat{D}:=\norm{\hat{\vect{r}}}{}$, respectively.
The goal now becomes driving the sum of left hand sides (for all microphones) to zero: $$\label{eqSphLS}
\epsilon_{\mathrm{sp}} = \sum\limits_{\ing{m'}=2}^{\ing{M}} \left( d_{\ing{1},\ing{m'}} \hat{D} + \transp{\vect{r}_{\ing{m'}}}\hat{\vect{r}} - b_{\ing{m'}} \right)^2, \; \hat{D}^2=\norm{\hat{\vect{r}}}{}^2,$$ which leads to the following (compactly written) constrained optimization problem [@beck2008exact]: $$\begin{aligned}
\label{eqConstrained}
\minim_{\hat{\vect{c}}} & \norm{\mtrx{\Phi} \hat{\vect{c}} - \vect{b}}{}^2 \\
\text{subject to} \; \transp{\hat{\vect{c}}} \left[
\begin{smallmatrix}
1 & \mtrx{0} \\
\vect{0} & -\mtrx{I}
\end{smallmatrix} \right] \hat{\vect{c}} & = 0 \; \text{and} \; \hat{c}_{(1)} \geq 0, \nonumber\end{aligned}$$ where $\mtrx{\Phi} = \left[
\begin{smallmatrix}
d_{\ing{1},\ing{1}} & \transp{\vect{r}_{\ing{1}}} \\
d_{\ing{1},\ing{2}} & \transp{\vect{r}_{\ing{2}}} \\
\hdots & \hdots \\
d_{\ing{1},\ing{M}} & \transp{\vect{r}_{\ing{M}}}
\end{smallmatrix} \right]$, $\hat{\vect{c}} = \left[ \begin{smallmatrix} \hat{D} \\ \hat{\vect{r}} \end{smallmatrix} \right]$, $\vect{b} = \left[ \begin{smallmatrix} b_{\ing{1}} \\ b_{\ing{2}} \\ \hdots \\ b_{\ing{M}} \end{smallmatrix} \right]$, and $\hat{c}_{(1)}$ denotes the first entry of the column vector $\hat{\vect{c}}$.
In the literature, the problem above is tackled as:
Unconstrained LS
: : by ignoring the constraints relating the position estimate $\hat{\vect{r}}$ and its magnitude $\hat{D}$, the problem admits a closed-form solution $\hat{\vect{c}}^* = \left(\transp{\mtrx{\Phi}} \mtrx{\Phi} \right)^{-1} \transp{\mtrx{\Phi}} \vect{b}$. As pointed in [@stoica2006lecture; @wei2008comments], several well-known estimation algorithms [@smith1987closed; @li2004least; @gillette2008linear] actually yield the unconstrained LS estimate. The minimum of $\ing{M}=5$ microphones (*i.e.* four RD measurements), in three dimensions, are required in order for $\left(\transp{\mtrx{\Phi}} \mtrx{\Phi} \right)^{-1}$ to be an invertible matrix.
Constrained LS
: : While the unconstrained LS is simple and computationally efficient, its estimate is known to have a large variance compared to the CRB [@wei2008comments], hence the interest for solving the constrained problem. Unfortunately, is non-convex due to quadratic constraints. To directly incorporate the constraint(s), a Lagrangian-based iterative method has been proposed in [@huang2001real], albeit without any performance guarantees.
Later, in their seminal paper [@beck2008exact], Beck and Stoica provided a closed-form *global* solution of the problem, and demonstrated that it gives orders of magnitude more accurate solution (at an increased computational cost) than the unconstrained LS estimator. Moreover, the results in [@wang2011importance] indicate that it is generally more accurate than the two-stage ML solution [@chan1994simple].
Conic LS
--------
In [@schmidt1972new], Schmidt has shown that (in two dimensions) the RDs of *three* known microphones define the major axis of a general conic[^2], on which the corresponding microphones lie. In addition, the source is positioned on its focus. In three dimensions, this axis becomes a plane containing the source. The fourth (non-coplanar) microphone is needed to infer the source position $\vect{r}$, by calculating the intersection coordinates of three such planes[^3]. Thus, the method attains the theoretical minimum for the required number of microphones for TDOA localization.
To illustrate the approach, let one such triplet of microphones be described by $(\vect{r}_{\ing{1}}, \vect{r}_{\ing{2}}, \vect{r}_{\ing{3}})$, and $(D_{\ing{1}}, D_{\ing{2}}, D_{\ing{3}})$ – their position vectors, and the distances to the source, respectively. For each pair $(\ing{i},\ing{j})$ of these microphones, we have the following expression for the product of the *range sum* $\Sigma_{\ing{i,j}}$ and the RD $d_{\ing{i,j}}$: $$\label{eqSumDif}
\Sigma_{\ing{i,j}}d_{\ing{i,j}} = (D_{\ing{i}} + D_{\ing{j}})(D_{\ing{j}} - D_{\ing{i}}) = \norm{\vect{r}_{\ing{j}}}{}^2 - \norm{\vect{r}_{\ing{i}}}{}^2 - 2 \transp{(\vect{r}_{\ing{j}} - \vect{r}_{\ing{i}} )} \vect{r}$$
By rearranging the terms in , and having $d_{\ing{k,i}}=\Sigma_{\ing{i,j}} - \Sigma_{\ing{j,k}}$, the range sums can be eliminated. Eventually, this gives the aforementioned plane equation $$\begin{gathered}
\label{eqPlane}
\transp{\left( d_{\ing{2,3}} \vect{r}_{\ing{1}} + d_{\ing{3,1}} \vect{r}_{\ing{2}} + d_{\ing{1,2}} \vect{r}_{\ing{3}} \right)}\vect{r} \\ = \frac{1}{2} \left( d_{\ing{1,2}}d_{\ing{2,3}}d_{\ing{3,1}} + d_{\ing{2,3}} \norm{\vect{r}_{\ing{1}}}{}^2 + d_{\ing{3,1}} \norm{\vect{r}_{\ing{2}}}{}^2 + d_{\ing{1,2}} \norm{\vect{r}_{\ing{3}}}{}^2 \right).\end{gathered}$$ This is a linear equation of three unknowns, thus the exact solution is obtained when three triplets (*i.e.* four non-coplanar microphones) are available. Browsing the literature, we found that exactly the same closed-form approach has been recently reinvented in the highly cited article [@bucher2002synthesizable], some $30$ years after Schmidt’s original paper.
For $\ing{M}$ microphones, one ends up with $\ing{M}\choose{3}$ such equations (in 3D) - the classical LS solution is to stack them into a matrix form, and calculate the position $\vect{r}$ by applying the Moore-Penrose pseudoinverse. Let $A_{\ing{pqr}}$, $B_{\ing{pqr}}$, $C_{\ing{pqr}}$ and $F_{\ing{pqr}}$ denote the coefficients and the right hand side of the expression , for the microphone triplet $\ing{m} \in \{\ing{p, q, r} \}$, respectively. For all such triplets, we have $$\underbrace{\left[
\begin{matrix}
A_{\ing{123}} & B_{\ing{123}} & C_{\ing{123}} \\
\vdots & \vdots & \vdots \\
A_{\ing{pqr}} & B_{\ing{pqr}} & C_{\ing{pqr}} \\
\vdots & \vdots & \vdots \\
\end{matrix}
\right]}_{\mtrx{\Psi}}
\vect{r} =
\underbrace{\left[
\begin{matrix}
F_{\ing{123}} \\
\vdots \\
F_{\ing{pqr}} \\
\vdots
\end{matrix}
\right]}_{\vect{\psi}},$$ where $$\begin{aligned}
A_{\ing{pqr}} &= d_{\ing{q,r}} r_{\ing{p}(1)} + d_{\ing{r,p}} r_{\ing{q}(1)} + d_{\ing{p,q}} r_{\ing{r}(1)}, \\
B_{\ing{pqr}} &= d_{\ing{q,r}} r_{\ing{p}(2)} + d_{\ing{r,p}} r_{\ing{q}(2)} + d_{\ing{p,q}} r_{\ing{r}(2)}, \\
C_{\ing{pqr}} &= d_{\ing{q,r}} r_{\ing{p}(3)} + d_{\ing{r,p}} r_{\ing{q}(3)} + d_{\ing{p,q}} r_{\ing{r}(3)} \; \text{and} \\
F_{\ing{pqr}} &= \frac{1}{2} \left( d_{\ing{p,q}}d_{\ing{q,r}}d_{\ing{r,p}} + d_{\ing{q,r}} \norm{\vect{r}_{\ing{p}}}{}^2 + d_{\ing{r,p}} \norm{\vect{r}_{\ing{q}}}{}^2 + d_{\ing{p,q}} \norm{\vect{r}_{\ing{r}}}{}^2 \right), \end{aligned}$$ as in . However, such LS solution is strongly influenced by the triplets having large $A_{\cdot}$, $B_{\cdot}$, $C_{\cdot}$ or $F_{\cdot}$ values. Instead, as proposed in [@schmidt1972new], the matrix $\mtrx{\Psi}$ needs to be preprocessed prior to computing the pseudoinverse - its rows should be scaled by $1/\sqrt{A_{\cdot}^2 + B_{\cdot}^2 + C_{\cdot}^2}$, as well as the corresponding entry of the vector $\vect{\psi}$.
Likewise, the presence of noise in the TDOA measurements $d_{\ing{i,j}}$ could seriously degrade the localization accuracy. In that case, the observation model contains an additive noise term, which varies accross different measurements, rendering them *inconsistent*. This means that the intrinsic redundancy within TDOAs does not hold, *e.g* $d_{\ing{i,k}} \neq d_{\ing{i,j}} + d_{\ing{j,k}}$. In the noiseless case, the vector $\vect{d}$ of concatenated TDOA measurements, lies in the range space of a simple first-order difference matrix [@schmidt1996least], specified by and the ordering of distances $D_{\ing{m}}$. Thus, the measurements could be preconditioned, by replacing them with the closest feasible TDOAs, in the LS sense. This is done by projecting the measured $\vect{d}$ onto the range space of the finite difference matrix, or, equivalently by the technique called “TDOA averaging” [@schmidt1996least].
Conclusion
==========
The RD-based localization has a long history, and the methods exploiting these are, in general, theoretically well-founded. However, as there is no (to the best knowledge of the author) comprehensive benchmark of the most performant algorithms of each class (ML, hyperbolic/spherical/conic LS), there is neither a clear winner. While the ML methods are closed to optimal in theory, they resort to various approximations in order to combat the intrinsic hardness of the localization problem. The LS approaches instead solve easier, but artificial optimization problems. On the other hand, some of them are computationally very efficient, and seemingly also work very well in practice. Hence, the choice of the method depends on the given use case - what is the type and the level of measurement noise, how important is the computational complexity, how many microphones comprise the array, and finally, whether the empirical results are at hand.
[^1]: Not to be confused with “direct” localization methods based on grid search, such as steered response power beamformer.
[^2]: A conic (section) is either a hyperbola, an elipse or a parabola.
[^3]: Hence the name *plane intersection method* in the literature [@smith1987closed].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we prove the conjecture of Molino that for every singular Riemannian foliation $(M,{\ensuremath{\mathcal{F}}})$, the partition ${\overline}{{\ensuremath{\mathcal{F}}}}$ given by the closures of the leaves of ${\ensuremath{\mathcal{F}}}$ is again a singular Riemannian foliation.'
address:
- |
Marcos M. Alexandrino Universidade de São Paulo\
Instituto de Matemática e Estatística, Rua do Matão 1010,05508 090 São Paulo, Brazil
- |
Marco Radeschi Mathematisches Institut\
WWU Münster, Einsteinstr. 62, Münster, Germany.
author:
- 'Marcos M. Alexandrino'
- Marco Radeschi
title: 'Closure of singular foliations: the proof of Molino’s conjecture'
---
Introduction
============
Given a Riemannian manifold $M$, a singular Riemannian foliation ${\ensuremath{\mathcal{F}}}$ on $M$ is, roughly speaking, a partition of $M$ into smooth connected and locally equidistant submanifolds of possibly varying dimension (the *leaves* of ${\ensuremath{\mathcal{F}}}$), which is spanned by a family of smooth vector fields. The precise definition, given in Section \[S:prelim\], was suggested by Molino, by combining the concepts of *transnormal system* of Bolton [@Bolton] and of *singular foliation* by Stefan and Sussmann [@Sussmann].
A typical example of a singular Riemannian foliation is the decomposition of a Riemannian manifold $M$ into the orbits of an isometric group action $G$ on $M$. Such a foliation is called *homogeneous*. Another example of foliation is given by the partition of an Euclidean vector bundle $E\to L$, endowed with a metric connection, into the *holonomy tubes* around the zero section (cf. Example \[E:foliation-vector-bundle\]). Such a foliation, which we call *holonomy foliation*, will be a sort-of prototype in the structural results that will appear later on. Holonomy foliations are in general not homogeneous (the zero section $L$ is always a leaf but in general not a homogeneous manifold), however they are *locally homogeneous*, in the sense that the infinitesimal foliation at every point of $E$ is homogeneous (cf. Sections \[SS:inf-foliation\] and \[SS:orbit-like\]). This construction is related to other important types of foliations, like polar foliations [@toeben] or Wilking’s dual foliation to the Sharafutdinov projection [@Wilking], see Remark \[R:eamples\]. In general, the leaves of a singular Riemannian foliation might not be closed, even in the simple cases defined above. In the homogeneous case, consider for example the foliation on the flat torus $T^2$ by parallel lines, of irrational slope. These are non closed orbits, of an isometric ${\mathbb{R}}$-action on $T^2$.
Given a (regular) Riemannian foliation $(M,{\ensuremath{\mathcal{F}}})$ with non-closed leaves, Molino proved that replacing the leaves of ${\ensuremath{\mathcal{F}}}$ with their closure yields a new singular Riemannian foliation ${\overline}{{\ensuremath{\mathcal{F}}}}$. Moreover, he conjectured that the same result should hold true if one starts with a singular Riemannian foliation, and this has become known, in the last decades, as *Molino’s Conjecture*.
Molino proved that the closure ${\overline}{{\ensuremath{\mathcal{F}}}}$ of a singular Riemannian foliation $(M,{\ensuremath{\mathcal{F}}})$ is a transnormal system [@Molino], thus leaving to prove that it is a singular foliation as well. Moreover, in [@Molino2] he suggested a strategy to prove the conjecture for the case of *orbit-like foliations*, i.e. foliations which, roughly speaking, are locally diffeomorphic to the orbits of some proper isometric group action around each point (cf. Section \[SS:orbit-like\]). A formal alternative proof in this case can be found in \[4\]. Molino’s conjecture was also proved for *polar foliations* and then *infinitesimally polar foliations* in [@Alexandrino-molino-polar] and [@Alexandrino-Lytchak], respectively.
These partial results do not cover every possible foliation. Since the Eighties there are examples of non orbit-like foliations, and in recent years it was shown the existence of a remarkably large class of “infinitesimal” foliations that are neither homogeneous nor polar, the so-called Clifford foliations [@Radeschi] (these infinitesimal foliations have been shown, however, to have an algebraic nature, cf. [@Lytchak-Radeschi]). Therefore, it is important to give a complete answer to the conjecture, to fully understand the semi-local dynamic of singular Riemannian foliations.\
The goal of this paper is to prove the full Molino’s conjecture.
(Molino’s Conjecture) Let $(M,{\ensuremath{\mathcal{F}}})$ be a singular Riemannian foliation on a complete manifold $M$, and let ${\overline}{{\ensuremath{\mathcal{F}}}}=\{{\overline}{L}\mid L\in {\ensuremath{\mathcal{F}}}\}$ be the partition of $M$ into the closures of the leaves of ${\ensuremath{\mathcal{F}}}$. Then $(M,{\overline}{{\ensuremath{\mathcal{F}}}})$ is a singular Riemannian foliation.
This result is in fact a direct consequence of the following.
Let $(M,{\ensuremath{\mathcal{F}}})$ be a singular Riemannian foliation, let $L$ be a (possibly not closed) leaf, and let $U$ be an $\epsilon$-neighbourhood around the closure of $L$. Then for $\epsilon$ small enough, there is a metric ${\ensuremath{ \mathrm{g} }}^\ell$ on $U$ and a singular foliation $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$, such that:
1. $(U,{\ensuremath{ \mathrm{g} }}^\ell,\widehat{{\ensuremath{\mathcal{F}}}}^\ell)$ is an orbit-like singular Riemannian foliation.
2. The foliation $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ coincides with ${\ensuremath{\mathcal{F}}}$ on ${\overline}{L}$.
3. The closure of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ is contained in the closure of ${\ensuremath{\mathcal{F}}}$.
In short, the foliation $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ is obtained by first constructing the *linearized foliation* ${\ensuremath{\mathcal{F}}}^\ell$ of ${\ensuremath{\mathcal{F}}}$ in $U$, which is a subfoliation of ${\ensuremath{\mathcal{F}}}$ spanned by the first order approximations, around $L$, of the vector fields tangent to ${\ensuremath{\mathcal{F}}}$ (see Section \[SS:lin-foliation\] for a precise definition). The foliation $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ is then obtained from ${\ensuremath{\mathcal{F}}}^\ell$ by taking the “local closure” of the leaves of ${\ensuremath{\mathcal{F}}}^\ell$. The foliations ${\ensuremath{\mathcal{F}}},\,{\ensuremath{\mathcal{F}}}^\ell,\,\widehat{{\ensuremath{\mathcal{F}}}}^\ell$, together with their closures, are then related by the following inclusions:
$\renewcommand{\arraystretch}{1.3}
\begin{array}[c]{ccccc}
\vspace{-0.1in} {\ensuremath{\mathcal{F}}}&\supseteq &{\ensuremath{\mathcal{F}}}^\ell&\subseteq&\widehat{{\ensuremath{\mathcal{F}}}}^\ell\\
\vspace{0.1in}\rotatebox{-90}{$\subseteq$}&&\rotatebox{-90}{$\subseteq$}&&\rotatebox{-90}{$\subseteq$}\\
{\overline}{{\ensuremath{\mathcal{F}}}}&\supseteq &{\overline}{{\ensuremath{\mathcal{F}}}}^\ell&=&{\overline}{\widehat{{\ensuremath{\mathcal{F}}}}}{}^\ell\\
\end{array}$
\[E:prototype\] Consider an Euclidean vector bundle $E$ over a complete Riemannian manifold $L$, with a metric connection $\nabla^{E}$ and a connection metric $g^{E}$ (cf. Example \[E:foliation-vector-bundle\]). Let $H_p$ denote the holonomy group of $(E,\nabla^E)$ at $p$, acting by isometries on the Euclidean fiber $E_p$, and let $(E_p,{\ensuremath{\mathcal{F}}}^0_p)$ be a singular Riemannian foliation preserved by the $H_p$-action. Finally, let $K_p$ be the maximal connected group of isometries of $E_{p}$ that fixes each leaf of ${\ensuremath{\mathcal{F}}}^{0}_{p}$ as a set.
Letting ${\ensuremath{\mathcal{F}}}$ the partition of $E$ into the holonomy translates of the leaves of ${\ensuremath{\mathcal{F}}}^0_p$ (i.e., for every leaf $\mathcal{L}\in {\ensuremath{\mathcal{F}}}^0_p$, $L_{\mathcal{L}}$ denotes the set of points in $E$ that can be reached via $\nabla^E$-parallel translation from a point in $\mathcal{L}$), then ${\ensuremath{\mathcal{F}}}$ is a singular Riemannian foliation. In this case, the linearized foliation ${\ensuremath{\mathcal{F}}}^\ell$ is the foliation by the holonomy translates of the $K_p$-orbits in $E_p$, and the local closure of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ is the foliation by the holonomy translates of the $\overline{K}_p$-orbits in $E_p$, where $\overline{K}_p$ denotes the closure of $K_p$ in ${\ensuremath{\mathsf{O}}}(E_p)$.
This can be restated in the language of groupoids: defining $H$ the holonomy groupoid of the connection $\nabla^E$, then ${\ensuremath{\mathcal{F}}}=\{H(\mathcal{L})\mid \mathcal{L}\in {\ensuremath{\mathcal{F}}}^0_p\}$, ${\ensuremath{\mathcal{F}}}^\ell$ is given by the orbits of $HK_p$, and its local closure $\hat{{\ensuremath{\mathcal{F}}}}^\ell$ is given by the orbits of $H\overline{K}_p$.
This paper is organized as follows: after a section of preliminaries (Section \[S:prelim\]) we show how the Molino’s Conjecture follows from the Main Theorem (Section \[S:Molino\]). In Section \[S:setup\] we fix the setup in which we work for the rest of the paper. In Section \[S:distributions\] we define three distributions of the tangent bundle $TU$. We first use these to obtain information on the local structure of ${\ensuremath{\mathcal{F}}}^\ell$ and define the local closure $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ (Section \[S:linearized\]) and then to define the metric ${\ensuremath{ \mathrm{g} }}^\ell$ used in the Main Theorem (Section \[S:proof\]). In this final section we also prove the Main Theorem.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors thank Prof. Lytchak and Prof. Thorbergsson for consistent support.
Preliminaries {#S:prelim}
=============
Given a Riemannian manifold $(M,{\ensuremath{ \mathrm{g} }})$, a partition ${\ensuremath{\mathcal{F}}}$ of $M$ into complete connected submanifolds (the *leaves* of ${\ensuremath{\mathcal{F}}}$) is called a *transnormal system* if geodesics starting perpendicular to a leaf stay perpendicular to all leaves, and a *singular foliation* if every vector tangent to a leaf can be locally extended to a vector field everywhere tangent to the leaves. A singular Riemannian foliation will be denoted by the triple $(M,{\ensuremath{ \mathrm{g} }},{\ensuremath{\mathcal{F}}})$. However, if the Riemannian metric of $M$ is understood, we will drop it and simply write $(M,{\ensuremath{\mathcal{F}}})$.
The following notation will be used throughout the rest of the paper. Given a point $p\in M$, the leaf of ${\ensuremath{\mathcal{F}}}$ through $p$ will be denoted by $L_p$. A small relatively compact open subset $P\subset L$ is called a *plaque*. The tangent and normal spaces to $L_p$ at $p$ are denoted by $T_pL_p$ and $\nu_pL_p$, respectively. Given some $\epsilon>0$, $\nu^\epsilon_pL_p$ denotes the set of vectors $x\in \nu_pL_p$ with norm $<\epsilon$. If $\epsilon$ is small enough that the normal exponential map $\exp:\nu_p^\epsilon L_p\to M$ is a diffeomorphism onto the image, such image is called a *slice* of $L_p$ at $p$, and it is denoted by $S_p$. The *slice foliation* ${\ensuremath{\mathcal{F}}}|_{S_p}$ denotes the partition of $S_p$ into the connected components of the intersections $L\cap S_p$, where $L\in {\ensuremath{\mathcal{F}}}$.
Vector fields of a singular Riemannian foliation
------------------------------------------------
We review here the main notations about vector fields of a singular Riemannian foliation.
A vector field $V$ is called *vertical* if it is tangent to the leaves at each point. The set of smooth vertical vector fields is a Lie algebra, which is denoted by $\mathfrak{X}(M,{\ensuremath{\mathcal{F}}})$.
A vector field $X$ is called *foliated* if its flow takes leaves to leaves or, equivalently, if $[X, V]\in \mathfrak{X}(M,{\ensuremath{\mathcal{F}}})$ for every $V\in \mathfrak{X}(M,{\ensuremath{\mathcal{F}}})$. Any vertical vector field is foliated, but there are other foliated vector fields. A vector field is called *basic* if it is both foliated and everywhere normal to the leaves.
Homothetic Transformation Lemma
-------------------------------
One of the most fundamental results in the theory of singular Riemannian foliations is the Homothetic Transformation Lemma. A deeper discussion of this lemma, with proof and applications, can be found in Molino [@Molino], Ch. 6, in particular Lemma 6.1 and Proposition 6.7.
Let $(M,{\ensuremath{\mathcal{F}}})$ be a singular foliation, let $L$ be a leaf of ${\ensuremath{\mathcal{F}}}$, and let $P\subset L$ a plaque. Let $\epsilon>0$ be such that the normal exponential map $\exp:\nu^\epsilon P\to M$ is a diffeomorphism onto its image $B_{\epsilon}(P)$. For any two radii $r_1, r_2=\lambda r_1$ in $(0,\epsilon)$, it makes sense to define the *homothetic transformation* $$h_{\lambda}:B_{r_1}(P)\to B_{r_2}(P),\qquad h_{\lambda}(\exp v)=\exp \lambda v.$$ The leaves of ${\ensuremath{\mathcal{F}}}$ intersect $B_{r_i}(P)$, $i=1,2$, in plaques that foliate $B_{r_i}(P)$. We call ${\ensuremath{\mathcal{F}}}|_{B_{r_i}(P)}$ the foliation of $B_{r_i}(P)$ into the path components of such intersections. One has then the following:
The homothetic transformation $h_{\lambda}$ takes the leaves of $(B_{r_1}, {\ensuremath{\mathcal{F}}}|_{B_{r_1}(P)})$ onto the leaves of $(B_{r_2}, {\ensuremath{\mathcal{F}}}|_{B_{r_2}(P)})$.
This result still holds, more generally, if we replace the plaque $P$ by an open subset ${B}$ of some submanifold $N\subset M$ which is a union of leaves of the same dimension. In this case we consider some $\epsilon>0$ such that $\exp:\nu^\epsilon {B}\to M$ is a diffeomorphism onto the image $B_{\epsilon}({B})$, and define the homothetic transformation *around $W$*, $h_{\lambda}:B_{r}({B})\to B_{\lambda r}({B})$, as before. In this case, an analogous version of the Homothetic Transformation Lemma applies.
Infinitesimal foliation {#SS:inf-foliation}
-----------------------
Let $(M,{\ensuremath{\mathcal{F}}})$ be a singular Riemannian foliation, $p\in M$ a point, and $S_p$ a slice at $p$.
The *infinitesimal foliation of ${\ensuremath{\mathcal{F}}}$ at $p$*, denoted by $(\nu_pL_p,{\ensuremath{\mathcal{F}}}_p)$ is defined as the partition of $\nu_pL_p$ whose leaf at $v\in \nu_pL_p$ is given by $$L_v=\{w\in \nu_pL_p\mid \exp_ptw\in L_{\exp_ptv}\quad \forall t>0\textrm{ small enough}\},$$ where $L_{\exp_ptv}$ denotes the leaf of $(S_p,{\ensuremath{\mathcal{F}}}|_{S_p})$ through $\exp_ptv$.
The leaf $L_v$ is well defined because, by the Homothetic Transformation Lemma, if $\exp_p{t_0w}$ belongs to the same leaf of $\exp_pt_0v$ for some small $t_0$, then $\exp_ptw$ belongs to the same leaf of $\exp_ptv$ for every $t\in (0,t_0)$. In the following proposition we collect the important facts about infinitesimal foliations that we will need.
Given a singular Riemannian foliation $(M,{\ensuremath{\mathcal{F}}})$ and a point $p\in M$ with infinitesimal foliation $(\nu_pL_p,{\ensuremath{\mathcal{F}}}_p)$, then:
1. The foliation $(\nu_pL_p,{\ensuremath{\mathcal{F}}}_p)$ is a singular Riemannian foliation with respect to the flat metric ${\ensuremath{ \mathrm{g} }}_p$ at $p$.
2. The normal exponential map $\exp_p:\nu_p^{\epsilon}L_p\to M$ sends the leaves of ${\ensuremath{\mathcal{F}}}_p$ to the leaves of $(S_p,{\ensuremath{\mathcal{F}}}|_{S_p})$.
3. $(\nu_pL_p,{\ensuremath{\mathcal{F}}}_p)$ is invariant under rescalings $r_{\lambda}:\nu_pL_p\to \nu_pL_p$, $r_{\lambda}(v)=\lambda v$.
1\) [@Molino], Prop. 6.5.\
2) Follows from the definition of infinitesimal foliation, and of slice foliation.\
3) Via the exponential map $\exp:\nu^{\epsilon}L_p\to S_p$, this corresponds to the Homothetic Transformation Lemma on $S_p$.
The following fact will come very useful.
\[P:differential-foliated\] Given singular Riemannian foliations $(M,{\ensuremath{\mathcal{F}}})$, $(M',{\ensuremath{\mathcal{F}}}')$ and a foliated diffeomorphism $\phi: U\to U'$, between open sets $U,U'$ of $M,M'$ respectively, sending a point $p\in U$ to $p'\in U'$, the differential of $\phi$ induces a linear, foliated isomorphism $\phi_*:(\nu_pL_p,{\ensuremath{\mathcal{F}}}_p)\to (\nu_{p'}L_{p'},{\ensuremath{\mathcal{F}}}'_{p'})$.
By substituting $(M',{\ensuremath{ \mathrm{g} }}',{\ensuremath{\mathcal{F}}}')$ with $(M,\phi^*g',{\ensuremath{\mathcal{F}}})$, the problem can be reduced to the case where $M=M'$, $\phi=id$, $p=p'$, and ${\ensuremath{\mathcal{F}}}={\ensuremath{\mathcal{F}}}'$ is a singular Riemannian foliation with respect to two metrics, ${\ensuremath{ \mathrm{g} }}$ and $\tilde{{\ensuremath{ \mathrm{g} }}}$. In the following, we will denote with a “tilde” ($\;\tilde{ }\;$) every geometric object related to the metric $\tilde{{\ensuremath{ \mathrm{g} }}}$, and without the tilde any geometric object related to ${\ensuremath{ \mathrm{g} }}$.
Let $S_p$ (resp. ${\widetilde}{S}_p$) denote a slice at $p$ with respect to ${\ensuremath{ \mathrm{g} }}$ (resp. $\tilde{{\ensuremath{ \mathrm{g} }}}$). Consider the set $\{X_1,\ldots X_k\}\subset \mathfrak{X}(M,{\ensuremath{\mathcal{F}}})$, $k=\dim L_p$, of vector fields such that $\{X_1(p),\ldots, X_k(p)\}$ is a basis of $T_pL_p$. Denote by $\Phi_i^t$ the flow of $X_i$, and define $\Phi_{(t_1,\ldots t_k)}=\Phi_k^{t_k}\circ\ldots \circ \Phi_1^{t_1}$.
Around $p$, both $S_p$ and ${\widetilde}{S}_p$ are transverse to $\textrm{span}(X_1,\ldots X_k)$ and, up to possibly replacing $S_p$ and ${\widetilde}{S}_p$ with smaller open subsets, we can assume that for every $q\in S_p$ there exists a unique $\tilde{q}\in {\widetilde}{S}_p$ of the form $\tilde{q}=\Phi_{(t_1,\ldots t_k)}(q)$. This gives rise to a map $H:S_p\to {\widetilde}{S}_p$, $H(q)=\tilde{q}$ which is differentiable and, since $q$ and $\tilde{q}$ belong to the same leaf of ${\ensuremath{\mathcal{F}}}$, sends the leaves of ${\ensuremath{\mathcal{F}}}|_{S_p}$ to the leaves of ${\ensuremath{\mathcal{F}}}|_{{\widetilde}{S}_p}$. In other words, there is a foliated diffeomorphism $H: (S_p,{\ensuremath{\mathcal{F}}}|_{S_p})\to ({\widetilde}{S}_p,{\ensuremath{\mathcal{F}}}|_{{\widetilde}{S}_p})$.
Consider the composition $\psi$ of foliated diffeomorphisms $$(\nu_p^\epsilon L_p,{\ensuremath{\mathcal{F}}}_p)\stackrel{\exp_p}{{\longrightarrow}}(S_p,{\ensuremath{\mathcal{F}}}|_{S_p})\stackrel{H}{{\longrightarrow}}({\widetilde}{S}_p,{\ensuremath{\mathcal{F}}}|_{{\widetilde}{S}_p})\stackrel{\tilde{\exp}_p^{-1}}{{\longrightarrow}} (\tilde{\nu}_p^\epsilon L_p,{\widetilde}{{\ensuremath{\mathcal{F}}}}_p).$$ For any $\lambda\in (0,1)$, one can define a new foliated diffeomorphism $$\psi_\lambda:(\nu_p^{\epsilon/\lambda}L_p,{\ensuremath{\mathcal{F}}}_p)\to (\tilde{\nu}_p^{\epsilon/\lambda}L_p,{\widetilde}{{\ensuremath{\mathcal{F}}}}_p), \qquad \psi_\lambda(v)={1\over \lambda}\psi(\lambda v).$$ As $\lambda\to 0$, the maps $\psi_\lambda$ converge to the differential $d_0\psi$ of $\psi$ at $0$. This is an invertible linear map (in particular a diffeomorphism) and, as a limit of foliated maps, it is itself foliated. Therefore, the map $$\phi_*:=d_0\psi:(\nu_pL,{\ensuremath{\mathcal{F}}}_p){\longrightarrow}(\tilde{\nu}_pL,\widetilde{{\ensuremath{\mathcal{F}}}}_p)$$ satisfies the statement of the proposition.
\[R:essential-part\] Given a singular Riemannian foliation $(M,{\ensuremath{\mathcal{F}}})$ and a submanifold $N\subset M$ which is a union of leaves of the same dimension, the infinitesimal foliation at a point $p\in M$ splits as a product $(\nu_p(L_p,N)\times \nu_pN,\{\textrm{pts.}\}\times{\ensuremath{\mathcal{F}}}_p|_{\nu_pN})$, where $\nu_p(L_p,N)=\nu_pL_p\cap T_pN$. In this case, the foliation $(\nu_pN,{\ensuremath{\mathcal{F}}}_p|_{\nu_pN})$ is the “essential part” of the infinitesimal foliation $(\nu_pL_p,{\ensuremath{\mathcal{F}}}_p)$. By abuse of notation, we will call the foliation $(\nu_pN,{\ensuremath{\mathcal{F}}}_p|_{\nu_pN})$ *infinitesimal foliation at $p$* as well, and denote it by ${\ensuremath{\mathcal{F}}}_p$.
Given a singular Riemannian foliation $(M,{\ensuremath{\mathcal{F}}})$ and a point $p\in M$, the infinitesimal foliation $(\nu_pL_p,{\ensuremath{\mathcal{F}}}_p)$ at $p$ contains the origin as a leaf of ${\ensuremath{\mathcal{F}}}_p$. Based on this fact, we make the following definition.
An *infinitesimal foliation* is a singular Riemannian foliation $(V,{\ensuremath{\mathcal{F}}})$ on an Euclidean vector space, with the origin $\{0\}$ being a 0-dimensional leaf.
Homogeneous and orbit like foliations {#SS:orbit-like}
-------------------------------------
A singular Riemannian foliation $(M,{\ensuremath{\mathcal{F}}})$ is called *homogeneous* (sometimes *Riemannian homogeneous*) if there exists a connected Lie group $G$ acting by isometries on $M$, whose orbits are precisely the leaves of ${\ensuremath{\mathcal{F}}}$. Furthermore, a singular Riemannian foliation $(M,{\ensuremath{\mathcal{F}}})$ is called *orbit-like* if at every point $p\in M$, the infinitesimal foliation $(\nu_pL_p,{\ensuremath{\mathcal{F}}}_p)$ is closed and homogeneous.
\[E:foliation-vector-bundle\] An example of orbit like foliation, which will be useful to keep in mind later on, can be constructed as follows. Consider a Riemannian manifold $L$, and an Euclidean vector bundle $E$ over $L$, that is, a vector bundle over $L$ with an inner product $\langle\,,\,\rangle_p$ on each fiber $E_p$, $p\in L$. Let $\nabla^E$ be a metric connection on $E$, i.e. a connection on $E$ such that, for every vector field $X$ on $L$ and sections $\xi, \eta$ of $E$, one has $$X\langle \xi, \eta \rangle=\langle\nabla^E_X\xi, \eta\rangle+\langle \xi, \nabla^E_X\eta\rangle.$$ Given $(E,\nabla^E)$, there is an induced Riemannian metric ${\ensuremath{ \mathrm{g} }}^E$ on $E$, called *connection metric*. Moreover, $\nabla^E$ induces a *parallel transport* on $E$: given $\xi\in E_p$ and a curve $\gamma:[0,1]\to L$ with $\gamma(0)=p$, there exists a unique lift $\xi(t)$, $t\in [0,1]$ with $\xi(0)=\xi$ such that $\nabla^{E}_{\gamma'(t)}\xi(t)=0$ for every $t\in [0,1]$. On $E$ one can now define a foliation ${\ensuremath{\mathcal{F}}}^E$, by declaring two vectors $\xi, \eta\in E$ in the same leaf if they can be connected to one another via a composition of parallel transports. The leaves of ${\ensuremath{\mathcal{F}}}^E$ are usually referred to as the *holonomy tubes* around the zero section $L\subset E$, and they define a singular Riemannian foliation on $(E,{\ensuremath{ \mathrm{g} }}^E)$. Moreover, the infinitesimal foliation at any point of $E$ is homogeneous: in fact, for any point $p$ along the zero section $L$, one can first construct the holonomy group $H_p$ of the connection $\nabla^E$, which acts by isometries on the fiber $E_p$ and whose orbits are precisely the leaves of the infinitesimal foliation of ${\ensuremath{\mathcal{F}}}^E$ at $p$. Similarly, the infinitesimal foliation at a point $\xi\in E_p$ is given by the orbits in $\nu_\xi L_\xi$ of the stabilizer $H_\xi\subset H_p$ of $\xi$. The foliation ${\ensuremath{\mathcal{F}}}^{E}$ coincides with its own linearization with respect to the zero section (see definition in Section \[SS:lin-foliation\]). Moreover, if the leaves of ${\ensuremath{\mathcal{F}}}^E$ are closed then $(E,{\ensuremath{ \mathrm{g} }}^E, {\ensuremath{\mathcal{F}}}^E)$ is an orbit-like foliation.
\[R:eamples\] When $L\subset M$ is a submanifold of somewhat special geometry, the holonomy foliation on the normal bundle $E$ of $L$, endowed with the Levi-Civita connection, induces via the normal exponential map a foliation on the whole of $M$. For example, if $L$ has *parallel focal structure*, then the induced foliation on $M$ is a *polar foliation* [@toeben]. If $M$ is a complete, non-compact manifold with sectional curvature $\geq 0$ and $L$ is a *soul* of $M$ [@Cheeger-Gromoll], then the induced foliation on $M$ is Wilking’s *dual foliation* to the Sharafutdinov projection [@Wilking].
Although in principle the property of being orbit-like might depend on the metric, the following proposition shows in fact that being orbit like is invariant under foliated diffeomorphisms.
\[P:orbit-like-invariant\] The following hold:
1. Given a foliated linear isomorphism $\varphi:(V,{\ensuremath{\mathcal{F}}})\to(V',{\ensuremath{\mathcal{F}}}')$ between infinitesimal foliations, $(V,{\ensuremath{\mathcal{F}}})$ is homogeneous if and only if $(V',{\ensuremath{\mathcal{F}}}')$ is homogeneous.
2. Given a foliated diffeomorphism $\phi:(M,{\ensuremath{\mathcal{F}}})\to (M',{\ensuremath{\mathcal{F}}}')$ between singular Riemannian foliations, $(M,{\ensuremath{\mathcal{F}}})$ is orbit-like if and only if $(M',{\ensuremath{\mathcal{F}}}')$ is orbit-like.
1\) By the symmetric roles of $V$ and $V'$, it is enough to show that if $(V,{\ensuremath{\mathcal{F}}})$ is homogeneous, so is $(V',{\ensuremath{\mathcal{F}}}')$. Suppose that $(V,{\ensuremath{\mathcal{F}}})$ is homogeneous, and therefore the foliation ${\ensuremath{\mathcal{F}}}$ is spanned by Killing fields. Recall that a vector field $X$ on an Euclidean space $(V,{\ensuremath{ \mathrm{g} }})$ is Killing if and only if is of the form $X(v)=Av$, where $A$ is a skew symmetric endomorphism of $V$, in the sense that ${\ensuremath{ \mathrm{g} }}(Av,v)=0$ for every $v\in V$. Letting $\{X_1,\ldots X_k\}$ denote a set of Killing fields on $V$ spanning the foliation ${\ensuremath{\mathcal{F}}}$, the set $\{Y_1,\ldots Y_k\}$ with $Y_i(v')=\varphi_*\big(X_i\big(\varphi^{-1}(v')\big)\big)$ spans the foliation ${\ensuremath{\mathcal{F}}}'$ as well. Since $\varphi$ is a linear map and the vector fields $X_i$ are linear, it follows that $Y_i$ can be written as $Y_i(v')=B_iv'$ for some endomorphism $B_i:V'\to V'$, $i=1\ldots k$. Since $(V',{\ensuremath{ \mathrm{g} }}', {\ensuremath{\mathcal{F}}}')$ is a singular Riemannian foliation, the leaf $L_{v'}$ through $v'$ lies in a distance sphere from the origin, and in particular ${\ensuremath{ \mathrm{g} }}'(T_{v'}L_{v'},v')=0$. Since $Y_i(v')$ is tangent to $L_{v'}$, it follows that $$0={\ensuremath{ \mathrm{g} }}'(Y_i(v'),v')={\ensuremath{ \mathrm{g} }}(B_iv',v')$$ In other words, $B_i$ is skew-symmetric and thus $Y_i$ is a Killing field as well. Therefore the foliation $(V',{\ensuremath{\mathcal{F}}}')$ is spanned by Killing vector fields, hence it is homogeneous as well.
2\) Up to exchanging the roles of $M$ and $M'$, it is enough to show that if $(M,{\ensuremath{\mathcal{F}}})$ is orbit-like, so is $(M',{\ensuremath{\mathcal{F}}}')$. Fixing a point $p\in M$, Proposition \[P:differential-foliated\] states that the foliated diffeomorphism $\phi$ induces a foliated linear isomorphism $\phi_*:(\nu_pL_p, {\ensuremath{\mathcal{F}}}_p){\longrightarrow}(\nu_{p'}L_{p'},{\ensuremath{\mathcal{F}}}_{p'})$, where $p'=\phi(p)$. Since $(M,{\ensuremath{\mathcal{F}}})$ is orbit-like, it follows that $(\nu_pL_p, {\ensuremath{\mathcal{F}}}_p)$ is closed and homogeneous. From the first point above it follows that $(\nu_{p'}L_{p'},{\ensuremath{\mathcal{F}}}_{p'})$ is homogeneous as well, and by the continuity of $\phi_*$ one has that $(\nu_{p'}L_{p'},{\ensuremath{\mathcal{F}}}_{p'})$ is closed. Since $p'$ was chosen arbitrarily, it follows that $(M',{\ensuremath{\mathcal{F}}}')$ is orbit-like.
Linearization, and linearized foliation {#SS:lin-foliation}
---------------------------------------
Let $(M,{\ensuremath{\mathcal{F}}})$ be a singular Riemannian foliation, $B\subset M$ a submanifold saturated by leaves, and $U\in M$ an $\epsilon$-tubular neighbourhood of $B$ with metric projection ${\ensuremath{\mathsf{p}}}:U\to B$. Given a vector field $V$ in $U$ tangent to the leaves of ${\ensuremath{\mathcal{F}}}$, it is possible to produce a new vector field $V^\ell$, called the *linearization of $V$ with respect to $B$*, as follows: $$V^\ell=\lim_{\lambda\to 0} (h_{\lambda}^{-1})_*(V|_{h_{\lambda}(U)})$$ where $h_\lambda:U\to U$ denotes the homothetic transformation around $B$. From [@Mendes-Radeschi], Prop. 5, the linearization $V^\ell$ is a smooth vector field invariant under the homothetic transformation $h_\lambda$, and it coincides with $V$ along $B$. On $U$, consider the module $\mathfrak{X}(U,{\ensuremath{\mathcal{F}}})^\ell$ given by the linearization, with respect to $B$, of the vector fields in $\mathfrak{X}(U,{\ensuremath{\mathcal{F}}})$: $$\mathfrak{X}(U,{\ensuremath{\mathcal{F}}})^\ell=\{V^\ell\mid V\in \mathfrak{X}(U,{\ensuremath{\mathcal{F}}})\}.$$ Let $\mathsf{D}$ the pseudogroup of local diffeomorphisms of $U$, generated by the flows of linearized vector fields, and let $(U,{\ensuremath{\mathcal{F}}}^\ell)$ the partition of $U$ into the orbits of diffeomorphisms in $\mathsf{D}$. By Sussmann [@Sussmann Thm. 4.1], such orbits are (possibly non-complete) smooth submanifolds of $M$. Moreover, as noted By Molino [@Molino Lem. 6.3], this foliation is spanned, at each point, by the vector fields in $\mathfrak{X}^\ell(U,{\ensuremath{\mathcal{F}}})$.
We call $(U, {\ensuremath{\mathcal{F}}}^\ell)$ the *linearized foliation of ${\ensuremath{\mathcal{F}}}$ with respect to $B$*. We will show, later, that the leaves of the linearized foliation are actually complete, and have a particularly nice local structure (cf. Section \[S:linearized\]).
Given a point $p\in B$, define $U_p={\ensuremath{\mathsf{p}}}^{-1}(p)\subset U$ and let ${\ensuremath{\mathcal{F}}}_p$ (resp. $({\ensuremath{\mathcal{F}}}^\ell)_p$) denote the partition of $U_p$ into the connected components of $L\cap U_p$, as $L$ ranges through the leaves of ${\ensuremath{\mathcal{F}}}$ (resp. ${\ensuremath{\mathcal{F}}}^\ell$). If $U_p$ is given the flat metric ${\ensuremath{ \mathrm{g} }}_p$ of $\nu_pB$ via the exponential map $\exp_p:\nu_p^{\epsilon}B\to U_p$, then ${\ensuremath{\mathcal{F}}}_p$ corresponds to the infinitesimal foliation at $p$ (cf. Remark \[R:essential-part\]) which justifies the notation of ${\ensuremath{\mathcal{F}}}_p$ for this foliation. Furthermore, as noted in [@Molino Sec. 6.4], $({\ensuremath{\mathcal{F}}}^\ell)_p$ is given by the linearization of $(U_p,{\ensuremath{ \mathrm{g} }}_p,{\ensuremath{\mathcal{F}}}_p)$ with respect to the origin. In other words, $({\ensuremath{\mathcal{F}}}^\ell)_p=({\ensuremath{\mathcal{F}}}_p)^\ell$ and it makes sense to denote this foliation simply by ${\ensuremath{\mathcal{F}}}_p^\ell$. Moreover, letting ${\ensuremath{\mathsf{O}}}({\ensuremath{\mathcal{F}}}_p)$ denote the Lie group of (linear) isometries of $(U_p,{\ensuremath{ \mathrm{g} }}_p)$ sending every leaf to itself, one has:
\[P:linearized-infinitesimal\] The foliation $(U_p,{\ensuremath{\mathcal{F}}}^\ell_p)$ is homogeneous, given by the orbits of the identity component $H_p$ of ${\ensuremath{\mathsf{O}}}({\ensuremath{\mathcal{F}}}_p)$.
We identify here $U_p$ with a neighbourhood of the origin in $\nu_pB$ via the exponential map, and we think of ${\ensuremath{\mathcal{F}}}_p^\ell$ as the linearization of ${\ensuremath{\mathcal{F}}}_p$.
Given a vector field $V\in \mathfrak{X}(U_p,{\ensuremath{\mathcal{F}}}_p)$, its linearization $V^\ell$ is linear, in the sense that $V^\ell_p=A\cdot p$ for some $A\in \textrm{End}(U_p)$. Since ${\ensuremath{\mathcal{F}}}_p$ is a singular Riemannian foliation, the leaves are tangent to the distance spheres around the origin and therefore perpendicular to the radial directions from the origin: $\langle V^\ell_p,p\rangle=0$. In other words, $V^\ell_p=A\cdot p$ with $A$ skew symmetric, which implies that the flow of $V^\ell$ is an isometry of $U_p$. Moreover, since $V^\ell$ is everywhere tangent to the leaves of ${\ensuremath{\mathcal{F}}}_p^\ell$, the flow of $V^\ell$ is a 1-parameter group in $H_p$, moving every leaf of $({\ensuremath{\mathcal{F}}}_p)^\ell$ to itself. In particular, the orbits of $H_p$ are contained in the leaves of $({\ensuremath{\mathcal{F}}}_p)^\ell$.
However, by definition of $H_p$, the tangent space of a $H_p$-orbit through a point $q\in U_p$ is given by $$T_q(H_p\cdot q)=\{W_q\mid W\textrm{ Killing vector field tangent to the leaves of }{\ensuremath{\mathcal{F}}}_p\}$$ and such vector fields coincide precisely with the vector fields in $\mathfrak{X}(U_p,{\ensuremath{\mathcal{F}}}_p)^\ell$. Therefore, $H_p\cdot q$ is the integral manifold of $\mathfrak{X}(U_p,{\ensuremath{\mathcal{F}}}_p)^\ell$ through $q$.
Molino’s conjecture, assuming the Main Theorem {#S:Molino}
==============================================
Before proving the Main Theorem, we show how Molino’s Conjecture follows from it as a corollary.
Let $(M,{\ensuremath{\mathcal{F}}})$ be a singular Riemannian foliation, and let ${\overline}{{\ensuremath{\mathcal{F}}}}$ denote the closure of ${\ensuremath{\mathcal{F}}}$. Molino himself proved that ${\overline}{{\ensuremath{\mathcal{F}}}}$ is a partition into complete smooth closed submanifolds, and that ${\overline}{{\ensuremath{\mathcal{F}}}}$ is a transnormal system. Therefore, in order to prove the conjecture, it is enough to show that for any leaf $L\in {\ensuremath{\mathcal{F}}}$ with closure ${\overline}{L}$ and any vector $v\in \nu(L,{\overline}{L}):=\nu L\cap T{\overline}{L}$, there exists a smooth extension of $v$ to a vector field $V$ everywhere tangent to the leaves of ${\overline}{{\ensuremath{\mathcal{F}}}}$.
Let $U$ be a tubular neighbourhood of ${\overline}{L}$, and let $(U, \widehat{{\ensuremath{\mathcal{F}}}}^\ell)$ be the foliation satisfying the Main Theorem. Since $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ coincides with ${\ensuremath{\mathcal{F}}}$ along ${\overline}{L}$, it follows that $L$ is a leaf of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ as well. Since $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ is an orbit-like foliation, by Theorem 1.6 of [@Alexandrino-Radeschi-smooth-flows], given $v\in \nu(L,{\overline}{L})$ there is a vector field $V$ extending $v$ which is tangent to the closure of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$. Since this closure is contained in ${\overline}{{\ensuremath{\mathcal{F}}}}$, it follows that $V$ is also tangent to ${\overline}{{\ensuremath{\mathcal{F}}}}$ and this ends the proof of the conjecture.
The setup {#S:setup}
=========
Fix a leaf $L$, and distance tube $U=B_{\epsilon}(L)$ around ${\overline}{L}$. Using the normal exponential map $\exp:\nu{\overline}{L}\to M$, $U$ can be identified with the $\epsilon$-tube $\nu^\epsilon {\overline}{L}$ around the zero section. By the Homothetic Transformation Lemma, the pull-back foliation $\exp^{-1}{\ensuremath{\mathcal{F}}}$ on $\nu^{\epsilon}{\overline}{L}$ is invariant under the rescalings $r_{\lambda}:\nu^{\epsilon}{\overline}{L}\to \nu^{\epsilon}{\overline}{L}$, $r_\lambda(p,v)=(p,\lambda v)$ for any $\lambda\in (0,1)$.
For this reason, in the following sections we will be considering the (slightly more general) setup:
- $U$ is the $\epsilon$-tube around the zero section of some Euclidean vector bundle $E\to {B}$ (in our case ${B}={\overline}{L}$), with projection ${\ensuremath{\mathsf{p}}}: U\to {B}$.
- ${\ensuremath{ \mathrm{g} }}$ is a Riemannian metric on $U$ with the same radial function as the Euclidean metric on each fiber of $E$.
- $(U,{\ensuremath{ \mathrm{g} }},{\ensuremath{\mathcal{F}}})$ is a singular Riemannian foliation on $U$, invariant under rescalings $r_\lambda$. In particular, the zero-section ${B}$ is saturated by leaves and the projection ${\ensuremath{\mathsf{p}}}$ sends leaves onto leaves.
- The restriction ${\ensuremath{\mathcal{F}}}_{B}={\ensuremath{\mathcal{F}}}|_{B}$ is a regular Riemannian foliation.
- For every leaf $L{\subseteq}{B}$ and any point $p\in L$, the normal exponential map $\nu_p^{\epsilon}L\to U$ is an embedding.
Three distributions {#S:distributions}
===================
Let $(U,{\ensuremath{ \mathrm{g} }},{\ensuremath{\mathcal{F}}})$, ${\ensuremath{\mathsf{p}}}:U\to {B}$ be as in Section \[S:setup\]. In order to prove the Main Theorem, it is first needed to produce a nicer metric on $U$, and for this we first need to split the tangent space of $U$ into three components. The first, $\mathcal{K}=\ker{\ensuremath{\mathsf{p}}}_*$, is the distribution tangent to the fibers of ${\ensuremath{\mathsf{p}}}$. For the remaining two notice that, since the foliation $({B},{\ensuremath{\mathcal{F}}}_{B})$ is regular, the tangent bundle $T{B}$ splits into a tangent and a normal part to the foliation: $T{B}=T{\ensuremath{\mathcal{F}}}|_{{B}}\oplus\nu{\ensuremath{\mathcal{F}}}|_{{B}}$. The last two distributions will be constructed as (appropriately chosen) extensions $\mathcal{T}$ and $\mathcal{N}$ of $T{\ensuremath{\mathcal{F}}}|_{{B}}$ and $\nu{\ensuremath{\mathcal{F}}}|_{{B}}$ respectively, to the whole of $U$.
The distribution $\mathcal{T}$
------------------------------
From [@Alexandrino-desingularization] there exists a distribution $\widehat{\mathcal{T}}$ of rank $\dim {\ensuremath{\mathcal{F}}}|_{B}$, which extends $T{\ensuremath{\mathcal{F}}}|_{{B}}$ and is everywhere tangent to the leaves of ${\ensuremath{\mathcal{F}}}$. The distribution $\mathcal{T}$ is simply defined as the *linearization of $\widehat{\mathcal{T}}$ with respect to ${B}$*, as follows: consider a family of vector fields $\{V_\alpha\}_{\alpha}$ spanning $\widehat{\mathcal{T}}$. Since $\widehat{\mathcal{T}}|_{{B}}$ is tangent to ${B}$, the vector fields $V_\alpha$ lie tangent to ${B}$ as well and therefore it makes sense to consider their linearization $V_\alpha^\ell$ with respect to ${B}$. By the properties of the linearization, these linearized vector fields still span a smooth distribution of the same rank as $\widehat{\mathcal{T}}$, which we call $\mathcal{T}$.
The distribution $\mathcal{N}$
------------------------------
At each point $q\in U$ with ${\ensuremath{\mathsf{p}}}(q)=p$, the slice $S_p=\exp_p(\nu_p^{\epsilon}L_p)$ contains $q$ as well as the whole ${\ensuremath{\mathsf{p}}}$-fiber $U_p$ through $p$. In particular, $\mathcal{K}_q$ lies tangent to $S_p$. Moreover, $S_p$ comes equipped with a flat metric ${\ensuremath{ \mathrm{g} }}_p$, inherited from the metric on $\nu_pL$ via the diffeomorphism $\exp_p:\nu^\epsilon L_p\to S_p$.
Define $\widehat{\mathcal{N}}_q$ as the subspace of $T_qS_p$ which is ${\ensuremath{ \mathrm{g} }}_p$-orthogonal to $\mathcal{K}_q$. Finally, define $\mathcal{N}$ as the *linearization* of $\widehat{\mathcal{N}}$, as defined in the previous section. The distributions $\widehat{\mathcal N}$ and $\mathcal N$ satisfy the following property:
\[P:extension\] For every smooth ${\ensuremath{\mathcal{F}}}_{B}$-basic vector field $X_0$ along a plaque $P$ in ${B}$ there exists a smooth extension $X$ to an open set of $U$ such that
1. $X$ is foliated and tangent to $\widehat{\mathcal N}$.
2. The linearization $X^{\ell}$ of $X$ with respect to ${B}$ is tangent to $\mathcal{N}$, and it is foliated with respect to both ${\ensuremath{\mathcal{F}}}$ and ${\ensuremath{\mathcal{F}}}^\ell$.
1\) Fix a leaf $L$ in ${B}$, a plaque $P\subset L$ and a parametrization $$\varphi:(-1,1)^k\to P\subset L,$$ where $k=\dim {\ensuremath{\mathcal{F}}}_B$. We first show that there exists a small neighbourhood of $P$ in $U$, on which any ${\ensuremath{\mathcal{F}}}_B$-basic vector field $X_0$ along $P$ can be extended to a foliated vector field $X'$, *whose restriction to ${\ensuremath{\mathsf{p}}}^{-1}(P)$ is tangent to $\widehat{\mathcal{N}}$*.
Let $\partial_{y_1},\ldots, \partial_{y_k}$ be coordinate vector fields on $P$, and let $Y_1, \ldots Y_k$ denote vector fields, linearized *with respect to $L$*, that extend $\partial_{y_1},\ldots \partial_{y_k}$ to a neighbourhood of $P$ in $U$. There is a foliated diffeomorphism $$\begin{aligned}
F: (P\times S_p, P\times {\ensuremath{\mathcal{F}}}_{S_p})&{\longrightarrow}(U,{\ensuremath{\mathcal{F}}})\\
(\varphi(y_1,\ldots y_k),q)&\longmapsto \Phi_k^{y_k}\circ\ldots\circ \Phi_1^{y_1}(q)\end{aligned}$$ where $\Phi_i^{y_i}$ is the flow of $Y_i$, after time $y_i$.
Furthermore, the foliation $(P\times S_p,P\times{\ensuremath{\mathcal{F}}}_{S_p})$ locally splits as $$(P\times \nu_p(L,{B})\times \nu_p{B},P\times \{\textrm{pts.}\}\times {\ensuremath{\mathcal{F}}}|_{\nu_p{B}})$$ where $\nu(L,{B})=\nu L \cap T{B}$. Moreover, if $S_p$ is endowed with the Euclidean metric ${\ensuremath{ \mathrm{g} }}_p$ on $\nu_pL$, the splitting $S_p=\nu_p(L,{B})\times \nu_p{B}$ is in fact Riemannian.
The map $F$ satisfies the following:
- The set $P\times \{0\}\times \nu_p^\epsilon{B}$ is sent to ${\ensuremath{\mathsf{p}}}^{-1}(P)=\nu^\epsilon{B}|_P$.
- The set $P\times \nu_p^{\epsilon}(L,{B})\times \{0\}$ is sent to a neighbourhood of $P$ in $B$.
- Since $F$ is defined via linearized vector fields, each fiber $\{p'\}\times S_p\subset P\times S_p$ is sent, via $F$, to the slice $S_{p'}$, isometrically with respect to the flat metrics on $S_p$ and $S_{p'}$ (cf. [@Mendes-Radeschi]).
From the last point, it follows that the distribution of $P\times \nu_p(L,{B})\times \nu_p{B}$ tangent to the second factor is sent, along $\nu^\epsilon{B}|_{P}$, precisely to the distribution $\widehat{\mathcal{N}}$.
Any ${\ensuremath{\mathcal{F}}}_B$-basic vector field $X_0$ along $P$ corresponds, via $F$, to a vector field along $P\times \{0\}\times \{0\}$ of the form $(0,x_0,0)$ where $x_0\in \nu_p(L,{B})$ is a fixed vector. One can clearly extend such a vector field to the foliated vector field $X'=F_*(0,x_0,0)$. Since $F$ is a foliated map, the vector field $X'$ is a foliated vector field, whose restriction to $B$ is tangent to $B$ by the second point above. Moreover, by the discussion above the restriction of $X'$ to ${\ensuremath{\mathsf{p}}}^{-1}(P)$ is tangent to $\widehat{\mathcal{N}}$.
This proves the first claim, made at the beginning of the proof. In particular, since the plaque $P$ was chosen arbitrarily, this shows that the distribution $\widehat{\mathcal{N}}$ is *foliated*: that is, given a vector $y$ tangent to $\widehat{\mathcal{N}}$ at a point $q$, there exists a foliated extension $Y$ along a plaque containing $q$ which is everywhere tangent to $\widehat{\mathcal{N}}$. It is easy to see that $\mathcal{K}$ and ${\mathcal{T}}$ are foliated as well. In particular, given the foliated vector field $X'$, the (unique) decomposition $$X'=X'_{_\mathcal{K}}+X'_{_\mathcal{T}}+X'_{_{\widehat{\mathcal{N}}}},\qquad X'_{_\mathcal{K}}\in \mathcal{K}, \quad X'_{_\mathcal{T}}\in \mathcal{T},\quad X'_{_{\widehat{\mathcal{N}}}}\in\widehat{\mathcal{N}}$$ produces three vector fields $X'_{_\mathcal{K}}, \, X'_{_\mathcal{T}},\,X'_{_{\widehat{\mathcal{N}}}}$ which are foliated. In particular, the vector field $X=(X')_{\widehat{\mathcal{N}}}$ is foliated, everywhere tangent to $\widehat{\mathcal{N}}$, and it extends $X_0=(X_0)_{_{\widehat{\mathcal{N}}}}$ to an open set of $U$, as we needed to show.
2\) Since $X$ is tangent to $\widehat{\mathcal{N}}$, its linearization $X^\ell$ is tangent to the linearization of $\widehat{\mathcal{N}}$, which is $\mathcal{N}$. Moreover, since $X$ is foliated and $r_\lambda:U\to U$ is a foliated map, $X^\ell=\lim_{\lambda\to 0}(r_\lambda^{-1})_*X\circ r_\lambda$ is foliated as well. Finally, since $X$ is foliated, for every vector field $V$ tangent to ${\ensuremath{\mathcal{F}}}$ one has that $[X,V]$ is also tangent to ${\ensuremath{\mathcal{F}}}$. Since $r_\lambda$ is a diffeomorphism, one computes $$\begin{aligned}
[X^\ell,V^\ell] &=\lim_{\lambda\to 0}[(r_\lambda^{-1})_*X\circ r_\lambda,(r_\lambda^{-1})_*V\circ r_\lambda]\\
&=\lim_{\lambda\to 0}(r_\lambda^{-1})_*[X,V]\circ r_\lambda\\
&=[X,V]^\ell.\end{aligned}$$ Since the linearization $V^\ell$ are precisely the vector fields generating ${\ensuremath{\mathcal{F}}}^\ell$, it follows from the equation above that $[X^\ell,V^\ell]$ is tangent to ${\ensuremath{\mathcal{F}}}^\ell$ whenever $V^\ell$ is, and therefore $X^\ell$ is foliated with respect to ${\ensuremath{\mathcal{F}}}^\ell$.
Structure of ${\ensuremath{\mathcal{F}}}^\ell$, and the local closure $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ {#S:linearized}
=================================================================================================================
Using the extensions $X^\ell$ defined in Proposition \[P:extension\], one can prove the following:
\[P:local-form\] Around any point $p\in {B}$ there is a neighbourhood $W$ of $p$ in $B$ such that $({\ensuremath{\mathsf{p}}}^{-1}(W),{{\ensuremath{\mathcal{F}}}}^\ell|_{{\ensuremath{\mathsf{p}}}^{-1}(W)})$ is foliated diffeomorphic to a product $$({\mathbb{D}}^k\times {\mathbb{D}}^{m-k}\times U_p, {\mathbb{D}}^k\times\{\textrm{pts.}\}\times {{\ensuremath{\mathcal{F}}}}^\ell_p)$$ where $k=\dim {\ensuremath{\mathcal{F}}}|_{B}$ and $m=\dim {B}$.
Let $W$ be a coordinate neighbourhood of ${B}$ around $p$, with a foliated diffeomorphism $\varphi:(W,{\ensuremath{\mathcal{F}}}|_W)\to ({\mathbb{D}}^k\times {\mathbb{D}}^{m-k},{\mathbb{D}}^k\times\{pts.\})$. Let ${\partial \over \partial y_1}, \ldots {\partial \over \partial y_k}$ denote a basis of vector fields in $W$ tangent to the leaves of ${\ensuremath{\mathcal{F}}}|_W$, and let $V_1,\ldots V_k$ denote vector fields on ${\ensuremath{\mathsf{p}}}^{-1}(W)$, linearized with respect to ${B}$, extending ${\partial \over \partial y_i}$, $i=1,\ldots k$ and spanning the foliation $\mathcal{T}$. Similarly, let ${\partial \over \partial x_1},\ldots {\partial \over \partial x_{m-k}}$ denote a basis of basic vector fields in $W$ normal to the leaves, and let $X^\ell_1,\ldots X^\ell_{m-k}$ denote linearized vector fields in $\pi^{-1}(W)$ defined as in Proposition \[P:extension\], extending the vectors ${\partial \over \partial x_i}$, $i=1,\ldots m-k$. Finally, define $\Phi_i^t$ and $\Psi_i^t$ the flows of $V_i$ and $X^\ell_i$ respectively, after time $t$, and let $$\begin{aligned}
G:{\mathbb{D}}^k\times {\mathbb{D}}^{m-k}\times U_p&\longrightarrow {\ensuremath{\mathsf{p}}}^{-1}(W)\\
((t_1,\ldots,t_k),(s_1,\ldots, s_{m-k}),q)&\longmapsto \Phi_k^{t_k}\circ\ldots\circ \Phi_1^{t_1}\circ \Psi_{m-k}^{s_{m-k}}\circ\ldots \circ \Psi_1^{s_1}(q).\end{aligned}$$ Since the $V_i$ and $X^\ell_i$ are linearized, they take fibers of ${\mathbb{D}}^k\times {\mathbb{D}}^{m-k}\times U_p\to {\mathbb{D}}^k\times {\mathbb{D}}^{m-k}$ to fibers of ${\ensuremath{\mathsf{p}}}:{\ensuremath{\mathsf{p}}}^{-1}(W)\to W$. Since the flows $\Psi_i$ send the leaves of ${\ensuremath{\mathcal{F}}}^\ell$ to leaves, and the flows $\Phi_i$ take leaves of ${\ensuremath{\mathcal{F}}}^\ell$ to themselves, the leaves of ${\mathbb{D}}^k\times ({\mathbb{D}}^{m-k},\{\textrm{pts.}\})\times (U_p, {\ensuremath{\mathcal{F}}}^\ell_p)$ are sent into the leaves of ${\ensuremath{\mathcal{F}}}^\ell$. Since the differential $dG$ is invertible at $(0,0,p)\in{\mathbb{D}}^k\times {\mathbb{D}}^{m-k}\times U_p$, it is a diffeomorphism around $G(0,0,p)=p$ and, by dimensional reasons, the leaves of $({\mathbb{D}}^k\times {\mathbb{D}}^{m-k}\times U_p, {\mathbb{D}}^k\times\{\textrm{pts.}\}\times {\ensuremath{\mathcal{F}}}^\ell_p)$ are mapped diffeomorphically onto the leaves of $({\ensuremath{\mathsf{p}}}^{-1}(W),{\ensuremath{\mathcal{F}}}^\ell|_{{\ensuremath{\mathsf{p}}}^{-1}(W)})$.
The local closure of ${\ensuremath{\mathcal{F}}}^\ell$ {#the-local-closure-of-ensuremathmathcalfell .unnumbered}
------------------------------------------------------
Even though $(U_p,{\ensuremath{\mathcal{F}}}^\ell_p)$ is homogeneous for every $p\in {B}$, it might be the case that its leaves are not closed, which happens when the group $H_p{\subseteq}{\ensuremath{\mathsf{O}}}(U_p)$ defined in Proposition \[P:linearized-infinitesimal\] is not closed. To obviate this problem we define a new foliation $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$, called the *local closure* of ${\ensuremath{\mathcal{F}}}^\ell$, such that ${\ensuremath{\mathcal{F}}}^\ell\subset \widehat{{\ensuremath{\mathcal{F}}}}^\ell\subset {\overline}{{\ensuremath{\mathcal{F}}}^\ell}$ and whose restriction $\widehat{{\ensuremath{\mathcal{F}}}}^\ell_p$ to each ${\ensuremath{\mathsf{p}}}$-fiber $U_p$ is homogeneous and closed.
Recall that ${\ensuremath{\mathcal{F}}}^\ell$ is defined by the orbits of the pseudogroup $\mathsf{D}$ of local diffeomorphisms, generated by the flows of linearized vector fields. For each $q\in U_p$, consider the closure ${\overline}{H}_p$ of $H_p$ in ${\ensuremath{\mathsf{O}}}(U_p)$, and define the $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$-leaf $\widehat{L}_q$ through $q$ to be the $\mathsf{D}$-orbit of ${\overline}{H}_p\cdot q$: $$\widehat{L}_q=\{q'=\Phi(h\cdot q)\mid \Phi\in \mathsf{D},\, h\in {\overline}{H}_p\}$$ Let $\sim$ denote the relation $q\sim q'$ if and only if $q'=\Phi(h\cdot q)$ for some $\Phi\in \mathsf{D}$ and $h\in {\overline}{H}_p$. In this way, the leaf of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ through $q$ can be rewritten as $\{q'\in U\mid q'\sim q\}$. As for the other foliations, for every $p\in {B}$ we define $(U_p,\widehat{{\ensuremath{\mathcal{F}}}}^\ell_p)$ to be the partition of $U_p$ into the connected components, of the intersections of $U_p$ with the leaves in $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$.
\[P:prop-loc-closure\] The following hold:
1. $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ is a well defined partition of $U$.
2. For every $p\in {B}$ the leaves of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell_p$ are the orbits of ${\overline}{H}_p$ on $U_p$.
1\. One must prove that the relation $\sim$ defined above is an equivalence relation. For this, notice that, since any $\Phi\in \mathsf{D}$ defines a foliated isometry between $(U_p,{\ensuremath{\mathcal{F}}}^\ell_p)$ and $(U_{\Phi(p)},{\ensuremath{\mathcal{F}}}^\ell_{\Phi(p)})$ for any $p\in {B}$, in particular it defines a foliated isometry between the respective closures $(U_p,{\overline}{H}_p)$ and $(U_{\Phi(p)},{\overline}{H}_{\Phi(p)})$. In particular, for any $h\in {\overline}{H}_p$ and $\Phi\in \mathsf{D}$, one has $h'=\Phi\circ h\circ \Phi^{-1}\in {\overline}{H}_{\Phi(p)}$.
- Reflexivity of $\sim$: if $q'\sim q$ then $q'=\Phi(h(q))$ for some $h\in {\overline}{H}_p$ and $\Phi\in \mathsf{D}$. Then $q'=h'(\Phi(q))$, where $h'=\Phi\circ h\circ \Phi^{-1}\in {\overline}{H}_{\Phi(p)}$, and therefore $q=\Phi^{-1}((h')^{-1}q')$, that means $q\sim q'$.
- Transitivity of $\sim$: if $q'\sim q$ and $q''\sim q'$ then $q'=\Phi(h(q))$ and $q''=\Psi(g(q'))$ for some $h\in {\overline}{H}_p$, $g\in {\overline}{H}_{\Phi(p)}$, and $\Phi,\Psi\in \mathsf{D}$. Then $q''=(\Psi\circ \Phi)((g' \circ h)(q))$, where $g'=\Phi^{-1}\circ g\circ \Phi\in {\overline}{H}_p$, and therefore $q''\sim q$.\
2. Let $L'$ denote a leaf of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$. From (1), the intersection of $L'$ with $U_p$ is a union of orbits of ${\overline}{H}_p$. On the other hand, we claim that the intersection $L'\cap U_p$ consists of countably many orbits of ${\overline}{H}_p$, so that each connected component of such intersection must consists of a single ${\overline}{H}_p$-orbit. From the definition of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$, it is enough to prove that the subgroup $\mathsf{D}_p\subset \mathsf{D}$ of diffeomorphisms fixing $p$ moves every ${\overline}{H}_p$-orbit in $L'\cap U_p$ to at most countably many orbits. For this, consider a piecewise smooth loop $\gamma:[0,1]\to L_p$ with $\gamma(0)=\gamma(1)=p$. Using linearized vector fields with $\gamma$ as integral curve, one can construct a continuous path $\Phi_t:[0,1]\to \mathsf{D}$ of diffeomorphisms such that $\Phi_0=id_U$ and $\Phi_t(p)=\gamma(t)$, as described in [@Mendes-Radeschi Cor. 7]. Fixing some ${\overline}{H}_p$-orbit $\mathcal{O}$ in $L'\cap U_p$, its image $\Phi_1(\mathcal{O})$ is again some ${\overline}{H}_p$-orbit, which only depends on the class $[\gamma]\in \pi_1(L_p,p)$ and not on the actual path $\gamma$, nor on the specific choice of $\Phi_t$. This gives a map $$\partial: \pi_1(L_p,p)\to \{{\overline}{H}_p\textrm{-orbits in }L'\cap U_p\}$$ This map admits a section, namely: for every orbit $\mathcal{O}'$ in $L'\cap U_p$, take a path $\gamma$ in $L'$ from a point in a (fixed) orbit $\mathcal{O}$ to a point in $\mathcal{O}'$. Under the projection ${\ensuremath{\mathsf{p}}}:U\to B$, the composition ${\ensuremath{\mathsf{p}}}\circ\gamma$ is a loop in $L_p$. The section of $\partial$ sends $\mathcal{O}'$ to $[{\ensuremath{\mathsf{p}}}\circ\gamma]\in \pi_1(L_p,p)$. In particular, the map $\partial$ is surjective, and therefore the set of ${\overline}{H}_p$-orbits in $L'\cap U_p$ has at most the cardinality of $\pi_1(L_p,p)$, which is at most countable since $L_p$ is a manifold.
As a corollary of Propositions \[P:prop-loc-closure\] and \[P:local-form\], one gets the following:
\[C:local-form\] Let $(U,{\ensuremath{\mathcal{F}}})$ be a singular Riemannian foliation as in Section \[S:setup\], let ${\ensuremath{\mathcal{F}}}^\ell$ be its linearized foliation and $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ the local closure. Then $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ is a singular foliation with complete leaves. Moreover, around each point $p\in B$ there is a neighbourhood $W$ of $p$ in $B$ such that $({\ensuremath{\mathsf{p}}}^{-1}(W),\widehat{{\ensuremath{\mathcal{F}}}}^\ell|_{{\ensuremath{\mathsf{p}}}^{-1}(W)})$ is foliated diffeomorphic to a product $$({\mathbb{D}}^k\times {\mathbb{D}}^{m-k}\times U_p,{\mathbb{D}}^k\times\{\textrm{pts.}\}\times \{\textrm{orbits of }{\overline}{H}_p{\subseteq}{\ensuremath{\mathsf{O}}}(U_p)\}),$$ which can be given the structure of a singular Riemannian foliation.
Once it is shown that $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ is also a transnormal system with respect to some metric, then by the corollary above it is globally a singular Riemannian foliation.
A new metric {#S:proof}
============
Let $\mathcal{T},\,\mathcal{N},\,\mathcal{K}$ be the distributions as in the previous section. Clearly, one has $TU=\mathcal{T}\oplus\mathcal{N}\oplus\mathcal{K}$. Define now the new metric ${\ensuremath{ \mathrm{g} }}^\ell$ on $U$, as the metric defined by the following properties:
- $\mathcal{T}\oplus\mathcal{N}$ and $\mathcal{K}$ are orthogonal with respect to ${\ensuremath{ \mathrm{g} }}^\ell$.
- ${\ensuremath{ \mathrm{g} }}^\ell|_{\mathcal{T}\oplus \mathcal{N}}={\ensuremath{\mathsf{p}}}^*{\ensuremath{ \mathrm{g} }}_{B}$, where ${\ensuremath{ \mathrm{g} }}_{B}$ denotes the restriction of the original metric on ${B}$. In particular, $\mathcal{T}$ and $\mathcal{N}$ are also orthogonal to one another.
- For any $q\in U_p$, recall that $\mathcal{K}_q=T_qU_p$, and define ${\ensuremath{ \mathrm{g} }}^\ell|_{\mathcal{K}_q}={\ensuremath{ \mathrm{g} }}_p$ the flat metric on $U_p$ induced from $\exp_p:\nu_p{B}\to U_p$.
These conditions characterize the metric ${\ensuremath{ \mathrm{g} }}^\ell$ uniquely. The most useful property of this metric is the following.
\[P:new-metric\] The triples $(U,{\ensuremath{ \mathrm{g} }}^\ell,{\ensuremath{\mathcal{F}}}^\ell)$ and $(U,{\ensuremath{ \mathrm{g} }}^\ell,\widehat{{\ensuremath{\mathcal{F}}}}^\ell)$ are singular Riemannian foliations.
The arguments for ${\ensuremath{\mathcal{F}}}^\ell$ and $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ are ideantical, therefore we will only check the Proposition for $(U,{\ensuremath{ \mathrm{g} }}^\ell,\widehat{{\ensuremath{\mathcal{F}}}}^\ell)$ (which is the only case we need for the Main Theorem anyway). Moreover, the statement is local in nature, therefore it is enough to prove the statement on certain open sets covering the whole of $U$. For any point $p\in {B}$, let $W$ denote a neighbourhood of $p$ in ${B}$ and ${\ensuremath{\mathsf{p}}}^{-1}(W)$ a neighbourhood of $p$ in $U$. We need to check that $({\ensuremath{\mathsf{p}}}^{-1}(W),{\ensuremath{ \mathrm{g} }}^\ell,\widehat{{\ensuremath{\mathcal{F}}}}^\ell)$ is a singular Riemannian foliation. To prove this, we apply Proposition 2.14 of [@Alexandrino-desingularization] which states that it is enough to check two conditions:
1. $({\ensuremath{\mathsf{p}}}^{-1}(W),g',\widehat{{\ensuremath{\mathcal{F}}}}^\ell)$ is a singular Riemannian foliation with respect to *some* Riemannian metric $g'$.
2. For every stratum $\Sigma\subset {\ensuremath{\mathsf{p}}}^{-1}(W)$ (i.e. union of leaves of the same dimension), the restriction of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ to $\Sigma$ is a (regular) Riemannian foliation.
The first condition is satisfied by Corollary \[C:local-form\]. The second condition is equivalent to checking that, for every leaf $\widehat{L}_q$ of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell|_{{\ensuremath{\mathsf{p}}}^{-1}(W)}$ and every basic vector field $X$ along $\widehat{L}_q$ tangent to the stratum through $\widehat{L}_q$, the norm $\|X\|_{{\ensuremath{ \mathrm{g} }}^\ell}$ is constant along $\widehat{L}_q$.
By definition of the metric ${\ensuremath{ \mathrm{g} }}^\ell$, the space $\nu \widehat{L}_q$ is given by $\mathcal{N}|_{\widehat{L}_q}\oplus (\nu \widehat{L}_q \cap \mathcal{K})$. From Proposition \[P:extension\], along $\widehat{L}_q$ the space $\mathcal{N}$ is spanned by linearized vector fields $X^\ell_i$, which are then ${\ensuremath{ \mathrm{g} }}^\ell$-basic (i.e., foliated *and* ${\ensuremath{ \mathrm{g} }}^\ell$-orthogonal to the leaves). In particular, any basic vector field ${\overline}{X}$ along $\widehat{L}_q$ splits as a sum ${\overline}{X}={\overline}{X}_1+{\overline}{X}_2$, where ${\overline}{X}_1$ is tangent to $\mathcal{N}$, ${\overline}{X}_2$ is tangent to $\mathcal{N}':=\nu \widehat{L}_q \cap \mathcal{K}$, and ${\ensuremath{ \mathrm{g} }}^\ell({\overline}{X}_1,{\overline}{X}_2)=0$. Therefore, it is enough to check independently that for every basic vector field ${\overline}{X}$ along $\widehat{L}_q$, tangent to either $\mathcal{N}$ or $\mathcal{N}'$, the norm of ${\overline}{X}$ is constant along $\widehat{L}_q$.
If ${\overline}{X}$ is tangent to $\mathcal{N}$, then by the construction in Proposition \[P:extension\] it projects to some basic vector field $X$ along $\widehat{L}_p\subset B$. Since $(B,{\ensuremath{ \mathrm{g} }}_B,{\ensuremath{\mathcal{F}}}|_B)$ is a Riemannian foliation, the norm $\|X\|_{{\ensuremath{ \mathrm{g} }}_B}$ is constant along $\widehat{L}_p$. By the construction of the metric ${\ensuremath{ \mathrm{g} }}^\ell$, one has $\|{\overline}{X}\|_{{\ensuremath{ \mathrm{g} }}^\ell}=\|X\|_{{\ensuremath{ \mathrm{g} }}_B}$ and, therefore, the norm of ${\overline}{X}$ is constant along $\widehat{L}_q$.
If ${\overline}{X}$ is tangent to $\mathcal{N}'$, then it is tangent to any fiber $U_{p'}$, $p'\in \widehat{L}_p$. The restriction ${\overline}{X}|_{U_{p'}}$ is a basic vector field of $(U_{p'},\widehat{{\ensuremath{\mathcal{F}}}}^\ell_{p'})$ along $\widehat{L}_{q}\cap U_{p'}$, and therefore the norm $\|{\overline}{X}|_{U_{p'}}\|_{{\ensuremath{ \mathrm{g} }}_{p'}}$ is locally constant along $\widehat{L}_{q}\cap U_{p'}$. By the construction of ${\ensuremath{ \mathrm{g} }}^\ell$, it follows that $\|{\overline}{X}|_{U_{p'}}\|_{{\ensuremath{ \mathrm{g} }}^\ell}$ is also locally constant along each $\widehat{L}_{q}\cap U_{p'}$. However, given two points $p', p''\in \widehat{L}_p$, and a vertical, foliated vector field $V^\ell$ whose flow $\Phi$ moves $p'$ to $p''$, one also has that $\Phi$ moves $U_{p'}$ isometrically to $U_{p''}$, and ${\overline}{X}|_{U_{p'}}$ to ${\overline}{X}|_{U_{p''}}$. In particular, $\|{\overline}{X}|_{U_{p'}}\|_{{\ensuremath{ \mathrm{g} }}^\ell}=\|{\overline}{X}|_{U_{p'}}\|_{{\ensuremath{ \mathrm{g} }}_{p'}}$ does not really depend on the point $p'\in \widehat{L}_p$, and it is actually constant along the whole leaf $\widehat{L}_q$.
With this in place, one can finally prove the Main Theorem:
Let $U$ be an $\epsilon$-tubular neighbourhood around the closure ${\overline}{L}$ of a leaf $L\in {\ensuremath{\mathcal{F}}}$. Letting $B={\overline}{L}$, we are under the assumptions of Section \[S:setup\]. In particular, it is possible to define the linearized foliation ${\ensuremath{\mathcal{F}}}^\ell$ on $U$, its local closure $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$, and the metric ${\ensuremath{ \mathrm{g} }}^\ell$ as in Proposition \[P:new-metric\]. It is clear by construction that $\widehat{{\ensuremath{\mathcal{F}}}}^\ell|_{{\overline}{L}}={\ensuremath{\mathcal{F}}}|_{L}$ and that the closure of $\widehat{{\ensuremath{\mathcal{F}}}}^\ell$ is contained in the closure of ${\ensuremath{\mathcal{F}}}$. Moreover, by Corollary \[C:local-form\] the foliation $(U,{\ensuremath{ \mathrm{g} }}^\ell,\widehat{{\ensuremath{\mathcal{F}}}}^\ell)$ is, locally around each point, foliated diffeomorphic to the orbit like foliation $({\mathbb{D}}^k\times {\mathbb{D}}^{m-k}\times U_p,{\mathbb{D}}^k\times\{\textrm{pts.}\}\times \widehat{{\ensuremath{\mathcal{F}}}}^\ell_p)$. By Proposition \[P:orbit-like-invariant\], the foliation $(U,{\ensuremath{ \mathrm{g} }}^\ell,\widehat{{\ensuremath{\mathcal{F}}}}^\ell)$ is orbit like as well, and this concludes the proof.
[10]{}
M. M. Alexandrino, *Proofs of conjectures about singular Riemannian foliations*, Geom. Dedicata **119** (2006) no. 1, 219–234.
M. M. Alexandrino *Desingularization of singular Riemannian foliation*, Geom. Dedicata **149** (2010) 397–416.
M. M. Alexandrino and A. Lytchak, *On smoothness of isometries between orbit spaces*, Riemannian geometry and applications–Proceedings RIGA (2011), 17–28, Ed. Univ. Bucureşti.
M. M. Alexandrino and M. Radeschi, *Smoothness of isometric flows on orbit spaces and applications to the theory of foliations*, Transf. Groups, DOI: 10.1007/s00031-016-9386-5 (2016), 1–26.
J. Bolton, *Transnormal systems*, Q. J. of Math. **24**, no.1 (1973) 385–395.
J. Cheeger and D. Gromoll, *On the Structure of Complete Manifolds of Nonnegative Curvature*, Ann. of Math. **96** (1972), no. 3, 413–443.
A. Lytchak and M. Radeschi, *Algebraic nature of singular Riemannian foliations in spheres*, to appear in J. Reine Ang. Math., DOI: https://doi.org/10.1515/crelle-2016-0010.
R. Mendes and M. Radeschi *Smooth basic functions*, preprint: arXiv:1511.06174 \[math.DG\].
P. Molino, *Riemannian foliations*, Progress in Mathematics **73**, Birkhäuser, Boston (1988).
P. Molino, *Orbit-like foliations*, Proceedings of Geometric study of Foliations, held in Tokyo, November 1993, ed. T. Mizutani et al. World Scientific, Singapore (1994), 97–119.
S. Kobayashi and K. Nomizu, *Foundations of Differential geometry, Vol. I*, Interscience tracts in pure and applied mathematics, Interscience Editor, New York (1963).
M. Radeschi, *Clifford algebras and new singular Riemannian foliations in spheres*, Geom. Funct. Anal **24** (2014) 5, 1660–1682.
H. J. Sussmann, *Orbits of families of vector fields and integrability of distributions*. Trans. of the Am. Math. Soc., **180** (1973), 171–188.
D. Toeben, *Parallel focal structure and singular Riemannian foliations*, Trans. Am. Math. Soc. **358**, 1677–1704 (2006).
B. Wilking, *A Duality Theorem for Riemannian Foliations in Nonnegative Sectional Curvature*, Geom. Funct. Anal. **17** (2007), no. 4, 1297–1320.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Photometric classification of supernovae (SNe) is imperative as recent and upcoming optical time-domain surveys, such as the Large Synoptic Survey Telescope (LSST), overwhelm the available resources for spectrosopic follow-up. Here we develop a range of light curve classification pipelines, trained on 513 spectroscopically-classified SNe from the Pan-STARRS1 Medium-Deep Survey (PS1-MDS): $357$ Type Ia, $93$ Type II, $25$ Type IIn, $21$ Type Ibc, and $17$ Type I SLSNe. We present a new parametric analytical model that can accommodate a broad range of SN light curve morphologies, including those with a plateau, and fit this model to data in four PS1 filters ($g_\mathrm{P1}r_\mathrm{P1}i_\mathrm{P1}z_\mathrm{P1}$). We test a number of feature extraction methods, data augmentation strategies, and machine learning algorithms to predict the class of each SN. Our best pipelines result in $\approx 90\%$ average accuracy, $\approx 70\%$ average purity, and $\approx 80\%$ average completeness for all SN classes, with the highest success rates for Type Ia SNe and SLSNe and the lowest for Type Ibc SNe. Despite the greater complexity of our classification scheme, the purity of our Type Ia SN classification, $\approx 95$%, is on par with methods developed specifically for Type Ia versus non-Type Ia binary classification. As the first of its kind, this study serves as a guide to developing and training classification algorithms for a wide range of SN types with a purely empirical training set, particularly one that is similar in its characteristics to the expected LSST main survey strategy. Future work will implement this classification pipeline on $\approx 3000$ PS1/MDS light curves that lack spectroscopic classification.'
author:
- 'V. A. Villar'
- 'E. Berger'
- 'G. Miller'
- 'R. Chornock'
- 'A. Rest'
- 'D. O. Jones'
- 'M. R. Drout'
- 'R. J. Foley'
- 'R. Kirshner'
- 'R. Lunnan'
- 'E. Magnier'
- 'D. Milisavljevic'
- 'N. Sanders'
- 'D. Scolnic'
bibliography:
- 'mybib.bib'
title: 'Supernova Photometric Classification Pipelines Trained on Spectroscopically Classified Supernovae from the Pan-STARRS1 Medium-Deep Survey'
---
Introduction
============
Optical time-domain astronomy has entered a new era of large photometric datasets thanks to current and upcoming deep and wide-field surveys, such as the Panoramic Survey Telescope and Rapid Response System (Pan-STARRS; @kaiser2010), the Asteroid Terrestrial-impact Last Alert System (ATLAS; @jedicke2012atlas), the Zwicky Transient Facility (ZTF; @kulkarni2018zwicky), the Large Synoptic Survey Telescope (LSST; @ivezic2011large), and the Wide Field Infrared Survey Telescope (WFIRST; @spergel2015). The current surveys are already discovering $\sim 10^4$ SNe per year, a hundred-fold increase over the rate of discovery only a decade ago. LSST will increase this discovery rate to $\sim 10^6$ SNe per year.
Supernovae have traditionally been classified based on their spectra [@filippenko1997optical]. In the early days this was accomplished through visual inspection, then with template-matching techniques (e.g., SNID; @blondin2007determining), and most recently with deep learning techniques (e.g., @muthukrishna2016deep). However, given the current discovery rate, and the anticipated LSST discovery rate, spectroscopic follow up is severely limited. The consequence of this fact is twofold. First, we need a way to effectively identify “needles” in the haystack – the events that will yield the greatest scientific return with detailed follow up observations (e.g., spectroscopy, radio, X-ray). Second, we need to devise methods to extract as much information and physical insight as possible from the “haystack” of SNe for which no spectroscopy or other data will be available. Here, we specifically focus on the latter issue and explore the question: Given complete optical light curves, can we classify SNe into their main spectroscopic classes (Ia, Ibc, IIP, etc.)?
[{width="\textwidth"}]{}
\[fig:redmag\]
Previous studies in this area have largely focused on the simpler task of separating thermonuclear Type Ia SNe from non-Type Ia SNe, motivated by the use of Type Ia SNe as standardizable cosmological candles, and taking advantage of their uniformity (e.g., @moller2016 [@kimura2017single]). Separating the classes of core-collapse SNe (CCSNe) is a broader and more challenging problem. First, unlike Type Ia SNe, CCSNe exhibit broad diversity between and within each class in terms of basic properties such as luminosity, timescale, and color (e.g., @drout2011first [@taddia2013carnegie; @sanders2015toward; @nicholl2017magnetar; @villar2017theoretical]). This is due to their wide variety of progenitor systems, energy sources, and circumstellar environments. Second, the overall diversity of CCSNe is less thoroughly explored, due to small sample sizes and few published uniform studies. As a consequence, most previous works on photometric classification of CCSNe have relied on simulated datasets to train and test classification algorithms (e.g., @richards2011semi [@charnock2017deep; @kimura2017single; @ishida2018optimizing]). Simulated datasets are based on strong assumptions about the underlying populations of each SN class and often do not reflect the true event diversity, or the effects of actual survey conditions.
Here, we approach the question of SN photometric classification using a large and uniform dataset of $513$ spectroscopically-classified SNe from the PS1-MDS. Importantly, the characteristics of this dataset in terms of filters, depth, and cadence are the closest available analogue to the LSST main survey design. We fit the observed light curves with a flexible analytical model that can accommodate all existing light curve shapes, using a Markov chain Monte Carlo (MCMC) approach. We then train and evaluate 24 classification pipelines that span different feature extraction, data augmentation, and classifications methods. We further use the posteriors of our MCMC fits to determine overall uncertainties on our classifications.
The paper is organized as follows. In §\[sec:data\] we introduce the PS1-MDS dataset utilized here. In §\[sec:model\] we describe our analytical light curve model and iterative MCMC fitting approach. In §\[sec:classification\] we describe the key components of our various classification pipelines, including feature extraction, data augmentation, and classification approaches. We present the results of our classifications in §\[sec:results\], compare to previous classifications efforts in §\[sec:comp\], and discuss limitations and future directions in §\[sec:future\].
Throughout this paper, we assume a flat $\Lambda$CDM cosmology with $\Omega_M=0.286$, $\Omega_\Lambda=0.712$ and $H_0 = 69.3$ km s$^{-1}$ Mpc$^{-1}$ [@hinshaw2013nine].
PS1-MDS Supernova Light Curves and Spectroscopic Classifications {#sec:data}
================================================================
Pan-STARRS1 (PS1) is a wide-field survey telescope with a 1.8 m diameter primary mirror located on Haleakala, Hawaii [@kaiser2010]. The PS1 1.4 gigapixel camera (GPC1) is an array of 60 $4800\times 4800$ pixel detectors with a pixel scale of $0.''258$ and an overall field of view of $7.1$ deg$^2$. The PS1 survey used five broadband filters, $g_\mathrm{P1}r_\mathrm{P1}i_\mathrm{P1}z_\mathrm{P1}y_\mathrm{P1}$. The details of the filters and the photometry system are given in @stubbs2010 and @tonry2012pan.
The PS1-MDS, conducted in $2010-2014$, consisted of ten single-pointing fields for a total area of about $70$ deg$^2$ [@chambers2016]. About $25\%$ of the overall survey observing time was dedicated to the MDS fields, which were observed with a cadence of about 3 days per filter in $g_\mathrm{P1}r_\mathrm{P1}i_\mathrm{P1}z_\mathrm{P1}$ to a $5\sigma$ depth of $\approx 23.3$ mag per visit. The typical sequence consisted of $g_\mathrm{P1}$ and $r_\mathrm{P1}$ on the same night, followed by $i_\mathrm{P1}$ and then z$_\mathrm{P1}$ on subsequent nights. Observations in $y_\mathrm{P1}$-band were concentrated near full moon with a shallower $5\sigma$ depth of $\approx 21.7$ mag; we do not use the $y_\mathrm{P1}$-band data in this study due to its significantly shallower depth and poorer cadence.
The reduction, astrometry, and stacking of the nightly images were carried out by the Pan-STARRS1 Image Processing Pipeline (IPP; @magnier2016a [@magnier2016b; @waters2016pan]). The nightly stacks were then transferred to the Harvard FAS Research Computing Odyssey cluster for a transient search using the [photpipe]{} pipeline, previously used in the SuperMACHO and ESSENCE surveys [@rest2005testing; @miknaitis2007essence] and described in detail in our previous analyses of PS1-MDS data [@rest2014; @scolnic2017complete; @jones2018measuring].
In the full PS1-MDS dataset we have identified 5235 likely SNe [@jones2017measuring; @jones2018measuring]. During the course of the survey, spectroscopic observations were obtained for over 500 events using the MMT 6.5-m telescope, the Magellan 6.5-m telescopes, and the Gemini 8-m telescopes. We further obtained spectroscopic host galaxy redshifts for 3147 SN-like transients. The transients spectroscopically and photometrically classified as Type Ia SNe were published in @jones2017measuring; the light curves and photometric classification of the remaining objects will be presented in future work. Similarly, the bulk of the Type IIP SNe (76 events) were published in @sanders2015toward, and the Type I SLSNe (17 events) were published in @lunnan2018. Here we focus on $513$ spectroscopically classified events, which were classified using the SNID software package [@blondin2007determining]. The sample contains $357$ Type Ia SNe, $93$ IIP/L SNe, $25$ Type IIn SNe, $21$ Type Ibc SNe, and $17$ Type I SLSNe[^1].
Our sample is limited events with high-confidence spectroscopic classifications with a statistically useful number of members in each class. As part of the PS1-MDS we discovered several other rare transients, including tidal disruption events [@gezari2012ultraviolet; @Chornock2014] and fast-evolving luminous transients [@drout2014rapidly], but the sample sizes for those are too small for inclusion in this study. It is possible that SNID misclassification exist in our dataset; e.g., low SNR events are more likely to match to Type Ia SNe [@blondin2007determining]. To partially counteract this, we check each member of our Type Ibc and SLSNe classes (our smallest classes) by eye to ensure high purities. Finally, we note that the magnitude limit for our spectroscopic follow up was generally shallower by about 1.5 mag relative to the PS1-MDS nominal per-visit depth. This does not affect our ability to test classifiers on the spectroscopic sample itself, but will be considered when extending our method to the full photometric dataset in future work (see §\[sec:comp\]).
The light curves range from a minimum of 3 to $\approx 150$ total data points in any filter with a signal-to-noise ratio of ${\rm S/N}>3$, with a median of about 30 data points in each light curve. We have only eliminated events with light curves that contain fewer than two $3\sigma$ detections in three or more filters, eliminating $7$ SNe from our sample[^2] (6 Type Ia SNe and 1 Type II SN) leaving 506 remaining SNe for our training set.
In Figure \[fig:redmag\] we plot the peak absolute $i_\mathrm{P1}$ magnitude versus redshift for our spectroscopic sample. The sample spans $M_i\approx -14.5$ to $-22.5$ and extends in redshift to $z\approx 1.6$, with only the brightest classes (SLSNe and Type IIn) being observed at $z\gtrsim 0.6$. Specifically, we find a range of $M_i\approx -14.5$ to $-18.5$ mag for the Type II SNe, $\approx -16.5$ to $-19.5$ mag for the Type Ibc SNe, $-16$ to $-20.5$ for the Type IIn SNe, and $\approx -20.5$ to $-22.5$ for the SLSNe.
Analytical Light Curve Model and Fitting {#sec:model}
========================================
[{width="\textwidth"}]{}
Rather than interpolating data points, a common method to standardize data is to fit a simple parametric model to the light curves (e.g., @bazin2009core [@newling2011statistical; @karpenka2013]). However, the majority of existing analytical light curve models are best-suited for Type Ia SNe and have limited flexibility for the full observed range of SN light curve shapes. Here we present and fit our data with a new parametric piecewise model that is designed to be flexible enough for a broad range of light curve morphologies:
$$\label{eqn:model}
F=\begin{cases}
\frac{A+\beta(t-t_0)}{1+e^{-(t-t_0)/\tau_\mathrm{rise}}} & t<t_1 \\
\frac{\left(A+\beta\left(t_1-t_0\right)\right)e^{-(t-t_1)/\tau_\mathrm{fall}}}{1+e^{-(t-t_0)/\tau_\mathrm{rise}}} & t\ge t_1
\end{cases}$$
The model contains seven free parameters, whose effects on the resulting light curves are shown in Figure \[fig:lc\_example\]. Although each parameter has a unique and interpretable effect, some degeneracies between the parameters exist. For example, the parameter $A$ affects the amplitude of the light curve, although its value does not exactly correspond to the peak flux. Similarly, $t_0$ acts as a temporal shift in the light curve, but does not directly correspond to the time of explosion or the time of peak. The parameters, $t_\mathrm{\rm rise}$, $t_1$, and $t_\mathrm{\rm fall}$ control the rise, plateau onset, and fall time of the light curve, respectively. For the purposes of fitting, we reparameterize $t_1$ into a new parameter $\gamma\equiv t_1-t_0$, which better represents the plateau duration of the light curve and results in fewer degeneracies when fitting. Finally, the parameter $\beta$ controls the slope of the plateau phase.
This functional form is similar to those presented in @bazin2009core (with five free parameters) and @karpenka2013 (with six free parameters), but incorporates a plateau component. In Figure \[fig:lc\_iip\] we show examples of fits to a Type IIP SN and a Type Ia SN with our model, the Bazin model and the Karpenka model. Our model provides a better fit to both the fast rise time and plateau phase of the Type IIP SN light curve, and is flexible enough to also fit the smoother light curve of a Type Ia SN. We note that @sanders2015toward presented a similar piecewise model with 11 free parameters to fit a sample of 76 PS1-MDS Type II SNe; however, @lochner2016photometric found that this model was not robust when fitting data without the use of informative priors, due to the large number of free parameters. Additionally, the sharp transitions between rise and decline in the Sanders model make it difficult to fit CCSNe with smooth peaks.
[ccc]{}\[t!\] $\tau_\mathrm{rise}$ (days) & Rise Time & U(0.01,50)\
$\tau_\mathrm{fall}$ (days) & Decline Time & U(1,300)\
$t_0$ (MJD) & “Start” Time & U($t_\mathrm{min}-50$, $t_\mathrm{max}+300$\
$A$ & Amplitude & U(3$\sigma$,100 $F_\mathrm{max}$)\
$\beta$ (flux/day) & Plateau slope & U($-F_\mathrm{max}/150$,0)\
$c$ (flux) & Baseline Flux & U($-3\sigma$,$3\sigma$)\
$\gamma$ (days) & Plateau duration & $(2/3)N(5,5)+(1/3)N(60,30)$
One common Type Ia SN light curve feature missing from our model is the second peak in the red light curves at about 1 month post-explosion (e.g., @kasen2006secondary [@mandel2011type; @dhawan2015near]). We find that this feature manifests itself as a “plateau” in our analytical model in the $i$- and $z$-bands. However, as we show in §\[sec:results\], our classification pipelines can reliably classify Type Ia SNe without explicitly including a second peak in our model.
We fit the light curves using PyMC (V2; @patil2010pymc), a Python module that implements a Metropolis-Hastings MCMC sampling algorithm. We assume uniform priors on all parameters with the exception of $\gamma$. We found that light curves typically fall in one of two solutions: light curves with a long plateau (in the case of Type IIP SNe) and light curves that lack a plateau (all other types). To best reflect this fact, we set the prior of $\gamma$ to a double Gaussian peaked at 5 and 60 days. This prior helps to remove a degeneracy in which a steep exponential decline can resemble a linear decline. The priors are listed in Table \[table:results1\]. We use a standard likelihood function, incorporating both the observational error and a scalar white noise scatter term added in quadrature.
We find that several of our model parameters are correlated (degenerate) with one another. In particular, the amplitude ($A$) is negatively correlated with both the rise time ($t_\mathrm{rise}$) and plateau duration ($\gamma$) but negatively correlated with the start time ($t_\mathrm{0}$). Additionally, duration is negatively correlated to both the rise time and start time, while the rise time is positively correlated with the start time.
We fit the light curve in each of the 4 filters independently, in the observer frame, but use an iterative fitting routine to incorporate combined information from all filters. We first run the MCMC to convergence on each filter independently with the same set of priors. We then combine the marginalized posteriors (i.e., we ignore parameter covariances) from each filter and use the combined posterior as a new prior for a second iteration of fitting. We can apply this process repeatedly, but we find that a single iteration is sufficient for the vast majority of events. Our iterative procedure is essential for fitting light curves in which some filters have significantly fewer data points, a situation that is common in photometric surveys due to differences in relative sensitivity, the intrinsic colors and color evolution of SNe, and varying observing conditions. An example of the best-fit solutions given by the first and second iterations is shown in Figure \[fig:lc\_its\]. In this example, the peak times in $g$- and $i$ are in disagreement with $r$- and $z$ due to poorly-sampled data in the former two filters. Following the second iteration, this disagreement is removed, leading to more realistic fits.
Representative light curves and their best fits are shown in Figure \[fig:lc\_examples\]. The solutions are constrained for well-sampled light curves (e.g., the Type Ia SN shown) but more poorly constrained for sparse light curves (e.g., the SLSN shown). Crucially, because we have access to the full posterior of light curve solutions, we can feed many samples of the posterior through our classification algorithm to quantify the classification uncertainty for each event.
Unless otherwise specified, we use the observer-frame light curve fits to extract features. We then include the redshift to transform to absolute magnitudes, including a cosmological $k$-correction: $M=m-5\log(d_L/10\mathrm{pc})+2.5\log(1+z)$, where $d_L$ is the luminosity distance. We do not apply $k$-corrections to account for the intrinsic spectral energy distribution of the various SN types.
Classification Pipelines {#sec:classification}
========================
For each SN, our MCMC fitting generates posterior distributions for the model light curve parameters. To train a classifier, we need to extract features from the light curves generated by the fitted parameters. We test several methods of feature extraction, data augmentation and classification. We describe each method in the following subsections, and we compare the algorithms in terms of classification purity, completeness, and accuracy in §\[sec:results\]. *Purity* (also called *precision*) is defined as the fraction of events in a predicted class that are correctly identified; for example, if our classifier predicts a total of 100 Type Ia SNe, but only 70 of those are spectroscopically-classified as Type Ia SNe, the purity would be 0.7. *Completeness* (also called *recall*) is defined as the fraction of events in an observed class that are correctly identified; for example, if our sample contains 100 spectrosopically-classified Type Ia SNe, but our classifier has only identified 70 of those events as Type Ia SNe, then our completeness would be 0.7. *Accuracy* is defined as the total fraction of events that are classified correctly as being a member or not a member of a given class; for example, if a sample of 100 SNe contains 70 spectrosopically-classified Type Ia SNe, and our classifier correctly identifies the 70 Type Ia SNe but incorrectly classifies 20 more CCSNe as Type Ia SNe, the overall accuracy is 0.8. The three terms are mathematically defined as follows: $$\begin{aligned}
\mathrm{Purity} &= \frac{\mathrm{TP}}{\mathrm{TP+FP}}\\ \mathrm{Completeness} &= \frac{\mathrm{TP}}{\mathrm{TP+FN}}\\
\mathrm{Accuracy} &= \frac{\mathrm{TP+TN}}{\mathrm{TS}},\end{aligned}$$ where TP (FP) is the number of true (false) positives, TN (FN) is the number true (false) negatives, and TS is the total sample size.
Feature Selection {#sec:features}
-----------------
Although our analytical model produces interpretable features for each light curve (albeit ones that are somewhat degenerate) we would like to explore various methods of feature extraction, based on the analytical fits. In particular, we explore the following four types of features:
- **Model Parameters (M):** We use the analytical model parameters as features, as well as the peak absolute magnitude in each filter, including a cosmological $k$-correction but no correction for intrinsic SN colors and color evolution.
- **Hand-Selected Features (HS):** We use hand-selected interpretable features: the peak absolute magnitude in each filter, including a cosmological $k$-correction but no correction for intrinsic SN colors and color evolution; and the rest-frame rise and fall times by 1, 2 and 3 mag relative to peak (where we do *not* correct the rise and fall times for cosmological time-dilation).
- **Principal Component Analysis (PCA):** We fit a PCA decomposition model to the full set of analytical model fits (without any redshift corrections) independently for each filter. We use the first 6 PCA components from each filter, corresponding to an explained variance within the light curves of $\sim 99.9$%. We also use the peak absolute magnitude, including a cosmological $k$-correction, in each filter in addition to the PCA components.
- **Light Curves (LC):** We use the model light curves as the features. We renormalize the flux of each light curve, correcting for luminosity distance; however, we find that neglecting time dilation corrections improves classification accuracy, and therefore we do not make these corrections. We down-sample each filter model to 10 observations logarithmic-spacing between $t_0$ and $t_0+300$ to decrease the number of features.
To provide some intuition, we highlight a sub-space of the hand-selected features ($M_{\rm peak}$ versus duration time to rise and fall by 2 mag) in Figure \[fig:dlps\]. We find that some SN classes, such as SLSNe versus Type II, or Type Ia versus Type IIn, easily separate in the duration-luminosity feature space. However, other classes, such as Type Ibc versus Type Ia and IIP, have substantial overlap in this space, regardless of filter. This highlights that while simple heuristics can be used as first-order classifiers for some SN classes, other classes are intrinsically difficult to disentangle from light curve information alone.
Data Augmentation {#sec:augmentation}
-----------------
Data augmentation is ubiquitous in machine learning applications, as a larger dataset can significantly improve the accuracy and generalizability of most classification algorithms. Data augmentation methods have already been utilized in the astrophysical context (e.g., @hoyle2015data).
[{width="\textwidth"}]{}
\[fig:smote\]
Here, we augment our training set with simulated events for two key reasons. First, our training set is unbalanced in terms of SN classes due to the differing observed rates of transients, with Type Ia SNe representing $\approx 70$% of our sample (and more generally, of any magnitude-limited optical survey). Classification algorithms trained on unbalanced training sets tend to over-classify all objects as the dominant class. This is because the algorithms can minimize the decision-making complexity by ignoring minority classes in favor of correctly classifying the majority class. In our case, a classification algorithm may preferentially label all objects as Type Ia SNe to achieve an overall high accuracy. Second, our training set is small in the context of machine learning, with the smallest class (SLSNe) containing just $17$ events.
One approach to overcome this in the context of our method is to augment our training set with many draws from the MCMC posteriors. However, this would lead to clustering of solutions in feature-space that may bias the training algorithms. Instead, we address the issue of a small and imbalanced training set by synthesizing more event samples using two techniques. First, we use the Synthetic Minority Over-sampling Technique (SMOTE; @chawla2002smote) to over-sample all the non-Type Ia SN classes to be equally represented as the Type Ia SNe. SMOTE creates synthetic samples in feature space by randomly sampling along line segments joining the $k$ nearest neighbors of a sample, where $k$ is a free parameter of the algorithm. Here we find that $k=5$ performs well for sampling the minority classes. An example of the SMOTE resampling algorithm is shown in Figure \[fig:smote\]. A key feature of SMOTE resampling is that it produces realistic samples within each class, but it cannot produce samples outside the extent of the original sample. While this prevents the generation of unphysical models, it may overly constrain the properties of classes with only a few samples (e.g., SLSNe).
Second, we augment the non-Type Ia SN classes by fitting the feature space of each class to a multivariate-Gaussian (MVG) and resampling from the fitted MVG. This is similar to the SMOTE algorithm in that it allows for the generation of new events that encompass a larger potential feature space. However, one key difference is that this method allows for synthesized events beyond the feature boundaries seen in the data. While this may lead to some unphysical models, it better reflects the potential spread in light curve parameters in poorly-sampled classes. An example of the MVG resampling is shown in Figure \[fig:smote\].
Both augmentation methods aim to increase our training set in a way which is representative of the set and therefore makes no attempt to correct for potential biases. This can potentially lead to increased misclassifications if our labelled training set is unrepresentative of a future test set; however, we expect no such effects within the training set of 513 objects.
Classification {#sec:classifiers}
--------------
Following the work of @lochner2016photometric, we test three classification algorithms: a support vector machine (SVM), a random forest (RF), and a multilayer perceptron (MLP). We optimize the hyperparameters of each algorithm independently using a grid search. Each algorithm and its tunable hyperparameters are described below. We use the [scikit-learn]{} python package throughout the classification portion of our pipeline.
### Support Vector Machine (SVM)
A SVM classifies the training set by finding the optimal hyperplane in feature space to minimize the number of misclassified samples. In particular, the SVM will select a hyperplane that maximizes the distance between class samples nearest the hyperplane (also known as the support vectors). In the majority of cases, the classes are not linearly separable within the feature space alone (i.e., there may be significant overlap between classes). Instead, the features are expanded into an infinite basis function using the so-called Kernel trick [@aizerman1964theoretical], allowing one to find a feature space in which the separating hyperplane is linear. We optimize the kernel and a regularization term using a coarse grid search, allowing the kernel size to logarithmically range from $\sigma=1$ to $\sigma=100$ and the normalization to logarithmically range from 1 to 1000. We find that a radial basis function kernel with width $\sigma=10$ typically results in optimal classification, with normalization values ranging depending on the pipeline.
### Random Forest (RF)
RF classifiers [@breiman2001random] are built on the idea of a decision tree, which is a model that generates a set of rules to map input features to classes. This mapping is based on a series of branching decisions based on feature values (e.g., “is the peak $g$-band magnitude brighter than $-19$?”). While single trees are theoretically sufficient for classification problems, they often lead to over-fitting due to specialized branching required for each class. Random forests overcome this problem by combining decision trees that are trained on different subsets of the training data and features. The ensemble of decision trees is then used as the classifier. There are a number of free parameters within a RF, including the number of decision trees, the number of nodes for each tree and the splitting rules for each node. Through a grid search of hyperparameters, we find that 100 decision trees utilizing the Gini impurity (the probability that a randomly chosen SN from a labelled class is misclassified) as a splitting criterion and allowing nodes to be split until all leaves are pure results in the highest accuracy.
### Multilayer Perceptron (MLP)
A fully-connected MLP is the simplest artificial neural network (e.g., @schmidhuber2015deep). It is composed of a series of layers of neurons, where each neuron is the dot product of the previous layer and a set of optimizable weights, passed through a nonlinear activation function. A “fully-connected” MLP means that each neuron is connected to all neurons in the preceding layer. The nonlinear activation function is what allows a MLP to model nonlinear mappings between the feature set and classes. MLPs have many tunable parameters, including the number of layers, the number of neurons within each layer, the learning rate and a regularization term. We optimize the hyperparameters using a grid search, finding that two layers with ten neurons each typically performs best, and use the [Adam]{} optimization algorithm [@adamopti2014] to train the MLP.
An example of a complete pipeline, excluding the MCMC fitting step, is available on GitHub[^3].
[{width="\textwidth"}]{}
\[fig:purcom\]
[{width="\textwidth"}]{}
\[fig:purcom\_hist\]
[{width="\textwidth"}]{}
\[fig:hist\_meth\]
Classification Results {#sec:results}
======================
[![Confusion matrix for one of our best performing classification pipelines (PCA feature extraction, MVG data augmentation, and RF classifier) calculated using the full posterior distributions for each SN. We show the confusion matrix for both the full SN sample of 513 objects ([*Top*]{}) and only for the 429 events with a high classification confidence probability of $p>0.8$ ([*Bottom*]{}).[]{data-label="fig:confusion_matrix"}](confusion_matrix.pdf "fig:"){width="45.00000%"}]{}
[{width="95.00000%"}]{}
We combine each of the four feature extraction methods (M, HS, PCA, and LC), two data augmentation methods (SMOTE and MVG), and three classification algorithms (SVM, RF, and MLP) to test a total of 24 classification pipelines. For each pipeline, we use the full dataset to find the hyperparameters which optimize overall accuracy for the classification method. We optimize the hyperparameters over a coarse grid, due to the computational costs of performing a large grid search. We then perform leave-one-out cross-validation by iteratively removing one object from the sample, performing data augmentation on the remaining dataset, and training a classifier on the new set. We then test the trained classifier on the median posterior values of the removed object and record the predicted label. Due to computational costs, we only utilize the full posteriors for classification error estimation using our optimal pipeline.
[{width="\textwidth"}]{}
General Trends
--------------
In Figures \[fig:purcom\] and \[fig:purcom\_hist\] we plot the purity, completeness and accuracy for each of the 24 pipelines and each of the 5 SN classes. Figure \[fig:purcom\] provides a matrix representation with the percentage score noted for each combination of pipeline and SN class, while Figure \[fig:purcom\_hist\] shows the same results in histogram format to aid in visualizing the range of completeness, purity, and accuracy values across the 24 pipelines for each SN class.
We find that SLSNe and Type Ia SNe are consistently the classes with the highest purity and completeness, reaching $\gtrsim 90$% for the best classification pipelines. This is due to the fact that SLSNe are easily separable from the other classes due to their high luminosity and longer durations (Figures \[fig:redmag\] and \[fig:dlps\]), while Type Ia SNe are tightly clustered in feature space due to their intrinstic uniformity.
In contrast, we find that Type Ibc SNe typically have the lowest purity and completeness, with $\approx 15-35\%$ and $\approx 25-65\%$, respectively, and a much wider spread in performance for the various pipelines. The lower classification success rate is due to broader diversity within Type Ibc SNe, as well as their significant overlap with Type Ia SNe (e.g., Figure \[fig:dlps\]).
For Type II SNe we find high values of purity and completeness of $\approx 65-85\%$ and $\approx 60-80\%$, respectively. This overall high success rate is mainly due to the presence of a plateau phase that helps to distinguish most Type II SNe from the other classes. However, the failed classifications are most likely due to the faster evolving Type II SNe (often called Type IIL), which tend to be misclassified as Type Ibc or Type Ia SNe due to overlap in light curve shapes (e.g., Figure \[fig:dlps\]).
Finally, for Type IIn SNe we find purity and completeness of $\approx 30-80\%$ and $\approx 45-70\%$, respectively, reflecting the broad diversity of light curve morphologies and luminosities, with some events overlapping similar areas in feature space with Type Ia and Ibc SNe (e.g., Figure \[fig:dlps\]). As for the Type Ibc SNe, we find quite a broad dispersion in performance between the various pipelines.
For the overall accuracy across the 5 SN classes, we find generally high values of $\approx 100\%$ for SLSNe, $\approx 95\%$ for Type IIn SNe, $\approx 90\%$ for Type II SNe, $\approx 85-95\%$ for Type Ibc SNe, and $\approx 85-90\%$ for Type Ia SNe. These values are essentially independent of the classification pipeline used.
To further explore the relative performance of the various pipelines, in Figure \[fig:hist\_meth\] we plot the distribution of completeness across the full dataset, grouping the classification pipelines by feature extraction method, classification method, and data augmentation method. We find that the classification method has the largest impact on completeness, with the RF classifiers performing noticeably better, and more uniformly, than the SVM and NN classifiers. In terms of feature extraction we find that use of the model parameters (M) and PCA are somewhat advantageous compared to hand-selected (HS) features and the LC approach, although the PCA extraction leads to a broader range of outcomes. Finally, the MVG augmentation method performs slightly better than SMOTE.
The top three pipelines in terms of purity, completeness and accuracy share RF classification and PCA feature extraction, with both MVG and SMOTE augmentation. Between these pipelines, the overall accuracy differs by $\lesssim 5\%$ across the 5 SN classes. In addition to performing well, the RF classifier also has the advantage of allowing us to measure the relative important of each feature. For example, we test the relative importance of our hand-selected and model features in the RF classification pipeline using the “gini importance”, a measure of the average gini impurity decrease across descending nodes [@leo1984classification]. We find that the peak magnitudes are the most important interpretable features, with durations and other parameters being roughly equally important.
For simplicity, below we focus on the results of our pipeline with the highest purity ($72\%$) and completeness ($78\%$) scores with an average accuracy of $93\%$ across the 5 SN classes. This pipeline consists of PCA feature extraction, MVG data augmentation, and RF classifier; however, we emphasize that this pipeline does not significantly outperform the others. In Figure \[fig:confusion\_matrix\] we present the final confusion matrix for this pipeline across the full training set. The confusion matrix is a quick-look visualization of how each class is correctly or incorrectly classified. We generate the confusion matrix using the full posteriors for each SN, so the probability densities have been effectively smoothed out across the matrix. To specifically assess the role of poor quality classifications, we show the confusion matrix for the full sample, as well as separately for classifications with a confidence of $p>0.8$ only (representing $\sim85$% of the original sample). In practice, one can optimize pipeline parameters to maximize sample purity, completeness or some other metric.
Assessing Misclassifications
----------------------------
Although the overall completeness for each SN class is high, we note several common misclassifications. First, Type II and Ia are the most likely classes to be misclassified as Type Ibc SNe. The Type II SNe that are misclassified as Type Ibc SNe are typically either poorly sampled or are rapidly evolving (the so-called IIL events). Second, Type Ibc, IIn, and II SNe are the most likely classes to be misclassified as Type Ia SNe. This is again due to specific events in those diverse classes that occupy the region in feature space that overlaps with the uniform Type Ia SNe. Finally, Type IIn and Ibc SNe are the most likely classes to be misclassified as Type II SNe, again due to overlaps in feature space. Comparing the full sample to the subset of events with high classification confidence ($p>0.8$) we find that the fraction of misclassified events indeed declines (most notably for Type Ibc and Ia SNe), indicating that some misclassifications are simply due to poorly sampled light curves. However, the overall trends for which classes are most likely to be misclassified as others remains the same, indicating that there is an inherent limitation to the classification success rate that is due to real overlaps in feature space.
We highlight several SNe that are misclassified, but with high confidence in Figure \[fig:incor\]. In these examples, a spectroscopic Type II SN with a rapid linear decline is misclassified as a Type Ibc SN; a slightly dim Type Ia SN is misclassified as a Type Ibc SN; and a fairly luminous Type Ibc SN is misclassified as a Type Ia SN. In each of these cases, the posterior of the fitted light curves is narrow, leading to little variability (i.e., a high confidence) in the final classification. These events indicate that even with good photometric data quality there is inherent overlap of SNe in feature space that leads to misclassification.
The misclassifications of SNe are further highlighted in Figure \[fig:con\_acc\]. Each panel in the top part of Figure \[fig:con\_acc\] represents a spectroscopically classified class, while in the bottom part each panel represents a photometrically assigned class. The misclassified events in both cases are labeled to provide insight into the most common misclassification. In all panels the ordinate represents the overall classification certainty, based on many draws from the posteriors of each event. In all cases, the majority of misclassifications occur at the low confidence end ($p<0.8$), but there are also high confidence misclassifications.
We explore the role of data quantity in Figure \[fig:info\_content\], where we plot the classification accuracy as a function of total light curve data points for all 5 SN classes. We again find that misclassifications are more likely in the regime of low number of data points, specifically $\lesssim 20$ data points. However, as noted above, there are also high confidence misclassifications for events with a large number of data points.
[![Classification accuracy as a function of number of light curve data points. The colors and shapes reflect the SN classes, and the black line represents a smoothed median to guide the eye. Filled symbols are SNe classified correctly, while open symbols are misclassified events. We find that misclassifications are more prevalent for light curves with fewer points, but also that some events are misclassified even with tens of data points, as also highlighted in Figure \[fig:incor\]. ](info_content.pdf "fig:"){width="50.00000%"}]{}
\[fig:info\_content\]
Comparison to Previous Photometric Classification Approaches {#sec:comp}
============================================================
The photometric classification of optical transients has been previously explored in the existing literature. Previous studies on machine learning methods have focused almost exclusively on the binary problem of Type Ia versus non-Type Ia SN classifications (e.g., @campbell2013cosmology [@ishida2013kernel; @jones2017measuring]), or have been trained and tested on simulated datasets (e.g., @kessler2010results [@tonry2012pan; @moller2016photometric; @charnock2017deep; @moller2019supernnova; @muthukrishna2016deep]). We highlight the strengths and weaknesses of both approaches (which we note are disjoint) compared to our methodology. We emphasize that classification pipelines should ideally be compared using the same dataset and set of labels. No machine learning method, including the one presented in this paper, can be applied to a new test set without retraining or careful consideration of training-vs-test set biases. This is especially crucial when comparing our method to those created for simulated datasets, which have known biases, uncertainties and simulated physics.
Identification of Type Ia SNe from photometric light curves is essential for precision cosmology in the era of large photometric surveys [@scolnic2014systematic; @jones2017measuring], which is why many studies have specifically focused on Type Ia SN classification. However, the binary problem of Type Ia vs non-Typa Ia SN classification is much narrower (and simpler) than full classification of CCSN classes. As standardizable candles, Type Ia SNe are fairly homogeneous with observational variations (excluding reddening) that are well described by two observable features: stretch and peak luminosity. As a result, it is easier to separate the small area of feature-space corresponding to Type Ia SNe from other transients. Studies that focus on this approach achieve a classification accuracy of $\gtrsim 0.95$ (e.g., @ishida2013kernel [@charnock2017deep; @jones2017measuring; @narayan2018machine; @pasquet2019pelican]). Although our pipeline is trained and tested on an empirical dataset for 5 distinct SN classes, we find that our achieved purity ($\approx 95\%$), completeness ($\approx 90\%$) and accuracy ($\approx 95\%$) for Type Ia SN classification are actually comparable to methods that specifically train on the binary classification. However, we note that studies such as @moller2016 achieve this high purity rate without redshift information, which our method currently requires.
The vast majority of previous photometric classification studies used simulated datasets to train classifiers. This is largely due to the fact that few homogeneous photometric datasets with large numbers of spectroscopically-classified SNe exist. Most studies that train on simulated datasets use the Supernova Photometric Classification Challenge (SNPCC) training set [@kessler2010results]. The SNPCC dataset consists of 20,000 simulated SNe with $griz$ light curves, generated from templates of Type Ia, Ibc, IIP and IIn SNe (they do not include SLSNe). This dataset was presented as a community-wide classification challenge in preparation for the Dark Energy Survey, and was widely successful, with the top algorithms reaching an average Type Ia SN classification purity of $\approx 80\%$ and completeness of $\approx 95\%$. Works such as @moller2019supernnovaand @moss2018improved have reported average classification accuracies of $\approx 90\%$ for CCSNe classes (similar to our reported accuracies here). Similarly, @lochner2016photometric report an average Type Ia classification accuracy of $\sim84$% using SALT2 light curve features. They further break down the CCSNe subclass into Type Ibc and Type II, where they report accuracies of $\sim63$% and $\sim93$%, respectively. We caution that the SNPCC dataset is not representative of the real diversity we encounter in on-going and future surveys, and should not be used as a benchmark for CCSN classification. In particular, to generate synthetic light curves, @kessler2010results fit well-sampled real light curves from each CCSN class with a Bazin function. Then they stretch Nugent CCSN templates[^4] to match the Bazin light curves. Variations within each class are included from both the sample of templates available and from random color variations derived from the Hubble scatter of Type Ia SNe and the peak luminosity derived from @richardson2002comparative. While the collection of simulated Type Ia SNe likely samples the full phase-space of light curves, the non-Type Ia templates used to build the model light curve were severely limited. For example, only 2 Type IIn SN templates were used to generate 800 template light curves, and only 16 Type Ibc SN templates were used to generate 3,200 light curves. Because of this, we can expect methods that rely on this dataset to overestimate the accuracy of classifications for CCSN classes.
A new classification challenge, PLAsTiCC [@allam2018photometric; @plasticc], is a more realistic simulated dataset that can be used as a benchmark for CCSN classification, although it too largely relies on theoretical models. Recent work by @muthukrishna2019 find an average completeness of $\approx 65\%$ over the five SN classes that we have classified here (although we note that the PLAsTiCC challenge combines Type IIP/L and Type IIn SNe into one class). Our average completeness is significantly higher, at $\approx 77\%$.
Limitations and Future Directions {#sec:future}
=================================
The challenge of photometric classification for optical transients is broad and cannot be solved with one classification method alone. Like all methods, our classification pipeline aims to solve a simplified version of this problem: Given a complete light curve, a redshift, and a list of SN classes, what is the type of a given transient? Here we highlight several improvements that can be made to our pipeline, and more broadly outline outstanding problems in the field of transient classification.
Our pipeline requires a redshift, which simplifies the problem of classification by anchoring the absolute magnitudes of every light curve. In our training set these redshifts were obtained from spectra of the transients and their host galaxies. However, in the on-going and future surveys we expect that spectroscopic redshifts (from the SNe or host galaxies) will be rare. On the other hand, LSST will provide photometric redshifts (photo-$z$) for all galaxies with $m<27.5$ mag, with an expected root-mean-square scatter of $\sigma_z/(1+z)\lesssim 0.05$ for galaxies with $m<25.3$ [@abell2009lsst], and a fraction of outliers of $<10\%$ [@graham2017photometric]. A classification algorithm that can associate a transient to its host galaxy will therefore be able to utilize the photo-$z$ value. We anticipate that the additional uncertainty in the model fits due to the photo-$z$ uncertainty will not be a dominant factor. We additionally note that by including redshift information as a feature (even when doing so indirectly) we have limited the use of our pipeline to surveys of similar depth.
Additionally, our classification pipelines best utilizes full light curves, and are thus most naturally applicable for after-the-fact classification. The most natural use is on the yearly samples of $\sim 10^6$ transients from LSST to enable large-scale population studies, as well as targeted studies of specific subsets (e.g., host galaxies of SLSNe). While our method can work on partial light curves for real-time classification, its performance in this context is yet to be evaluated. Several studies that have explored the specific issue of real-time classification have found that recurrent neural networks perform well for this purpose (e.g., @charnock2017deep [@moller2019supernnova; @muthukrishna2019rapid]).
Our algorithm currently relies exclusively on information derived from the transient light curves (other than the redshift). However, useful contextual information about a SN can be extracted from the host galaxy. For example, SLSNe prefer low metallicity, dwarf galaxies [@lunnan2014hydrogen], other CCSN classes span a wide range of star forming galaxies, and Type Ia SNe are found in both star forming and elliptical galaxies. Simple galaxy features, such as Hubble type, color, and SN offset can be easily incorporated into the classification pipeline (e.g.,@foley2013classifying). This will be explored in follow-up work.
Furthermore, our algorithm is limited to classification within known SN classes (in this case 5 classes). To add additional classes under our current framework, we would need to incorporate new data into the training set and retrain the classification algorithms. Our pipeline is amenable to rapid training, so it is feasible to incorporate more classes in this way. For a more complex classification pipeline (e.g., one involving a large neural network), one could incorporate new classes cheaply using “one-shot” learning [@lv2006camera], in which a classifier learns the characteristics of a new class using very limited examples. However, the addition of new classes will not solve the issue of how to identify unforeseen classes of transients and entirely new phenomena. Such a classifier is challenging to train, since outlier events are (by definition) rare.
Because our original training set is imbalanced and small, we needed to augment our dataset with simulated events drawn from the observed populations. For completeness, we test our best classification pipeline (PCA feature extraction and RF classifier) on the original training set without data augmentation. As expected, we find that we can classify classes with the most samples (Type Ia and II SNe) or those that are well-separated in feature space (SLSNe), as well as or better than our classification pipeline with data augmentation. However, the completeness of the minority classes, like Type Ibc and IIn SNe, falls by $20-40\%$. This is a good indication that data augmentation in the extracted feature space is a potential solution to the imbalanced classes.
Our method neglects the possibility of a biased spectroscopic sample. For example, if the spectroscopic samples contains only the brighter end of the luminosity function for rare transients. In our presumed classification case in which we have access to the full light curves, one can use the full dataset to detect and minimize the effects of selection bias without knowing the true underlying distribution. For example, one can re-weight the importance of each SN in the spectroscopic training sample to better reflect the distribution of features from the full dataset (using, e.g., @huang2007correcting and @cortes2008sample). A detailed study of the effect of observational biases on transient classification is essential, but beyond the scope of this work.
Finally, we note that classification is only the first step in understanding the uncovered transients. Even for the currently rare SLSN class, LSST will discover $\sim 10^4$ events per year [@villar2018superluminous]. Additional data cuts that remove light curves with a minimal *information content* (or those from which we cannot extract physical parameters) may be necessary in order to realistically fit a representative set of light curves.
Conclusions {#sec:conc}
===========
Given increasingly large datasets and limited spectroscopic resources, photometric classification of SNe is a pressing problem within the wide scope of time-domain astrophysics. Here we used the PS1-MDS spectroscopiccaly classified SNe dataset (513 events) to test a number of classification pipelines, varying the features extracted from each light curve, the augmentation method to bolster the training set, and the classification algorithms. We used a flexible analytical model with an iterative MCMC process to model the g$_\mathrm{P1}$r$_\mathrm{P1}$i$_\mathrm{P1}$z$_\mathrm{P1}$ light curves of each event, and to generate posterior distributions. We find that several pipelines (e.g., PCA feature extraction, MVG resampling, and RF classifier) perform well across the 5 relevant SN classes, achieving an average accuracy of about $90\%$ and a Type Ia SN purity of about $95\%$.
Our study is the first to use an empirical dataset to classify multiple classes of SNe, rather than just Type Ia versus non-Type Ia SN classification. Our overall results rival similar pipelines trained on simulated SN datasets, as well as those that utilize only a binary classification. This indicates that we can utilize this approach to generate robust samples of both common and rare SN type (e.g., Type IIn, SLSNe) from LSST.
Finally, we highlight several areas for future exploration and improvement of our classification approach, including the use of contextual information and the possible application to real-time classification. We plan to extend this work and other classification approaches to the full set of PS1-MDS SN photometric light curves in future work.
The Berger Time-Domain Group is supported in part by NSF grant AST-1714498 and NASA grant NNX15AE50G. V.A.V. acknowledges support by the National Science Foundation through a Graduate Research Fellowship. The UCSC team is supported in part by NASA grant NNG17PX03C; NSF grants AST–1518052 and AST–1815935; the Gordon & Betty Moore Foundation; the Heising-Simons Foundation; and by a fellowship from the David and Lucile Packard Foundation to R.J.F. R.L. is supported by a Marie Skłodowska-Curie Individual Fellowship within the Horizon 2020 European Union (EU) Framework Programme for Research and Innovation (H2020-MSCA-IF-2017-794467). Some of the computations in this paper were run on the Odyssey cluster supported by the FAS Division of Science, Research Computing Group at Harvard University. The Pan-STARRS1 Surveys (PS1) and the PS1 public science archive have been made possible through contributions by the Institute for Astronomy, the University of Hawaii, the Pan-STARRS Project Office, the Max-Planck Society and its participating institutes, the Max Planck Institute for Astronomy, Heidelberg and the Max Planck Institute for Extraterrestrial Physics, Garching, The Johns Hopkins University, Durham University, the University of Edinburgh, the Queen’s University Belfast, the Harvard-Smithsonian Center for Astrophysics, the Las Cumbres Observatory Global Telescope Network Incorporated, the National Central University of Taiwan, the Space Telescope Science Institute, the National Aeronautics and Space Administration under Grant No. NNX08AR22G issued through the Planetary Science Division of the NASA Science Mission Directorate, the National Science Foundation Grant No. AST-1238877, the University of Maryland, Eotvos Lorand University (ELTE), the Los Alamos National Laboratory, and the Gordon and Betty Moore Foundation.
[^1]: Three of the 17 SLSNe (PS1-12cil, PS1-10ahf, and PS1-13or) do not have spectroscopic host redshift measurements. @lunnan2018 estimated their redshifts (0.32, 1.10 and 1.52, respectively) from strong rest-frame UV features for the $z>1$ objects and SN Ic-like post-peak features for PS1-12cil.
[^2]: For completeness, we ran our final classifier on these light curves as well and found that 5 of the 6 Type Ia SNe, as well as the one Type II SN were actually correctly classified, albeit with a low classification confidence.
[^3]: <https://github.com/villrv/ps1ml>
[^4]: <https://c3.lbl.gov/nugent/nugent_templates.html>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an analysis of simultaneous particle and field measurements from the ARTEMIS spacecraft which demonstrate that quasi-parallel whistler waves in the solar wind can be generated locally by a bulk flow of halo electrons (whistler heat flux instability). ARTEMIS observes quasi-parallel whistler waves in the frequency range $\sim 0.05 - 0.2 f_{ce}$ simultaneously with electron velocity distribution functions that are a combination of counter-streaming core and halo populations. A linear stability analysis shows that the plasma is stable when there are no whistler waves, and unstable in the presence of whistler waves. In the latter case, the stability analysis shows that the whistler wave growth time is from a few to ten seconds at frequencies and wavenumbers that match the observations. The observations clearly demonstrate that the temperature anisotropy of halo electrons crucially affects the heat flux instability onset: a slight anisotropy $T_{\parallel}/T_{\perp}>1$ may quench the instability, while a slight anisotropy $T_{\parallel} / T_{\perp}<1$ may significantly increase the growth rate. These results demonstrate that heat flux inhibition is strongly dependent on the microscopic plasma properties.'
author:
- Yuguang Tong
- 'Ivan Y. Vasko'
- Marc Pulupa
- 'Forrest S. Mozer'
- 'Stuart D. Bale'
- 'Anton V. Artemyev'
- Vladimir Krasnoselskikh
title: Whistler wave generation by halo electrons in the solar wind
---
Introduction
============
The mechanisms controlling the heat flux in collisionless or weakly-collisional plasmas are of high interest in astrophysics [@Cowie77; @Pistinner1998a; @Roberg-Clark2018a]. In-situ measurements in the solar wind indicate that the heat flux is generally different from the classical Spitzer-H[ä]{}rm prediction [@Feldman75; @Scime94; @Bale13] and apparently constrained by a threshold dependent on local plasma parameters [@Gary1999b; @Gary2000a; @Tong18]. Such observations have motivated many studies on the detailed physics of heat flux inhibition in the solar wind.
In the slow solar wind ($v_{sw}\lesssim 400$ km/s) the electron velocity distribution function can often be approximated by a bi-Maxwellian thermal dense core and a tenuous, suprathermal halo [@Feldman75; @Maksimovic97]. The heat flux is predominantly parallel to the magnetic field and carried by suprathermal electrons. Linear stability analysis shows that the counter-streaming core and halo electrons are capable of driving whistler waves propagating quasi-parallel to the bulk flow of the halo population via the so-called heat flux instability [@Gary75; @Gary1994a; @Gary2000a]. The quasi-linear theory [@Gary77; @Pistinner1998a] and numerical simulations [@Roberg-Clark2018a; @komarov_2018; @Roberg-Clark:2018b] suggest that the scattering of halo electrons by the whistler waves should suppress the heat flux below some threshold value that is in general agreement with the heat flux constraints observed in the solar wind [@Gary1994a; @Gary1999b; @Tong18]. However, the aforementioned experimental studies did not provide measurements of whistler waves accompanying the electron heat flux measurements, and are therefore insufficient to firmly establish the heat flux inhibition by whistler waves in the solar wind.
It is not until recently that careful studies of whistler waves presumably generated by the heat flux instability in freely expanding solar wind have been reported with measurements on Cluster and ARTEMIS spacecraft. [@Lacombe14] reported whistler waves observed along with the heat flux values close to the theoretical threshold given by [@Gary1999b]. [@Stansby16] presented observations of similar whistler waves on ARTEMIS and determined the dependence of the whistler wave dispersion relation on $\beta_{e}$. However, neither study showed that the whistler waves were indeed generated by the heat flux instability in the local plasma, leaving the possibility that whistler waves were generated in a very different plasma by an alternative mechanism and propagated to the spacecraft location. We note that whistler waves in the solar wind can be associated with shocks and stream interaction regions [@LengyelFrey96; @Lin1998a; @Breneman10; @Wilson13], while we focus on whistler waves in the freely expanding solar wind.
In this study we analyze simultaneous particle and wave measurements for data intervals presented by [@Stansby16] and carry out linear stability analysis on electron velocity distribution functions. We find that the observed whistler waves are indeed generated locally by the heat flux instability on a time scale of a few seconds. In this letter we present one of those events, which also demonstrates crucial features of the heat flux instability.
Observations
============
We consider observations of ARTEMIS [@Angelopoulos2011a] on November 9, 2010 for ten minutes around 10:17:00 UT as the spacecraft was in the pristine solar wind about 40 Earth radii upstream of the Earth’s bow shock. We use measurements of the following instruments aboard ARTEMIS: the magnetic fields with 3 second resolution provided by the Flux Gate Magnetometer [@Auster08], the electron velocity distribution function (32 log-spaced energy bins from a few eV up to 25 keV and 88 angular bins) and particle moments (density, bulk velocity and temperatures) with 3 second time resolution provided by the Electrostatic Analyzer [@McFadden08], measurements of three magnetic and electric field components at 128 Hz sampling rate provided by the Search Coil Magnetometer [@LeContel08] and Electric Field Instrument [@Bonnell08].
{width="1.0\linewidth"}
Figure \[fig1\] shows that the solar wind was streaming at about $v_{sw}\sim 320$ km/s, the quasi-static magnetic field was gradually decreasing from $B_0\sim 5$ nT to 3 nT, the plasma density was $n_0\sim 5$ cm$^{-3}$, and the electron temperature was $T_{e}\sim 15$ eV. The ion temperature was $T_i\sim 5$ eV (from the OMNI dataset and not shown). The electron cyclotron frequency $f_{ce}$ was varying from 150 to 90 Hz, the Alfv[é]{}n speed $v_{A}=B_0/(4\pi n_0 m_i)^{1/2}$ from 90 to 30 km/s, while $\beta_{i,e}=8\pi n_0 T_{i,e}/B_0^2\sim 0.5-2$. Over the ten minute interval, continuous electric and magnetic field measurements at 128 samples per second were available. Panel (e) presents the wavelet power spectrum of one of the magnetic field components perpendicular to the quasi-static magnetic field. The enhancement of spectral power density from a few Hz up to about 0.2 $f_{ce}$ corresponds to whistler waves [@Stansby16]. Since the power spectra above 64 Hz cannot be obtained from the search coil magnetic field time series, we also checked the on board FFT power spectra of search coil magnetic fields (not shown) covering 8 Hz-4 kHz and verified that there was no significant power between 64 Hz and $f_{ce}$. Panel (f) presents the spectral coherence between the two magnetic field components perpendicular to the quasi-static magnetic field and indicates a high coherence of the whistler waves. We carry out a spectral polarization analysis following @Santolik03 to determine the obliqueness of whistler waves to the quasi-static magnetic field. Panel (g) presents the cosine of the propagation angle and confirms that whistler waves propagate almost parallel or anti-parallel to the quasi-static magnetic field in accordance with the conclusions of [@Stansby16].
{width="0.53\linewidth"} {width="0.4\linewidth"}
Figure \[fig3\] presents an example of the processed electron velocity distribution function (VDF). The raw electron VDF measured around 10:17:49 UT is corrected for the effect of spacecraft potential and transformed from the spacecraft frame into the solar wind frame using the ion bulk velocity measurements. Panel (a) shows the processed VDF $f(v_\parallel, v_\perp)$ averaged over the gyrophase, where $v_{||}$ and $v_{\perp}$ correspond to velocities parallel and perpendicular to the background magnetic field. The VDF is asymmetric in the direction parallel to the magnetic field with opposite asymmetries below and above a few thousand km/s, indicating counter-streaming of cold and hot electrons. Panel (b) shows VDF cuts $f_{||}$, $f_{\perp}$ and $f_{||}^-$ corresponding to electrons streaming parallel (pitch angles $\alpha\sim 0^{\circ}$), perpendicular ($\alpha\sim 90^\circ$) and anti-parallel ($\alpha\sim 180^\circ$) to the quasi-static magnetic field. Below $\sim$30 eV, $f_{||}^{-}>f_{||}$, consistent with core electrons streaming anti-parallel to the magnetic field. At higher energies, $f_{||}>f_{||}^{-}$ shows that the hotter electrons are streaming in the opposite direction.
The counter-streaming cold and hot electrons persist through the whole ten minutes in Figure \[fig1\]. Is this plasma indeed capable of generating the observed whistler waves? How fast is the instability? What controls the absence of whistler waves before 10:16:00 UT and their later appearance? To address these questions we fit the processed electron VDFs and carry out a linear kinetic stability analysis using the previously developed numerical code [@Tong15].
Analysis
========
During this slow solar wind interval, the electron VDFs are well described by a combination of core and halo populations $f=f_c+f_h$. The core and halo are modelled respectively with drifting bi-Maxwellian and bi-kappa distributions $$\begin{aligned}
f_c&=&A_c\exp\left[-\frac{m_e\left(v_{||}-v_{0c}\right)^2}{2T_{||c}} - \frac{m_e v_{\perp}^2}{2T_{\perp c}}\right],\\
f_h&=&A_h B_{\kappa}\left[1 + \frac{m_e(v_{||}-v_{0h})^2}{(2\kappa-3)T_{||h}}+\frac{m_ev_{\perp}^2}{(2\kappa-3)T_{\perp h}}\right]^{-(\kappa+1)},\end{aligned}$$ where $A_{s}=n_s (m_e/2\pi T_{\perp s}^{2/3}T_{||s}^{1/3})^{3/2}$, $B_{\kappa}=\Gamma(\kappa+1)/(\kappa-3/2)^{3/2}\Gamma(\kappa-1/2)$ and $n_{s}$, $v_{0s}$, $T_{\perp s}$, $T_{||s}$ are densities, bulk velocities and temperatures (parallel and perpendicular to the quasi-static magnetic field ${\bf B}_0$) of the core and halo populations ($s=c,h$). These parameters are estimated by fitting the model to VDF cuts $f_{||}$, $f_{\perp}$ and $f_{||}^-$ using the standard $\chi^2$ minimization method. Following [@Feldman75] the electron current in the solar wind frame is kept zero by restricting the parameters to $n_cv_{0c}+n_hv_{0h}=0$.
Figures \[fig:example\_fits\] (a) and (c) illustrate the fitting procedure, using an electron VDF measured in absence of whistler wave activity at 10:12:11 UT and another VDF in presence of whistler waves at 10:17:49 UT. Panels (a) and (c) present the VDF cuts, the model fits and the best fit parameters. Only data points above the one count level have been used in the fitting procedure. Core electrons make up about 80-85% of the total electron density, the bulk velocity is 100-200 km/s (anti-parallel to ${\bf B}_0$ in the solar wind frame), or about four times larger than the local Alfv[é]{}n speed, the temperature is around 9 eV, and the parallel and perpendicular temperatures are slightly different, $T_{\perp c}/T_{|| c}\sim 1.06$. The halo bulk velocity is about 500-1000 km/s (parallel to ${\bf B}_0$) and the temperature is about 30 eV. The halo population is rather anisotropic in (a) with $T_{\perp h}/T_{\parallel,h}\sim 0.8$ and essentially isotropic in (c) with $T_{\perp h}/T_{\parallel,h}\sim 1.0$.
We address the whistler wave generation by carrying out a linear stability analysis. In the computations we use the model electron VDF with the best fit parameters and isotropic Maxwellian protons with a temperature of 5 eV. The precise shape of the ion distribution function is not critical, because thermal ions do not interact resonantly with the observed whistler waves. We have restricted computations to parallel propagating whistler waves, because counter-streaming core and halo electrons with parameters realistic to the solar wind are known to generate whistler waves propagating only quasi-parallel to the bulk flow of the halo population [@Gary75; @Gary1994a], which is parallel to ${\bf B}_0$ in our case.
Figures \[fig:example\_fits\] (b) and (d) present growth rates and dispersion curves of parallel propagating whistler waves computed for electron VDFs in (a) and (c). In agreement with observations we find whistler waves to be stable for VDF (a), but unstable for VDF (c). In the latter case the linear stability analysis predicts the fastest growing whistler waves at the frequency of 0.05 $f_{ce}$. Although it is in general agreement with the whistler wave spectrum in Figure \[fig1\]e, a careful comparison requires Doppler shifting the plasma frame frequency of 0.05 $f_{ce}$ into the spacecraft frame (see below). Panel (d) shows that the maximum growth rate is about 10$^{-3}f_{ce}$ or $0.5$ s$^{-1}$ in physical units, which corresponds to an e-folding time of about a second. During this time, whistler waves can only propagate a few hundred kilometers, because the phase velocity of the whistler waves is about $c (f/f_{ce})^{1/2}f_{ce}/f_{pe}\sim $500 km/s, where $f$ and $f_{pe}$ are whistler and plasma frequencies (see also [@Stansby16]). This indicates that the observed whistler waves were likely generated locally.
In order to uniquely identify the free energy source driving the whistler waves, we computed growth rates for electron VDFs (a) and (c), but with either 1) core and halo bulk velocities set to zero or 2) temperature-isotropic core and halo. Panels (b) and (d) show that the electron VDFs with zero bulk velocities (blue curves) are stable and can not generate whistler waves. The free energy driving the observed whistler waves is hence provided by the bulk motions of the core and halo or, in other words, by the electron heat flux. The assumption of isotropic core and halo makes VDF (a) unstable, demonstrating thereby that $T_{||h}/T_{\perp h}>1$ acts to suppress and possibly quench the instability [@Gary77].
Figures \[fig:stability\_analysis\] (a)-(d) summarize the results of the fitting of all 183 electron VDFs available over the ten-minute interval. Panel (a) demonstrates that the total electron density derived from the fitting matches (within 5%) the calibrated electron moment densities shown previously in Figure \[fig1\] (a). Panel (b) shows that the core and halo parallel temperatures are steady. Panel (c) demonstrates that the core temperature anisotropy $T_{\perp c}/T_{||c}$ is steady and around 1.1, while the halo is temperature-anisotropic with $T_{\perp h}/T_{||h}\sim 0.8$ before 10:15:00 UT, gradually becoming isotropic at 10:16:00 UT, and remaining nearly isotropic until the end of the interval. Panel (d) shows that the bulk velocity of the core population varies between 2 and 7$v_{A}$. We perform the linear stability analysis on every electron VDF and determine the growth rate $\gamma_{m}$, frequency $f_{m}$ and wavenumber $k_{m}$ of the fastest growing whistler wave. In the spacecraft frame the whistler wave will be observed at a Doppler-shifted frequency $f_{m} + {\bf k}_{m}\mathbf{\bf v}_{sw}$, where ${\bf k}_{m}$ is parallel to the quasi-static magnetic field ${\bf B}_0$.
Panel (e) demonstrates that the Doppler-shifted frequency of the fastest growing whistler wave indeed traces the observed whistler waves. There are no whistler waves before about 10:16:00 UT, while the plasma is stable. Whistler waves suddenly appear around 10:16:00 UT, when the plasma becomes unstable. Around 10:21:00 UT the plasma is stable for a short time interval, and the coherent whistler waves disappear over this interval. The strong correlation between whistler waves and the local plasma stability/instability indicates that the whistler waves are indeed generated locally. Panel (f) strengthens this conclusion by demonstrating that the e-folding time $\gamma_{m}^{-1}$ of the fastest growing whistler wave is from 1 to 10 seconds.
{width="0.45\linewidth"} {width="0.45\linewidth"}
{width="\linewidth"}
The abrupt transition from stable to unstable plasma around 10:16:00 UT coincides with the halo population becoming more isotropic. As we demonstrated in Figure \[fig:example\_fits\], the reason is that the halo temperature anisotropy quenches the whistler heat flux instability. The crucial role of the temperature anisotropy is further demonstrated in Figure \[fig:heatflux\]. Panel (a) presents the electron heat flux $q_e$ normalized to the free streaming heat flux $q_0=1.5 n_e T_e (T_e/m_e)^{1/2}$ versus $\beta_{c||}=8\pi n_cT_{||c}/B_0^2$. At any given $\beta_{c||}$ the heat flux is clearly below a threshold given by $q_{e}/q_0\sim 1/\beta_{c||}$, that is similar to the marginally stable values in literature [@Gary1999a; @Pistinner1998a; @Roberg-Clark2018a; @komarov_2018; @Roberg-Clark:2018b]. However, at a given $q_{e}/q_0$ both stable and unstable VDFs are observed, indicating thereby that some other parameter controls the onset of the whistler wave generation. Panel (b) shows that the halo temperature anisotropy separates stable and unstable VDFs with a similar heat flux value. This re-emphasizes the crucial effect of the halo temperature anisotropy on the heat flux constraints in the solar wind.
{width="\linewidth"}
Discussion and Conclusion
=========================
In-situ observations indicated that whistler waves generated by the heat flux instability highly likely constrain the heat flux in the solar wind [@Feldman75; @Gary1999b; @Tong18]. However, there have been no previous analyses that would prove that whistler waves in the solar wind are actually produced locally by the whistler heat flux instability. In this letter we have presented a careful analysis of simultaneous particle and wave measurements for one of the time intervals in [@Stansby16]. We have performed similar analysis for other [@Stansby16] time intervals and confirmed that whistler waves are generated locally by the heat flux instability in those intervals as well. The presented event has shown that the e-folding growth time of whistler waves can be as short as one second and clearly demonstrated the crucial effect of the halo temperature anisotropy $T_{\perp h}/T_{||h}<1$. In some of the [@Stansby16] events the halo population has $T_{\perp h}/T_{||h}>1$. The linear stability analysis has shown that even a slight $T_{\perp h}/T_{||h}>1$ significantly enhances the growth rate of the heat flux instability, but we stress that the observed temperature anisotropies are insufficient to drive whistler waves purely via the temperature-anisotropy (without core and halo bulk motion) [@Sagdeev60; @Kennel66]. The presented analysis indicates that the sporadic occurrence of whistler waves in the solar wind pointed out by [@Lacombe14] may be due to an interplay between the electron heat flux and the halo temperature anisotropy that may easily quench or enhance the instability. Future statistical studies should carefully address the halo temperature anisotropy in any analysis of the source of whistler waves in the solar wind.
Up to this point we have been focused on the electron heat flux constrained by wave-particle interactions. In fact, Coulomb electron-electron collisions can also affect solar wind electrons and constrain the electron heat flux [@Salem2003; @Bale13; @Pulupa:2014; @Landi:2014a]. The Knudsen number for the observed solar wind $K_n\sim 1 - 1.5$ falls into the collisionless regime [c.f. Figure 2 in @Bale13]. Consistently, the observed heat flux is 30-50% lower than the Spitzer-H[ärm]{} prediction. This implies that the observed heat flux constraint and deviation from the Sptizer-H[ä]{}rm prediction are due to electron scattering by the whistler waves.
Finally, the presented whistler waves are observed in the slow solar wind, where the electron VDF is satisfactorily described by counter-streaming core and halo [@Feldman75; @Maksimovic97]. In the fast solar wind there is an additional anti-sunward propagating strahl population [@Pilipp87; @Stverak09] that do not directly interact with parallel whistler waves driven by whistler heat flux instabilities. Hence we expect the whistler heat flux instabilities to operate in the fast wind as well.
We acknowledge the THEMIS team for the use of data. We thank T. A. Bowen, J. W. Bonnel, J. M. McTiernan and A. Hull for useful discussions. Y. T. and S. D. B. were supported in part by NASA contract NNN06AA01C. I. V. and F. M. were supported by Johns Hopkins University/Applied Physics Lab Contract No. 922613 (Radiation Belt Storm Probes-Electric Fields and Waves).
natexlab\#1[\#1]{}\[1\][[\#1](#1)]{} \[1\][doi: [](http://doi.org/#1)]{} \[1\][[](http://ascl.net/#1)]{} \[1\][[](https://arxiv.org/abs/#1)]{}
, V. 2011, , 165, 3,
, H. U., [Glassmeier]{}, K. H., [Magnes]{}, W., [et al.]{} 2008, , 141, 235,
, S. D., [Pulupa]{}, M., [Salem]{}, C., [Chen]{}, C. H. K., & [Quataert]{}, E. 2013, , 769, L22,
, J. W., [Mozer]{}, F. S., [Delory]{}, G. T., [et al.]{} 2008, , 141, 303,
, A., [Cattell]{}, C., [Schreiner]{}, S., [et al.]{} 2010, Journal of Geophysical Research (Space Physics), 115, A08104,
, L. L., & [McKee]{}, C. F. 1977, , 211, 135,
, W. C., [Asbridge]{}, J. R., [Bame]{}, S. J., [Montgomery]{}, M. D., & [Gary]{}, S. P. 1975, , 80, 4181,
, S. P., & [Feldman]{}, W. C. 1977, , 82, 1087,
, S. P., [Feldman]{}, W. C., [Forslund]{}, D. W., & [Montgomery]{}, M. D. 1975, , 80, 4197,
, S. P., & [Li]{}, H. 2000, , 529, 1131,
, S. P., [Neagu]{}, E., [Skoug]{}, R. M., & [Goldstein]{}, B. E. 1999, , 104, 19843,
, S. P., [Scime]{}, E. E., [Phillips]{}, J. L., & [Feldman]{}, W. C. 1994, , 99, 23,
, S. P., [Skoug]{}, R. M., & [Daughton]{}, W. 1999, Physics of Plasmas, 6, 2607,
, C. F., & [Petschek]{}, H. E. 1966, , 71, 1,
Komarov, S., Schekochihin, A. A., Churazov, E., & Spitkovsky, A. 2018, Journal of Plasma Physics, 84, 905840305,
, C., [Alexandrova]{}, O., [Matteini]{}, L., [et al.]{} 2014, , 796, 5,
Landi, S., Matteini, L., & Pantellini, F. 2014, The Astrophysical Journal Letters, 790, L12
, O., [Roux]{}, A., [Robert]{}, P., [et al.]{} 2008, , 141, 509,
, D., [Hess]{}, R. A., [MacDowall]{}, R. J., [et al.]{} 1996, , 101, 27555,
, N., [Kellogg]{}, P. J., [MacDowall]{}, R. J., [et al.]{} 1998, , 103, 12023,
, M., [Pierrard]{}, V., & [Riley]{}, P. 1997, , 24, 1151,
, J. P., [Carlson]{}, C. W., [Larson]{}, D., [et al.]{} 2008, , 141, 477,
, W. G., [Miggenrieder]{}, H., [Montgomery]{}, M. D., [et al.]{} 1987, , 92, 1075,
, S. L., & [Eichler]{}, D. 1998, , 301, 49,
, M. P., [Bale]{}, S. D., [Salem]{}, C., & [Horaites]{}, K. 2014, Journal of Geophysical Research (Space Physics), 119, 647,
Roberg-Clark, G. T., Drake, J. F., Reynolds, C. S., & Swisdak, M. 2018, Phys. Rev. Lett., 120, 035101,
Roberg-Clark, G. T., Drake, J. F., Swisdak, M., & Reynolds, C. S. 2018, The Astrophysical Journal, 867, 154
, R. Z., & [Shafranov]{}, V. D. 1960, Soviet JETP, 39, 181
Salem, C., Hubert, D., Lacombe, C., [et al.]{} 2003, The Astrophysical Journal, 585, 1147
, O., [Parrot]{}, M., & [Lefeuvre]{}, F. 2003, Radio Science, 38, 1010,
, E. E., [Bame]{}, S. J., [Feldman]{}, W. C., [et al.]{} 1994, , 99, 23,
, D., [Horbury]{}, T. S., [Chen]{}, C. H. K., & [Matteini]{}, L. 2016, , 829, L16,
, Y., [Bale]{}, S. D., [Chen]{}, C. H. K., [Salem]{}, C. S., & [Verscharen]{}, D. 2015, , 804, L36,
, Y., [Bale]{}, S. D., [Salem]{}, C., & [Pulupa]{}, M. 2018, ArXiv e-prints.
, [Š]{}., [Maksimovic]{}, M., [Tr[á]{}vn[í]{}[č]{}ek]{}, P. M., [et al.]{} 2009, Journal of Geophysical Research (Space Physics), 114, A05104,
, L. B., [Koval]{}, A., [Szabo]{}, A., [et al.]{} 2013, Journal of Geophysical Research (Space Physics), 118, 5,
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work, we propose a class of $SU(N)$ Yang-Mills models, with adjoint Higgs fields, that accept BPS center vortex equations. The lack of a *local* magnetic flux that could serve as an energy bound is circumvented by including a new term in the energy functional. This term tends to align, in the Lie algebra, the magnetic field and one of the adjoint Higgs fields. Finally, a reduced set of equations for the center vortex profile functions is obtained (for $N=2,3$). In particular, $Z(3)$ BPS vortices come in three colours and three anticolours, obtained from an ansatz based on the defining representation and its conjugate.'
author:
- |
L. E. Oxman\
\
Instituto de Física, Universidade Federal Fluminense,\
Campus da Praia Vermelha, Niterói, 24210-340, RJ, Brazil.
title: 'BPS center vortices in nonrelativistic $SU(N)$ gauge models with adjoint Higgs fields'
---
**Keywords**:\
**Pacs**: 11.15.-q, 11.10.Lm, 11.15.Kc
Introduction
============
Topological solitons are present in many areas of Physics. Some well-known examples are: kinks in polyacetylene [@Sch], vortices in type II superconductors, skyrmions in magnetic systems [@Fradkin], and skyrmions to describe baryons in flavour symmetric models [@skyrme; @skyrme-rational]. To gain information about these objects, it is important to identify a critical point where a BPS bound is obtained. Namely, a point where the energy can be written as a sum of squares plus the topological charge of the field configuration. As continuous field deformations cannot modify this charge, setting the squares to zero leads to a set of (BPS) equations whose solutions are absolute minima in the given topological sector. In this process, the equations are reduced to first order, which facilitates analytical and numerical studies of these systems. In addition, BPS multisoliton solutions with a given total charge have the same energy, so the forces between BPS solitons vanish. For these reasons, the critical point provides a nice reference to introduce perturbations and study the soliton dynamics [@Manton].
Topological solitons are also important in effective descriptions of the strong interactions. Abelian Higgs models have been proposed to describe the $q\bar{q}$ potential [@Baker; @Baker1] and the interaction among three quarks [@MSu]-[@C1]. In refs. [@deVega]-[@HV], center vortices were accommodated in $SU(N)$ Yang-Mills models with $N$ adjoint Higgs fields; these objects can describe the N-ality properties of the confining string [@Konishi-Spanu]. Recently, we proposed a class of flavour symmetric models supporting not only the confining string between a $q\bar{q}$ colourless pair of external quarks, but also other possible excited states [@conf-qg]. Among them, $qg\bar{q}'$ hybrid mesons [@hybrids]-[@DuE], formed for example by a red/anti-green pair of quarks bound by an anti-red/green valence gluon. While the normal string is a center vortex of the effective model, the excited string is formed by a pair of center vortices interpolated by a monopole, which is identified with a confined valence gluon.
The topology and classification of center vortices have been analyzed in ref. [@Konishi-Spanu], when a general compact gauge group $G$ is broken down to its center. The roots of the Lie algebra and the weights of their representations play an important role, as occurs when characterizing non Abelian monopoles [@GNO]. BPS equations for non Abelian vortices have been obtained in refs. [@David1]-[@David], for a review, see refs. [@vortices-d; @notes].
In ref. [@conf-qg], we proposed a Lorentz invariant flavour symmetric model that is expected to contain center vortices, as it possesses the proper SSB pattern and topology. The problem is that the field equations are mathematically difficult. They can only be solved by following numerical methods. With the aim of exploring the usual tools to understand topological objects, in this work we shall look for models accepting BPS center vortices, governed by [*first order*]{} field equations. As an intermediate step, we shall simplify the content of the flavour symmetric model, which is based on $N^{2}-1$ adjoint Higgs fields that form a local Lie basis at the nontrivial vacua of the Higgs potential. Observing that the essential features of the Lie algebra can be captured by a reduced set of fields and conditions, labelled by the simple roots, a Lorentz invariant model that for $N\geq 3$ has a simplified field content will be obtained. The Higgs potential will be such that its minimization returns a set of conditions that essentially define a Chevalley basis. This model has the same SSB pattern and topology than the former. Next, we shall make an extension to obtain a model that accepts BPS center vortices. In this process, we can anticipate some peculiarities. Generally, BPS equations are derived by working on the energy functional to obtain a bound (for an alternative approach, see ref. [@B-GP]). For $U(1)$ vortices, the bound is given by the magnetic flux. This is a topological term that can be written locally, by means of a flux density, so it can indeed arise by working on the energy, which is a local functional. On the other hand, for center vortices, the flux concept is replaced by the asymptotic behavior of the gauge invariant Wilson loop, a nonlocal object that may not appear in the energy calculation. Then, the search for BPS equations in $SU(N)\to Z(N)$ SSB models led us to consider the introduction of a nonrelativistic interaction term that tends to align, in the Lie algebra, the magnetic field along one of the adjoint Higgs fields. This in turn implied a different type of bound. After completing the squares, the energy is always greater than or equal to zero. Thus, BPS center vortices are nonrelativistic objects characterized by an exact compensation between the positive definite part of the energy functional (kinetic energy plus Higgs potential) and the Lie algebra alignment contribution.
In this regard, two comments are in order: i) The reason for considering the intermediate step is that, for $N \geq 3$, the direct inclusion of the alignment term in the flavour symmetric model would lead, after completing the squares, to too many (possibly incompatible) conditions to saturate the bound. ii) Because of Lorentz symmetry breaking, and rotational symmetry breaking in $3+1$ dimensions, the BPS models are not directly physically relevant. However, the presence of a BPS point in the extended parameter space could serve as a check for the numerical analysis, when moving away from the physically relevant non BPS Lorentz invariant confining models.
The general BPS solution will be written in terms of a set of profile functions and a mapping $R(S)$ in the adjoint representation of $SU(N)$. The mapping $S\in SU(N)$ contains information about the asymptotic Wilson loop and the possible defects at the vortex guiding centers, which determine the profile behaviours. Because of the model’s topology, a given phase $S_{0}(\varphi)$, defined close to and around a vortex guiding center, can be extended to different asymptotic phases $S_{a}(\varphi)$, where $S_{a}(\varphi+2\pi)=e^{i2\pi z_{a}/N}S_{a}(\varphi)$, $z_{a}\in\mathbb{Z}$. The $Z(N)$ charge is due to the fact that the different extensions are related by $(z_{a'}-z_{a})/N\in\mathbb{Z}$. For the same reason, a given $S_{a}$ can be matched with different phases $S_{0}$, with their respective pointlike defects. When leaving the critical point, by lowering the alignment interaction term, some of these extensions will become unstable. For example, for vanishing $Z(N)$-charge the defect can be avoided, and the lowest energy solution will simply correspond to a trivial regular gauge transformation of the SSB vacua. For $Z(N)$ charge $\pm1$, we shall discuss the BPS solutions that are expected to be related to the stable $Z(2)$ and $Z(3)$ noncritical center vortices.
The article is organized as follows. In section 2, we construct the simplified $SU(N)$ model and discuss the possible vacua. In section 3, we obtain the bounds and the set of BPS equations (for $N=2,3$). Some properties of the field parametrization are discussed in section 4. Section 5 is devoted to obtaining information about the BPS solutions and discussing the BPS center vortex. Finally, in section 6, we present our conclusions.
Models with $SU(N)\to Z(N)$ SSB
===============================
In order to support center vortices, we are interested in driving a phase where the gauge symmetry is spontaneously broken down to $Z(N)$. For example, in ref. [@conf-qg], we introduced a model displaying a flavour symmetry. That is, we considered the energy functional[^1], $$\begin{aligned}
E & = & \int d^{3}x\,\left(\frac{1}{2}\langle B_{i}\rangle^{2}+\frac{1}{2}\langle D_{i}\psi_{A}\rangle^{2}+V_{{\rm Higgs}}(\psi_{A})\right)\;,\label{mflavour}\end{aligned}$$ where $B_i$ is the chromomagnetic field, $D_i=\partial_i-ig[A_i,~]$, and the potential for the hermitian adjoint Higgs fields $\psi_{A}$, $A=1,\dots,d=N^2-1$, is given by, $$\begin{aligned}
V_{{\rm Higgs}} & = & c+\frac{\mu^{2}}{2}\,\langle\psi_{A},\psi_{A}\rangle + \frac{\kappa}{3}\, f_{ABC}\langle (-i)[\psi_{A}, \psi_{B}],\psi_{C}\rangle+\frac{\lambda}{4}\,\langle [\psi_{A},\psi_{B}],[\psi_{A},\psi_{B}]\rangle\;,\nonumber \\
\label{Vf}\end{aligned}$$ where $f_{ABC}$ are structure constants of the $\mathfrak{su}(N)$ Lie algebra. At $\mu^{2}=\frac{2}{9}\frac{\kappa^{2}}{\lambda}$, $\kappa<0$, we can write, $$\begin{aligned}
V_{{\rm Higgs}} & = & \frac{\lambda}{4}\,\langle\Psi_{AB}\rangle^{2}\makebox[.5in]{,}\Psi_{AB}=f_{ABC}\, v_{c}\,\psi_{C}+i[\psi_{A},\psi_{B}]\;,\end{aligned}$$ $$v_{c}=-\frac{\kappa}{2\lambda}\pm\sqrt{\left(\frac{\kappa}{2\lambda}\right)^{2}-\frac{\mu^{2}}{\lambda}}=-\frac{2\kappa}{3\lambda}\;,$$ after adjusting $c$, so that the potential energy for vacuum configurations vanishes. The space of vacua ${\cal M}$ is obtained from the conditions $\Psi_{AB}=0$, i.e., $$[\psi_{A},\psi_{B}]=if_{ABC}\, v_{c}\,\psi_{C}\;.$$ This encompasses the trivial symmetric point $\psi_{A}=0$, separated by a potential barrier from the nontrivial points. Of course, starting from a nontrivial point $\psi_{A}\in{\cal M}$, we can generate a continuum $S\psi_{A}S^{-1}$, $S\in SU(N)$, that is also in ${\cal M}$. In addition, the only transformations that leave these points invariant are $S\in Z(N)$, so they correspond to $SU(N)\to Z(N)$ SSB vacua. For $N\geq3$, the SSB points can be divided into a pair of distinct sets, separated by a potential barrier, corresponding to the defining representation and its conjugate, $$\psi_{A}=ST_{A}S^{-1}\makebox[.5in]{,}\psi_{A}=S(-T_{A})^{\ast}S^{-1}\;.$$ For $N=2$, this pair collapses into a single component, as a matrix $S_{c}\in SU(2)$ exists such that $(-T_{A})^{\ast}=S_{c}T_{A}S_{c}^{-1}$, $A=1,2,3$.
Although the model in eq. (\[mflavour\]) contains center vortices, we did not succeed in taking it as a starting point to obtain BPS equations (for $N\geq 3$). The presence of too many fields ultimately leads to incompatible conditions to saturate the bound. For this reason, in the next section, we shall look for a simplified model. Instead of the previous $N^2-1$ hermitian fields, we introduce $N-1$ hermitian and $N-1$ complex adjoint Higgs fields. For $N \geq 3$, this will result in a simpler set of fields and conditions to define the SSB vacua. This, together with the “alignment” term introduced in section \[BPSmodels\], will finally lead to a set of compatible BPS equations.
Simplified model
----------------
Let us consider hermitian variables, $\psi_{q}$, $q=1,\dots,r=N-1$, and complex variables $\zeta_{\alpha}$, labelled by the positive simple roots $\vec{\alpha}_{q}$ ($\vec{\alpha}_{1}<\vec{\alpha}_{2}\dots<\vec{\alpha}_{r}$). The conditions, $$[\psi_{q},\psi_{p}]=0\makebox[.5in]{,}v_{c}\,\vec{\alpha}|_{q}\,\zeta_{\alpha}-[\psi_{q},\zeta_{\alpha}]=0\;,\label{cond1}$$ contain most of the relevant structure of the Lie algebra. For nontrivial fields $\zeta_{\alpha}$, we can imply,
- i\) The fields $\psi_{q}$ are nontrivial, as their sizes are fixed by the eigenvalues $v_{c}\,\vec{\alpha}|_{q}$. They are also linearly independent: if there is a combination $\gamma^{q}\,\psi_{q}=0$, then using eq. (\[cond1\]) we get $\vec{\gamma}\cdot\vec{\alpha}=0$, for every simple root, so that $\vec{\gamma}=0$.
- ii\) $[\zeta_{\alpha},\zeta_{\alpha}^{\dagger}]$ is in the Cartan subalgebra generated by the fields $\psi_{q}$.
- iii\) As the positive (negative) roots can be written as a linear combination of the simple roots, with nonnegative (nonpositive) integer coefficients, any root vector is proportional to an appropriate chain of operations of the form $[\zeta_{\alpha},\zeta_{\alpha'}]$ ($[\zeta_{\alpha}^{\dagger},\zeta_{\alpha'}^{\dagger}]$).
- iv\) As the difference of a pair of positive simple roots cannot be a root, we have $[\zeta_{\alpha},\zeta_{\alpha'}^{\dagger}]=0$.
However, considering a potential whose minimization only leads to the conditions in eq. (\[cond1\]) would not be the desired one. Given a nontrivial solution $(\psi_{q}$, $\zeta_{\alpha})$, the replacement $\zeta_{\alpha}\to t\,\zeta_{\alpha}\,$, $t\in\mathbb{R}$, would also lead to a solution. Then, the interesting SSB initial point could be continuously moved to $(\psi_{q},0)$, and then $\psi_{q}$ could be continuously moved to $0$, always staying in the space of vacua ${\cal M}$. That is, there would be no potential barrier between the interesting configurations and the trivial one. This will be corrected by including a term in the potential to avoid, after minimization, the possibility of moving the fields $\zeta$ to zero, when we start with a SSB point. For this purpose, let us consider the additional condition, $$\sum_{q}s_{q}\,\left(v_{c}\,\vec{\alpha}_{q}\cdot\vec{\psi}-[\zeta_{\alpha_{q}},\zeta_{\alpha_{q}}^{\dagger}]\right)=0\makebox[.5in]{,}\vec{\alpha}\cdot\vec{\psi}=\vec{\alpha}|_{q'}\psi_{q'}\;,\label{addc}$$ where $s_{q}$ takes values $+1$ or $-1$. Now, we consider a solution to eq. (\[cond1\]), and recall that given linearly independent fields $\psi_{q}$ it is always possible to introduce unique elements $\mathscr{H}_{\alpha}$ such that, $$\langle\mathscr{H}_{\alpha},\psi_{q}\rangle=v_{c}\,\vec{\alpha}|_{q}\;.\label{varia}$$ As is well known [@Humphreys]-[@Giorgi], these variables satisfy, $$[\zeta_{\alpha},\zeta_{\alpha}^{\dagger}]=\langle\zeta_{\alpha},\zeta_{\alpha}\rangle\,\mathscr{H}_{\alpha}\;.\label{Hal1}$$ Using this information in eq. (\[addc\]) and projecting with $\mathscr{H}_{\alpha_{p}}$, we get, $$\begin{aligned}
\lefteqn{\sum_{q}s_{q}\left(v_{c}\,\vec{\alpha}_{q}|_{q'}\,\langle\mathscr{H}_{\alpha_{p}},\psi_{q'}\rangle-\langle\zeta_{\alpha_{q}},\zeta_{\alpha_{q}}\rangle\,\langle\mathscr{H}_{\alpha_{p}},\mathscr{H}_{\alpha_{q}}\rangle\right)=}\nonumber \\
& & =\sum_{q}s_{q}\left(v_{c}^{2}\,\vec{\alpha}_{q}\cdot\vec{\alpha}_{p}-\langle\zeta_{\alpha_{q}},\zeta_{\alpha_{q}}\rangle\,\langle\mathscr{H}_{\alpha_{p}},\mathscr{H}_{\alpha_{q}}\rangle\right)=0\;.\label{term1}\end{aligned}$$
From the Lie algebra internal product and the mapping $\vec{\alpha}\to\mathscr{H}_{\alpha}$, an internal product on the root space can be defined [@Humphreys]-[@Giorgi], $$\langle\vec{\alpha},\vec{\alpha}\,'\rangle\equiv\langle\mathscr{H}_{\alpha},\mathscr{H}_{\alpha'}\rangle\;.$$ These quantities are strongly constrained. In particular, $$2\,\langle\vec{\alpha},\vec{\alpha}\,'\rangle/\langle\vec{\alpha}\,',\vec{\alpha}\,'\rangle\in\mathbb{Z}\;,\label{Ci}$$ are the so-called Cartan integers, which determine the geometry of the root lattice. They do not depend on the Cartan basis, coinciding with, $$2\,(\vec{\alpha}\cdot\vec{\alpha}\,')/(\vec{\alpha}\,'\cdot\vec{\alpha}\,')\;,$$ which corresponds to (\[Ci\]), when computed with an orthogonal basis $\psi_{q}$, $\langle\psi_{q},\psi_{p}\rangle=v_{c}^{2}\,\delta_{qp}$. Note that in this case, $\mathscr{H}_{\alpha}=(1/v_{c})\,\vec{\alpha}|_{q}\psi_{q}$. In addition, for $\mathfrak{su}(N)$, the lengths of the roots are equal, $\langle\vec{\alpha},\vec{\alpha}\rangle=c$, $\vec{\alpha}\cdot\vec{\alpha}=1/N$. Then, using this information, eq. (\[term1\]) implies, $$\sum_{q}s_{q}\,\left(v_{c}^{2}\,\vec{\alpha}_{q}\cdot\vec{\alpha}_{p}-Nc\,\langle\zeta_{\alpha_{q}},\zeta_{\alpha_{q}}\rangle\,\vec{\alpha}_{q}\cdot\vec{\alpha}_{p}\right)=0\;.$$ This is valid for any basis element $\vec{\alpha}_{p}$, that is, $$\sum_{q}s_{q}\,\left(v_{c}^{2}-Nc\,\langle\zeta_{\alpha_{q}},\zeta_{\alpha_{q}}\rangle\right)\,\vec{\alpha}_{q}=0\;,$$ and as the simple roots $\vec{\alpha}_{q}$ are linearly independent, we get, $$\langle\zeta_{\alpha_{q}},\zeta_{\alpha_{q}}\rangle=\frac{v_{c}^{2}}{Nc}\;.\label{prot}$$ This means that if we define the space of vacua ${\cal M}$ by means of the conditions (\[cond1\]) and (\[addc\]), a nontrivial SSB point $(\psi_{q},\zeta_{\alpha})\in{\cal M}$ cannot be continuously moved to the trivial solution $(0,0)$, always staying in ${\cal M}$. In effect, starting with nontrivial fields $\zeta_{\alpha}$ implies linearly independent fields $\psi_{q}$, and this in turn leads to nontrivial $\mathscr{H}_{\alpha}$ that protect the size of $\zeta_{\alpha}$ through eq. (\[prot\]). In other words, a potential whose minimization gives the conditions (\[cond1\]) and (\[addc\]) has a barrier between the SSB vacua and the trivial one.
It is convenient to introduce a model without referring to a particular convention for the simple roots. The field $\vec{\psi}$ with $r=N-1$ components, such that $\vec{\psi}|_{q}=\psi_{q}$, can be expanded either in terms of the simple roots $\vec{\alpha}_{q}$ or the $\vec{\lambda}^{q}$ basis satisfying, $$\vec{\alpha}_{q}\cdot\vec{\lambda}^{p}=\delta_{q}^{\, p}\makebox[.5in]{,}\vec{\alpha}_{q}|_{p}\,\vec{\lambda}^{q}|^{p'}=\delta_{p}^{\, p'}\;,\label{alam}$$ $$\vec{\psi}=\vec{\alpha}_{q}\,\phi^{q}\makebox[.8in]{{\rm with,}}\phi^{q}=\vec{\lambda}^{q}\cdot\vec{\psi}\;,$$ $$\vec{\psi}=\vec{\lambda}^{q}\,\phi_{q}\makebox[.8in]{{\rm with,}}\phi_{q}=\vec{\alpha}_{q}\cdot\vec{\psi}\;,$$ $$\psi_{p}=\vec{\alpha}_{q}|_{p}\,\phi^{q}=\vec{\lambda}^{q}|_{p}\,\phi_{q}\;.$$ For $\mathfrak{su}(N)$, the $\vec{\lambda}^{q}$ basis is given by, $$\vec{\lambda}^{q}=2N\,\vec{\Lambda}^{q}\;,$$ where $\vec{\Lambda}^{q}$ are the fundamental weights (see appendix A).
The relation between the different components is, $$\phi_{q}=A_{qp}\,\phi^{p}\makebox[.8in]{,}\phi^{q}=A^{qp}\,\phi_{p}\;,$$ $$A_{qp}=\vec{\alpha}_{q}\cdot\vec{\alpha}_{p}\makebox[.5in]{,}A^{qp}=\vec{\lambda}^{q}\cdot\vec{\lambda}^{p}\;.$$ For $\mathfrak{su}(N)$, the quantities $C_{qp}=2NA_{qp}$ are the elements of the Cartan matrix, which define the natural product in the root space.
With these definitions, together with $\zeta_{q}=\zeta_{\alpha_{q}}$, the conditions in eqs. (\[cond1\]) and (\[addc\]) become, $$[\phi_{q},\phi_{p}]=0\makebox[.5in]{,}v_{c}\,\delta_{(p)}^{q}\,\zeta_{(p)}-[\phi^{q},\zeta_{p}]=0\;,\label{conda}$$ $$\sum_{q}s_{q}\,\left(v_{c}\,\phi_{q}-[\zeta_{(q)},\zeta_{(q)}^{\dagger}]\right)=0\;,$$ which can be obtained by minimizing the Higgs potential, $$V_{{\rm Higgs}}=\frac{1}{2}\langle\Phi,\Phi\rangle+\langle Z_{q},Z_{q}\rangle+{\cal R}\;,\label{potsim}$$ where, $$\Phi=\sum_{q}s_{q}\Phi_{q}\makebox[.5in]{,}\Phi_{q}=\sqrt{\gamma}\left(v_{c}\,\phi_{q}-[\zeta_{(q)},\zeta_{(q)}^{\dagger}]\right)\;,$$ $$Z_{q}=\sqrt{\gamma_{z}}\left(v_{c}\,\zeta_{q}-[\phi^{(q)},\zeta_{(q)}]\right)\;,$$ $${\cal R}=\gamma_{r}\sum_{q\neq p}\left(\frac{1}{2}\langle[\phi_{q},\phi_{p}]\rangle^{2}+\langle[\phi^{q},\zeta_{p}]\rangle^{2}\right)\;.$$ Now, noting that, $$\langle D_{i}\psi_{q},D_{i}\psi_{q}\rangle=A^{qp}\langle D_{i}\phi_{q},D_{i}\phi_{p}\rangle\;,$$ we initially propose the model, $$E=\int d^{3}x\,\left(\frac{1}{2}\langle B_{i}\rangle^{2}+\frac{1}{2}\langle D_{i}\phi_{q},D_{i}\phi^{q}\rangle+\langle D_{i}\zeta_{q},D_{i}\zeta_{q}\rangle+V_{{\rm Higgs}}\right)\;.\label{Emod}$$ Here, the space of vacua ${\cal M}$ is given by the trivial point $\phi_{q}=0$, $\zeta_{q}=0$, separated by a potential barrier from the SSB points. For $N\geq3$, the latter can be separated into the sets, $$\phi_{q}=S(\vec{\alpha}_{q}\cdot\vec{H})S^{-1}\makebox[.5in]{,}\zeta_{q}=SE_{\alpha_{q}}S^{-1}\label{defi}$$ $$\phi_{q}=S(-\vec{\alpha}_{q}\cdot\vec{H})S^{-1}\makebox[.5in]{,}\zeta_{q}=S(-E_{\alpha_{q}}^{T})S^{-1}\;,\label{conju}$$ where $E_{\alpha}^{T}$ is the transpose of $E_{\alpha}$, and $H_{q}$, $E_{\alpha}$ are elements of a Cartan basis (see appendix A). For $N=2$, the sets in (\[defi\]) and (\[conju\]) are equal.
Nonrelativistic models with BPS center vortex equations {#BPSmodels}
=======================================================
As is well known, center vortices are characterized by a center element $$\mathfrak{z}=e^{i2\pi z/N}I\in SU(N)$$ such that, for a path linking the vortex and contained in the asymptotic region, the Wilson loop gives, $$W[A]=\mathfrak{z}\;.\label{Wzeta}$$ The center vortex has a $Z(N)$ charge given by $z$, defined modulo $N$. In particular, this is the case when in an asymptotic region $r>r_{m}$ the gauge field is given by, $$A_{i}=\frac{1}{g}\partial_{i}\varphi\,\vec{\beta}\cdot\vec{H}\makebox[.5in]{,}e^{i2\pi\,\vec{\beta}\cdot\vec{H}}=\mathfrak{z}\;,\label{poss-b}$$ where $r$ and $\varphi$ are polar coordinates with respect to the vortex axis. The possible magnetic weights $\vec{\beta}$ satisfy $$\vec{\beta}\cdot\vec{\alpha}\in Z\;,\label{qc}$$ for every root $\vec{\alpha}$. The solutions to eq. (\[qc\]) are [@GNO], [@Fidel2; @Konishi-Spanu], $$\vec{\beta}=2N\vec{w}\;,\label{solution}$$ where $\vec{w}$ are the weights of the different representations. The minimum charge center vortices ($z=\pm1$) can be labelled by the weights of the defining representation and its conjugate [@Konishi-Spanu].
In the asymptotic region, if a Higgs field takes the value $X_{0}$, $\langle X_{0},X_{0}\rangle=v_{c}^{2}$ at $\varphi=0$ then, on the circle at infinity, the non Abelian phase will accompany the $A_{i}$ behaviour in eq. (\[poss-b\]) as follows[^2], $$X=SX_{0}S^{-1}\makebox[.5in]{,}S=e^{i\varphi\,\vec{\beta}\cdot\vec{T}}\;.\label{asyr}$$
Now, we would like to propose models accepting BPS center vortex equations for $SU(2)$ and $SU(3)$. To simplify the discussion, let us consider planar systems, replacing $d^{3}x\to d^{2}x=dx_{1}dx_{2}$, $B_{3}\to B$, and taking $B_{1}=B_{2}=0$. Initially, we note that the type of models we have discussed so far cannot accept a BPS bound. Indeed, this would be the case in any model whose energy functional only vanishes for vacuum configurations, while on the space of field configurations $\{A,X\}_{\mathfrak{z}}$, with a given nontrivial asymptotic behaviour labelled by $\mathfrak{z}$, it is strictly positive. In this case, to obtain BPS center vortex equations, the energy functional should be bounded by a nonzero term verifying: i) gauge invariance, ii) it assumes a fixed value on the space$\{A,X\}_{\mathfrak{z}}$, that only depends on $\mathfrak{z}$ (topological), iii) as the bound would be derived by working on the energy density, it should have the form $\int d^{2}x\,\rho$ (locallity). While the Wilson loop verifies i) and ii), it is a nonlocal object that cannot arise in the calculation. On the other hand, while $$\int d^{2}x\,\langle\eta,B\rangle\makebox[.5in]{,}B=\partial_{1}A_{2}-\partial_{2}A_{1}-ig[A_{1},A_{2}]\;,\label{termc}$$ with $\eta$ an adjoint field, satisfies i) and iii), it does not satisfy ii). This would be a boundary term for homogeneous $\eta$ and those Abelian-like fields in $\{A,X\}_{\mathfrak{z}}$ such that $B=\partial_{1}A_{2}-\partial_{2}A_{1}$ on the *whole* plane. Then, the search for BPS center vortices should consider a modified class of models where configurations in $\{A,X\}_{\mathfrak{z}}$ do not necessarily have strictly positive energy.
For example, we will see that the model (\[Emod\]) could be reorganized as a sum of squares plus a term of the form (\[termc\]). As this term does not satisfy ii), setting the squares to zero will not produce, in a given sector $\{A,X\}_{\mathfrak{z}}$, solutions to the field equations associated with (\[Emod\]). Then, it is natural to try a modified model, $$E=\int d^{2}x\,\left(\frac{1}{2}\langle B\rangle^{2}+\frac{1}{2}\langle D_{i}\phi_{q},D_{i}\phi^{q}\rangle+\langle D_{i}\zeta_{q},D_{i}\zeta_{q}\rangle+V_{{\rm Higgs}}-\langle\eta,B\rangle\right)\;.\label{energy}$$ Here, we have included the gauge invariant $\langle\eta,B\rangle$-interaction, that tends to align $B$ along $\eta$ in the Lie algebra. The field $\eta$ will be an appropriate combination of the adjoint Higgs fields, to be determined in order for the model to accept BPS center vortex equations. At the critical point, we shall see that in spite of the last term in eq. (\[energy\]), this energy functional satisfies $E\geq0$. For BPS solutions, the contribution originated from the positive definite terms will be exactly compensated by the energy lowering due to the Lie algebra alignment between magnetic and Higgs fields. Thus, the topologically nontrivial BPS center vortices will have vanishing energy. We note that with this term the planar model becomes nonrelativistic although it continues to be isotropic in $2+1$ dimensions[^3].
Let us derive the fundamental property to discuss BPS bounds. Using the ciclicity of the internal product, $$\langle X,[Y,Z]\rangle=\langle[X,Z^{\dagger}],Y\rangle\;,\label{mix1}$$ as $A_{i}$ is hermitian, we have, $$\begin{aligned}
\langle D_{i}X,Y\rangle & = & \langle\partial_{i}X-ig[A_{i},X],Y\rangle\nonumber \\
& = & \partial_{i}\langle X,Y\rangle-\langle X,\partial_{i}Y\rangle+ig\langle[A_{i},X],Y\rangle\nonumber \\
& = & \partial_{i}\langle X,Y\rangle-\langle X,D_{i}Y\rangle\;,\label{porpart}\end{aligned}$$ $$\begin{aligned}
\langle D_{i}X,Y\rangle+\langle X,D_{i}Y\rangle & = & \partial_{i}\langle X,Y\rangle\;.\label{psipsi}\end{aligned}$$ Now, defining, $$D=D_{1}+iD_{2}\;,$$ we note that, $$\begin{aligned}
& & \langle DX,DX\rangle=\langle D_{1}X+iD_{2}X,D_{1}X+iD_{2}X\rangle\nonumber \\
& & =\langle D_{1}X,D_{1}X\rangle+\langle D_{2}X,D_{2}X\rangle-i\langle D_{2}X,D_{1}X\rangle+i\langle D_{1}X,D_{2}X\rangle\;.\end{aligned}$$ In addition, as $B$ is hermitian, $$\langle X,[B,X]\rangle=\langle[X,X^{\dagger}],B\rangle=\langle B,[X,X^{\dagger}]\rangle\;.$$ This together with eq. (\[porpart\]) and, $$\left[D_{\mu},D_{\nu}\right]X=-ig[F_{\mu\nu},X]\;,$$ which is obtained from the Jacobi identity, we get, $$\begin{aligned}
\lefteqn{\langle D_{2}X,D_{1}X\rangle-\langle D_{1}X,D_{2}X\rangle=}\nonumber \\
& & =-\langle X,D_{2}D_{1}X\rangle+\langle X,D_{1}D_{2}X\rangle+\partial_{2}\langle X,D_{1}X\rangle-\partial_{1}\langle X,D_{2}X\rangle\nonumber \\
& & =-ig\langle X,[F_{12},X]\rangle+\partial_{2}\langle X,D_{1}X\rangle-\partial_{1}\langle X,D_{2}X\rangle\nonumber \\
& & =-ig\langle B,[X,X^{\dagger}]\rangle+\partial_{2}\langle X,D_{1}X\rangle-\partial_{1}\langle X,D_{2}X\rangle\;.\end{aligned}$$ Therefore, $$\begin{aligned}
\lefteqn{\langle D_{i}X,D_{i}X\rangle=}\nonumber \\
& & =\langle DX,DX\rangle+g\langle B,[X,X^{\dagger}]\rangle+\partial_{2}\langle X,iD_{1}X\rangle-\partial_{1}\langle X,iD_{2}X\rangle\;,\label{pI}\end{aligned}$$ and similarly, $$\begin{aligned}
\lefteqn{\langle D_{i}X,D_{i}X\rangle=}\nonumber \\
& & =\langle\bar{D}X,\bar{D}X\rangle-g\langle B,[X,X^{\dagger}]\rangle-\partial_{2}\langle X,iD_{1}X\rangle+\partial_{1}\langle X,iD_{2}X\rangle\;,\label{pIbar}\end{aligned}$$ $$\bar{D}=D_{1}-iD_{2}\;.$$
$SU(2)$ model {#Bsu2}
-------------
For $SU(2)$, there is simply a one component positive root, $\alpha_{1}=\frac{1}{\sqrt{2}}$, and $A_{11}=\frac{1}{2}$, $A^{11}=2$. Naming $\phi_{1}=\phi$, $\zeta_{1}=\zeta$, the model in eq. (\[energy\]) is, $$E=\int d^{2}x\,\left(\frac{1}{2}\langle B\rangle^{2}+\langle D_{i}\phi\rangle^{2}+\langle D_{i}\zeta\rangle^{2}+V_{{\rm Higgs}}-\langle\eta,B\rangle\right)\;.\label{ESU2}$$ The Higgs potential can be written as, $$\begin{aligned}
V_{{\rm Higgs}}=\frac{1}{2}\langle\Phi\rangle^{2}+\langle Z\rangle^{2}\;,\end{aligned}$$ $$\Phi=\sqrt{\gamma}\left(v_{c}\,\phi-[\zeta,\zeta^{\dagger}]\right)\makebox[.5in]{,}Z=\sqrt{\gamma_{z}}\left(v_{c}\,\zeta-2\,[\phi,\zeta]\right)\;.\label{ZSU2}$$ Now, using $$\begin{aligned}
& & \langle B\rangle^{2}+\langle\Phi\rangle^{2}=\langle\Phi-B\rangle^{2}+2\langle\Phi,B\rangle\;.\end{aligned}$$ and the property (\[pI\]), for $X=\zeta$, namely, $$\langle D_{i}\zeta\rangle^{2}=\langle D\zeta\rangle^{2}+g\langle[\zeta,\zeta^{\dagger}],B\rangle+\partial_{2}\langle\zeta,iD_{1}\zeta\rangle-\partial_{1}\langle\zeta,iD_{2}\zeta\rangle\;,$$ we obtain, $$\begin{aligned}
E & = & \int d^{2}x\,\left(\langle D_{i}\phi\rangle^{2}+\langle D\zeta\rangle^{2}+\langle Z\rangle^{2}\right.\nonumber \\
& & \left.+\frac{1}{2}\langle\Phi-B\rangle^{2}+\langle\Phi+g\,[\zeta,\zeta^{\dagger}]-\eta\,,B\rangle\right)\;.\end{aligned}$$ Here, we have used a boundary condition at $(x^{1},x^{2})\to\infty$, $$D_{i}\zeta\to0\;.\makebox[.5in]{{\rm for}}(x^{1},x^{2})\to\infty\;.$$
Then, at $\gamma=g^{2}$ and taking the Lie algebra element, $$\eta=gv_{c}\,\phi\;,$$ we get, $$\begin{aligned}
E & = & \int d^{2}x\,\left(\langle D_{i}\phi\rangle^{2}+\langle D\zeta\rangle^{2}+\langle Z\rangle^{2}+\frac{1}{2}\langle\Phi-B\rangle^{2}\right)\;.\nonumber \\\end{aligned}$$ The bound is saturated when, $$D_{i}\phi=0\;,\label{q23}$$ $$v_{c}\,\zeta-2\,[\phi,\zeta]=0\;,\label{withoutF}$$
$$D\zeta=0\;,\label{mas}$$
$$B=g\left(v_{c}\,\phi-[\zeta,\zeta^{\dagger}]\right)\;.\label{pe1}$$
At the critical point, and taking $\gamma_{z}=g^{2}/2$, we can write, $$\begin{aligned}
V_{{\rm Higgs}}=\frac{g^{2}}{2}\left[\langle v_{c}\,\phi-[\zeta,\zeta^{\dagger}]\rangle^{2}+\langle v_{c}\,\zeta-2\,[\phi,\zeta]\rangle^{2}\right]\;.\end{aligned}$$ Using $\phi=\alpha_{1}\psi_{1}=\frac{1}{\sqrt{2}}\,\psi_{1}$, and defining, $$\psi_{2}=\frac{\zeta+\zeta^{\dagger}}{\sqrt{2}}\makebox[.5in]{,}\psi_{3}=\frac{\zeta-\zeta^{\dagger}}{\sqrt{2}i}\makebox[.5in]{,}\sigma=\frac{v_{c}}{\sqrt{2}}\;,\label{p12}$$ the model accepting BPS solutions is given by, $$E=\int d^{2}x\,\left(\frac{1}{2}\langle B\rangle^{2}+\frac{1}{2}\langle D_{i}\psi_{A}\rangle^{2}+V_{{\rm Higgs}}-gv_{c}\,\langle\phi,B\rangle\right)\;.\label{m1}$$
$$\begin{aligned}
V_{{\rm Higgs}}=\frac{g^{2}}{2}\left[\langle\sigma\,\psi_{1}+i[\psi_{2},\psi_{3}]\rangle^{2}+\langle\sigma\,\psi_{2}+i[\psi_{3},\psi_{1}]\rangle^{2}+\langle\sigma\,\psi_{3}+i[\psi_{1},\psi_{2}]\rangle^{2}\right]\;,\nonumber \\\end{aligned}$$
which is a modified version of the flavour symmetric model in eqs. (\[mflavour\]), (\[Vf\]).
$SU(3)$ model {#Bsu3}
-------------
The Higgs potential is, $$V_{{\rm Higgs}}=\frac{1}{2}\langle\Phi,\Phi\rangle+\langle Z_{q},Z_{q}\rangle+{\cal R}\;,\label{H3}$$ $$\Phi=s_{1}\,\Phi_{1}+s_{2}\,\Phi_{2}\makebox[.5in]{,}\Phi_{q}=\sqrt{\gamma}\left(v_{c}\,\phi_{q}-[\zeta_{(q)},\zeta_{(q)}^{\dagger}]\right)\;,$$ $${\cal R}=\gamma_{r}\left(\langle[\phi_{1},\phi_{2}]\rangle^{2}+\langle[\phi^{1},\zeta_{2}]\rangle^{2}+\langle[\phi^{2},\zeta_{1}]\rangle^{2}\right)\;.$$ To obtain a set of BPS equations, we initially diagonalize the $\phi_{q}$-kinetic term in eq. (\[energy\]). Note that any quantity of the form $\langle X_{q},X^{q}\rangle$ can be written as, $$\langle X_{q},X^{q}\rangle=\frac{1}{2}\langle X_{2}+X_{1},X^{2}+X^{1}\rangle+\frac{1}{2}\langle X_{2}-X_{1},X^{2}-X^{1}\rangle\;.$$ On the other hand, the Cartan matrix for $SU(3)$ is, $$\mathbb{C}=6\,\mathbb{A}=\left(\begin{array}{rr}
2 & -1\\
-1 & 2
\end{array}\right)\;,$$ $\mathbb{C}|_{qp}=C_{qp}$, $\mathbb{A}|_{qp}=A_{qp}$. Therefore, $$X_{2}+X_{1}=(A_{22}+A_{12})\, X^{2}+(A_{21}+A_{11})\, X^{1}=\frac{1}{6}\,(X^{2}+X^{1})\;,$$ $$X_{2}-X_{1}=(A_{22}-A_{12})\, X^{2}+(A_{21}-A_{11})\, X^{1}=\frac{1}{2}\,(X^{2}-X^{1})\;.$$ That is, $$\langle X_{q},X^{q}\rangle=3\,\langle X_{+},X_{+}\rangle+\langle X_{-},X_{-}\rangle\makebox[.3in]{,}X_{+}=X_{2}+X_{1}\makebox[.3in]{,}X_{-}=X_{2}-X_{1}\;,$$ and the energy functional in eq. (\[energy\]) results, $$\begin{aligned}
E & = & \int d^{2}x\,\left(\frac{1}{2}\langle B\rangle^{2}+\frac{3}{2}\langle D_{i}\phi_{+}\rangle^{2}+\frac{1}{2}\langle D_{i}\phi_{-}\rangle^{2}+\langle D_{i}\zeta_{q},D_{i}\zeta_{q}\rangle\right)\nonumber \\
& & +\int d^{2}x\,\left(\frac{1}{2}\langle\Phi\rangle^{2}+\langle Z_{q},Z_{q}\rangle+{\cal R}-\langle\eta,B\rangle\right)\;.\label{ene3}\end{aligned}$$ Next, similarly to the $SU(2)$ case, using, $$\begin{aligned}
& & \langle B\rangle^{2}+\langle\Phi\rangle^{2}=\langle\Phi-B\rangle^{2}+2\langle\Phi,B\rangle\;,\end{aligned}$$ and properties (\[pI\]), (\[pIbar\]) for $X=\zeta_{2}$, $\zeta_{1}$, respectively, $$\langle D_{i}\zeta_{2},D_{i}\zeta_{2}\rangle=\langle D\zeta_{2},D\zeta_{2}\rangle+g\langle[\zeta_{2},\zeta_{2}^{\dagger}],B\rangle+\partial_{3}\langle\zeta_{2},iD_{2}\zeta_{2}\rangle-\partial_{2}\langle\zeta_{2},iD_{3}\zeta_{2}\rangle\;,$$ $$\langle D_{i}\zeta_{1},D_{i}\zeta_{1}\rangle=\langle\bar{D}\zeta_{1},\bar{D}\zeta_{1}\rangle-g\langle[\zeta_{1},\zeta_{1}^{\dagger}],B\rangle-\partial_{3}\langle\zeta_{1},iD_{2}\zeta_{1}\rangle+\partial_{2}\langle\zeta_{1},iD_{3}\zeta_{1}\rangle\;,$$ we obtain, $$\begin{aligned}
E & = & \int d^{2}x\,\left(\frac{3}{2}\langle D_{i}\phi_{+}\rangle^{2}+\frac{1}{2}\langle D_{i}\phi_{-}\rangle^{2}+\langle\bar{D}\zeta_{1}\rangle^{2}+\langle D\zeta_{2}\rangle^{2}+\langle Z_{q},Z_{q}\rangle\right)\nonumber \\
& & +\int d^{2}x\,\left(\frac{1}{2}\langle\Phi-B\rangle^{2}+\langle\Phi+g[\zeta_{2},\zeta_{2}^{\dagger}]-g[\zeta_{1},\zeta_{1}^{\dagger}]-\eta\,,B\rangle\right)\nonumber \\
& & +\int d^{2}x\,\gamma_{r}\left(\langle[\phi_{1},\phi_{2}]\rangle^{2}+\langle[\phi^{1},\zeta_{2}]\rangle^{2}+\langle[\phi^{2},\zeta_{1}]\rangle^{2}\right)\;,\end{aligned}$$ where we have used that the system is in a local vacuum at $(x^{1},x^{2})\to\infty$.
Then, taking $s_{1}=-1$, $s_{2}=1$, $\gamma=g^{2}$, which gives, $$\Phi=gv_{c}\,\phi_{-}-g\,\left([\zeta_{2},\zeta_{2}^{\dagger}]-[\zeta_{1},\zeta_{1}^{\dagger}]\right)\makebox[.5in]{,}\phi_{-}=\phi_{2}-\phi_{1}\;,\label{Phi-}$$ and the Lie algebra element, $$\eta=gv_{c}\,\phi_{-}\;,$$ we obtain, $$\begin{aligned}
E & = & \int d^{2}x\,\left(\frac{3}{2}\langle D_{i}\phi_{+}\rangle^{2}+\frac{1}{2}\langle D_{i}\phi_{-}\rangle^{2}+\langle\bar{D}\zeta_{1}\rangle^{2}+\langle D\zeta_{2}\rangle^{2}+\langle Z_{q},Z_{q}\rangle\right)\\
& & +\int d^{2}x\,\left(\frac{1}{2}\langle\Phi-B\rangle^{2}+\gamma_{r}\left(\langle[\phi_{1},\phi_{2}]\rangle^{2}+\langle[\phi^{1},\zeta_{2}]\rangle^{2}+\langle[\phi^{2},\zeta_{1}]\rangle^{2}\right)\right)\;.\end{aligned}$$ Therefore, the BPS equations are, $$D_{i}\phi_{-}=0\makebox[.5in]{,}D_{i}\phi_{+}=0\;,\label{q23n}$$ $$v_{c}\,\zeta_{2}-[\phi^{2},\zeta_{2}]=0\makebox[.5in]{,}\phi^{2}=3\phi_{+}+\phi_{-}\;,\label{withoutF1}$$ $$v_{c}\,\zeta_{1}-[\phi^{1},\zeta_{1}]=0\makebox[.5in]{,}\phi^{1}=3\phi_{+}-\phi_{-}\;,$$ $$[\phi_{1},\phi_{2}]=0\makebox[.3in]{,}[\phi^{1},\zeta_{2}]=0\makebox[.3in]{,}[\phi^{2},\zeta_{1}]=0\;.\label{tr}$$
$$D\zeta_{2}=0\makebox[.5in]{,}\bar{D}\zeta_{1}=0\;,\label{mas1}$$
$$B=g\left(v_{c}\,\phi_{-}-\left([\zeta_{2},\zeta_{2}^{\dagger}]-[\zeta_{1},\zeta_{1}^{\dagger}]\right)\right)\;.\label{pe1n}$$
Center vortex ansatz
====================
In order to propose a center vortex ansatz, it would be useful having a parametrization analogous to the simple $U(1)$ case, where evidencing the modulus and the phase of the complex Higgs field, $\rho\, e^{i\chi}$, accompanied by the gauge field, $a\,\partial_{i}\chi$, permits the implementation of boundary conditions. For this purpose, we could initially determine whether the asymptotic vacua are of the form given in eq. (\[defi\]) or (\[conju\]), and then look for the mapping $S$ (the non Abelian phase) such that, $$S(\vec{\alpha}_{q}\cdot\vec{H})S^{-1}\makebox[.5in]{,}SE_{\alpha_{q}}S^{-1}\;,\label{locb}$$ respectively $$S(-\vec{\alpha}_{q}\cdot\vec{H})S^{-1}\makebox[.5in]{,}S(-E_{\alpha_{q}}^{T})S^{-1}\;,\label{locb1}$$ is the closest local basis to the field configuration $\phi_{q}$, $\zeta_{q}$. The “polar” decomposition is then, $$\phi_{q}=SF_{q}S^{-1}\makebox[.5in]{,}\zeta_{q}=SZ_{q}S^{-1}\;.\label{cansatz}$$
The notion of closest mapping can be obtained by following similar steps to those used when defining adjoint Laplacian center gauges [@deFP]. For example, in $SU(2)$, we can take $\psi_{1}=\sqrt{2}\,\phi_{1}$, together with $\psi_{2}$, $\psi_{3}$ (obtained from $\zeta_{1}$ using eq. (\[p12\])), and expand these fields in the $T_{A}$ basis, $$\psi_{A}=\psi_{AB}\, T_{B}\makebox[.5in]{,}\psi_{AB}=\langle\psi_{A},T_{B}\rangle\;,$$ $A=1,2,3$. The real elements $\psi_{AB}$ form a $3\times3$ matrix $\Psi$, for which a polar decomposition exists, $$\Psi=Q\, R\;,$$ where $R\in SO(3)$ and $Q$ is real symmetric and positive semidefinite. The closest orthogonal matrix to $\Psi$ is $R$, then the closest orthonormal basis to $\{\psi_{A}\}$ is given by, $$n_{A}=R_{AB}\, T_{B}=ST_{A}S^{-1}\;,\label{lf}$$ where $S$ is defined up to a global center element. That is, for $SU(2)$ adjoint Higgs fields, the “modulus and phase” decomposition is, $$\psi_{A}=S(Q_{AB}\, T_{B})S^{-1}\;,$$ which can be translated back to $\phi_{1}$, $\zeta_{1}$-language.
With regard to the gauge field, we note that on any simply connected region, which does not contain the pointlike defects of the local basis, the Higgs field ansatz looks as a gauge transformation. Therefore, in that region, the field equations would be simplified by representing the smooth $A_{i}$ as a gauge transformation of a vector field ${\cal A}_{i}$. However, in the defining representation, $S$ is in general discontinuous on some curves, as it changes by a center element when we go around a center vortex. Therefore, on $R^{2}-\{{\rm pointlike~defects}\}$, the ansatz, $$S{\cal A}_{i}S^{-1}+\frac{i}{g}\, S\partial_{i}S^{-1}\;,$$ cannot work, as it contains a contribution ($I_{i}$) concentrated at the points where $S^{-1}$ is discontinuous. There are three equivalent possibilities to circumvent this problem.
- proceed as in ref. [@engelhardt1; @reinhardt], proposing the parametrization, $$A_{i}=S{\cal A}_{i}S^{-1}+\frac{i}{g}\, S\partial_{i}S^{-1}-I_{i}\;,\label{an1}$$
- proceed as in [@lucho; @Lucho2], to write $$A_{i}=({\cal A}_{i}^{A}-C_{i}^{A})\, n_{A}\;,\label{an2}$$ where $C_{i}^{A}$ only depends on the local colour frame (\[lf\]).
- work with the fields mapped into the adjoint representation[^4], $$Ad(A_{i})=RAd({\cal A}_{i})R^{-1}+\frac{i}{g}R\partial_{i}R^{-1}\makebox[.5in]{,}Ad({\cal A}_{i})={\cal A}_{i}^{A}\, M_{A}\;,\label{gad}$$ $$Ad(\psi_{q})=RAd(P_{q})R^{-1}\makebox[.5in]{,}Ad(\zeta_{\alpha})=RAd(P_{\alpha})R^{-1}\;,\label{anad}$$
where we used, $$R_{AB}\, M_{B}=RM_{A}R^{-1}\;.$$ Here, we shall use the third possibility. The advantage of the second and third options is that $n_{A}$ and $R$ contain at most pointlike defects, as they are always single-valued when we go around a loop. Then, the $R\partial_{i}R^{-1}$ term does not introduce delta distributions concentrated on curves, and a smooth $Ad(A_{i})$ ansatz can be implemented with $Ad({\cal A}_{i})$ satisfying appropriate boundary conditions at the vortex guiding centers.
It is important to underline that, in the ansatz (\[gad\]), $Ad(A_{i})$ *is not* a gauge transformation of $Ad({\cal A}_{i})$. The magnetic field $B$ is given by, $$Ad(B)=RAd({\cal B})R^{-1}+\frac{i}{g}R[\partial_{1},\partial_{2}]R^{-1}\;,\label{adja1}$$ $${\cal B}=\partial_{1}{\cal A}_{2}-\partial_{2}{\cal A}_{1}-ig[{\cal A}_{1},{\cal A}_{2}]\;,\label{field-st}$$ where the last term in eq. (\[adja1\]) is concentrated at the vortex guiding centers. The profiles ${\cal A}_{i}$, $F_{q}$ and $Z_{q}$, must be such that $B$ and the Higgs fields be well-defined and smooth everywhere, and satisfy the desired asymptotic behaviour. For a single center vortex, with charge $z$ modulo $N$, we can impose in the asymptotic region, $${\cal A}_{i}\to0\makebox[.5in]{,}S\to e^{i\varphi\,\vec{\beta}\cdot\vec{T}}\;,$$ where $\vec{\beta}$ satisfies eq. (\[poss-b\]). When minimizing the energy, the extension of $R=Ad(S)$ from the asymptotic region to the vortex core should not only contemplate keeping $R$ along a Cartan direction but also other possibilities. In this regard, note that for $$\vec{\beta}-\vec{\beta}_{0}=2N\vec{\gamma}\makebox[.5in]{,}\vec{\gamma}\in\Lambda(Ad(SU(N)))\;,\label{c0}$$ where $\Lambda(\dots)$ represents the lattice of weights of the adjoint representation (or root lattice), it is always possible to obtain a map $R(r,\varphi)$ verifying, $$R(r,\varphi)=\left\{ \begin{array}{ll}
e^{i\varphi\,\vec{\beta}\cdot\vec{M}}, & r>r_{m}\\
e^{i\varphi\,\vec{\beta}_{0}\cdot\vec{M}}, & r<r_{0}\;,
\end{array}\right.\label{beha}$$ that is smooth for $r\geq r_{0}$. This map can be constructed as $R=e^{i\varphi\,\vec{\beta}_{0}\cdot\vec{M}}R_{0}$, with $$R_{0}(r,\varphi)=\left\{ \begin{array}{ll}
e^{i\varphi\,2N\vec{\gamma}\cdot\vec{M}}, & r>r_{m}\\
I, & r<r_{0}\;.
\end{array}\right.$$ Note that $R_0$ always exists as $e^{i\varphi\,2N\vec{\gamma}\cdot\vec{T}}$ is a closed path in $SU(N)$, and therefore $e^{i\varphi\,2N\vec{\gamma}\cdot\vec{M}}$ is topologically trivial in $Ad(SU(N))$. Different magnetic weights $\vec{\beta}_{0}$ imply different types of defect and profile function behaviours at $r\to0$. For example, an asymptotic behaviour with $z=0$ is described by any $\vec{\beta}\in2N\,\Lambda(Ad(SU(N)))$. All these values can be extended to $\vec{\beta}_{0}=0$. For this choice, $R(r,\varphi)$ contains no defect at the origin and the minimization process will simply return a trivial result, corresponding to a pure gauge transformation of the vacuum configuration. For $z=\pm1$, i.e. $\vec{\beta}=2N\,\vec{w}+2N\,\vec{\gamma}$, where $\vec{w}$ is a weight of the defining representation or its conjugate, there is no manner to avoid a defect at $r\to0$. The energy is expected to be minimized by $\vec{\beta}_{0}=2N\,\vec{w}$, as in this case some of the basis components will only give one turn when we go around a small circle centered at $r=0$.
BPS center vortices
===================
At the critical point, to solve the $SU(2)$ and $SU(3)$ BPS equations, it will be enough to consider, $$\phi_{q}=v_{c}\, S(\vec{\alpha}_{q}\cdot\vec{H})S^{-1}\makebox[.5in]{,}\zeta_{q}=u\, SE_{\alpha_{q}}S^{-1}\;,$$ (and a similar expression for the conjugate sector). The possible non Abelian phases $S$ are such that $R=Ad(S)$ behaves as in eq. (\[beha\]). As we will see, the $A_{i}$ parametrization in terms of ${\cal A}_{i}$ and $S$ together with the BPS equations imply, $${\cal A}_{i}=c_{i}\,\vec{\delta}\cdot\vec{H}\;,$$ where $H_{q}$ denote the Cartan generators. Then, from eqs. (\[an1\])-(\[gad\]), for $r<r_{0}$ the gauge field is, $$A_{i}=c_{i}\,\vec{\delta}\cdot\vec{H}+\frac{1}{g}\partial_{i}\varphi\,\vec{\beta}_{0}\cdot\vec{H}\;,\label{beh0}$$ and in order to obtain a regular magnetic field, we must have $\vec{\delta}=\pm\vec{\beta}_{0}$ and $c_{i}\to\mp\frac{1}{g}\partial_{i}\varphi$, when $r\to0$.
$\mathfrak{su}(2)$
------------------
For nonzero $\zeta$, eq. (\[withoutF\]) implies, $$\phi=\frac{v_{c}}{\sqrt{2}}\, SH_{1}S^{-1}\makebox[.5in]{,}\zeta=u\, SE_{\alpha_{1}}S^{-1}\;,$$ where the possible magnetic weights $\beta$ in eq. (\[beha\]) are $\beta=q\sqrt{2}$, $q\in\mathbb{Z}$. Now, using any of the parametrizations (\[an1\])-(\[gad\]), eq. (\[q23\]) gives, $$D_{i}({\cal A})(H_{1})=0\makebox[.9in]{{\rm or},}[{\cal A}_{i},H_{1}]=0\;,$$ whose solution is, $${\cal A}_{i}=c_{i}\,\beta_{0}H_{1}\;,$$ (the case $\delta=-\beta_{0}$ is discussed at the end). Similarly, eq. (\[mas\]) becomes, $$D({\cal A})(uE_{\alpha_{1}})=0\makebox[.9in]{{\rm or},}(\partial_{1}+i\partial_{2})u\, E_{\alpha_{1}}-igu\,[{\cal A}_{1}+i{\cal A}_{2},E_{\alpha_{1}}]=0\;.$$ Thus, joining this information, we obtain, $$\frac{\beta_{0}}{\sqrt{2}}\,(c_{1}+ic_{2})=\frac{1}{2g}\,(\partial_{2}h-i\partial_{1}h)\makebox[.5in]{,}u=v_{c}\, e^{h/2}\;.$$ $$Ad({\cal B})=(\partial_{1}c_{2}-\partial_{2}c_{1})\,\beta_{0}Ad(H_{1})=-\frac{1}{\sqrt{2}g}(\partial_{1}^{2}+\partial_{2}^{2})h\, Ad(H_{1})\;,$$ where we have changed the variables from $u$ to $h=2\ln(u/v_{c})$, as is usually done in the $U(1)$ case. Therefore, eqs. (\[pe1\]) and (\[adja1\]) imply, $$\left((\partial_{1}^{2}+\partial_{2}^{2})h+g^{2}v_{c}^{2}\,(1-e^{h})\right)RAd(H_{1})R^{-1}=i\sqrt{2}\, R[\partial_{1},\partial_{2}]R^{-1}\;.$$ The second member is obtained from eq. (\[beha\]), $$\begin{aligned}
i\, R[\partial_{1},\partial_{2}]R^{-1}=\beta_{0}\,[\partial_{1},\partial_{2}]\varphi\, Ad(H_{1})\;.\end{aligned}$$ As is well-known, although for $j_{i}=\partial_{i}\varphi$ the quantity $\partial_{2}j_{3}-\partial_{3}j_{2}=[\partial_{2},\partial_{3}]\varphi$ seems to vanish, it is in fact concentrated at $x^{2}=x^{3}=0$, where $e^{i\varphi}$ contains a defect. Namely, $$\partial_{1}j_{2}-\partial_{2}j_{1}=2\pi\,\delta^{(2)}(x^{1},x^{2})\;.$$ This can be checked using Stokes’ theorem. Then, we get, $$(\partial_{1}^{2}+\partial_{2}^{2})\, h+g^{2}v_{c}^{2}\,(1-e^{h})=2\pi\sqrt{2}\beta_{0}\,\delta^{(2)}(x^{1},x^{2})\;.\label{anzeq}$$ For $q$ even, the asymptotic behaviour $S=e^{iq\sqrt{2}\,\varphi H_{1}}$, $R=e^{iq\sqrt{2}\,\varphi Ad(H_{1})}$, on the circle $r\to\infty$, can be continuously changed to a behaviour characterized by $\beta_{0}=0$, as $r$ is varied from $\infty$ to $0$. The abscence of defects will lead to a trivial pure gauge solution for the BPS equations. On the other hand, for $q$ odd, the asymptotic behavior can be changed to $\beta_{0}=+\sqrt{2}$, as well as $\beta_{0}=-\sqrt{2}$. For these values, the frame components $ST_{\alpha_1} S^{-1}$, $ST_{\bar{\alpha}_1}S^{-1}$, $$T_{\alpha}=\frac{1}{\sqrt{2}}(E_{\alpha}+E_{-\alpha})\makebox[.5in]{,}T_{\bar{\alpha}}=\frac{1}{\sqrt{2}i}(E_{\alpha}-E_{-\alpha})\;,\label{Tes}$$ rotate only once, when we go close to and around the origin. The solution to eq. is well-defined for $\beta_{0}=+\sqrt{2}$, while it is ill-defined for $\beta_{0}=-\sqrt{2}$. In the latter case, the well-defined solution is obtained using the conjugate ansatz (see a similar discussion in \[su3e\]), $$\phi=\frac{v_{c}}{\sqrt{2}}\, S(-H_{1})S^{-1}\makebox[.5in]{,}\zeta=u\, S(-E_{\alpha_{1}}^{T})S^{-1}\makebox[.5in]{,}{\cal A}_{i}=c_{i}\,(-\sqrt{2})H_{1}\;.$$ In $SU(2)$, both the vortex and its conjugate satisfy, $$S(\varphi+2\pi)=-S(\varphi)\;,$$ so they are equivalent objects.
$\mathfrak{su}(3)$ {#su3e}
------------------
The equations (\[withoutF1\])-(\[tr\]) imply, $$\phi_{q}=v_{c}\, S(\vec{\alpha}_{q}\cdot\vec{H})S^{-1}\makebox[.5in]{,}\zeta_{q}=u_{q}\, SE_{\alpha_{q}}S^{-1}\label{ang}$$ (q=1,2) so that imposing eq. (\[q23n\]), we obtain, $$D_{i}({\cal A})(\vec{\alpha}_{q}\cdot\vec{H})=0\makebox[.9in]{{\rm or},}[{\cal A}_{i},\vec{\alpha}_{q}\cdot\vec{H}]=0\;.$$ This means that ${\cal A}_{i}$ is in the Cartan subalgebra. Taking $\vec{\delta}=+\vec{\beta}_{0}$, $${\cal A}_{i}=c_{i}\,\vec{\beta}_{0}\cdot\vec{H}\;,$$ eq. (\[mas1\]) gives, $$\begin{aligned}
& & (\partial_{1}+i\partial_{2})u_{2}\, E_{\alpha_{2}}-ig\,(c_{1}+ic_{2})u_{2}\,[\vec{\beta}_{0}\cdot\vec{H},E_{\alpha_{2}}]=0\;,\end{aligned}$$ $$\begin{aligned}
& & (\partial_{1}-i\partial_{2})u_{1}\, E_{\alpha_{1}}-ig\,(c_{1}-ic_{2})u_{1}\,[\vec{\beta}_{0}\cdot\vec{H},E_{\alpha_{1}}]=0\;.\end{aligned}$$ Then, we get, $$(\vec{\beta}_{0}\cdot\vec{\alpha}_{2})\,(c_{1}+ic_{2})=\frac{1}{2g}\,(\partial_{2}h_{2}-i\partial_{1}h_{2})\;,\label{cm}$$ $$(\vec{\beta}_{0}\cdot\vec{\alpha}_{1})\,(c_{1}-ic_{2})=-\frac{1}{2g}\,(\partial_{2}h_{1}+i\partial_{1}h_{1})\;,\label{cc}$$ $${\cal B}=-\frac{1}{2g}(\partial_{1}^{2}+\partial_{2}^{2})h\,\frac{\vec{\beta}_{0}\cdot\vec{H}}{\vec{\beta}_{0}\cdot\vec{\alpha}_{2}}\;,\label{3i}$$ where $u_{i}=v_{c}\, e^{h_{i}/2}$. In addition, eq. (\[pe1n\]) reads, $$\begin{aligned}
\lefteqn{iR[\partial_{1},\partial_{2}]R^{-1}=}\\
& & =R\, Ad\left(-g\,{\cal B}+g^{2}v_{c}^{2}\,(\vec{\alpha}_{2}-\vec{\alpha}_{1})\cdot\vec{H}-g^{2}v_{c}^{2}\,(\vec{\alpha}_{2}\, e^{h_{2}}-\vec{\alpha}_{1}\, e^{h_{1}})\cdot\vec{H}\right)R^{-1}\;,\end{aligned}$$ while for a single vortex, eq. (\[beha\]) implies, $$\begin{aligned}
i\, R[\partial_{1},\partial_{2}]R^{-1}=[\partial_{1},\partial_{2}]\varphi\, Ad(\vec{\beta}_{0}\cdot\vec{H})=2\pi\,\delta^{(2)}(x^{1},x^{2})\, Ad(\vec{\beta}_{0}\cdot\vec{H})\;.\end{aligned}$$ Putting this information together, we arrive at, $$-g\,{\cal B}+g^{2}v_{c}^{2}\,(\vec{\alpha}_{2}-\vec{\alpha_{1}})-g^{2}v_{c}^{2}\,(\vec{\alpha}_{2}\, e^{h_{2}}-\vec{\alpha}_{1}\, e^{h_{1}})=2\pi\,\delta^{(2)}(x^{1},x^{2})\,\vec{\beta}_{0}\;.\label{pt}$$ Let us consider the case where $\vec{\beta}_{0}$ is associated with a weight of the defining representation. Noting that ${\cal B}=(\partial_{1}c_{2}-\partial_{2}c_{1})\,\vec{\beta}_{0}\cdot\vec{H}$ and $\vec{\alpha}_{2}-\vec{\alpha}_{1}=\frac{1}{2}\,\vec{\beta}_{2}$, in order to have a nontrivial solution we are led to $\vec{\beta}_{0}=\pm\vec{\beta}_{2}$. For these cases, $\vec{\beta}_{0}\cdot\vec{\alpha}_{2}=-\vec{\beta}_{0}\cdot\vec{\alpha}_{1}$, the equations (\[cm\]) and (\[cc\]) give $h_{1}=h_{2}=h$, $u_{i}=u=e^{h}$, and both sides of eq. (\[pt\]) turn out to be oriented along the same direction. Under these conditions, we obtain, $$(\vec{\beta}_{0}\cdot\vec{\alpha}_{2})^{-1}\,(\partial_{1}^{2}+\partial_{2}^{2})h\,\vec{\beta}_{0}+g^{2}v_{c}^{2}\,(1-e^{h})\,\vec{\beta}_{2}=4\pi\,\delta^{(2)}(x^{1},x^{2})\,\vec{\beta}_{0}\;.\label{3ii}$$ That is, for $\vec{\beta}_{0}=+\vec{\beta}_{2}$ ($\vec{\beta}_{0}\cdot\vec{\alpha}_{2}=+1$), $$(\partial_{1}^{2}+\partial_{2}^{2})h+g^{2}v_{c}^{2}\,(1-e^{h})=4\pi\,\delta^{(2)}(x^{1},x^{2})\;.\label{well}$$ On the other hand, the choice $\vec{\beta}_{0}=-\vec{\beta}_{2}$ would imply, $$(\partial_{1}^{2}+\partial_{2}^{2})h+g^{2}v_{c}^{2}\,(1-e^{h})=-4\pi\,\delta^{(2)}(x^{1},x^{2})\;.\label{ill}$$ The second choice does not lead to well-defined Higgs fields. In effect, while close to the origin eq. (\[well\]) gives $h\sim2\ln r$, $u=e^{h/2}\sim r$, producing single-valued Higgs fields (and $c_{i}\sim-\frac{1}{g}\,\partial_{i}\varphi$), eq. (\[ill\]) gives $h\sim-2\ln r$, $u=e^{h/2}\sim1/r$. However, it is easy to see that the new ansatz obtained from (\[ang\]) by the replacement, $$\vec{\alpha}_{q}\cdot\vec{H}\to-\vec{\alpha}_{q}\cdot\vec{H}\makebox[.5in]{,}E_{\alpha_{q}}\to-E_{\alpha_{q}}^{T}\;,\label{conjug}$$ solves the BPS equations with a well-defined $h$ satisfying eq. (\[well\]), provided we choose $\vec{\beta}_{0}=-\vec{\beta}_{2}$. Other weights can be obtained by replacing in eq. (\[ang\]) (resp. eq. (\[conjug\])), $$\vec{\alpha}_{q}\to\vec{\alpha}_{q}^{\, W}\;,$$ where $W$ is a Weyl transformation. The solutions will be characterized by the gauge field behaviour (\[beha\]), with $\vec{\beta}_{0}\to\vec{\beta}_{0}^{\, W}=\vec{\beta}_{2}^{\, W}$ (resp. $\vec{\beta}_{0}\to\vec{\beta}_{0}^{\, W}=-\vec{\beta}_{2}^{\, W}$) (and $\vec{\beta}\to\vec{\beta}^{\, W}$). Then, these solutions are characterized by the weights of the defining representation, $\vec{\beta}_{1},\vec{\beta}_{2},\vec{\beta}_{3}$, and their conjugates, $-\vec{\beta}_{1},-\vec{\beta}_{2},-\vec{\beta}_{3}$. As the mappings $S$ satisfy, $$S(\varphi+2\pi)=e^{\pm i2\pi/3}\, S(\varphi)\;,$$ they correspond to center vortices with the minimum charges $z=\pm1$.
Conclusions
===========
In this article we presented Yang-Mills-Higgs nonrelativistic models with $SU(N)\to Z(N)$ SSB pattern that accept BPS center vortex equations (for $N=2,3$).
For this purpose, we initially proposed a class of $SU(N)$ Lorentz invariant models containing real and complex adjoint Higgs fields, that can be labelled by the simple roots of the $\mathfrak{su}(N)$ Lie algebra. The Higgs potential is such that its minimization returns a set of conditions that essentially define a Chevalley basis. The space of vacua also contains a trivial symmetry preserving point, where the Higgs fields vanish, separated from the SSB points by a potential barrier.
Next, we introduced a nonrelativistic interaction term so as to obtain a set of BPS equations. This is a term that tends to align, in the Lie algebra, the magnetic field and one of the Higgs fields. Finally, we obtained some solutions. For example, the $Z(3)$ vortices come in three colours (the weights of the defining representation), which are physically equivalent, and three anticolours, obtained from an ansatz based on the conjugate representation.
Generally, BPS equations are derived by working on the energy functional, which is a local object, and obtaining a bound that only depends on some topological charge. For $U(1)$ vortices, the bound is given by the magnetic flux. This is a topological term that can be written locally, by means of a flux density. On the other hand, for center vortices, the flux concept is given by the asymptotic behaviour of the gauge invariant Wilson loop, a nonlocal object that may not arise in the calculation. For this reason, the search for BPS equations led us to consider the alignment interaction. After completing the squares, the energy is always greater than or equal to zero. Thus, BPS center vortices are characterized by an exact compensation between the positive definite part of the energy functional (kinetic energy plus Higgs potential) and the contribution originated from alignment.
Similarly to the minima of the Higgs potential, the BPS equations have trivial solutions with vanishing Higgs fields (and pure gauge fields) and a sector where the asymptotic fields are in SSB vacua. Although the BPS solutions have vanishing energy, no finite energy configurations continuously interpolating the center vortex ($z=\pm1$) and the trivial configuration ($z=0$) exist. In other words, there is an energy barrier for the continuous deformation of one configuration into the other. The general solution to the BPS equations was written in terms of a reduced set of profile functions and a mapping $R(S)$ in the adjoint representation of $SU(N)$. The mapping $S\in SU(N)$, contains information about the asymptotic Wilson loop and the set of possible defects at the vortex guiding centers, which determine the behaviour of the profile functions.
In spite of the Abelian looking profile functions obtained, we would like to underline two important differences. As the number of BPS center vortices is increased, the energy continues to vanish. This is in contrast to the $U(1)$ case, where the energy increases linearly with the number of vortices, a property that is modified below and above the critical coupling, implying either attractive or repulsive forces. In addition, the topological properties of the adjoint representation of $SU(N)$ modify the relation between asymptotic phases and defects. A $U(1)$ asymptotic phase implies a unique type of pointlike defect, and a unique order for the zero of the corresponding Higgs profile function. On the other hand, for an asymptotic non Abelian phase, many extensions to reach a pointlike defect are possible, with corresponding conditions on the profile functions. For $Z(N)$-charge equal to $\pm1$, a defect is always present, while for vanishing $Z(N)$-charge the defects in $R(S)$ can be avoided, and the lowest energy solution simply corresponds to a regular gauge transformation of the SSB vacua. When leaving the critical point, the new energetics, topology, and field content are expected to modify the forces between center vortices, as compared with the $U(1)$ case. This may be a possibility worth exploring.
Summarizing, the search for BPS bounds is among the preferred analytical tools to understand topological objects. In this manuscript, we showed what would be the situation in the context of center vortex $(2+1)$d models: they become nonrelativistic. Then, although BPS center vortices are not directly physically relevant, they could provide a useful concept when embarking on numerical simulations. The presence of a BPS point in the extended parameter space could serve as a check of the numerical analysis when moving away from the physically relevant non BPS Lorentz invariant confining models.
Acknowledgements {#acknowledgements .unnumbered}
================
The Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq-Brazil), the Fundação de Amparo a Pesquisa do Estado do Rio de Janeiro (FAPERJ), and the Proppi-UFF are acknowledged for their financial support.
Appendix A: Cartan decomposition of $\mathfrak{g}$ {#appendix-a-cartan-decomposition-of-mathfrakg .unnumbered}
==================================================
A compact connected simple Lie algebra $\mathfrak{g}$ can be decomposed in terms of hermitian Cartan generators $H_{q}$, $q=1,\dots,r$, which generate a Cartan subgroup $H$, and off-diagonal generators $E_{\alpha}$, or root vectors. The latter are labelled by a system of roots $\vec{\alpha}=(\alpha_{1},\dots,\alpha_{r})$. They satisfy, $$[H_{q},H_{p}]=0\makebox[.3in]{,}[H_{q},E_{\alpha}]=\alpha_{q}\, E_{\alpha}\makebox[.3in]{,}[E_{\alpha},E_{\alpha}^{\dagger}]=\langle E_{\alpha},E_{\alpha}\rangle\, H_{\alpha}\;,\label{algebrag}$$ where, for every root $\vec{\alpha}$, $H_{\alpha}$ is defined by, $$\langle H_{\alpha},H_{q}\rangle=\vec{\alpha}|_{q}\;.$$ The rank of $\mathfrak{su}(N)$ is $r=N-1$, and its dimension is $d=N^{2}-1$. The weights of the defining representation, can be ordered according to, $$\vec{w}_{1}>\vec{w}_{2}>\dots>\vec{w}_{N}\;,$$ so that the positive and simple roots are, respectively, $$\vec{\alpha}_{qp}=\vec{w}_{q}-\vec{w}_{p}\makebox[.3in]{,} q<p
\makebox[.5in]{,} \vec{\alpha}_{q}=\vec{w}_{q}-\vec{w}_{q+1}
\label{difw}\;.$$ Finally, recalling that the fundamental weights $\vec{\Lambda}^{q}$ are defined by, $$\frac{2\,\vec{\alpha}_{q}\cdot\vec{\Lambda}^{p}}{\vec{\alpha}_{(q)}\cdot\vec{\alpha}_{(q)}}=\delta_{q}^{\, p}\;,$$ the $\vec{\lambda}^{q}$ basis in eq. (\[alam\]) can be written as, $$\vec{\lambda}^{q}=2N\,\vec{\Lambda}^{q}\;,$$ with, $$\vec{\Lambda}^{1}=\vec{w}_{1}\makebox[.3in]{,}\vec{\Lambda}^{2}=\vec{w}_{1}+\vec{w}_{2}\makebox[.3in]{,}\vec{\Lambda}^{3}=\vec{w}_{1}+\vec{w}_{2}+\vec{w}_{3}\makebox[.3in]{,}\dots$$
[10]{} W. P. Su, J. R. Schrieffer, and A. J. Heeger, Phys. Rev. Lett. **42** (1979) 1698.
E. Fradkin, Field Theories of Condensed Matter Physics (2nd edition, Cambridge University Press, 2013).
T. H. R. Skyrme, Proc. R. Soc. Lond. **260** (1961) 127.
C. J. Houghton, N. S. Manton, and P. M. Sutcliffe, Nuc. Phys. **B510** (1998) 507.
N. Manton and P. Sutcliffe, Topological Solitons (Cambridge University Press, 2004).
M. Baker, J. S. Ball and F. Zachariasen, Phys. Rev. **D51** (1995) 1968.
M. Baker, J. S. Ball, N. Brambilla, G. M. Prosperi and F. Zachariasen, Phys. Rev. **D54** (1996) 2829, erratum, ibid. **D56** (1997) 2475.
S. Maedan and T. Suzuki, Progr. Theor. Phys. **81** (1989) 229.
H. Shiba, S. Kamisawa, Y. Matsubara and T. Suzuki, Nucl. Phys. **B389** (1993) 563.
M. N. Chernodub and D. A. Komarov, JETP Lett. **68** (1998) 117.
K. Konishi and L. Spanu, Int. J. Mod. Phys. **A18** (2003) 249.
H. J. de Vega, Phys. Rev. **D18** (1978) 2932.
H.J. de Vega and F.A. Schaposnik, Phys. Rev. Lett. **56** (1986) 2564.
H. J. de Vega and F. A. Schaposnik, Phys. Rev. **D34** (1986) 3206.
J. Heo and T. Vachaspati, Phys. Rev. **D58** (1998) 065011.
L. E. Oxman, J. High Energy Phys. **03** (2013) 038.
B. Ketzer, PoS (QNP2012) 025.
J. J. Dudek, R. G. Edwards, M. J. Peardon, D. G. Richards and C. E. Thomas, Phys. Rev. Lett. **103** (2009) 262001.
J. J. Dudek and R. G. Edwards, arXiv:1201.2349.
P. Goddard, J. Nyuts and D. Olive, Nucl. Phys. **B125** (1977) 1.
A. Hanany and D. Tong, JHEP **0307** (2003) 037.
R. Auzzi, S. Bolognesi, J. Evslin, K. Konishi and A. Yung, Nucl. Phys. B 673 (2003) 187.
D. Tong, Phys. Rev. **D69** (2004) 065003.
D. Tong, Ann. of Phys. **324** (2009) 30.
K. Konishi, Lect. Notes Phys. **737** (2008) 471.
S.B. Bradlow, O. Garcia-Prada, Non-abelian monopoles and vortices, arXiv:alg-geom/9602010.
J. E. Humphreys, *Introduction to Lie Algebras and Representation Theory* (Springer, New York, 1972).
B. C. Hall, *Lie groups, Lie Algebras and Representations* (Springer, New York, 2003).
H. Giorgi, *Lie Algebras in Particle Physics*, Frontiers in Physics.
Ph. de Forcrand and M. Pepe, Nucl. Phys. **B598** (2001) 557.
M. Engelhardt, H. Reinhardt, Nucl. Phys. **B567** (2000) 249.
H. Reinhardt, Nucl. Phys. **B628** (2002) 133.
L. E. Oxman, JHEP **12** (2008) 089.
L. E. Oxman JHEP **07** (2011) 078.
[^1]: We are using the inner product $\langle X,Y\rangle=Tr\left(Ad(X)^{\dagger}Ad(Y)\right)$, where $Ad(\cdot)$ is a linear map into the adjoint representation.
[^2]: we use $X$ to denote any of the Higgs fields $\phi_q$, $\zeta_q$.
[^3]: In $3+1$ dimensions, this type of model would also break rotation symmetry.
[^4]: The matrices $M_{A}$ are generators of the adjoint representation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper we present a machine vision system to efficiently monitor, analyze and present visual data acquired with a railway overhead gantry equipped with multiple cameras. This solution aims to improve the safety of daily life railway transportation in a two-fold manner: (1) by providing automatic algorithms that can process large imagery of trains (2) by helping train operators to keep attention on any possible malfunction. The system is designed with the latest cutting edge, high-rate visible and thermal cameras that observe a train passing under an railway overhead gantry. The machine vision system is composed of three principal modules: (1) an automatic wagon identification system, recognizing the wagon ID according to the UIC classification of railway coaches; (2) a temperature monitoring system; (3) a system for the detection, localization and visualization of the pantograph of the train. These three machine vision modules process batch trains sequences and their resulting analysis are presented to an operator using a multitouch user interface.
We detail all technical aspects of our multi-camera portal: the hardware requirements, the software developed to deal with the high-frame rate cameras and ensure reliable acquisition, the algorithms proposed to solve each computer vision task, and the multitouch interaction and visualization interface. We evaluate each component of our system on a dataset recorded in an ad-hoc railway test-bed, showing the potential of our proposed portal for train safety assessment.
author:
- Giuseppe Lisanti
- Svebor Karaman
- Daniele Pezzatini
- Alberto Del Bimbo
bibliography:
- 'SISSI\_mva.bib'
date: 'Received: date / Accepted: date'
title: 'A Multi-Camera Image Processing and Visualization System for Train Safety Assessment'
---
[example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore
Introduction
============
In the last years train safety got the attention of media and public opinion after several disastrous train accidents, as those that happened in Italy in 2009 [@ViareggioIncident] and in France in 2013 [@FrenchTrainCrash]. Train accidents may be a result of either a problem on the railway tracks, as it was the case for the French accident, or some issues with the train itself. The analysis of the railway tracks requires the installation of sensors on board of a train that should travel on the tracks that have to be inspected. Several proposals have been made in this sense [@camargo2011emerging; @edwards2009advancements]. This work focuses on the safety assessment of the train itself.
A train is composed of a locomotive and multiple wagons, any of its components can be a risk for the train safety. A single wagon failure can trigger the derailment of several wagons and have dramatic consequences. The Viareggio accident [@ViareggioIncident] is believed to be the consequences of an axle failure on a tank wagon, the wagon hit the platform of the station and overturned to the left and several following wagons also overturned, exploded and caught fire. Another important aspect of train safety is temperature monitoring especially when the train is approaching a tunnel where escape in case of fire can be difficult. For example, the Kaprun disaster [@schupfer2001fire] was due to an electric fan heater that caught fire. Hence monitoring an abnormal temperature on any part of the train can provide an early notice of an issue and thus prevent its potential dramatic outcome. A train can be considered safe for transit if all wagons are adapted for the transit on the railway, the locomotive and all wagons exhibit nominal temperatures and no out of shape elements are present. Failure to fulfill any of these requirements may indicate a risk situation. The analysis of each train status may be done by stopping and analyzing each train in a offtrack location before being allowed to travel. This would induce serious delays on the train traffic. Hence, more interest have been put on portal based system that could be installed on some important keypoints of the railway network (e.g. before a tunnel or before entering a train station) in order to asses on-the-fly all the safety requirements. This solution has the clear advantage of not requiring to stop the train to run the analysis and it is also possible to install multiple sensors on a single portal providing a thorough analysis of the train status at once. However, such portal based approaches require the monitoring system to be able to capture all required signals even for a train running at full speed.
This paper depicts our proposed multi-camera portal for train safety assessment developed in the context of the Integrated Intermodal System for Security and Signaling on Rail (SISSI) project, funded by Regione Toscana (Italy). Our system relies on high-speed and thermal cameras to monitor several aspects of the train. The acquired signals are processed by computer vision methods to extract meaningful information. Finally, all the information is provided to an operator through a touch-based user interface. We first review in the next section the state-of-the-art of computer vision based system for train safety and of touch-based interaction for control rooms and train safety. We then give an in-depth presentation of our proposed system in section \[sec:oursystem\], specifying the hardware and giving an overview of the software developed to obtain reliable data acquisition from all sensors. In section \[sec:train\_analysis\], we detail how we solve each target task of the train analysis,namely the automatic wagon identification, the temperature monitoring, and the detection and localization of the pantograph of the train. We then describe how all the results are provided to the operator on our multitouch interface. Finally in section \[sec:evaluation\], we give an evaluation of each sub-system of our multi-camera portal on a dataset recorded in an ad-hoc railway test-bed.
Related work
============
Computer vision based systems for train safety
----------------------------------------------
We can distinguish in the literature the approaches that target safety assessment of the train surroundings or the train itself. Some approaches focus on a single aspect of train safety, while multi-modal portals tries to analyze at the same times multiple safety features.
Many railway accidents happen at railway crossing where an object such as a car is stopped on the railway, hence one common use of computer vision is to detect if an obstacle is obstructing the railway. Machine vision was used in [@pu2014study] to detect moving obstacle in these locations. A 3D vision system for obstacle detection is proposed in [@weichselbaum2013accurate]. Train stations are also a risk environment, in [@delgado2014automatic] the authors present a method for automatically detecting people jumping or falling off a train platform.
Another aspect to consider to assess train safety is the proper configuration of the train itself. In [@sacchipavisys], a system to detect misalignment of a train pantograph is proposed. The authors of [@fumagallimultifunction] proposed a multi-function portal similar in spirit to ours with the main objectives of detecting misalignment of carriage or abnormal temperature so as to be able to stop a train before it enters a tunnel. They rely on line-scan cameras to obtain the train image in the visible domain, pyroelectric line cameras for thermal imaging and a distributed time-of-flight telemeter for the train shape analysis. The evaluation targets mostly the sensors performance and only qualitative results of out-of-shape detections are given.
Touch-based interface for control rooms and train safety
--------------------------------------------------------
Operators in control rooms are often asked to monitor multiple safety characteristics and have to perform crucial security operations in a short amount of time. This is the main reason why information visualization and touch-based interaction play a key role when developing a monitoring tool for a control room. Several studies [@kin2009determining; @Forlines:2007:DVM:1240624.1240726] have been done to assess the benefits of the adoption of a multitouch workstation for tasks that require interaction with multiple visual cues. Results have shown that multitouch interaction can be twice as fast as a mouse based one. Furthermore, multitouch interaction is often preferred by the users due to the direct manipulation of graphical elements offered, resulting in a more natural and effective approach to carry out the requested tasks. Touch-based interaction has been exploited in control and security process since the early seventies. In 1973, Beck and Stumpe [@beck1973two] proposed a prototype of touchscreen device to control the new CERN accelerator. In recent years, studies were conducted to propose and evaluate good practices in the design process of touch-based interfaces for security operators. Zahler [@zahler2008design] proposes multiple patterns for the design of touch-based user interface for railways security and other safety-critical applications. In [@bjorneseth2012assessing], the authors investigate the effectiveness of direct manipulation in multitouch interfaces for safety-critical situations in maritime control room. Results showed that direct manipulation of interface elements can enhance situational awareness of users.
Evaluation and testing safety-critical interfaces is crucial to show whether a novel developed system actually fulfills its goals. Authors of [@stelzer2014evaluating] propose a method for the evaluation of user interface for safety in railway based on a high-fidelity simulator of an interlocking systems. Although many standardized usability evaluation methods exist and are commonly used for general purpose systems, some specific methods have been defined for the evaluation of safety critical interactive systems [@thimbleby2007interaction].
Contribution
------------
Our proposal is to use high-speed cameras mounted on a railway overhead gantry to monitor multiple aspects of the train. In particular, we have designed a machine vision system that coordinate different sensors with different speed to: (1) automatically segment the wagon identifier according to the UIC classification of railway coaches; (2) extract the wagon temperature to prevent fires and flames on board; and (3) detect the pantograph passage.
We propose a touch-based interface that adopts interaction metaphors like direct manipulation and multitouch gestures. The goal of the touch-based user interface is to give operators a quick and efficient way to interact with results of the video sequences analysis. Manipulating all the output of the machine vision system, the operator is able to efficiently control the train’s safety requirements.
![The portal equipped with all cameras. The visual matrix camera (HM-640) is on top in the center of the gantry, two thermal cameras (256L) are positioned on each side of the portal, and two visual line cameras (Spyder 4K) are positioned on the same side but one on top of the other.[]{data-label="fig:portal"}](portal){width="1\columnwidth"}
Our multi-sensors portal {#sec:oursystem}
========================
In this section we describe the physical structure and the sensors characteristics of our multi-camera portal. We then detail the acquisition manager we designed to manage all the sensors together and deal with their high-rate acquisition.
{width="100.00000%"}
Architecture of the portal and sensors involved
-----------------------------------------------
The portal is built over a single rail, has a height of 8.5 meters and is 6 meters large. The distance from the train side is around 1.5 meters. This gantry is equipped with a total of 5 sensors, the \[fig:portal\] illustrate our portal configuration.
We used three different types of high rate sensors. We summarize the characteristics of each sensor in Table \[table:sensors\]. Two of this sensors work in the visual spectrum (one linear and the other matricial) while the last one works in the thermal spectrum (with linear acquisition).
In particular, the matrix camera operating in the visual spectrum (Teledyne Dalsa HM-640) acquire 300 grayscale images per second at a resolution of 640x480 pixels. This camera is positioned on the top and at the center of the portal and is used as a general overview of the train passing by and for the visual analysis of the state of the pantograph.
The linear camera operating in the visual spectrum (Teledyne Dalsa Spyder 4K) acquires 18500 grayscale lines per second with a height of 4096 pixels. We have positioned two linear cameras on the side of the portal: one is used to observe the bottom and central part of the train and is the entry signal of the wagon identification system, see section \[sec:wagonID\]; the second linear camera is positioned higher on the side of the portal to capture the top of the train and it will be used for the detection of the pantograph detailed in section \[sec:pantograph\].
Finally, the linear camera operating in the thermal spectrum (PYROLINE 256L) acquires 512 lines of 256px each with a temperature range of $[30^{\circ} ... 800^{\circ}]$. One of this thermal sensor is positioned on the top of each side of the portal, and they are used to monitor the temperature anywhere in the passing train as explained in section \[sec:temp\].
Acquisition framework
---------------------
Full synchronization of all the sensors can be really difficult due to the high and different rates of acquisition. However, each sensor being devoted to a specific function, we don’t need full and perfect synchronization of the acquisition. We designed a specific hardware/software solution that allows us to obtain a coarse synchronization between all the sensors, enabling a meaningful and easily interpretable playback of the acquisition for the operator.
As regards hardware, we designed three separate servers to deal with the large amount of data induced by the high-rate cameras mounted on the portal. Specifically, there are 2 servers acquiring data from one linear and one thermal camera, while the third server deals only with the matricial camera as shown in the overview of our system architecture given in \[fig:acqsystem\]. These servers have 8 SAS disk in RAID-01 to obtain the sufficient speed needed (about 270MB/s) to write all the data generated, while maintaining sufficient reliability.
Concerning the software, we designed an effective solution that allows us to contemporary control each sensor focusing on the optimization of CPU, memory and disk usage. In particular, we designed a frame-grabber for each type of sensor: matrix, linear, visual or thermal; exploiting the respective SDK given by the cameras vendor.
As shown in \[fig:acqsystem\] each frame grabber is controlled by a HTTP server (AcquisitionServer). For each camera this server implements some acquisition primitives (e.g. start, stop, pause) as well as some specific function depending on the camera model, for example, focus control for the thermal cameras.
To contemporary control each AcquisitionServer (of each sensor) we designed another HTTP server called the AcquisitionManager. Each AcquisitionServer registers to the AcquisitionManager and periodically send its state. Once an AcquisitionServer is registered to the AcquisitionManager it is possible, using a simple web interface we developed, to control the IP address, the state of the grabber, and all the primitives expected for the relative sensor.
The acquisition primitives in common between all sensors are shown as a unique button in the web interface of the AcquisitionManager, while the primitives specific to a sensor are shown only for the registered sensor that can use them. Having a common interface showing the state of all sensors is particularly useful, for example, to prevent starting a new acquisition or stopping an ongoing acquisition if some of the AcquisitionServers is still saving some recently acquired data. When an AcquisitionServer is closed it unregisters from the AcquisitionManager. This software design offers the advantage that a sensor can be easily added, activated or deactivated for a specific acquisition.
Train analysis {#sec:train_analysis}
==============
The aim of the proposed system is to analyze both the visual and thermal data extracted using the high-rate sensors, described in the previous section. In particular, we developed three sub-systems in order to recognize the wagon identifier, monitor the temperature, and detect the pantograph. Finally, we will give an overview of the multitouch user interface we designed to enable an operator to interact with the processing results in the control room.
Wagon identification {#sec:wagonID}
--------------------
The *Wagon identifier subsystem* aims to identify the wagon by segmenting its unique international identification number from the image acquired with the Visual Line Camera 1 positioned at the bottom right of the portal, see \[fig:portal\]. From this identifier multiple characteristics can be extracted (type of wagon/locomotive, owner and country for example) and thus one can understand if the wagon is expected and allowed to transit on the monitored railway section. Due to the huge dimension of the image and the presence of noise we need to apply a robust identifier segmentation method. The whole method, described in Algorithm \[alg:text\], relies on image processing and geometric analysis to obtain the position of the identification number in the image.
Compute $\tau_O = AdaptiveThreshold(\mathbf{I})$;\
Extract edge $\mathbf{I}_e = CannyEdgeDetector(\mathbf{I},\tau_O)$;\
Perform morphological dilation $\mathbf{I}_D = (\mathbf{I_e} \oplus disk(r_D)$);\
Perform hole filling $\mathbf{I}_f = fill(\mathbf{I}_D)$;\
Extract Connected Components $\mathbf{c}_{bbox} = extractCC(\mathbf{I}_f);$\
Initialize votes $\mathbf{v} \leftarrow \mathbf{0}$;\
Initialize $j \leftarrow 0$;\
Initialize $k \leftarrow 0$;\
$\mathbf{\widehat{b}} = SegmentSalientRegions(\mathbf{I},\mathbf{c}_{bbox},\mathbf{v}); $\
Given an image of a wagon $\mathbf{I}$ of width $w$ and height $h$, we first apply an adaptive thresholding method to find the optimal threshold that separate foreground from background pixels [@1979:ots], such as: $$\begin{aligned}
\tau_O = AdaptiveThreshold(\mathbf{I})\end{aligned}$$ The Otsu adaptive thresholding algorithm assumes that the image contains two classes of pixels (e.g. foreground and background) then calculates the optimum threshold separating these two classes in order to minimize intra-class variance. After that, the threshold $\tau_O$ is used as input for the Canny edge detector [@Canny:1986:CAE:11274.11275] to segment the contour of the foreground elements present in the image. $$\mathbf{I}_e = CannyEdgeDetector(\mathbf{I},\tau_O).$$ The regions defined by connected edges are filled first by using the morphological operation of dilation and then with a fill operation to definitely close small holes: $$\begin{aligned}
\mathbf{I}_D & = & (\mathbf{I_e} \oplus disk(r_D)),\\
\mathbf{I}_f & = & fill(\mathbf{I}_D),\end{aligned}$$ where $r_D$ represent the disk ray size used for the dilation operation. Once those regions are filled, a connected components labelling algorithm is used to define the bounding boxes containing the blob regions previously segmented: $$\centering
\mathbf{c}_{bbox} = extractCC(\mathbf{I}_f).$$ Since the connected components generally correspond to the foreground objects in the image, we can say that after the labelling we are able to know how many foreground objects are contained in the image and what are the pixels that belong to each object. The question remaining to solve is which of these objects are characters of the wagon identifier.
In order to identify the connected components corresponding to characters we apply a voting procedure based on the sliding-window paradigm. In particular, given all the bounding boxes of the connected components we can infer that the characters of the identifier are close to each other and mostly aligned along a line. For this purpose, we apply a sliding-window procedure (with sampling step of $s$ pixels) to the image and for each sub-window, of dimension $d\times d$ pixels, we estimate a line through the RANSAC algorithm [@Fischler:1981:RSC:358669.358692] considering only the bottom-right points of the bounding boxes present in that sub-windows: $$\begin{aligned}
\mathbf{\widehat{c}}_{bbox} & = & SelectCC(\mathbf{c}_{bbox},j,k,d),\\
\mathbf{in} & = & RansacFitLine(\mathbf{\widehat{c}}_{bbox}),\end{aligned}$$ where $j$ and $k$ represent the top right coordinate of the sub-windows considered. In each sliding-window, the points $\mathbf{in}$ selected by RANSAC as inliers accumulate a vote. At the end of this procedure the points with the most votes will represent the bounding box with a higher probability of containing a character: $$\mathbf{v} = \mathbf{v} + Voting(\mathbf{c}_{bbox}, \mathbf{in}).$$
Finally, to obtain the identifier which is composed of 12 characters we selected the sub-region of the image $\mathbf{\widehat{b}}$ containing the most voted and aligned foreground objects. The alignment is estimated by computing the distances on the x-axis $D_x$ and y-axis $D_y$ for the 20 most important foreground regions according to $\mathbf{v}$. Then we take the exponential of the negative of these distances and we weight the votes previously obtained with those matrices separately. $$\mathbf{v}_w = \texttt{exp}(-D_x)*\mathbf{v}+\texttt{exp}(-D_y)*\mathbf{v}.$$ All those regions with a weighted vote $\mathbf{v}_w$ greater than zero represents a character of the ID. We then take a crop of the original image as the region containing the set of ID characters, see the example in \[fig:id\_sample\]. From this image, any Optical Character Recognition (OCR) method can be applied to obtain the identifier. This identifier segmentation step is necessary as wagon image have an average size of $4096 \times 80000$ and cannot be processed as is by an OCR.
![Example of wagon id segmentation.[]{data-label="fig:id_sample"}](sample1){width="\columnwidth"}
Temperatures segmentation {#sec:temp}
-------------------------
The *Thermal monitoring subsystem* acquires two thermal maps of each wagon and compare them to nominal operating temperatures in order to issue an alarm in case of fire risk due to abnormally high temperatures. The sensors involved in this subsystem are the Thermal Line Camera 1 and 2 positioned at the top right and left of the portal, see \[fig:portal\].
Each thermal camera is connected and managed by a different server in order to ensure a higher robustness of the thermal subsystem through duplication. The mosaic image obtained from all the acquired lines concatenation, see examples in \[fig:thermal\_samples\], is divided into subregions of fixed size and for each subregion both the mean and maximum temperatures are calculated. The minimum and maximum temperatures coming from one camera are compared with those extracted from the other camera in order to validate their output and ensure that both the servers and sensors are working correctly.
An important phenomenon to be considered is the distortion of the temperature values caused by the perspective in the image. In particular, the pixels furthest from the center of the sensor will be subject to a high distortion caused by the perspective between the sensor and the observed wagon (obviously this phenomenon depends also on the distance from the sensor to the wagon). For this reason we used two cameras observing the wagon from two different viewpoints.
Pantograph detection {#sec:pantograph}
--------------------
The *Pantograph detection subsystem* detects the passage of the pantograph in order to avoid false positive cuts in a laser based system[^1] that analyze the shape of each wagon. The pantograph detection is run on the data extracted from the Visual Line Camera 2 positioned at the middle right of the portal. The obtained segmented high resolution image can also be analyzed by an operator to determine if there is any anomaly in the pantograph shape.
![SIFT point extracted from the pantograph image template.[]{data-label="fig:pantograph"}](pantograph){width="1\columnwidth"}
The proposed solution is composed by an offline phase and an online phase. Offline, we extract SIFT [@Lowe:2004:DIF:993451.996342] keypoints from an image of the pantograph used as template $\mathbf{T}$, see \[fig:pantograph\]: $$\mathbf{D}_T = ExtractSIFT(\mathbf{T}).$$ The SIFT descriptors are representations of image regions highly discriminative and invariant to changes in brightness, scale and rotations. We store the extracted descriptors in a KD-Tree [@muja_flann_2009] in order to speedup the matching process: $$\mathbf{K}_\mathbf{T} = KDTree(\mathbf{D}_T).$$ Online, once the mosaic image of the train side is obtained, the proposed algorithm \[alg:pantograph\] extracts SIFT keypoints from that image: $$\mathbf{D}_c = ExtractSIFT(\mathbf{I}_c), \\$$
The KD-Tree nearest neighbor search provides the identifier of the closest descriptor, however, due to the curse of dimensionality, descriptors neighbors in $\mathbb{R}^{128}$, could be not visually similar. For this reason a second filtering is introduced. The distance between the first and the second more similar descriptor is measured and the match are discarded according to $\frac{\mathbf{d}_m^1}{\mathbf{d}_m^2} \leq \tau_d$, where $\tau_d = 0.67$, as in [@Brown03]: $$\mathbf{D}_m = MatchDescriptors(\mathbf{T},\mathbf{D}_c,\tau_d).$$ However, these matches are not guaranteed to be correct, this can occur for various reasons, for example repeated structures in the image or points with similar SIFT descriptors. For this reason a third validation step is performed by applying a geometric robust validation following a projective model transformation. This is obtained by exploiting the RANSAC algorithm [@Fischler:1981:RSC:358669.358692] to fit a projective model and successively by applying a consistency check algorithm to the estimated homography in order to determine if the fitted model is correct:
$$\begin{aligned}
[\mathbf{H},~\mathbf{in}] & = & RansacFitProj(\mathbf{D}_m);\\
\mathbf{p}_{bbox} & = & CheckGeomConsistency(\mathbf{H},\mathbf{in}).\end{aligned}$$
In this way it is possible to establish the presence of the pantograph and it is also possible to have an indication of its location in the image and segment the relative sub-image to be shown to the operator.
**Offline:**\
$\mathbf{D}_T = ExtractSIFT(\mathbf{T})$;\
$\mathbf{K}_\mathbf{T} = KDTree(\mathbf{D}_T)$;\
**Online:**\
$\mathbf{D}_c = ExtractSIFT(\mathbf{I}_c)$;\
$\mathbf{D}_m = MatchDescriptors(\mathbf{T},\mathbf{D}_c,\tau_d)$;\
$[\mathbf{H},~ \mathbf{in}] = RansacFitProj(\mathbf{D}_m)$;\
$\mathbf{p}_{bbox} = CheckGeomConsistency(\mathbf{H},\mathbf{in})$;
Touch-based user interface
--------------------------
To enable an operator to visualize and interact intuitively with the results of the wagon analysis we developed a touch-based graphical user interface (GUI) based on multitouch interactions. The aim of the GUI is twofold. On one hand, it is used to exhibit all the results of the acquisition and analysis to the security operator in a simple way so he can quickly get an overview of the train status. On the other hand it provides the operator with several tools for a direct and easy manipulation of all the data necessary to assess the train safety requirements.
An operator first loads a session of a processed wagon analysis in the interface. A session is composed by (i) a frontal video of the train obtained from the matrix camera positioned at the top of our portal, (ii) two thermal images that are the output of the thermal monitoring subsystem, (iii) a high-resolution image of the train acquired by the visual linear camera 1. The frontal video can be played and scanned through a timeline visualizer. The timeline of the video is synchronized with visual markers on both the thermal and linear imagery in order to give a visual time reference on all results.
Thermal images are displayed with a false-color scale obtained from the temperature values. By default the scale is based on the *max* and the *min* values of the image, but the operator can manually change the range of colors in order to enhance the visualization of specific temperature values. Left and right thermal images can be activated with a selector, so that only one image at a time is visualised in the interface.
The linear camera acquisition result is a high resolution image of the train. In order to allow a fluid and smooth manipulation of this image we adopted a multi-resolution tiling technique [@y2008methods]. Acquired images are pre-processed in order to have a set of downscaled versions of high-definition ones. Each downscaled version is then decomposed in tiles of 256x256 pixels. The rendering engine of the GUI loads and display only the tiles required for the current zoom level and portion of the image visualized by the operator, instead of loading the entire high-definition image. The operator can activate graphical overlays on the train image in order to visualize the results obtained by the *pantograph detection subsystem* and the *wagon identifier subsystem*.
\[fig:table\_sample\] shows an overview of the interface and some phases of the interaction of a control operator with the interactive GUI, like checking temperature of an area of the wagon or visualising an high-definition image of the train. The set of functionalities provided by the GUI allows the operator to have a quick overview of the image processing analysis results and to perform punctual and precise controls through direct manipulation using multitouch gestures.
System evaluation {#sec:evaluation}
=================
To evaluate our proposed system we recorded a dataset of sequences using the portal depicted in \[fig:portal\] in Poland (Zmigrod). We acquired 36 sequences of a train composed of one locomotive and one wagon on a test-bed railway track of 1 Km. The train passed under the multi-camera gantry at different time of the day, at different speeds and with different weather conditions. To register these sequences we used the system architecture and the web interface previously described in section \[sec:oursystem\]. In this section we will evaluate each of the sub-systems of our proposed approach.
Wagon identification {#wagon-identification}
--------------------
![Acquisition sample of a full HD video taken with a standard camera.[]{data-label="fig:bad"}](bad){width="1\columnwidth"}
As shown in \[fig:bad\] it would be really difficult to segment the identifier with a standard camera. Indeed, a high motion blur due to the train speed affects the readability of several characters. The use of a high-rate linear camera is hence necessary to be able to properly segment the wagon identifier. To evaluate the wagon identifier segmentation performance, we estimated the accuracy of both the full train ID and the single characters segmentation for each wagon. In particular, for the case of characters segmentation we count as true positive every detected region that contains a character of the wagon ID, as false negative every missed character of the wagon ID and as false positive every region classified as part of the ID but non containing a character of the wagon ID. While for the full ID segmentation, we count a true positive every time all the characters of the wagon ID are recognized, a false negative every time at least one character of the wagon ID is missed and as false positive all the regions classified as positive but that do not contain a character of the wagon ID. For the full ID segmentation evaluation, there is exactly one target detection by wagon making it easier to obtain higher false positive rate as any region that do not contains the ID will be counted as a false positive.
As it can be observed from Table \[table:textacc\_id\] the accuracy of the system is very high for full ID segmentation in the case of wagon 1 while for the wagon 2 we are always able to detect the full train ID. We can also appreciate that both false positives and false negatives are limited for the full train ID segmentation of each wagon. When evaluating in terms of character segmentation, see Table \[table:textacc\], the results are even better with really low false negative and false positive rates. False positives are mainly caused by the fact that sometimes small character in the train ID are merged together and a region close to the train ID can be considered part of it, as shown in Fig. \[fig:text\_fp\].
![Example of false positive when two character regions (here 0 and 3) are merged.[]{data-label="fig:text_fp"}](sissi_text_fp2){width="49.00000%"}
In \[fig:text\] we show a qualitative sample of how the proposed solution segment the train identifier, for both the locomotive and the wagon. One can observe how the train identifier is a very small part of the initial image and appreciate how our method successfully detects it.
{width="\textwidth"}
Pantograph detection {#pantograph-detection}
--------------------
The pantograph was observed only in the afternoon test session, so for 18 (out of 34) sequences of the dataset. However, for each one of the 18 sequences the pantograph is correctly detected by the proposed solution.
In \[fig:pantograph\_samples\] we report some samples of the pantograph matching working under very different illumination conditions.
Evaluation of the user interface
--------------------------------
Cognitive Walkthrough (CW) [@wharton1994cognitive] is a usability inspection method whose objective is to identify usability problems, focusing on how easy it is for new users to accomplish predefined tasks. It is a technique that aims at detecting errors in design that would interfere with the performance of users while using the interface. CW is usually carried out by specialists in the field of interface development and usability experts. As the walkthrough proceeds, comments of the users are recorded. We conducted an usability inspection of the proposed touch-based interface to assess its possible usability issues. For this purpose we defined the following different tasks:
T1
: Load the most recent analysed sequence
T2
: Position the video on the sequence corresponding to the pantograph detection
T3
: Visualize areas of the wagon which temperature is higher than $50^{\circ}$
T4
: Visualize the wagon ID number using the analysis results
For each task we defined a sequence of actions with details about specific task flow from beginning to end. We asked 5 examiners to perform the defined task using the so-called *think aloud* technique in order to record failure in the interaction and design suggestions. Previous studies on usability testing [@nielsen1993mathematical] showed that the number of usability problems $\mathbf{U_p}$ found in a usability test is: $$\mathbf{U_p} = \mathbf{N} (1-(1- \mathbf{L} )^\mathbf{n} )$$ where $\mathbf{n}$ is the number of users, $\mathbf{N}$ is the total number of usability problems in the design and $\mathbf{L}$ is the proportion of usability problems discovered while testing a single user. The typical value of $\mathbf{L}$ is 31%, suggesting that 85% of usability issues can be found with 5 testers.
All the examiners were able to complete assigned tasks, reporting usability and user interface related problems while performing the evaluation. Feedbacks from examiners allowed us to identify and correct minor but important usability issues, mostly regarding sizes and positions of objects on the screen or ambiguities in the use of textual labels.
Conclusions
===========
In this paper we introduced a multi-camera portal for train safety assessment. Our proposal is able to perform the analysis of multiple safety requirements of each train passing under the gantry without requiring the train to be stopped. We detailed the hardware used and the software developed to robustly acquire data from multiple high-rate sensors. Image processing and computer vision methods are applied on each data stream to extract meaningful information. We also presented our multitouch interface that enables an operator to quickly observe and simply interact with the processed data. The evaluation has shown the good performances of the analysis and the usability of the interface.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported by the Integrated Intermodal System for Security and Signaling on Rail (SISSI) project, funded by Regione Toscana - Italy.
[^1]: This system is composed of three infrared laser mounted on the portal. This proprietary solution was developed by Thales Italia and cannot be discussed in 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.